Introduction
When do local household shocks, such as losses of income or wealth, move aggregate credit prices? The purpose of this paper is to derive conditions under which local shocks move aggregate credit prices, not to classify shocks by their innate properties. I show that the financial importance of such shocks depends on two features, where the effects travel over time and whose investment demand changes. I build an object that captures these conditions: it follows a local shock along social ties to the responses of households in every county at every horizon, and weights each response by how strongly households there change their desired credit holdings. I call it the financially weighted propagation operator , and I call its weights the financial weights. My main result shows when household measurements are sufficient to recover this financially relevant response without reconstructing the entire propagation system.
I develop the mechanism in four steps. First, a local disturbance, such as a storm, lowers income, wealth, and spending in affected counties. Second, its effects travel along social ties to friends and relatives in other counties through short-lived news and changes in household spending and saving that can persist for months. The non-affected households may send money to those who lost homes or jobs, or save more in case of a similar loss, so spending falls in counties the disturbance never hit. Third, households there change their demand for corporate bonds, which many hold through mutual funds and retirement accounts. For example, a family that sends money to their affected relatives may redeem fund shares, and its thinner cushion above usual spending can make risky holdings less attractive [7 ] . Precautionary behavior can reinforce or offset this change since a cautious household may move from bonds into safer balances or from stocks into bonds. The financial weights account for how many households in each county hold bonds and how strongly their demand responds [28 , 31 ] . Fourth, intermediaries and other investors adjust their holdings as household bond demand changes. If the combined effects of transfers and precautionary behavior reduce household demand on net, these investors would require higher expected returns to hold the additional bonds, and vice-versa. The return response would be larger when intermediary balance sheets are tight. In the model, this adjustment changes the excess bond premium (EBP), the part of corporate bond spreads not explained by expected default [15 ] . A local disturbance could therefore move national credit prices months later.
I devise a law of transmission to track these effects. Studies using the Social Connectedness Index (SCI) of Bailey et al. [6 ] show that households spend less after COVID-19 outbreaks in socially connected counties [27 ] and purchase more flood insurance after floods affecting distant friends (Hu, 2022). Similarly, Gao et al. [14 ] study household financial outcomes after shale-income shocks, and social ties also shape investment and the diffusion of news [26 , 20 ] . For credit prices, this evidence leaves two questions open. The first is what additional financial information is needed to translate household responses into a change in credit demand. The SCI records social ties, not the resources or financial claims exchanged through them. Its weights therefore do not directly measure the strength of economic transmission or households’ bond-demand responses.
The second question concerns timing. A regression on a fixed exposure score, such as an SCI-weighted average of shocks in connected counties, lets the response change with the horizon but holds counties’ relative exposure fixed. The question is whether one such score can reproduce locations’ relative credit-price effects at every horizon. Estimating the law of transmission county by county would answer both questions, but it is not feasible without strong restrictions on it, since it links every county to every other at every horizon, far more than a few dozen storms can pin down. I therefore develop a method that needs only the part of the law that matters for credit prices.
I model transmission with two channels on one social graph. Each channel assigns a value to every county, and this value spreads over time to socially connected counties. I call such a channel a field. Both fields spread along the same social ties and differ in when they start to respond and how quickly they fade. The fast field carries news and attention, which react at once and die out quickly. The slow field carries changes in household demand, which start a period later and last longer. Building on the spectral approach to networks of Galeotti et al. [13 ] , I write every response as a sum of regional patterns of the graph, each fading at its own speed. Under the diffusion law, patterns that vary sharply across strong social ties decay faster than broad ones.
The slow field could thus lower household bond demand, and markets would then need to price that shift. Building on demand-system asset pricing [24 ] (Koijen and Yogo 2026) and intermediary risk-bearing capacity [4 , 16 ] (He and Krishnamurthy 2013; He et al. 2017), I derive a conditional equilibrium in which lower household demand raises the compensation required to hold credit. When intermediaries can still expand their holdings, lower capital increases the premium response to the same household demand shift. In the model, an explicit measurement relation links this premium to the EBP. The operator supplies the demand shift, and the equilibrium determines its price impact.
I then aggregate household responses across counties using the financial weights to derive their effect on the credit premium. Gabaix [11 ] , Acemoglu et al. [1 ] (2012), Carvalho et al. [8 ] show how firm size and production linkages shape the aggregate effects of microeconomic disturbances, and Herskovic [19 ] connects the production-network structure to equilibrium asset prices. In these models, the weights are tied to firms’ and sectors’ sales. Building on these insights, I construct financial weights that measure how strongly households in each county change their desired credit holdings. Holding intermediary capital and payoff moments fixed, the household-demand part of the premium therefore depends on the propagation system only through one financially weighted sum of responses at each horizon, given in equation (14) . Consequently, two propagation systems with the same sum imply the same response of this part of the premium, however differently they move individual counties.
I answer the first question with Proposition 6, which gives the exact condition under which household measurements are sufficient to recover this sum. Suppose two patterns of adjustment look identical in the measurements but imply different changes in credit demand. The measurements then cannot distinguish their price implications. Given how each response shows up in the measurements and independently measured financial sensitivities, recovery holds exactly when this cannot happen for the shocks and horizons studied. Household measurements can therefore support a credit-price prediction without identifying the propagation system itself. Consumption measurements, measurements of desired portfolio holdings at given prices, or both together can meet this condition. Additionally, estimates of household portfolio demand supply the financial sensitivities and their signs, which spending alone cannot reveal, since a household can absorb a loss by cutting spending or by drawing down assets. The condition holds whether households lower or raise their bond demand. An error bound separates error in predicting the measured responses from financial information the measurements miss.
I also show that the response common to all counties needs outside information, since comparisons across counties within an event reveal only how responses differ ([29 , 33 ] ). The network GIV of Chodorow-Reich et al. [9 ] addresses this problem with granular shocks. Here the common level matters for prices whenever the financial sensitivities do not sum to zero, because credit demand adds up all counties. Without restrictions on how responses propagate, recovering the aggregate response then requires an independent measurement of the common level or a bound on it.
I answer the second question with Theorem 1 and in general the answer is no. A location’s relative importance for credit prices generically changes with the horizon when the regional patterns or channels carrying the financially relevant response fade at different speeds. A location can account for a large share of the response soon after a shock and a smaller share later, even though its social ties have not changed. One fixed score then cannot reproduce these relative magnitudes at every horizon, except in the special case where the patterns’ horizon profiles are proportional. The diffusion model, applied to the observed social graph with proxy financial weights, implies that counties with nearly the same total connectedness can even reverse their relative importance (Figure 2 ).
Figure 2. Horizon-specific importance on the observed network. Panel A covers 1,590 counties. The pair distribution in Panel B uses counties with populations of at least 50,000 each. Its initial upper quartile (6.62) lies above the displayed limit of 4. Results are model-implied for equal initial field inputs in symmetric coordinates and dividend-proxy weights. Converting network spreading τ = κ D ( h − 1 ) \tau= \kappa_D(h - 1) τ = κ D ( h − 1 ) into weeks requires the rate κ D \kappa_D κ D . Appendix E.5 gives the selection rules and additional comparisons.
I begin the empirical analysis with the spending step of the mechanism, using the SCI and weekly county spending, with storms as localized disturbances. Spending is lower weeks later in counties more socially connected to storm-damaged places, although these counties lie outside the storms’ recorded footprints, after accounting for their own weather and for geographic and production links to the damaged places. This gradient remains similar across alternative records of where and when damage occurred, and it gives the propagation analysis its empirical starting point.
Finally, I use simulations to show that the method can work on the actual network. I use the SCI graph and candidate storm footprints and simulate the household accounts, aggregate measurements and fund data that recovery requires. When the data follow the specified diffusion law, the operator predicts the financially weighted responses to held-out storms more accurately than exposure through direct ties to the hit counties, with near-nominal interval coverage. Moreover, it keeps this advantage when the scale of the financial sensitivities and the price impact must be estimated. Adão et al. (2025) similarly judge quantitative models by the causal predictions that a particular economic question requires.
§ 2 develops the propagation model and § 3 derives the pricing kernel. § 4 and § 5 present the spending evidence and its dynamic counterpart, § 6 financial geography, EBP associations and fund demand, and § 7 the recovery program and its simulations.
Propagation on a social graph
Two fields on the social network
The SCI measures friendship links between counties. I use these links to model how a shock in one county affects households elsewhere. Households may respond briefly to news that friends were hit. They may also change spending and saving for months as they help relatives or prepare for future losses. I represent these responses with a fast attention field and a slow demand field. I use the slow field for the household bond-demand channel in § 3 .
I represent the links among n C n_C n C counties, indexed by i = 1 , … , n C i = 1, \ldots, n_C i = 1 , … , n C , by symmetric weights A = ( W + W ′ ) / 2 A = (W + W^{\prime})/2 A = ( W + W ′ ) /2 , where W ∈ R n C × n C W \in\mathbb{R}^{n_C \times n_C} W ∈ R n C × n C contains the retained, nonnegative SCI links, normalized so that W 1 = 1 W\mathbf{1} = \mathbf{1} W 1 = 1 . Here 1 ∈ R n C \mathbf{1} \in\mathbb{R}^{n_C} 1 ∈ R n C is the vector of ones. I let D ≡ diag ( A 1 ) D \equiv\operatorname{diag}(A\mathbf{1}) D ≡ diag ( A 1 ) record each county’s total connectedness, which is positive, and use the normalized Laplacian L ≡ Id n C − D − 1 / 2 A D − 1 / 2 \mathcal{L} \equiv\operatorname{Id}_{n_C} - D^{-1/2} A D^{-1/2} L ≡ Id n C − D − 1/2 A D − 1/2 to model diffusion through these links, with Id n C \operatorname{Id}_{n_C} Id n C denoting the n C × n C n_C \times n_C n C × n C identity matrix.
Let ε t ∈ R n C \varepsilon_t \in\mathbb{R}^{n_C} ε t ∈ R n C denote a fixed vector of local shocks at date t t t , with one entry for each county, such as the damage a storm causes in the counties it hits. I compare two paths with the same pre-shock history that differ only by this shock at date t t t . I report consumption and field responses relative to the path without the shock. Time is discrete, with h = 0 , 1 , … h = 0, 1, \ldots h = 0 , 1 , … measuring periods after the shock.
Let c t ∈ R n C \mathbf{c}_t \in\mathbb{R}^{n_C} c t ∈ R n C collect the direct component of county log-consumption responses, excluding contributions carried by the fields, with Δ c t ≡ c t − c t − 1 \Delta\mathbf{c}_t \equiv\mathbf{c}_t - \mathbf{c}_{t-1} Δ c t ≡ c t − c t − 1 . Its growth response is Δ c t + h = Φ h ε t \Delta\mathbf{c}_{t+h} = \Phi_h \boldsymbol{\varepsilon}_t Δ c t + h = Φ h ε t , where Φ h ∈ R n C × n C \Phi_h \in\mathbb{R}^{n_C \times n_C} Φ h ∈ R n C × n C . I feed the impact response, ν t ≡ Φ 0 ε t \boldsymbol{\nu}_t \equiv\Phi_0 \boldsymbol{\varepsilon}_t ν t ≡ Φ 0 ε t , into the slow field. The same shock generates a signal s t = B ε t \mathbf{s}_t = B\boldsymbol{\varepsilon}_t s t = B ε t , with B ∈ R n C × n C B \in\mathbb{R}^{n_C \times n_C} B ∈ R n C × n C , which enters the fast field. I assume the following dynamics.2
Assumption 1 (Two fields on one graph). The slow demand field d t ∈ R n C \mathbf{d}_t \in\mathbb{R}^{n_C} d t ∈ R n C and the fast attention field b t ∈ R n C \mathbf{b}_t \in\mathbb{R}^{n_C} b t ∈ R n C evolve as
d t + 1 = A D d t + η D ν t , A D ≡ ρ D e − κ D L , b t = A a t t b t − 1 + η a t t s t , A a t t ≡ ρ a t t e − κ a t t L , \begin{aligned}
\mathbf{d}_{t+1} &= A_D\mathbf{d}_t + \eta_D\boldsymbol{\nu}_t, & A_D &\equiv\rho_D e^{-\kappa_D\mathcal{L}},
\\
\mathbf{b}_t &= A_{\mathrm{att}}\mathbf{b}_{t-1} + \eta_{\mathrm{att}}\mathbf{s}_t, & A_{\mathrm{att}} &\equiv\rho_{\mathrm{att}}e^{-\kappa_{\mathrm{att}}\mathcal{L}},
\end{aligned} d t + 1 b t = A D d t + η D ν t , = A att b t − 1 + η att s t , A D A att ≡ ρ D e − κ D L , ≡ ρ att e − κ att L ,
with ρ D ∈ ( 0 , 1 ) \rho_D \in(0,1) ρ D ∈ ( 0 , 1 ) , ρ a t t ∈ [ 0 , 1 ) \rho_{\mathrm{att}} \in[0,1) ρ att ∈ [ 0 , 1 ) and κ D , κ a t t , η D , η a t t > 0 \kappa_D, \kappa_{\mathrm{att}}, \eta_D, \eta_{\mathrm{att}} > 0 κ D , κ att , η D , η att > 0 .
The shared pre-shock history gives initial field responses d t = b t − 1 = 0 \mathbf{d}_t = \mathbf{b}_{t-1} = 0 d t = b t − 1 = 0 . The attention field responds on impact, while the demand field begins one period later. The coefficients η D \eta_D η D and η a t t \eta_{\mathrm{att}} η att determine how strongly each input enters its field. I use ρ m \rho_m ρ m to govern persistence and κ m \kappa_m κ m to govern diffusion through social links, m ∈ { D , a t t } m \in\{D, \mathrm{att}\} m ∈ { D , att } .
I decompose county responses into spatial patterns called modes, the orthonormal eigenvectors U = ( u 1 , … , u n C ) U = (u_1, \ldots, u_{n_C}) U = ( u 1 , … , u n C ) of L \mathcal{L} L , ordered so that their eigenvalues satisfy 0 = λ 1 ≤ ⋯ ≤ λ n C 0 = \lambda_1 \le\cdots\le\lambda_{n_C} 0 = λ 1 ≤ ⋯ ≤ λ n C . Each mode gives counties’ relative movements, with counties of the same sign moving together. Its eigenvalue λ k \lambda_k λ k , the graph frequency of mode k k k , measures how sharply the pattern differs across social ties. Smooth modes have low graph frequency and assign similar degree-adjusted values D − 1 / 2 u k D^{-1/2}u_k D − 1/2 u k to strongly connected counties. Figure 1 contrasts a smooth mode, which separates the Upper Midwest and Plains from Appalachia, the Mid-Atlantic and the Carolinas, with a rough mode that alternates in sign among neighboring counties of the Upper Midwest.
Figure 1. Two eigenvectors of the SCI Laplacian on the sampled counties, plotted at county centroids. Loadings are D − 1 / 2 u k D^{-1/2}u_k D − 1/2 u k , scaled to the 97th percentile of their absolute value. Signs are arbitrary.
The transition satisfies A m u k = r m , k u k A_m u_k = r_{m,k}u_k A m u k = r m , k u k , where r m , k ≡ ρ m e − κ m λ k r_{m,k} \equiv\rho_m e^{-\kappa_m\lambda_k} r m , k ≡ ρ m e − κ m λ k . Without further inputs, each mode therefore keeps its shape and retains a fraction r m , k r_{m,k} r m , k of its amplitude each period. Diffusion makes rougher patterns fade faster when ρ m > 0 \rho_m > 0 ρ m > 0 (Appendix § 12.8 .§ 4 gives the implied half-lives). The first mode has λ 1 = 0 \lambda_1 = 0 λ 1 = 0 and is proportional to square-root connectedness, u 1 ∝ D 1 / 2 1 u_1 \propto D^{1/2}\mathbf{1} u 1 ∝ D 1/2 1 . Let’s call it the stationary mode. I assume the graph is connected, meaning chains of friends link every two counties. The stationary mode is then the only mode with λ = 0 \lambda= 0 λ = 0 , and it survives at rate ρ m \rho_m ρ m . The persistence parameter therefore sets how long the broadest pattern lasts. Calling the demand field slow presumes r D , k > r a t t , k r_{D,k} > r_{\mathrm{att},k} r D , k > r att , k for the modes carrying the response.
The regressions in § 4 measure exposure through direct SCI links. The dynamic model allows transmission along longer paths and links responses across horizons, so a one-hop coefficient is not an estimate of κ D \kappa_D κ D (Appendix § 13 .§ 3 ).
I also allow a storm’s economic effects to spread through production links. I represent this channel by a slow field d t P ∈ R n C d_t^P \in\mathbb{R}^{n_C} d t P ∈ R n C on a graph built from county industry shares and the BEA input–output matrix. It follows the slow law d t + 1 P = A D P d t P + η D P ν t d_{t+1}^P = A_D^P d_t^P + \eta_D^P \nu_t d t + 1 P = A D P d t P + η D P ν t , with initial response d t P = 0 d_t^P = 0 d t P = 0 and its own transition A D P A_D^P A D P and input coefficient η D P \eta_D^P η D P (Appendix § 13 .§ 2 ). I include production exposure as a competing channel because production disruptions can affect household resources and intermediary capital.
The response operator
To compare channels and counties, I combine the direct consumption effect and the fields into one response per county, the local valuation v t ∈ R n C v_t \in\mathbb{R}^{n_C} v t ∈ R n C ,
v t ≡ − γ Δ c t − a D d t − a a t t b t − a D P d t P , γ , a D , a a t t > 0 , a D P ≥ 0. (1) v_t \equiv-\gamma\Delta c_t - a_D d_t - a_{\mathrm{att}} b_t - a_D^P d_t^P,\qquad\gamma, a_D, a_{\mathrm{att}} > 0,\ a_D^P \ge0. \tag*{(1)} v t ≡ − γ Δ c t − a D d t − a att b t − a D P d t P , γ , a D , a att > 0 , a D P ≥ 0. (1)
Under power utility with curvature γ \gamma γ , the direct consumption response contributes − γ c t -\gamma c_t − γ c t to the log-marginal-utility response, and the first term records the change in this contribution between periods. I choose the signs so that negative direct consumption-growth responses and adverse field responses raise the index. I interpret a higher index as payoffs becoming more valuable to households in that county, though not necessarily to the investors who set prices. I combine county valuations with fixed weights π ∈ R n C \pi\in\mathbb{R}^{n_C} π ∈ R n C . Holding these weights fixed lets me ask whether a source county’s relative importance changes across horizons. The weights can have either sign, so the results below also apply to the financial loadings that § 3 derives for credit prices.
Definition 1 (Response operator and valuation projection). For the one-time shock ε t \varepsilon_t ε t of § 2.1 , the local response operator is O h ≡ ∂ v t + h / ∂ ε t ′ ∈ R n C × n C \mathcal{O}_h \equiv\partial v_{t+h}/\partial\varepsilon_t' \in\mathbb{R}^{n_C \times n_C} O h ≡ ∂ v t + h / ∂ ε t ′ ∈ R n C × n C , h ≥ 0 h \ge0 h ≥ 0 . The valuation projection is the row vector
Ω h = π ′ O h ∈ R 1 × n C . \Omega_h = \pi^{\prime}\mathcal{O}_h \in\mathbb{R}^{1 \times n_C}. Ω h = π ′ O h ∈ R 1 × n C .
Entry ( i , j ) (i,j) ( i , j ) of O h \mathcal{O}_h O h is the response of county i i i ’s valuation h h h periods after a unit shock in county j j j , relative to the path without the shock. Because the propagation system is linear, O h \mathcal{O}_h O h does not depend on the size of the shock. Entry j j j of Ω h \Omega_h Ω h adds these responses across recipient counties with weights π \pi π , so it measures the importance of source county j j j at horizon h h h .
Lemma 1 (Closed form). Under Assumption 1 , O h = − γ Φ h + O ~ h S C I + O ~ h P \mathcal{O}_h = -\gamma\Phi_h + \widetilde{\mathcal{O}}^{SCI}_h + \widetilde{\mathcal{O}}^{P}_h O h = − γ Φ h + O h S C I + O h P , where
O ~ h S C I = − a D η D 1 { h ≥ 1 } A D h − 1 Φ 0 − a a t t η a t t A a t t h B , O ~ h P = − a D P η D P 1 { h ≥ 1 } ( A D P ) h − 1 Φ 0 . (2) \widetilde{\mathcal{O}}^{SCI}_h = -a_D\eta_D\mathbf{1}\{h \ge1\}A_D^{h-1}\Phi_0 - a_{\mathrm{att}}\eta_{\mathrm{att}}A_{\mathrm{att}}^hB,\qquad\widetilde{\mathcal{O}}^{P}_h = -a_D^P\eta_D^P\mathbf{1}\{h \ge1\}(A_D^P)^{h-1}\Phi_0. \tag*{(2)} O h S C I = − a D η D 1 { h ≥ 1 } A D h − 1 Φ 0 − a att η att A att h B , O h P = − a D P η D P 1 { h ≥ 1 } ( A D P ) h − 1 Φ 0 . (2) The first term is the direct effect through local consumption. The second carries demand and attention through social ties, and the third carries persistent demand through production links. In the regressions, I control for measured own weather and include production exposure alongside social exposure (§ 4 ). The rest of this section studies the social part O ~ h S C I \widetilde{\mathcal{O}}^{SCI}_h O h S C I and its projection Ω ~ h S C I ≡ π ′ O ~ h S C I \widetilde{\Omega}^{SCI}_h \equiv\pi^{\prime}\widetilde{\mathcal{O}}^{SCI}_h Ω h S C I ≡ π ′ O h S C I . § 3 uses the slow field to derive the household-demand component of the credit-price response (equation (14) ). All proofs are in Appendix § 12.8 .
Spectral form and identification
I now ask whether one fixed exposure score can describe a source county’s importance at every horizon. As modes decay at different rates, the mixture of responses changes even though the social ties and county weights stay fixed. The same modal form then shows what one channel’s decay pattern identifies.
For a shock shaped like mode k k k , ε t = u k \varepsilon_t = u_k ε t = u k , the response of mode k k k itself is C k ( h ) ≡ u k ′ O ~ h S C I u k C_k(h) \equiv u_k^{\prime}\widetilde{\mathcal{O}}^{SCI}_h u_k C k ( h ) ≡ u k ′ O h S C I u k . Lemma 1 and A m u k = r m , k u k A_m u_k = r_{m,k}u_k A m u k = r m , k u k give C k ( h ) = C D , k ( h ) + C a t t , k ( h ) C_k(h) = C_{D,k}(h) + C_{\mathrm{att},k}(h) C k ( h ) = C D , k ( h ) + C att , k ( h ) , with
C D , k ( h ) ≡ 1 { h ≥ 1 } C D , k i n i t r D , k h − 1 , C a t t , k ( h ) ≡ C a t t , k i n i t r a t t , k h , (3) C_{D,k}(h) \equiv\mathbf{1}\{h \ge1\}C^{\mathrm{init}}_{D,k}r_{D,k}^{h-1},\qquad C_{\mathrm{att},k}(h) \equiv C^{\mathrm{init}}_{\mathrm{att},k}r_{\mathrm{att},k}^h, \tag*{(3)} C D , k ( h ) ≡ 1 { h ≥ 1 } C D , k init r D , k h − 1 , C att , k ( h ) ≡ C att , k init r att , k h , (3)
where C D , k i n i t ≡ − a D η D u k ′ Φ 0 u k C^{\mathrm{init}}_{D,k} \equiv-a_D\eta_D u_k^{\prime}\Phi_0u_k C D , k init ≡ − a D η D u k ′ Φ 0 u k and C a t t , k i n i t ≡ − a a t t η a t t u k ′ B u k C^{\mathrm{init}}_{\mathrm{att},k} \equiv-a_{\mathrm{att}}\eta_{\mathrm{att}}u_k^{\prime}Bu_k C att , k init ≡ − a att η att u k ′ B u k . The coefficient C m , k i n i t C^{\mathrm{init}}_{m,k} C m , k init sets channel m m m ’s initial response, and r m , k r_{m,k} r m , k sets its decay per period. The slow channel enters at h = 1 h = 1 h = 1 and the fast channel at h = 0 h = 0 h = 0 .
A shock shaped like one mode can also move other modes, because the direct response Φ 0 \Phi_0 Φ 0 and the signal loading B B B need not follow the network’s patterns. Let K ∈ { 1 , … , n C } K \in\{1,\ldots,n_C\} K ∈ { 1 , … , n C } be the number of retained modes and U K = ( u 1 , … , u K ) ∈ R n C × K U_K = (u_1,\ldots,u_K) \in\mathbb{R}^{n_C \times K} U K = ( u 1 , … , u K ) ∈ R n C × K .
Assumption 2 (Retained-mode diagonality). On the first K K K modes , U K ′ Φ 0 U K U_K^{\prime}\Phi_0U_K U K ′ Φ 0 U K and U K ′ B U K U_K^{\prime}BU_K U K ′ B U K are diagonal up to small errors, and the loadings between retained and discarded modes are small (norms in Appendix § 13 .§ 1 ).
I write O ~ h S C I = U diag ( C k ( h ) ) U ′ + E h , K \widetilde{\mathcal{O}}^{SCI}_h = U\operatorname{diag}(C_k(h))U^{\prime} + \mathcal{E}_{h,K} O h S C I = U diag ( C k ( h )) U ′ + E h , K , where E h , K \mathcal{E}_{h,K} E h , K collects the off-diagonal terms of O ~ h S C I \widetilde{\mathcal{O}}^{SCI}_h O h S C I in the modal basis, and Assumption 2 makes those involving the first K K K modes small. The diagonal part of the projection loads on mode k k k with coefficient ψ k ( h ) = ω k C k ( h ) \psi_k(h) = \omega_k C_k(h) ψ k ( h ) = ω k C k ( h ) , where ω k = π ′ u k \omega_k = \pi^{\prime}u_k ω k = π ′ u k . In the exact diagonal benchmark, E h , K = 0 \mathcal{E}_{h,K} = 0 E h , K = 0 and Ω ~ h S C I = ∑ k ψ k ( h ) u k ′ \widetilde{\Omega}^{SCI}_h = \sum_k \psi_k(h)u_k^{\prime} Ω h S C I = ∑ k ψ k ( h ) u k ′ . In this benchmark, I call channel m m m active when C m , k i n i t ≠ 0 C^{\mathrm{init}}_{m,k} \ne0 C m , k init = 0 for some mode k k k . With financial weights, a persistent pricing component therefore needs a mode on which the weights load, ω k ≠ 0 \omega_k \ne0 ω k = 0 , and which decays slowly. Within a channel, smoother modes decay more slowly. A rough mode can still matter when the financial weights load on it.
Theorem 1 (Nonseparability of horizon-specific loadings). In the exact diagonal benchmark, let C ˉ m , k i n i t ≡ ω k C m , k i n i t \bar{C}^{\mathrm{init}}_{m,k} \equiv\omega_k C^{\mathrm{init}}_{m,k} C ˉ m , k init ≡ ω k C m , k init , so that ψ k ( h ) = 1 { h ≥ 1 } C ˉ D , k i n i t r D , k h − 1 + C ˉ a t t , k i n i t r a t t , k h \psi_k(h) = 1\{h \ge1\}\bar{C}^{\mathrm{init}}_{D,k}r_{D,k}^{h-1} + \bar{C}^{\mathrm{init}}_{\mathrm{att},k}r_{\mathrm{att},k}^{h} ψ k ( h ) = 1 { h ≥ 1 } C ˉ D , k init r D , k h − 1 + C ˉ att , k init r att , k h . Suppose two modes k ≠ ℓ k \ne\ell k = ℓ have nonzero coefficient pairs ( C ˉ D , k i n i t , C ˉ a t t , k i n i t ) (\bar{C}^{\mathrm{init}}_{D,k},\bar{C}^{\mathrm{init}}_{\mathrm{att},k}) ( C ˉ D , k init , C ˉ att , k init ) and ( C ˉ D , ℓ i n i t , C ˉ a t t , ℓ i n i t ) (\bar{C}^{\mathrm{init}}_{D,\ell},\bar{C}^{\mathrm{init}}_{\mathrm{att},\ell}) ( C ˉ D , ℓ init , C ˉ att , ℓ init ) and distinct decay pairs, and both channels are active. Then, generically, no fixed s f i x ∈ R n C s^{\mathrm{fix}} \in\mathbb{R}^{n_C} s fix ∈ R n C and scalars a h a_h a h satisfy Ω ~ h S C I = a h ( s f i x ) ′ \widetilde{\Omega}^{\mathrm{SCI}}_h = a_h(s^{\mathrm{fix}})' Ω h SCI = a h ( s fix ) ′ at every horizon h ≥ 0 h \ge0 h ≥ 0 . Precisely, such a representation exists if and only if ψ k ( h ) ψ ℓ ( j ) − ψ ℓ ( h ) ψ k ( j ) = 0 \psi_k(h)\psi_\ell(j) - \psi_\ell(h)\psi_k(j) = 0 ψ k ( h ) ψ ℓ ( j ) − ψ ℓ ( h ) ψ k ( j ) = 0 for all modes k , ℓ k,\ell k , ℓ and horizons h , j h,j h , j .
A fixed score requires the nonzero weighted modal profiles ψ k ( h ) \psi_k(h) ψ k ( h ) to be proportional across horizons. One contributing mode satisfies this condition. Different decay rates can break it within a single channel, while a second channel supplies another route. Appendix A.11 defines the error of the best fixed-score approximation, and held-out performance measures what additional dimensions add.
The theorem concerns relative magnitudes, and a reversal of rankings is its most visible form. Figure 2 shows such reversals on the observed SCI graph. With county shares of IRS dividend income w w w as proxy financial weights and equal unit slow-field inputs in symmetric coordinates, the slow-channel pricing loading of source county j j j at diffusion time τ = κ D ( h − 1 ) \tau= \kappa_D(h - 1) τ = κ D ( h − 1 ) is I j ( τ ) = [ e − τ L w ] j I_j(\tau) = [e^{-\tau\mathcal{L}}w]_j I j ( τ ) = [ e − τ L w ] j , up to a common factor, with τ = 0 \tau= 0 τ = 0 at h = 1 h = 1 h = 1 . Each fixed score in Panel A tracks only part of the path. Among the 27,169 county pairs whose total SCI connectedness is within 1 percent, 24.5 percent reverse their initial ordering by τ = 4 \tau= 4 τ = 4 . Panel B follows a subset of these pairs with an initial loading ratio of at least two, and their median ratio falls from 3.73 to 1.32 by τ = 4 \tau= 4 τ = 4 . For example, Loudon, Tennessee, and Washington, New York, reverse their ranking (selection rules in Appendix § 15 .§ 5 ).
Theorem 2 (Bilinear restriction). Under a single active channel with nonzero retained-mode coefficients,
log ∣ C D , k ( h ) ∣ = log ∣ C D , k i n i t ∣ + ( h − 1 ) log ρ D − κ D ( h − 1 ) λ k , h ≥ 1 , (4) \log|C_{D,k}(h)| = \log|C^{\mathrm{init}}_{D,k}| + (h-1)\log\rho_D - \kappa_D(h-1)\lambda_k,\qquad h \ge1, \tag*{(4)} log ∣ C D , k ( h ) ∣ = log ∣ C D , k init ∣ + ( h − 1 ) log ρ D − κ D ( h − 1 ) λ k , h ≥ 1 , (4) log ∣ C a t t , k ( h ) ∣ = log ∣ C a t t , k i n i t ∣ + h log ρ a t t − κ a t t h λ k , h ≥ 0 , 0 < ρ a t t < 1. (5) \log|C_{\mathrm{att},k}(h)| = \log|C^{\mathrm{init}}_{\mathrm{att},k}| + h\log\rho_{\mathrm{att}} - \kappa_{\mathrm{att}}h\lambda_k,\qquad h \ge0,\ 0 < \rho_{\mathrm{att}} < 1. \tag*{(5)} log ∣ C att , k ( h ) ∣ = log ∣ C att , k init ∣ + h log ρ att − κ att h λ k , h ≥ 0 , 0 < ρ att < 1. (5) Under two active channels, log ∣ C D , k ( h ) + C a t t , k ( h ) ∣ \log|C_{D,k}(h) + C_{\mathrm{att},k}(h)| log ∣ C D , k ( h ) + C att , k ( h ) ∣ generally violates this form.
With one channel, each mode’s log response magnitude follows a straight line in the horizon from its first response onward, with slope log ρ m − κ m λ k \log\rho_m - \kappa_m\lambda_k log ρ m − κ m λ k , so rougher modes decay faster at a rate set by κ m \kappa_m κ m . When both channels contribute to a mode with different rates, they generally bend its profile (Appendix § 12.8 .§ 3 ).
Proposition 1 (What modal decay identifies). Suppose one channel m m m is active, ρ m > 0 \rho_m > 0 ρ m > 0 , and two modes with distinct eigenvalues have nonzero responses observed at two distinct horizons each. Call such modes and their eigenvalues informative. Then horizon variation identifies their decay factors, and
κ m = − log r m , k − log r m , ℓ λ k − λ ℓ , ρ m = r m , k e κ m λ k . (6) \kappa_m = -\frac{\log r_{m,k} - \log r_{m,\ell}}{\lambda_k - \lambda_\ell},\qquad\rho_m = r_{m,k}e^{\kappa_m\lambda_k}. \tag*{(6)} κ m = − λ k − λ ℓ log r m , k − log r m , ℓ , ρ m = r m , k e κ m λ k . (6) With only one distinct informative eigenvalue, ( ρ m , κ m ) (\rho_m,\kappa_m) ( ρ m , κ m ) cannot be separated. At λ k = 0 \lambda_k = 0 λ k = 0 only ρ m \rho_m ρ m is identified. Two close eigenvalues can make otherwise identified parameters weakly determined.
Separating two channels is a different problem, which weakens when their rates nearly coincide. § 5 applies these results to the storm responses, where the measurement of spending adds its own restriction.
Aggregation to states
I next consider responses observed only as state aggregates. County responses of opposite signs can cancel within a state, even when their financially weighted sum is nonzero. The rough mode in Panel B of Figure 1 , for example, nearly cancels within states. I ask when a few modes still approximate the state response.
Let M ∈ R n S × n C M \in\mathbb{R}^{n_S \times n_C} M ∈ R n S × n C map county values into the values of n S n_S n S states, and let R ∈ R n C × n S R \in\mathbb{R}^{n_C \times n_S} R ∈ R n C × n S satisfy M R = I d n S MR = \mathrm{Id}_{n_S} M R = Id n S . The matrix R R R assigns a specified within-state county profile to each state shock. The state-level network operator is O ~ h ( S ) ≡ M O ~ h S C I R \widetilde{O}^{(S)}_h \equiv M\widetilde{O}^{SCI}_hR O h ( S ) ≡ M O h S C I R , and its K K K -mode truncation is O ~ h , K ( S ) ≡ M U K diag ( C k ( h ) ) k ≤ K U K ′ R \widetilde{O}^{(S)}_{h,K} \equiv MU_K\operatorname{diag}(C_k(h))_{k\le K}U_K'R O h , K ( S ) ≡ M U K diag ( C k ( h ) ) k ≤ K U K ′ R .
Proposition 2 (Coarse-graining bound). Under Assumption 2, with ∥ ⋅ ∥ 2 \lVert\cdot\rVert_2 ∥ ⋅ ∥ 2 the spectral norm,
∥ O ~ h ( S ) − O ~ h , K ( S ) ∥ 2 ≤ ∥ M ∥ 2 ∥ R ∥ 2 max k > K ∣ C k ( h ) ∣ + ∥ E h , K ( S ) ∥ 2 , E h , K ( S ) ≡ M E h , K R , (7) \left\lVert\widetilde{O}^{(S)}_h-\widetilde{O}^{(S)}_{h,K}\right\rVert_2 \le\lVert M\rVert_2\lVert R\rVert_2\max_{k>K}|C_k(h)|+\left\lVert\mathcal{E}^{(S)}_{h,K}\right\rVert_2,\qquad\mathcal{E}^{(S)}_{h,K}\equiv M E_{h,K}R, \tag*{(7)} O h ( S ) − O h , K ( S ) 2 ≤ ∥ M ∥ 2 ∥ R ∥ 2 k > K max ∣ C k ( h ) ∣ + E h , K ( S ) 2 , E h , K ( S ) ≡ M E h , K R , (7) where the maximum is zero when K = n C K=n_C K = n C . In the exact diagonal benchmark, the remainder E h , K ( S ) \mathcal{E}^{(S)}_{h,K} E h , K ( S ) vanishes.
The first term is the largest response among the discarded modes, scaled by how much aggregation and allocation can amplify it. The second is the aggregated off-diagonal remainder. Assumption 2 keeps the part of the remainder that involves retained modes small, while the block among discarded modes stays unrestricted. A few modes therefore suffice when both terms in the bound are small.
The bound covers shocks whose within-state pattern is the one R R R assigns. For an actual county shock, the approximation error also depends on its within-state distribution. The additional response-error term is M O ~ h S C I ( I d n C − R M ) ε M\widetilde{O}^{SCI}_{h}(\mathrm{Id}_{n_C}-RM)\varepsilon M O h S C I ( Id n C − R M ) ε . Propagation also reshapes the within-state pattern, so repeating the state-level transition M A D R M A_D R M A D R need not reproduce the aggregated county path. After two periods, the gap is
M A D 2 R − ( M A D R ) 2 = M A D ( I d n C − R M ) A D R , M A_D^2 R-(M A_D R)^2=M A_D(\mathrm{Id}_{n_C}-RM)A_D R, M A D 2 R − ( M A D R ) 2 = M A D ( Id n C − R M ) A D R ,
and Appendix § 12.8 .12 gives the full decomposition. The same logic applies to a graph restricted to sampled counties, since omitted counties carry transmission paths as well as outcomes.
From propagation to credit prices
This section derives a conditional, within-period equilibrium, taking intermediary capital and payoff moments as given. It identifies the demand weights and projection of the slow operator that determine the credit-premium response. This benchmark isolates the slow household channel. Let z t ≡ − d t z_t\equiv-d_t z t ≡ − d t , so that positive values represent adverse household conditions. With B D ≡ − η D Φ 0 B_D\equiv-\eta_D\Phi_0 B D ≡ − η D Φ 0 ,
z t + 1 = A D z t + B D ε t . (8) z_{t+1}=A_Dz_t+B_D\varepsilon_t. \tag*{(8)} z t + 1 = A D z t + B D ε t . (8)
Household demand
Let P P P be the dollar price of one credit claim, V V V its next-period dollar payoff, R f R_f R f the gross risk-free rate and q ≡ E [ V ] − R f P q\equiv E[V]-R_fP q ≡ E [ V ] − R f P the claim’s expected excess dollar payoff, which is the premium. Participating households in county i i i have the local interior objective q x i − β i z i x i − ( x i − x ˉ i ) 2 / ( 2 ζ i ) qx_i-\beta_i z_i x_i-(x_i-\bar{x}_i)^2/(2\zeta_i) q x i − β i z i x i − ( x i − x ˉ i ) 2 / ( 2 ζ i ) for holdings x i x_i x i , with premium sensitivity ζ i > 0 \zeta_i>0 ζ i > 0 and sensitivity β i \beta_i β i to the local field. This sensitivity can have either sign, and the resource-loss benchmark below has β i ≥ 0 \beta_i\ge0 β i ≥ 0 . The first-order condition gives x i = x ˉ i + ζ i q − ζ i β i z i x_i=\bar{x}_i+\zeta_iq-\zeta_i\beta_i z_i x i = x ˉ i + ζ i q − ζ i β i z i , so x ˉ i \bar{x}_i x ˉ i is the holding at a zero premium and neutral field. Define
ζ ≡ ∑ i ζ i , X ˉ ≡ ∑ i x ˉ i , l i ≡ ζ i β i , δ ≡ ∑ i l i , S ≡ l ′ z , (9) \zeta\equiv\sum_i\zeta_i,\qquad\bar{X}\equiv\sum_i\bar{x}_i,\qquad l_i\equiv\zeta_i\beta_i,\qquad\delta\equiv\sum_i l_i,\qquad S\equiv l'z, \tag*{(9)} ζ ≡ i ∑ ζ i , X ˉ ≡ i ∑ x ˉ i , l i ≡ ζ i β i , δ ≡ i ∑ l i , S ≡ l ′ z , (9)
so that aggregate household holdings are X = X ˉ + ζ q − S X=\bar{X}+\zeta q-S X = X ˉ + ζ q − S . Here x i x_i x i , X X X , S S S are quantities of claims, and S S S is the desired reduction in holdings at a fixed premium, not executed sales. A negative S S S is a desired increase. Multiplying by a reference price P ˉ \bar{P} P ˉ converts the shift into dollars. The financial loading is precisely l i = − ∂ x i / ∂ z i ∣ q l_i=-\partial x_i/\partial z_i|_q l i = − ∂ x i / ∂ z i ∣ q . It weights marginal responses, not simply who is wealthy or owns bonds. When δ ≠ 0 \delta\ne0 δ = 0 , the normalized weights are π i H ≡ l i / δ \pi_i^H\equiv l_i/\delta π i H ≡ l i / δ , so that S = δ ( π H ) ′ z S=\delta(\pi^H)'z S = δ ( π H ) ′ z . In the benchmark, δ > 0 \delta>0 δ > 0 and π H \pi^H π H is a distribution over counties. Loadings of opposite sign can sum to zero while S S S still responds to a geographically uneven field. The vector l l l is therefore the general object, and displays written with π H \pi^H π H assume δ ≠ 0 \delta\ne0 δ = 0 .
Where do the weights then come from? Let s i > 0 s_i > 0 s i > 0 be households’ surplus wealth above the present value of habit or subsistence commitments. In a local mean–variance approximation with risk-aversion parameter γ i > 0 \gamma_i > 0 γ i > 0 , absolute risk aversion γ i / s i \gamma_i/s_i γ i / s i and claim-payoff variance σ i 2 \sigma_i^2 σ i 2 , claim demand is x i = s i q / ( γ i σ i 2 ) x_i = s_iq/(\gamma_i\sigma_i^2) x i = s i q / ( γ i σ i 2 ) . Suppose the adverse field lowers surplus from its baseline s ˉ i \bar{s}_i s ˉ i , so s i = s ˉ i − m i z i s_i = \bar{s}_i - m_i z_i s i = s ˉ i − m i z i with m i ≥ 0 m_i \ge0 m i ≥ 0 . Linearizing at a premium q ˉ > 0 \bar{q} > 0 q ˉ > 0 gives
ζ i = s ˉ i γ i σ i 2 , β i = q ˉ m i s ˉ i , ζ i β i = q ˉ m i γ i σ i 2 . (10) \zeta_i = \frac{\bar{s}_i}{\gamma_i \sigma_i^2}, \qquad\beta_i = \frac{\bar{q}m_i}{\bar{s}_i}, \qquad\zeta_i\beta_i = \frac{\bar{q}m_i}{\gamma_i\sigma_i^2}. \tag*{(10)} ζ i = γ i σ i 2 s ˉ i , β i = s ˉ i q ˉ m i , ζ i β i = γ i σ i 2 q ˉ m i . (10)
A smaller surplus cushion raises local risk aversion and lowers desired risky holdings. This is a reduced-form surplus-wealth illustration inspired by the habit mechanism of Campbell and Cochrane [7 ] , not a derivation of that model county by county. Within this local approximation l i = q ˉ m i / ( γ i σ i 2 ) l_i = \bar{q}m_i/(\gamma_i\sigma_i^2) l i = q ˉ m i / ( γ i σ i 2 ) , so households with the same holdings can have different marginal responses. Nonparticipation is a separate restriction l i = 0 l_i = 0 l i = 0 . Taking s ˉ i → 0 \bar{s}_i \to0 s ˉ i → 0 in the interior formula does not produce it, because β i \beta_i β i then diverges. Participation matters [28 , 31 ] , but participation data alone do not identify m i m_i m i , risk tolerance or payoff exposure. For any proxy w w w , ( π H ) ′ z = w ′ z + ( π H − w ) ′ z (\pi^H)^\prime z = w^\prime z + (\pi^H - w)^\prime z ( π H ) ′ z = w ′ z + ( π H − w ) ′ z , so a small spending-weighted average need not mean a small demand shift.
Selling claims is also only one way to absorb a loss. With positive quantities for the named flows, a household’s financing identity is
S a l e s = L Y − L C − I t r − B n e w − D c a s h . (11) \mathrm{Sales} = L_Y - L_C - I_{tr} - B_{new} - D_{cash}. \tag*{(11)} Sales = L Y − L C − I t r − B n e w − D c a s h . (11)
Here lost income L Y L_Y L Y is met by cutting consumption (L C L_C L C ), receiving insurance or transfers (I t r I_{tr} I t r ), borrowing (B n e w B_{new} B n e w ), drawing down cash (D c a s h D_{cash} D c a s h ) or selling assets. A spending cut therefore substitutes for liquidity-driven asset sales, holding the other financing sources fixed. These dollar flows differ from the claim-quantity shift S S S .
Loss of resources and precaution against background risk are two forces behind S S S . In the resource-loss benchmark, credit demand rises with available resources, so lost income, current transfers or future assistance commitments lower desired holdings. A current transfer lowers spending by less than the amount sent. Precaution can reinforce this reduction or offset it. Appendix § 16.4 .§ 1 derives both directions for a household with a safe asset, credit and equity. Under its three-state conditions, a small independent increase in income risk tilts the household away from the claim whose losses are concentrated in the worst state. It therefore lowers holdings of credit that pays in full outside that state, whereas credit holdings can rise with other payoff profiles or at a portfolio boundary.3 Lower current commitments can instead reduce spending while raising surplus and credit demand. Transfers redistribute resources, so net selling pressure requires donors’ reductions in desired holdings to exceed recipients’ increases. Thus the net loading l i l_i l i can have either sign, and its sign cannot be inferred from a spending decline.
Intermediary absorption
Intermediaries have capital N > 0 N > 0 N > 0 , post margin χ > 0 \chi> 0 χ > 0 per claim and bear a risk cost Γ > 0 \Gamma> 0 Γ > 0 . They choose holdings θ \theta θ to maximize θ q − Γ θ 2 / ( 2 N ) \theta q - \Gamma\theta^{2}/(2N) θ q − Γ θ 2 / ( 2 N ) subject to 0 ≤ θ ≤ N / χ 0 \le\theta\le N/\chi 0 ≤ θ ≤ N / χ . With outstanding supply B ˉ \bar{B} B ˉ , the net supply that households do not hold at q = 0 q = 0 q = 0 is Q ≡ B ˉ − X ˉ + S > 0 Q \equiv\bar{B} - \bar{X} + S > 0 Q ≡ B ˉ − X ˉ + S > 0 , and clearing requires ζ q + θ = Q \zeta q + \theta= Q ζ q + θ = Q . The domain also requires E [ V ] > q E[V] > q E [ V ] > q for a positive claim price.
Proposition 3 (Conditional equilibrium). The clearing premium is
q = G ( S , N ) = { Γ Q / ( N + ζ Γ ) , Q ≤ ( N + ζ Γ ) / χ ( slack ) , ( Q − N / χ ) / ζ , Q ≥ ( N + ζ Γ ) / χ ( binding ) . (12) q = G(S,N) =
\begin{cases}
\Gamma Q/(N+\zeta\Gamma), & Q \le(N+\zeta\Gamma)/\chi\quad(\text{slack}), \\
(Q-N/\chi)/\zeta, & Q \ge(N+\zeta\Gamma)/\chi\quad(\text{binding}).
\end{cases} \tag*{(12)} q = G ( S , N ) = { Γ Q / ( N + ζ Γ ) , ( Q − N / χ ) / ζ , Q ≤ ( N + ζ Γ ) / χ ( slack ) , Q ≥ ( N + ζ Γ ) / χ ( binding ) . (12) Strictly inside the slack region, G S = Γ / ( N + ζ Γ ) > 0 G_S = \Gamma/(N+\zeta\Gamma) > 0 G S = Γ/ ( N + ζ Γ ) > 0 , G N = − Γ Q / ( N + ζ Γ ) 2 < 0 G_N = -\Gamma Q/(N+\zeta\Gamma)^2 < 0 G N = − Γ Q / ( N + ζ Γ ) 2 < 0 and G S N = − Γ / ( N + ζ Γ ) 2 < 0 G_{SN} = -\Gamma/(N+\zeta\Gamma)^2 < 0 G S N = − Γ/ ( N + ζ Γ ) 2 < 0 . Strictly inside the binding region, G S = 1 / ζ G_S = 1/\zeta G S = 1/ ζ , G N = − 1 / ( ζ χ ) G_N = -1/(\zeta\chi) G N = − 1/ ( ζ χ ) and G S N = 0 G_{SN} = 0 G S N = 0 .
G S G_S G S is the price impact of the household demand shift, measured as the rise in the premium. In the slack region it is larger when capital is small, because intermediaries with less capital to bear inventory risk require more compensation for each claim they absorb, as He and Krishnamurthy (2013). When the margin cap binds, intermediaries absorb no additional claims, so the premium must rise by 1 / ζ 1/\zeta 1/ ζ per unit of the shift for households collectively to retain them. The premium is continuous at the boundary, but G S G_S G S jumps there by the factor ( N + ζ Γ ) / ( ζ Γ ) = 1 + N / ( ζ Γ ) (N+\zeta\Gamma)/(\zeta\Gamma) = 1 + N/(\zeta\Gamma) ( N + ζ Γ ) / ( ζ Γ ) = 1 + N / ( ζ Γ ) , which is larger when intermediaries normally absorb a large share of the shift. There is no two-sided derivative at the kink.
Clearing also shows why the fixed-price shift and executed selling differ, since
− d X d S = d θ d S = 1 − ζ G S = { N / ( N + ζ Γ ) , slack , 0 , binding . (13) -\frac{dX}{dS} = \frac{d\theta}{dS} = 1-\zeta G_S =
\begin{cases}
N/(N+\zeta\Gamma), & \text{slack}, \\
0, & \text{binding}.
\end{cases} \tag*{(13)} − d S d X = d S d θ = 1 − ζ G S = { N / ( N + ζ Γ ) , 0 , slack , binding . (13)
In the slack region intermediaries absorb part of the shift, while the higher premium induces households collectively to retain the rest. At the cap, aggregate household holdings do not fall further, although their composition can change. Fund redemptions and changes in security holdings therefore differ from the fixed-price shift. A muted price response alone does not establish intermediary absorption. Similarly, a small observed flow alone does not identify a small fixed-price shift. With B ˉ \bar{B} B ˉ , X ˉ \bar{X} X ˉ and ζ \zeta ζ fixed, clearing gives d S = ζ d q + d θ dS = \zeta\,dq + d\theta d S = ζ d q + d θ , so recovering the shift requires premium changes, aggregate executed flows and an estimate of ζ \zeta ζ .
The pricing kernel
The EBP is the part of corporate bond spreads not explained by expected default ([15 , 10 ] ). I link it to the model’s premium with an assumed local measurement relation, E t B P = α E + b E q t + u t E E^{BP}_t = \alpha_E + b_E q_t + u^E_t E t B P = α E + b E q t + u t E with b E > 0 b_E > 0 b E > 0 . Holding capital, payoff moments, demand parameters and u E u^E u E fixed within one smooth regime,
∂ E t + h B P ∂ ε t ′ = b E G S l ′ A D h − 1 B D , h ≥ 1. (14) \frac{\partial E^{\mathrm{BP}}_{t+h}}{\partial\varepsilon'_t}=b_EG_Sl'A_D^{h-1}B_D,\qquad h\geq1. \tag*{(14)} ∂ ε t ′ ∂ E t + h BP = b E G S l ′ A D h − 1 B D , h ≥ 1. (14)
From right to left, B D B_D B D maps the storm into local conditions and A D h − 1 A_D^{h-1} A D h − 1 spreads that response over the social graph. The loadings l l l then form a signed weighted sum of the propagated field, which is the fixed-price shift in claim demand. When δ ≠ 0 \delta\ne0 δ = 0 this sum equals δ ( π H ) ′ A D h − 1 B D \delta(\pi^H)'A_D^{h-1}B_D δ ( π H ) ′ A D h − 1 B D . With nonnegative loadings, π H \pi^H π H forms a loading-weighted average that the total sensitivity δ \delta δ scales into the demand shift. In turn, G S G_S G S converts that shift into a change in the premium, which b E b_E b E maps into the EBP. The scalar multiplier b E G S b_EG_S b E G S scales the row vector l ′ A D h − 1 B D l'A_D^{h-1}B_D l ′ A D h − 1 B D , which carries the network geometry.
The total response also contains capital dynamics, b E G N ∂ N t + h / ∂ ε t ′ b_EG_N\partial N_{t+h}/\partial\varepsilon'_t b E G N ∂ N t + h / ∂ ε t ′ , which can offset or reinforce the kernel (Appendix § 13 .§ 5 ). If attention shifts desired holdings as well, the kernel gains a term in A a t t h B A_{\mathrm{att}}^hB A att h B , and the nonseparability argument of Theorem 1 extends to channel-specific financial loadings under its modal and nondegeneracy conditions (Appendix § 12.8 .10). Because the EBP is an estimated spread residual, not a pure observation of this model’s q q q , an empirical test must distinguish the kernel from the measurement error u E u^E u E , expected default, liquidity and other risk-premium channels. Bank-capital evidence can inform absorption capacity or an alternative transmission route, but cannot substitute for the missing household loading.
Levels and changes
The kernel is a level effect. The monthly regressions instead relate changes in the EBP to lagged states and to their interaction with intermediary tightness, and the two can have opposite signs. A positive effect on the EBP level can shrink over time, so its change is negative. Let T ≡ N ‾ − N \mathcal{T}\equiv\overline{N}-N T ≡ N − N be the capital shortfall relative to a reference level N ‾ \overline{N} N , and let C N ≡ N ‾ + ζ Γ C_N\equiv\overline{N}+\zeta\Gamma C N ≡ N + ζ Γ , Q 0 ≡ B ‾ − X ‾ Q_0\equiv\overline{B}-\overline{X} Q 0 ≡ B − X and q 0 ≡ Γ Q 0 / C N q_0\equiv\Gamma Q_0/C_N q 0 ≡ Γ Q 0 / C N , with the reference point strictly inside the slack region.
Proposition 4 (Levels versus changes). Strictly inside the slack region,
q = q 0 + A S S + A T T + A S T S T + A T T T 2 + O ( ∥ ( S , T ) ∥ 3 ) , (15) q=q_0+A_SS+A_{\mathcal{T}}\mathcal{T}+A_{S\mathcal{T}}S\mathcal{T}+A_{\mathcal{T}\mathcal{T}}\mathcal{T}^2+O(\lVert(S,\mathcal{T})\rVert^3), \tag*{(15)} q = q 0 + A S S + A T T + A S T S T + A TT T 2 + O (∥( S , T ) ∥ 3 ) , (15) with A S = Γ / C N A_S=\Gamma/C_N A S = Γ/ C N , A T = Γ Q 0 / C N 2 A_{\mathcal{T}}=\Gamma Q_0/C_N^2 A T = Γ Q 0 / C N 2 , A S T = Γ / C N 2 A_{S\mathcal{T}}=\Gamma/C_N^2 A S T = Γ/ C N 2 and A T T = Γ Q 0 / C N 3 A_{\mathcal{T}\mathcal{T}}=\Gamma Q_0/C_N^3 A TT = Γ Q 0 / C N 3 . Let S t = l ′ z t S_t=l'z_t S t = l ′ z t , z t + 1 = A D z t + η t + 1 z_{t+1}=A_Dz_t+\eta_{t+1} z t + 1 = A D z t + η t + 1 and T t + 1 = ρ K T t + e t + 1 \mathcal{T}_{t+1}=\rho_K\mathcal{T}_t+e_{t+1} T t + 1 = ρ K T t + e t + 1 , where ρ K \rho_K ρ K is the persistence of the shortfall and the innovations are conditionally mean-zero. For this quadratic approximation
E t Δ q t + 1 = A S l ′ ( A D − Id n C ) z t + A T ( ρ K − 1 ) T t + A S T { T t l ′ ( ρ K A D − Id n C ) z t + l ′ Cov t ( η t + 1 , e t + 1 ) } + A T T { ( ρ K 2 − 1 ) T t 2 + Var t ( e t + 1 ) } . (16) \begin{aligned}
E_t\Delta q_{t+1}={}&A_Sl'(A_D-\operatorname{Id}_{n_C})z_t+A_{\mathcal{T}}(\rho_K-1)\mathcal{T}_t\\
&+A_{S\mathcal{T}}\{\mathcal{T}_tl'(\rho_KA_D-\operatorname{Id}_{n_C})z_t+l'\operatorname{Cov}_t(\eta_{t+1},e_{t+1})\}\\
&+A_{\mathcal{T}\mathcal{T}}\{(\rho_K^2-1)\mathcal{T}_t^2+\operatorname{Var}_t(e_{t+1})\}. \tag*{(16)}
\end{aligned} E t Δ q t + 1 = A S l ′ ( A D − Id n C ) z t + A T ( ρ K − 1 ) T t + A S T { T t l ′ ( ρ K A D − Id n C ) z t + l ′ Cov t ( η t + 1 , e t + 1 )} + A TT {( ρ K 2 − 1 ) T t 2 + Var t ( e t + 1 )} . (16)
Within this primitive linear-demand model, scarcer capital amplifies the premium response to a given demand shift in the slack regime. With fixed capital, nonnegative financial loadings and B D ε ≥ 0 B_D\varepsilon\geq0 B D ε ≥ 0 , the level response is also nonnegative, although its difference from the no-diffusion response need not be. For changes, a scalar mean-reversion interpretation additionally requires l ′ A D = r l ′ l^{\prime}A_D = rl^{\prime} l ′ A D = r l ′ , so that the aggregate shift follows S t + 1 = r S t + l ′ η t + 1 S_{t+1} = rS_t + l^{\prime}\eta_{t+1} S t + 1 = r S t + l ′ η t + 1 , even when county responses combine several decay rates. This restriction is not generic, however, and the vector expression supplies no sign theorem for the monthly partial regression coefficient. Capital dynamics, regime crossings, payoff news and observation timing require separate restrictions (Appendices § 12.8 .§ 7 and B.5).
What aggregation can reveal
Three results describe what household measurements and credit prices can reveal about (14) .
Aggregation can remove sensitivity to diffusion. Because A D A_D A D is symmetric, A D = U diag ( r D , k ) U ′ A_D = U\operatorname{diag}(r_{D,k})U^{\prime} A D = U diag ( r D , k ) U ′ , and the stationary mode decays at r D , 1 = ρ D r_{D,1} = \rho_D r D , 1 = ρ D . Hence
( π H ) ′ A D h − 1 B D = ( ( π H ) ′ u 1 ) ρ D h − 1 u 1 ′ B D ⏟ filtered incidence + ∑ k ≥ 2 ( ( π H ) ′ u k ) r D , k h − 1 u k ′ B D ⏟ nonstationary exposure . (17) (\pi^H)^{\prime}A_D^{h-1}B_D =
\underbrace{\left((\pi^H)^{\prime}u_1\right)\rho_D^{h-1}u_1^{\prime}B_D}_{\text{filtered incidence}}
+
\underbrace{\sum_{k\geq2}\left((\pi^H)^{\prime}u_k\right)r_{D,k}^{h-1}u_k^{\prime}B_D}_{\text{nonstationary exposure}}. \tag*{(17)} ( π H ) ′ A D h − 1 B D = filtered incidence ( ( π H ) ′ u 1 ) ρ D h − 1 u 1 ′ B D + nonstationary exposure k ≥ 2 ∑ ( ( π H ) ′ u k ) r D , k h − 1 u k ′ B D . (17)
The first term weights the initial field input B D ε B_D\varepsilon B D ε by square-root degree and decays at ρ D \rho_D ρ D , independently of the diffusion rate κ D \kappa_D κ D . In contrast, the second term weights nonstationary modes, but is not itself a diffusion effect. It generally remains nonzero when κ D = 0 \kappa_D = 0 κ D = 0 . If π H \pi^H π H were proportional to u 1 u_1 u 1 , the second term would vanish. More generally, any weight vector with w ′ A D = r D , k w ′ w^{\prime}A_D = r_{D,k}w^{\prime} w ′ A D = r D , k w ′ gives w ′ A D h − 1 B D = r D , k h − 1 w ′ B D w^{\prime}A_D^{h-1}B_D = r_{D,k}^{h-1}w^{\prime}B_D w ′ A D h − 1 B D = r D , k h − 1 w ′ B D for every h h h , a filtered input with no trace of the other modes. These are the aggregation weights that restore the single-exposure representation discussed in Theorem 1. The same split holds with l l l in place of π H \pi^H π H , which covers loadings of either sign.
To isolate what diffusion adds, hold B D B_D B D , l l l and c E = b E G S c_E = b_EG_S c E = b E G S fixed. Equation (14) can then be expressed as K h = c E l ′ A D h − 1 B D K_h = c_El^{\prime}A_D^{h-1}B_D K h = c E l ′ A D h − 1 B D . Let K h 0 = c E ρ D h − 1 l ′ B D K_h^0 = c_E\rho_D^{h-1}l^{\prime}B_D K h 0 = c E ρ D h − 1 l ′ B D be the heterogeneous no-diffusion pricing kernel. The diffusion increment is
Δ K h d i f f = K h − K h 0 = c E ρ D h − 1 l ′ [ e − κ D ( h − 1 ) L − I d C ] B D . (18) \Delta K_h^{\mathit{diff}} = K_h - K_h^0 = c_E\rho_D^{h-1}l^{\prime}\left[e^{-\kappa_D(h-1)\mathcal{L}} - I_{d_C}\right]B_D. \tag*{(18)} Δ K h diff = K h − K h 0 = c E ρ D h − 1 l ′ [ e − κ D ( h − 1 ) L − I d C ] B D . (18)
It is zero when κ D = 0 \kappa_D = 0 κ D = 0 and when h = 1 h = 1 h = 1 . Setting κ D = 0 \kappa_D = 0 κ D = 0 removes dynamic diffusion, not any nonlocal mechanism already built into B D B_D B D . On a two-county graph with eigenvalues 0, 2, let π H = ( p , 1 − p ) ′ \pi^H = (p, 1-p)^{\prime} π H = ( p , 1 − p ) ′ , B D = I d 2 B_D = I_{d_2} B D = I d 2 and normalize c E δ = 1 c_E\delta= 1 c E δ = 1 . The responses to unit inputs are
K 1 , 2 ( h ) = ρ D h − 1 2 [ 1 ± ( 2 p − 1 ) e − 2 κ D ( h − 1 ) ] . (19) K_{1,2}(h) = \frac{\rho_D^{h-1}}{2}\left[1 \pm(2p-1)e^{-2\kappa_D(h-1)}\right]. \tag*{(19)} K 1 , 2 ( h ) = 2 ρ D h − 1 [ 1 ± ( 2 p − 1 ) e − 2 κ D ( h − 1 ) ] . (19)
At p = 0.8 p = 0.8 p = 0.8 , the initial 0.8 versus 0.2 contrast is financial heterogeneity before diffusion, which diffusion subsequently narrows. A test must distinguish that change from the heterogeneous no-diffusion response, not merely control stationary incidence.
Within-event contrasts leave the common response unidentified. Let the county responses to event e e e be y e = 1 a e + G h ε e + u e y_e = \mathbf{1}a_e + G_h\varepsilon_e + u_e y e = 1 a e + G h ε e + u e , with response kernel G h G_h G h and an event-level common component a e a_e a e . A within-event comparison uses a contrast T T T with T 1 = 0 T\mathbf{1} = 0 T 1 = 0 , and for every row vector α ′ \alpha^{\prime} α ′ ,
T ( G h + 1 α ′ ) = T G h , w ′ ( G h + 1 α ′ ) = w ′ G h + α ′ if w ′ 1 = 1. (20) T(G_h+\mathbf{1}\alpha')=TG_h,\qquad w'(G_h+\mathbf{1}\alpha')=w'G_h+\alpha'
\quad\text{if }w'\mathbf{1}=1. \tag*{(20)} T ( G h + 1 α ′ ) = T G h , w ′ ( G h + 1 α ′ ) = w ′ G h + α ′ if w ′ 1 = 1. (20)
The common response thus drops out of every within-event contrast but shifts each normalized aggregate one-for-one. This is the missing-intercept problem of regional and network designs [29 , 33 ] , which the network GIV of Chodorow-Reich et al. [9 ] is designed to solve with granular shocks. The common response may include aid or insurance, as well as rebuilding. Unless the propagation law pins down the missing level, point recovery of an aggregate that the common response moves requires an independent anchor identifying the aggregate household response. A bound yields an interval instead. If C Y = I d n C C_Y = I_{d_n C} C Y = I d n C and ker T = span { 1 } \ker T = \operatorname{span}\{\mathbf{1}\} ker T = span { 1 } , l l l lies in row ( T ) \operatorname{row}(T) row ( T ) only if l ′ 1 = 0 l'\mathbf{1}=0 l ′ 1 = 0 . For unrestricted county responses, within-event spending contrasts therefore cannot recover the demand shift whenever the loadings do not sum to zero, however precisely they are estimated. Nonnegative nonzero loadings always fall in this case. Appendix § 12.8 .13 gives sharp bounds conditional on bounds on the common response. § 7 develops a relative-asset prediction that cancels it.
Household measurements recover the pricing functional only under a spanning condition. Let household measurements satisfy y t = − C Y z t + J Y ε t + υ t y_t=-C_Yz_t+J_Y\varepsilon_t+\upsilon_t y t = − C Y z t + J Y ε t + υ t , with signed measurement loading C Y C_Y C Y and direct input map J Y J_Y J Y , and let fixed-price demand pressure be S t = l ′ z t S_t=l'z_t S t = l ′ z t , with loadings l l l of either sign.
Proposition 5 (Observability of the pricing functional). For a given l l l , known C Y C_Y C Y and unrestricted z z z , the functional l ′ z l'z l ′ z is identified from noiseless C Y z C_Yz C Y z if and only if l ∈ row ( C Y ) l\in\operatorname{row}(C_Y) l ∈ row ( C Y ) .
The condition applies to consumption or fixed-price portfolio-demand measurements, separately or stacked together. For spending, the financing identity (11) explains why precision alone cannot reveal the financial response, since spending cuts and asset sales are alternative ways to absorb a loss. Given independently measured l l l , the measurements must retain the combination that changes claim demand. Within-event contrasts T C Y TC_Y T C Y can remove that combination (Appendix § 12.8 .9). § 7 extends this recovery condition to the relevant shock-response space, and § 6 examines the variation retained by specified proxy weights on the sample graph.
Local spending and social exposure
Data and design
The design compares counties outside a storm’s recorded footprint that differ in their social exposure to the affected counties. The county SCI [6 ] supplies the ties, and geography and production links provide competing exposure measures. FEMA incident dates and designated counties link the declarations to NOAA property-damage reports. The common sample contains 48 declarations, 37 onset weeks and 1,574 recipients. The local weights W l o c W^{\mathrm{loc}} W loc retain each recipient’s 100 largest out-of-county SCI-times-source-population links, normalized by row. The later spectral analysis uses a separate 1,590-county graph. Spending coverage is fixed across the local comparisons. Inference also groups declarations by physical episode and calendar block.
Spending is assigned by cardholder residence in the Opportunity Insights tracker ([30 ] ), and its mapping to consumption depends on the type of purchase (Appendix § 12.8 .10). I select Sunday seven-day observations through June 5, 2022 and published weekly observations thereafter, rather than average overlapping daily lookbacks. Let L j , t = log ( 1 + i n d e x j , t ) L_{j,t}=\log(1+\mathrm{index}_{j,t}) L j , t = log ( 1 + index j , t ) . Every retained county–event pair has actual, unclipped levels from week − 13 -13 − 13 through 12, without interpolation. Weeks − 13 -13 − 13 through − 9 -9 − 9 supply predated controls. The main post-event outcome is Y j e h = 100 ( L j , t e + h − L j , t e − 1 ) Y_{jeh}=100(L_{j,t_e+h}-L_{j,t_e-1}) Y j e h = 100 ( L j , t e + h − L j , t e − 1 ) , where week t e t_e t e contains the incident’s start. For leads and a common-baseline path I use week − 9 -9 − 9 instead. Table 1 reports the distribution of the exposure indices and the four-week spending change.
Table 1. Summary statistics in the county–declaration sample. Notes: Exposures are in index points, and spending is in log percentage points from week − 1 -1 − 1 to week 4. All rows use the same recipient–declaration observations across 48 declarations. Dispersion is unconditional.
Let D r e ( 4 ) D_{re}(4) D r e ( 4 ) be recorded damage in eligible source county r r r , dated within the declared incident and assigned no later than the end of week 4. The primary measure uses county-coded reports and their start dates. Let
s r e ( 4 ) = D r e ( 4 ) ∑ q D q e ( 4 ) , E j e S C I = 100 ∑ r W j r l o c s r e ( 4 ) . (21) s_{re}^{(4)}=\frac{D_{re}(4)}{\sum_q D_{qe}(4)},\qquad E_{je}^{SCI}=100\sum_r W_{jr}^{loc}s_{re}^{(4)}. \tag*{(21)} s r e ( 4 ) = ∑ q D q e ( 4 ) D r e ( 4 ) , E j e S C I = 100 r ∑ W j r l oc s r e ( 4 ) . (21)
The index weights links by damage shares, rather than counting the share of friends affected. Normalizing the source vector therefore locates the disturbance rather than preserving the event’s total monetary size. Set s r e ( 4 ) = 0 s_{re}^{(4)}=0 s r e ( 4 ) = 0 when the recorded total is zero. Coverage flags distinguish absent records from assessed zero loss. Appendix § 16.3 compares raw-dollar and income-scaled inputs, retains unknown loss coverage, and dates continuing destruction as additional source injections.
For each horizon, holding the week-4 damage shares fixed, I estimate
Y j e h = α j , h + μ e , h + β h S C I E j e S C I + β h G E j e G + β h u p E j e u p + β h d o w n E j e d o w n + Z j e ′ ϕ h + u j e h . (22) Y_{jeh}=\alpha_{j,h}+\mu_{e,h}+\beta_h^{SCI}E_{je}^{SCI}+\beta_h^{G}E_{je}^{G}+\beta_h^{up}E_{je}^{up}+\beta_h^{down}E_{je}^{down}+Z_{je}^{\prime}\phi_h+u_{jeh}. \tag*{(22)} Y j e h = α j , h + μ e , h + β h S C I E j e S C I + β h G E j e G + β h u p E j e u p + β h d o w n E j e d o w n + Z j e ′ ϕ h + u j e h . (22)
Gravity and payroll-weighted input–output exposures use the same source input. Z Z Z contains distance bins, a same-hit-state indicator, spending controls dated no later than week − 9 -9 − 9 and measured recipient weather in weeks − 9 -9 − 9 to 12 (Appendix § 16.3 ). County effects absorb persistent recipient differences, while event effects absorb common calendar shocks and the storm-wide response. Identification comes from differences among recipients within an event, conditional on competing exposures. Interpreting β S C I \beta^{SCI} β S C I causally requires social ties not to proxy for direct weather, evacuation, shared news or other recipient-specific shocks. Under that exclusion, the within-event comparison identifies a spatial response contrast, the part of propagation that differs across recipients. Recovering its common component is the aggregation task of § 7 . The spatial gradient is a household-response measurement, not a per-dollar source loading or a financial-demand sensitivity.
Spending responses and source construction
The four-week gradient is stable across the loss-record constructions in Table 2 . With incident-aligned county reports in Columns (3)–(4), a one-percentage-point increase in the damage-share-weighted link index is associated with 0.353 percent lower spending. Similarly, the same one-percentage-point increase is associated with 0.325 percent lower spending when NOAA zone reports that map uniquely to eligible source counties are included. Using report end dates instead of start dates gives a corresponding spending decline of 0.332 percent. Both retain the negative gradient under declaration and episode inference.
Table 2. Four-week spending responses across loss-record constructions
Notes: Coefficients are interpreted as the change in spending from week −1 to week 4, in log percentage points, associated with a 1-percentage-point (one-index-point) increase in the damage-share-weighted link index. Controls include own-weather, gravity and production-network measures. Decl. and Episode denote declaration- and physical-episode-clustered wild-bootstrap standard errors using 9,999 Webb draws [32 ] . Columns (1)–(2) use onset-week reports, while the others use incident reports dated through week 4, with matched competing exposures. Calendar-block checks preserve the negative signs (Appendix § 16.3 ). *p < 0.10 p < 0.10 p < 0.10 , **p < 0.05 p < 0.05 p < 0.05 , ***p < 0.01 p < 0.01 p < 0.01 .
All 48 leave-one-declaration-out estimates are negative, ranging from −0.404 to −0.309 percent per index point. Restricting onsets to December 2020 onward gives −0.319 percent, while omitting own-weather controls on the same full sample gives −0.312 percent. Thus the four-week result survives both a change in the observation period and recipient-weather adjustment. The immediate pre-event comparisons show little difference in spending changes by social exposure. Appendix § 16.3 reports the corresponding source-sensitivity results, the full event-time profile and the joint lead tests.
Table 3 dates repeated damage inputs explicitly. Columns (5)–(8) exclude counties with their own weather report at the starting week, asking how households respond when damage arrives through their connections while their own county has no recorded weather event that week. In this cohort, the four-week gradient is −0.457 percent without controls for subsequent damage and −0.441 percent with them. The four-week gradient and joint weekly path remain significant under both calendar-block checks. Following counties through later hits therefore preserves the local pattern while addressing additional shock arrivals. Appendix § 16.3 gives the weekly and accumulated paths and the conditions needed to interpret future-input controls as isolating a causal pulse. Taken together, the spending gradient is stable across loss records and after accounting for later hits. This relationship between disruptions in socially connected places and local spending is the empirical starting point for the propagation analysis.
Table 3. Four-week spending responses with dated, repeated damage inputs
Notes: Coefficients are interpreted as the change in spending from week −1 to week 4, in log percentage points, associated with a 1-percentage-point increase in the damage-share-weighted link index. SE types denote calendar-block lengths for the wild bootstrap using 9,999 Webb draws ([32 ] ). All models include county and week effects, own weather, competing inputs, recent spending and damage history. Future damage controls cover weeks 1–12. The origin cohort retains counties hit later. Joint tests cover weeks 0–12. Appendix § 16.3 reports the origin-cohort paths. * p p p < 0.10, ** p p p < 0.05, *** p p p < 0.01.
Notes: Coefficients are interpreted as changes in the monthly spending index, in percentage points, associated with a one-standard-deviation increase in exposure. Exposures are standardized within each horizon’s sample. Social exposure uses source states at least 300 km away and is residualized against production, distance and home-state payments within each round. All regressions include county and month effects. Two-way standard errors cluster by county and calendar month. Sample moments and round leave-outs are in Appendix § 14 .§ 4 . * p p p < 0.10, ** p p p < 0.05, *** p p p < 0.01.
A complementary test asks whether spending rises when connected places receive cash. Economic Impact Payments were federal stimulus checks paid in three rounds, aligned here to April 2020, January 2021 and March 2021. I map payments per resident in distant source states through SCILinks, separating this exposure from home-state payments while accounting for geography and production links. Table 4 therefore asks whether connections to other places receiving transfers predict spending beyond the county’s own payment exposure.
Table 4. Spending and social exposure to federal stimulus payments
A one-standard-deviation increase in residual social exposure is associated with a spending index 0.386 percentage points higher in the payment month and 0.359 points higher one month later. Both estimates remain positive and significant when any payment round is omitted (Appendix § 14 .§ 4 ). With three rounds, identification requires separating remote exposure from concurrent policy and pandemic channels. Spending falls with social exposure to storm damage and rises with social exposure to payments, in line with resources lost or received in the source places. Both exercises supply household-response measurements on the social geometry, while allowing each input its own response strength and timing. § 7 shows how those measurements, combined with financial sensitivities, yield credit-price predictions. Appendix § 16.3 examines monetary storm loadings, and Appendix § 14 .§ 6 reports the historical wage comparison.
From local exposure to dynamic responses
Estimating responses across modes and horizons
The local regressions measure how spending varies with social exposure. The operator also describes how this response differs across the network’s spatial patterns and evolves over time. With a single active channel, Theorem 2 makes each latent mode’s log response magnitude linear in the horizon, with a slope that falls with graph frequency.
Applying this to spending requires an observation restriction. Here B D = − η D Φ 0 B_D=-\eta_D\Phi_0 B D = − η D Φ 0 maps shocks into the adverse field z = − d z=-d z = − d , and C Y C_Y C Y maps that field into measured spending losses (§ 3.5 ). The slow-field component of the spending-loss level response is H h L = C Y A D h − 1 B D H_h^L=C_YA_D^{h-1}B_D H h L = C Y A D h − 1 B D for h ≥ 1 h\ge1 h ≥ 1 , with H 0 L = 0 H_0^L=0 H 0 L = 0 . Its first differences are H 1 Δ = C Y B D H_1^\Delta=C_YB_D H 1 Δ = C Y B D and H h Δ = C Y ( A D − Id n C ) A D h − 2 B D H_h^\Delta=C_Y(A_D-\operatorname{Id}_{n_C})A_D^{h-2}B_D H h Δ = C Y ( A D − Id n C ) A D h − 2 B D for h ≥ 2 h\ge2 h ≥ 2 . Before clipping or other outcome transformations, the corresponding spending-growth component, which the regressions below use, is − H h Δ -H_h^\Delta − H h Δ . A decaying level therefore need not produce growth of the same sign. Continued inputs, omitted counties and measurement mixing can also create several apparent decay rates from a single field. Appendix § 12.8 .11 states sufficient conditions under which masked, filtered and residualized outcome coefficients inherit the decay law. These conditions concern the measurement, which Assumption 2 does not cover.
For event e e e , let b e r a w b_e^{raw} b e r a w contain in-support county shares of its onset-week recorded damage and E j k e = u j k ( u k ′ b e r a w ) E_{jke}=u_{jk}(u_k'b_e^{raw}) E j k e = u j k ( u k ′ b e r a w ) . I estimate the nonstationary mode coefficients jointly:
g j , t + h = α j , h + μ e , h + ξ t + h , h + ∑ k ∈ K θ k Y ( h ) E j k e + Z j , e ′ ϕ h + u j , e , h . (23) g_{j,t+h}=\alpha_{j,h}+\mu_{e,h}+\xi_{t+h,h}+\sum_{k\in\mathcal{K}}\theta_k^Y(h)E_{jke}+Z_{j,e}'\phi_h+u_{j,e,h}. \tag*{(23)} g j , t + h = α j , h + μ e , h + ξ t + h , h + k ∈ K ∑ θ k Y ( h ) E j k e + Z j , e ′ ϕ h + u j , e , h . (23)
where K \mathcal{K} K contains 34 nonstationary modes and g g g is clipped weekly log spending growth. The source construction and controls also differ. Estimating responses separately across the retained modes and horizons asks more of the storms than estimating the pooled gradient, and the decay law is not pinned down by these data alone (Appendix § 14.2 ). Credit pricing, however, depends on the financially weighted combination of responses rather than on every modal path. That is why the recovery program targets this combination, using restrictions across horizons and additional household and financial measurements. Modal errors then matter only through the financial loadings l ′ u k l^{\prime}u_k l ′ u k (§ 7.5 ).
What the transport test measures
The transport exercise asks whether responses estimated on some storms predict spending differences across states after other storms. It averages residual weekly spending growth over the observed non-hit counties within each state, after removing county, event and week effects. The exercise builds on the modal responses above rather than on the level comparison of § 4 . Five purged folds estimate response coefficients and select tuning settings on training storms, then evaluate them on held-out storms. Held-out predictive performance is close to the zero-response benchmark for these residual state contrasts (Appendix § 14.2 .§ 3 ). Population-weighted state spending averages and credit demand combine county responses differently. For credit pricing, the relevant weights reflect how strongly households adjust their desired credit holdings. § 6 examines which spatial responses these financial weights retain. § 7 then evaluates recovery of the financially weighted response in simulation, using an aggregate anchor or restrictions on the common response.
Financial geography and the credit-pricing margin
Stationary and nonstationary exposure
Figure 2 shows that diffusion changes the relative importance of individual sources. How much of that variation survives aggregation over storm footprints? For each of the 67 storms with damage inside the 1,590-county support, I compute the weighted aggregate w ′ e − τ L b e = ∑ j b e j I j ( τ ) w^{\prime}e^{-\tau L}b_e=\sum_j b_{ej}I_j(\tau) w ′ e − τ L b e = ∑ j b e j I j ( τ ) , a damage-weighted sum of Figure 2 ’s loadings, using all modes of the retained graph, and compare it with the incidence term ( w ′ u 1 ) ( u 1 ′ b e ) (w^{\prime}u_1)(u_1^{\prime}b_e) ( w ′ u 1 ) ( u 1 ′ b e ) . Here b e b_e b e is normalized to sum to one over that storm’s in-support counties, and input loadings are taken proportional to damage. The factor ρ D h − 1 \rho_D^{h-1} ρ D h − 1 multiplies both terms and cancels, so the comparison needs no estimate of ρ D \rho_D ρ D . Because κ D \kappa_D κ D is not pinned down by these data, Figure 3 ’s horizontal axis measures diffusion time rather than calendar weeks. Appendix § 15 .§ 5 reports source coverage, the dollar-scaled comparison and an implementation check.
Figure 3. Model-implied pressure relative to filtered incidence. Panel A shows the median pressure-to-incidence ratio across storms. Shading gives the interquartile range for dividend weights. Panel B shows the across-storm correlation between pressure and incidence. Eighteen candidate storms have no in-support source mass, and 48 retained footprints are partly covered (Appendix § 15 .§ 5 ).
Before diffusion (τ = 0 \tau= 0 τ = 0 ), dividend-weighted pressure is 0.13 of incidence for the median storm, reflecting the proxy’s low weight on the initial footprint. Under the assumed diffusion, the median ratio rises to 0.56 at τ = 1 \tau= 1 τ = 1 and 0.82 at τ = 4 \tau= 4 τ = 4 (Figure 3 ). For unit in-support footprints, the correlation with incidence rises from 0.81 to 0.84 over that interval. Scaling each footprint by its in-support damage gives correlations of 0.97 before diffusion and 0.99 at τ = 2 \tau= 2 τ = 2 . Population weights and weights confined to the top 5 percent of dividend counties also converge toward incidence (Appendix § 15 .§ 5 ). The calculation shows why financial geography matters. The dividend proxy places financial weight away from the initial disturbance, so diffusion changes weighted exposure even when aggregate pressure is highly correlated with incidence. Measuring marginal demand sensitivities determines which of these spatial responses survives the projection into credit demand.
Monthly EBP associations
A separate monthly analysis summarizes county spending responses by the amplitude of a low-rank SCI state and examines its association with EBP changes as intermediary capital tightens. The regression relates E t + 1 E B P − E t E B P E_{t+1}^{EBP}-E_t^{EBP} E t + 1 E B P − E t E B P to the lagged state, lagged tightness (the negative of the standardized He–Kelly–Manela capital ratio), their product, the analogous production state, macro and COVID controls, and two lags of the outcome, over 49 months of the 2020–2024 sample. With both states standardized, the SCI interaction is +0.116 percentage points of EBP, compared with −0.010 for the production interaction. The positive interaction indicates that the association between the SCI amplitude and EBP changes is stronger when intermediary capital is lower. It also appears in the component of the amplitude not explained by active-storm counts. Because the states are built on the full sample and the specification was selected retrospectively, these are in-sample associations, and their partial coefficients are distinct from Proposition 4’s conditional benchmark (Appendix § 15 ). An aggregate association of this kind cannot by itself separate propagation from common storm timing. In the model, comparing the pricing kernel with its no-diffusion counterpart, holding inputs and other parameters fixed, isolates the increment due to diffusion (18) .
Fund demand and the credit-pricing margin
Fund data let us examine the market end of the household-demand mechanism. Table 5 links quarterly subscriptions to corporate holdings in Panel A and monthly inflows to EBP changes in Panel B. The pricing instrument uses granular fund-flow residuals, controlling for two common factors and two lagged EBP changes.
Table 5. Fund subscriptions, corporate-bond holdings and credit pricing
Panel A asks whether money entering funds is reflected in their corporate-bond portfolios. Each dollar subscribed accompanies $0.635 of approximate corporate-bond purchases, or $0.642 using external cash. At the pricing margin, inflows of 0.01% of nonfinancial corporate debt securities coincide with an EBP decline of 4.06 basis points, or 3.96 using external cash. Together, the two panels document corporate-portfolio adjustment and associated credit repricing at the fund end of the mechanism. Appendices § 15 .§ 8 and § 16.5 .§ 7 state the conditions for a structural reading.
Households change desired holdings at a fixed price, whereas subscriptions measure realized flows after prices adjust. The table records inflows, and the pricing map below is written for selling. From (13) , selling at reference price P ˉ \bar{P} P ˉ satisfies d F s e l l = P ˉ ( 1 − ζ G S ) d S dF^{\mathrm{sell}} = \bar{P}(1-\zeta G_S)dS d F sell = P ˉ ( 1 − ζ G S ) d S . Within a smooth regime with positive absorption, letting ϑ F = d E B P B P / d F s e l l \vartheta_F = dEBP^{BP}/dF^{\mathrm{sell}} ϑ F = d E B P B P / d F sell gives
K h = ϑ F P ˉ ( 1 − ζ G S ) l ′ A D h − 1 B D . (24) K_h = \vartheta_F \bar{P}(1-\zeta G_S)l' A_D^{h-1}B_D. \tag*{(24)} K h = ϑ F P ˉ ( 1 − ζ G S ) l ′ A D h − 1 B D . (24)
The operator and household sensitivities locate desired-demand changes. The executed share and price impact convert them into a premium response. Panel A relates two observed fund quantities. Household positions and response sensitivities connect those quantities to the model’s desired-demand shift.
Notes: Panel A uses approximate return-adjusted corporate-bond purchases and subscriptions, scaled by initial assets, with asset weights, quarter effects and fund-clustered standard errors (495 funds). Panel B reports EBP changes in basis points per inflow of 0.01% of nonfinancial corporate debt securities, using corporate-holdings weights. The pricing instrument follows Gabaix and Koijen (2024) (Appendix § 15 .§ 8 ). HAC (3) uses three-month Bartlett weights, conditional on estimated flow factors. * p < 0.10 p < 0.10 p < 0.10 , ** p < 0.05 p < 0.05 p < 0.05 , *** p < 0.01 p < 0.01 p < 0.01 .
Appendix § 16.5 .§ 7 states the demand-feedback and exclusion restrictions and compares granular and network moments ([9 ] . The 57-month sample leaves the separation of price impact from demand feedback weakly identified [5 ] . Changes in desired holdings generate flows. Cumulating them under static clearing gives (24) , while market persistence adds a temporal filter (Appendix § 16.5 .§ 7 ). § 7.4 demonstrates in simulation how separately estimated household and market responses combine into a credit-price prediction.
Recovering the financially relevant operator
Which household measurements recover the projection that shifts credit demand? The following program targets that conditional contribution, holding capital and payoff moments fixed. A total EBP prediction also needs restrictions on capital and competing channels.
From household responses to a pricing projection
Let the slow field’s impulse response be R h = A D h − 1 B D R_h=A_D^{h-1}B_D R h = A D h − 1 B D , h ≥ 1 h\ge1 h ≥ 1 , and the measured household responses be H h = C ∗ R h H_h=C_*R_h H h = C ∗ R h . The known map C ∗ C_* C ∗ specifies how household conditions enter each measurement, including its units and sign. Spending-loss measurements, i.e., can stack county contrasts T T T and an independent aggregate with weights w w w : C ∗ = ( ( T C Y ) ′ , ( w ′ C Y ) ′ ) ′ C_*=\left((TC_Y)^\prime,(w^\prime C_Y)^\prime\right)^\prime C ∗ = ( ( T C Y ) ′ , ( w ′ C Y ) ′ ) ′ . Consumption and fixed-price portfolio-demand measurements use their own maps and can be combined (Appendix § 12.8 .10). In each case, valid restrictions must separate the field’s response from direct damage, aid, migration and other channels. An aggregate measurement must likewise distinguish the household channel from a competing common response. Thus merely appending national spending doesn’t identify the household channel.
Proposition 6 (Recovery of the pricing projection). Let R \mathcal{R} R contain the field responses to the shocks and horizons being tested, with basis R R R_{\mathcal{R}} R R . Given l l l , C ∗ C_* C ∗ and c E ≡ b E G S > 0 c_E\equiv b_EG_S>0 c E ≡ b E G S > 0 , the target l ′ z l^\prime z l ′ z is recoverable from C ∗ z C_*z C ∗ z for every z ∈ R z\in\mathcal{R} z ∈ R if and only if there is an a a a satisfying l ′ R R = a ′ C ∗ R R l^\prime R_{\mathcal{R}}=a^\prime C_*R_{\mathcal{R}} l ′ R R = a ′ C ∗ R R . In that case the pricing kernel is
K h = c E l ′ R h = c E a ′ H h . (25) K_h=c_El^\prime R_h=c_Ea^\prime H_h. \tag*{(25)} K h = c E l ′ R h = c E a ′ H h . (25)
Under this condition, the pricing projection can be recovered without separately identifying the propagation and input matrices (proof in Appendix § 12.8 .10). This also holds for mixed-sign loadings. The recovered projection is then a net response, which is zero when household responses offset. Recovery weights a a a can be constructed from the measurement map, independently measured l l l and the maintained response subspace. Among valid weights, a precision-aware choice minimizes projected measurement variance. Response estimates supply H h H_h H h , and the weights translate these into changes in desired credit holdings. Market evidence supplies c E c_E c E . Event contrasts need an anchor when the common response affects the target and is not fixed by the propagation law. When household responses are measured as weekly growth, they must first be accumulated consistently. If county dollar losses satisfy y i ∗ = C i Y z i y_i^*=C_i^Y z_i y i ∗ = C i Y z i with known C i Y ≠ 0 C_i^Y\ne0 C i Y = 0 , then a i = l i / C i Y a_i=l_i/C_i^Y a i = l i / C i Y , giving the recovered response the marginal demand weight per dollar lost.
For an approximate representation, define r l = l − C ∗ ′ a r_l=l-C_*^\prime a r l = l − C ∗ ′ a and assume ∥ R h ε ∥ ≤ z ˉ h \lVert R_h\varepsilon\rVert\le\bar{z}_h ∥ R h ε ∥ ≤ z ˉ h . With a a a and c E c_E c E fixed, the forecast p ^ h = c E a ′ H ^ h ε \widehat{p}_h = c_Ea'\widehat{H}_h\varepsilon p h = c E a ′ H h ε obeys
∣ p ^ h − K h ε ∣ ≤ c E ∥ a ∥ ∥ ( H ^ h − H h ) ε ∥ + c E ∥ r l ∥ z ‾ h . (26) \left|\widehat{p}_h-K_h\varepsilon\right| \le c_E\lVert a\rVert\left\lVert\left(\widehat{H}_h-H_h\right)\varepsilon\right\rVert+c_E\lVert r_l\rVert\overline{z}_h. \tag*{(26)} ∣ p h − K h ε ∣ ≤ c E ∥ a ∥ ( H h − H h ) ε + c E ∥ r l ∥ z h . (26)
The first term scales household-response prediction error by the recovery weights. The second captures a financially relevant direction missing from the measurements, for which only the projection of r l r_l r l onto R \mathcal{R} R matters. More precise measurements can reduce the first. Additionally, portfolio-demand measurements can span the missing direction. Estimated a a a and c E c_E c E add a further error term (Appendix § 12.8 .10). § 7.4 estimates household loading scales and the market map from simulated observations.
Measuring the household margin
The financial loading combines changes in investable resources with changes in desired portfolio shares. At fixed asset price P a P_a P a , let W i W_i W i be county households’ investable wealth and s i a s_{ia} s ia their desired share in credit asset a a a , so desired dollar holdings are s i a W i s_{ia}W_i s ia W i . Define χ i W = − ∂ W i / ∂ z i \chi_i^W=-\partial W_i/\partial z_i χ i W = − ∂ W i / ∂ z i and ψ i a p o r t = − ∂ s i a / ∂ z i \psi_{ia}^{\mathrm{port}}=-\partial s_{ia}/\partial z_i ψ ia port = − ∂ s ia / ∂ z i . Differentiating desired dollar holdings gives:
l i a $ ≡ − ∂ ( s i a W i ) ∂ z i = s i a χ i W + W i ψ i a p o r t , l i a = l i a $ P a . (27) l_{ia}^{\$}\equiv-\frac{\partial(s_{ia}W_i)}{\partial z_i}=s_{ia}\chi_i^W+W_i\psi_{ia}^{\mathrm{port}},\qquad l_{ia}=\frac{l_{ia}^{\$}}{P_a}. \tag*{(27)} l ia $ ≡ − ∂ z i ∂ ( s ia W i ) = s ia χ i W + W i ψ ia port , l ia = P a l ia $ . (27)
The first term captures changes in resources, including accumulated saving. The second is portfolio rebalancingin either direction, which can occur without changing current consumption. Spending alone identifies neither margin nor its sign. Account flows and return-adjusted holdings, together with transfers, borrowing and liquid balances, distinguish them (Appendices § 16.4 .§ 1 and § 16.4 .§ 3 ). Financing assistance requires access to resources, whereas rebalancing can occur within an account without a withdrawal. Holdings measure s i a W i s_{ia}W_i s ia W i , not its derivative, so equal holdings can have different responses. Executed trades also respond to prices, requiring an identified demand system or valid price controls.
There is a useful intermediate prediction even when the common household response is missing. After normalization, a response common to all counties moves both assets equally. Their difference therefore retains the effect of their different investor geography. For two assets a a a and b b b whose totals δ a $ = ∑ i l i a $ \delta_a^{\$}=\sum_i l_{ia}^{\$} δ a $ = ∑ i l ia $ and δ b $ \delta_b^{\$} δ b $ are both nonzero, let π a = l a $ / δ a $ \pi_a=l_a^{\$}/\delta_a^{\$} π a = l a $ / δ a $ , and define π b \pi_b π b likewise. Let the market map c a > 0 c_a>0 c a > 0 convert fixed-price dollar demand pressure into a measured premium, net of capital and payoff news. Here Δ p a \Delta p_a Δ p a is the premium-level response relative to the no-shock path at the horizon being studied. Let z ⊥ z_{\perp} z ⊥ collect cross-county differences around the common response α \alpha α . Since both π \pi π ’s sum to one,
Δ p a c a δ a $ − Δ p b c b δ b $ = ( π a − π b ) ′ z = ( π a − π b ) ′ z ⊥ , z = z ⊥ + α 1. (28) \frac{\Delta p_a}{c_a\delta_a^{\$}}-\frac{\Delta p_b}{c_b\delta_b^{\$}}=(\pi_a-\pi_b)'z=(\pi_a-\pi_b)'z_{\perp},\qquad z=z_{\perp}+\alpha\mathbf{1}. \tag*{(28)} c a δ a $ Δ p a − c b δ b $ Δ p b = ( π a − π b ) ′ z = ( π a − π b ) ′ z ⊥ , z = z ⊥ + α 1 . (28)
Thus the common component α \alpha α cancels, although the household measurement map must still span this contrast as in Proposition 6. An asset whose loadings sum to zero needs no such difference, because the common component already cancels from its own demand shift. In the equal-degree two-county example, take π a = ( 0.8 , 0.2 ) ′ \pi_a=(0.8,0.2)' π a = ( 0.8 , 0.2 ) ′ and π b = ( 0.2 , 0.8 ) ′ \pi_b=(0.2,0.8)' π b = ( 0.2 , 0.8 ) ′ . A unit shock in county 1 gives a normalized premium gap 0.6 ρ D h − 1 e − 2 κ D ( h − 1 ) 0.6\rho_D^{h-1}e^{-2\kappa_D(h-1)} 0.6 ρ D h − 1 e − 2 κ D ( h − 1 ) , versus 0.6 ρ D h − 1 0.6\rho_D^{h-1} 0.6 ρ D h − 1 without diffusion. Here the prediction is a shrinking gap, not a ranking reversal. Given independently measured demand sensitivities and market maps, the operator predicts how the normalized premium gap evolves across horizons. An absolute aggregate premium response additionally requires the common household component and its market map.
An operator estimation program tied to that target
The estimation criterion fits household responses while emphasizing the combination that matters for credit demand. Let Y e h ∗ Y_{eh}^{*} Y e h ∗ collect the proposed household measurements, Z e h ϕ h Z_{eh}\phi_h Z e h ϕ h the separately identified competing channels, and H h ( ϑ ) = C ∗ A t r ( ϑ ) h − 1 B i n ( ϑ ) H_h(\vartheta) = C_{*}A_{\mathrm{tr}}(\vartheta)^{h-1}B_{\mathrm{in}}(\vartheta) H h ( ϑ ) = C ∗ A tr ( ϑ ) h − 1 B in ( ϑ ) a parsimonious candidate with transition A t r A_{\mathrm{tr}} A tr and input map B i n B_{\mathrm{in}} B in . Define v e h ( ϑ , ϕ ) = Y e h ∗ − H h ( ϑ ) ε e − Z e h ϕ h v_{eh}(\vartheta,\phi) = Y_{eh}^{*} - H_h(\vartheta)\varepsilon_e - Z_{eh}\phi_h v e h ( ϑ , ϕ ) = Y e h ∗ − H h ( ϑ ) ε e − Z e h ϕ h . Within each training fold, the criterion is
( ϑ ^ , ϕ ^ ) = argmin ϑ ∈ Θ , ϕ ∑ ( e , h ) ∈ t r a i n v e h ′ ( V e h − 1 + ω a a ′ ) v e h + λ P ( ϑ ) , ω ≥ 0. (29) (\widehat{\vartheta},\widehat{\phi}) = \underset{\vartheta\in\Theta,\phi}{\operatorname{argmin}} \sum_{(e,h)\in\mathrm{train}} v_{eh}^{\prime}\left(V_{eh}^{-1}+\omega aa^{\prime}\right)v_{eh}+\lambda\mathcal{P}(\vartheta), \qquad\omega\ge0. \tag*{(29)} ( ϑ , ϕ ) = ϑ ∈ Θ , ϕ argmin ( e , h ) ∈ train ∑ v e h ′ ( V e h − 1 + ω a a ′ ) v e h + λ P ( ϑ ) , ω ≥ 0. (29)
Here V e h V_{eh} V e h is the measurement-error covariance, P \mathcal{P} P regularizes the operator, and Θ \Theta Θ imposes stability and a predeclared low-dimensional input-loading family. For diffusion, 0 < ρ D < 1 0 < \rho_D < 1 0 < ρ D < 1 and κ D ≥ 0 \kappa_D \ge0 κ D ≥ 0 . The V e h − 1 V_{eh}^{-1} V e h − 1 component rewards accurate prediction of household measurements. The ω a a ′ \omega aa^{\prime} ω a a ′ component gives additional weight to errors in the financial combination recovered by (25) , with ω \omega ω putting that loss in compatible units. Set ω = 0 \omega= 0 ω = 0 until a a a , and hence a a a , is independently measured. Proxy choices are sensitivity exercises, and no financial outcome selects the operator’s loadings. Training also chooses regularization. Held-out storms compare the dynamic operator with one-hop on identical shock sources, recipient screens and competing exposures. This focus on the financially relevant response connects to Adão et al. (2025), who evaluate quantitative models against the causal predictions needed for a particular economic question.
Implementation on the sample design
The simulations put the recovery conditions to work on the observed SCI graph and all candidate storm declarations, first with known financial maps and then with estimated ones. They use a separately calibrated damage-intensity input and spending-noise design rather than the estimated event-study or county-calendar coefficients. The target is the financially weighted response of observed recipient households to each held-out storm over one to twelve weeks (Appendix § 16.5 ). The diffusion-kernel operator is estimated on training storms only and projected with county-resolution loadings. Its fit uses the measurement-precision weights of (29) , without the additional financial-loss weight or regularization penalty (ω = λ P = 0 \omega= \lambda_P = 0 ω = λ P = 0 ). A parametric bootstrap that re-estimates it gives pointwise and twelve-week path bands. Aggregate measurements give each training storm’s mean recipient spending response, including both propagation and the common component, which the model separates. Under the simulation’s measurement and exclusion assumptions, county spending contrasts identify the spatial differences governed by the operator, and financial loadings then translate the spatial and common responses into credit-demand pressure.
The anchored design shows that household measurements support recovery. Its simulated common response is proportional to damage, and financial maps come from hypothetical auxiliary measurements, with investment-income geography supplying the assumed, nonnegative loading shape. With known maps, the operator recovers held-out responses with normalized error 0.17, compared with 0.22 for one-hop and 0.30 for the anchor alone (Table 6 , row 1). The common response alone therefore does not exhaust the predictive information in household measurements. A flexible graph polynomial captures much of this information, with error 0.18, but its nominal 90 percent intervals cover 72 percent of true responses, against 90 percent for the operator (Appendix Table I.2).
Table 6. Notes: Error relative to the root mean square of the true response of observed recipient households, horizons 1–12 (0.72 bp with a common response, rows 1–2, and 0.26 bp without, row 3), with 90 percent bands. Without an anchor, the law generates the common response. Row 1 has noise calibrated to the denoised panel and rows 2–3 design-based noise, so the rows differ in more than one factor. Rows 1 and 3 use known loadings and market map, and noise multipliers scale standard deviations. There are 100 replications on fresh draws, with Monte Carlo s.e. at most 0.03.
Without an anchor, the restrictions must also supply the common response. In the more informative design of row 3, with less noise and twice the storms, operator error is 0.32, against 0.71 for the graph polynomial and 0.92 for one-hop. This is the structural route to recovery, in which the law links responses across counties and horizons and thereby determines their common component. It relies on that law generating the entire common response (Appendix § 16.5 .§ 4 examines misspecification).
In Figure 4 , the anchored prediction learns the response level from training-storm aggregates and tracks the true path. The no-anchor prediction instead recovers the path’s shape from the propagation law’s restrictions. Together, these results show that the framework can turn household responses into a financial prediction for a storm outside the estimation sample.
Figure 4. Simulated recovery for the held-out storm with the largest true response in one replication. Panel A uses training-storm anchors and known financial maps, with noise calibrated to the denoised panel (Table 6 , row 1). Panel B uses no anchors, one-quarter of the design-based noise standard deviation and twice as many storms (row 3). Event details are in Appendix § 16.5 .
Figure 5 asks whether this prediction advantage survives when the household and market maps must also be learned. It keeps the first environment of Table 6 and its target, the held-out desired-selling path valued at the conditional market loading and fixed capital. Household loading scales are learned from hypothetical account flows, with the within-group shape still assumed. The market map is learned from an independent simulated fund panel without the held-out storms, using an instrument built from simulated fund flows rather than the true demand innovation. The price impact of executed flows is a scenario set at Table 5 ’s column (1) estimate, not an identified structural value, so the market stage’s precision is conditional on this calibration.
Figure 5. Held-out prediction with estimated financial maps. Bars show mean normalized error across 200 simulated replications. Whiskers give 95 percent Monte Carlo intervals for the mean error. Each error is normalized by its target’s own root mean square over weeks 1–12. Panel B excludes the common response added in the simulation.
The operator retains a prediction advantage when the maps are estimated. At 57 simulated market months, its nominal 90 percent pointwise bands cover 88 percent of true total responses, against 49 percent for one-hop. The largest gain concerns the propagated component, where operator error is about 30 percent lower than one-hop’s. Under the simulated diffusion law, the dynamic operator thus adds predictive information about the propagated component beyond a static exposure measure. With 228 months, its error falls further as the market map is estimated more precisely.4 Empirically, spending panels supply the household responses H h H_h H h ,5 household accounts the financial sensitivities and market data the market map c E c_E c E .
What the operator contributes to the financial prediction
Diffusion affects credit prices by moving household responses across counties whose financial sensitivities differ. In the diffusion-kernel benchmark, r D , k = ρ e − κ D λ k r_{D,k}=\rho\mathrm{e}^{-\kappa_D\lambda_k} r D , k = ρ e − κ D λ k , so
∂ K h ∂ r D = − c E ( h − 1 ) ∑ k ≥ 2 λ k ( l ′ u k ) r D , k h − 1 u k ′ B D . (30) \frac{\partial K_h}{\partial r_D}=-c_E(h-1)\sum_{k\geq2}\lambda_k(l'u_k)r_{D,k}^{h-1}u'_kB_D. \tag*{(30)} ∂ r D ∂ K h = − c E ( h − 1 ) k ≥ 2 ∑ λ k ( l ′ u k ) r D , k h − 1 u k ′ B D . (30)
Only nonstationary modes enter this derivative. Changing the diffusion rate therefore affects prices when shocks excite those modes, financial sensitivities load on them, and the horizon permits propagation. Its effect need not be monotone, since diffusion can move pressure toward or away from marginal investors. Recovery requires precision in the combinations selected by the financial weights l ′ u k l'u_k l ′ u k , not equal precision in every mode (Appendix § 12.8 .10). As § 3 shows, the magnitude depends on how much households change desired credit holdings per dollar of spending shortfall and how readily others absorb that shift (Appendix § 16.4 .§ 3 ). As in § 3.5 , a test of the kernel’s diffusion-specific EBP restriction must separate the increment (18) from the no-diffusion component and competing channels (Appendix § 16.5 .§ 6 ).
Conclusion
The main contribution is a recovery condition for the household-demand component of credit prices. Given independently measured financial sensitivities and a known measurement map, household measurements recover the pricing projection exactly when they span its financial direction on the relevant shock responses. Thus the full propagation system can remain unidentified while its pricing projection is recoverable. Theorem 1 explains why the projection’s relative weights change with the horizon, so a single fixed exposure score generally cannot summarize them. In the empirical application, a percentage point of damage-share-weighted social exposure is associated with about 0.35 percent lower spending four weeks later. Stable across loss records and after accounting for later hits, this gradient gives the propagation analysis a credible starting point. In simulations on the observed network, the operator reduces error on the propagated credit-premium component by about 30 percent relative to one-hop under the specified diffusion law, with household loading scales and the market map learned from simulated accounts and fund flows. The next empirical step is to measure portfolio adjustment alongside spending, replacing proxy geography with measured demand sensitivities and their signs, and adding the aggregate and market measurements that recovery requires.
Appendix
Equation, theorem, table and figure numbers without a letter refer to the main text. Appendix § 12.8 proves the main-text results. Within it, Appendix § 12.8 .10 derives the pricing-projection recovery condition, its error bound and a minimum-variance representation. Appendix § 13 collects model details that the main text states briefly. Appendices § 14 –§ 15 report supplementary evidence. Within them, Appendix § 15 .§ 5 details the shock normalization used in the aggregation exercise and Appendix § 15 .§ 6 presents the municipal fund-flow pilot. Appendix § 16 evaluates a longer spending design using public data and synthetic experiments. Appendix § 16.3 documents recorded-loss inputs, source timing, measured-weather controls and event-level inference. Appendix § 16.4 derives the financial margin and examines the coverage and identifying variation available for stronger weather exclusions. Appendix § 16.5 documents the simulated recovery. Within it, Appendix § 16.5 .§ 7 estimates the financial maps from hypothetical household accounts and simulated fund observations.
Synthetic estimator experiments
Recorded losses, spending measurement and source timing
The source ledger and the exposure unit
Recipient weather and regression inference
Monetary severity and the household source response
Continuing destruction and response age
Calendar-pattern measurement and operator comparisons
Cross-disaster scope and supporting household finance
Local-response diagnostics
Financial measurement and a physical-weather design
Household demand: a measurable derivative
Relative identification and the role of the operator
Conditional financial visibility
Physical-weather coverage and recipient measurement
Warning geography and the cost of long exclusion windows
Simulated recovery on the sample design
Design
Estimator, comparisons, inference and the planning diagnostic
Measurement-noise models
Results
Development history
A diffusion-specific EBP restriction
Recovering the credit-premium response with an estimated market map
Proofs
Lemma 1
§ 2.1 gives the direct consumption response Δ c t + h = Φ h ε t \Delta c_{t+h} = \Phi_h\varepsilon_t Δ c t + h = Φ h ε t , h ≥ 0 h \ge0 h ≥ 0 . Combining it with the valuation vector (1) gives the direct term − γ Φ h -\gamma\Phi_h − γ Φ h . The slow field receives the impact innovation ν t = Φ 0 ε t \nu_t = \Phi_0\varepsilon_t ν t = Φ 0 ε t at t + 1 t+1 t + 1 with coefficient η D \eta_D η D , by (1). Propagating it h − 1 h-1 h − 1 further periods through A D A_D A D gives η D A D h − 1 Φ 0 ε t \eta_D A_D^{h-1}\Phi_0\varepsilon_t η D A D h − 1 Φ 0 ε t for h ≥ 1 h \ge1 h ≥ 1 and nothing at h = 0 h = 0 h = 0 . The fast field receives the signal B ε t B\varepsilon_t B ε t at t t t , by (2), and propagating it h h h periods gives η a t t A a t t h B ε t \eta_{\mathrm{att}} A_{\mathrm{att}}^h B\varepsilon_t η att A att h B ε t . The production field receives ν t \nu_t ν t at t + 1 t+1 t + 1 and contributes η D P ( A D P ) h − 1 Φ 0 ε t \eta_D^P(A_D^P)^{h-1}\Phi_0\varepsilon_t η D P ( A D P ) h − 1 Φ 0 ε t for h ≥ 1 h \ge1 h ≥ 1 . Each component enters v t v_t v t with coefficient − γ -\gamma − γ , − a D -a_D − a D , − a a t t -a_{\mathrm{att}} − a att or − a D P -a_D^P − a D P . Collecting terms gives (2) . The fields receive inputs only at the shock date, ν t \nu_t ν t for the slow and production fields and s t s_t s t for the fast field. Later direct responses Φ j ε t \Phi_j\varepsilon_t Φ j ε t , j ≥ 1 j \ge1 j ≥ 1 , do not feed them, so the operators are exact rather than approximations to the convolution η D ∑ j = 0 h − 1 A D h − 1 − j Φ j \eta_D\sum_{j=0}^{h-1} A_D^{h-1-j}\Phi_j η D ∑ j = 0 h − 1 A D h − 1 − j Φ j . □ \square □
Theorem 1
Theorem 1. Suppose Ω ~ h S C I = a h ( s f i x ) ′ \widetilde{\Omega}^{SCI}_h = a_h(\mathbf{s}^{\mathrm{fix}})' Ω h S C I = a h ( s fix ) ′ for all h h h . Under the spectral form, the modal loadings then satisfy ψ k ( h ) = a h s k f i x \psi_k(h) = a_h s_k^{\mathrm{fix}} ψ k ( h ) = a h s k fix with s k f i x ≡ ( s f i x ) ′ u k s_k^{\mathrm{fix}} \equiv(\mathbf{s}^{\mathrm{fix}})'u_k s k fix ≡ ( s fix ) ′ u k , so for any two modes k , ℓ k,\ell k , ℓ ,
ψ k ( 1 ) ψ ℓ ( 2 ) − ψ ℓ ( 1 ) ψ k ( 2 ) = 0. (31) \psi_k(1)\psi_\ell(2) - \psi_\ell(1)\psi_k(2) = 0. \tag*{(31)} ψ k ( 1 ) ψ ℓ ( 2 ) − ψ ℓ ( 1 ) ψ k ( 2 ) = 0. (31) From (3) , ψ k ( 1 ) = C ˉ D , k i n i t + C ˉ a t t , k i n i t r a t t , k \psi_k(1) = \bar{C}^{\mathrm{init}}_{D,k} + \bar{C}^{\mathrm{init}}_{\mathrm{att},k}r_{\mathrm{att},k} ψ k ( 1 ) = C ˉ D , k init + C ˉ att , k init r att , k and ψ k ( 2 ) = C ˉ D , k i n i t r D , k + C ˉ a t t , k i n i t r a t t , k 2 \psi_k(2) = \bar{C}^{\mathrm{init}}_{D,k}r_{D,k} + \bar{C}^{\mathrm{init}}_{\mathrm{att},k}r_{\mathrm{att},k}^2 ψ k ( 2 ) = C ˉ D , k init r D , k + C ˉ att , k init r att , k 2 , and similarly for ℓ \ell ℓ . Substituting into (A.1) gives
C ˉ D , k i n i t C ˉ D , ℓ i n i t ( r D , ℓ − r D , k ) + C ˉ a t t , k i n i t C ˉ a t t , ℓ i n i t r a t t , k r a t t , ℓ ( r a t t , ℓ − r a t t , k ) + C ˉ D , k i n i t C ˉ a t t , ℓ i n i t r a t t , ℓ ( r a t t , ℓ − r D , k ) + C ˉ a t t , k i n i t C ˉ D , ℓ i n i t r a t t , k ( r D , ℓ − r a t t , k ) = 0 , (32) \begin{aligned}
&\bar{C}^{\mathrm{init}}_{D,k}\bar{C}^{\mathrm{init}}_{D,\ell}(r_{D,\ell}-r_{D,k})+\bar{C}^{\mathrm{init}}_{\mathrm{att},k}\bar{C}^{\mathrm{init}}_{\mathrm{att},\ell}r_{\mathrm{att},k}r_{\mathrm{att},\ell}(r_{\mathrm{att},\ell}-r_{\mathrm{att},k})\\
&+\bar{C}^{\mathrm{init}}_{D,k}\bar{C}^{\mathrm{init}}_{\mathrm{att},\ell}r_{\mathrm{att},\ell}(r_{\mathrm{att},\ell}-r_{D,k})+\bar{C}^{\mathrm{init}}_{\mathrm{att},k}\bar{C}^{\mathrm{init}}_{D,\ell}r_{\mathrm{att},k}(r_{D,\ell}-r_{\mathrm{att},k})=0, \tag*{(32)}
\end{aligned} C ˉ D , k init C ˉ D , ℓ init ( r D , ℓ − r D , k ) + C ˉ att , k init C ˉ att , ℓ init r att , k r att , ℓ ( r att , ℓ − r att , k ) + C ˉ D , k init C ˉ att , ℓ init r att , ℓ ( r att , ℓ − r D , k ) + C ˉ att , k init C ˉ D , ℓ init r att , k ( r D , ℓ − r att , k ) = 0 , (32) a polynomial in the coefficients and decay factors. It is not identically zero: at C ˉ a t t , k i n i t = C ˉ a t t , ℓ i n i t = 0 \bar{C}^{\mathrm{init}}_{\mathrm{att},k} = \bar{C}^{\mathrm{init}}_{\mathrm{att},\ell} = 0 C ˉ att , k init = C ˉ att , ℓ init = 0 , C ˉ D , k i n i t = C ˉ D , ℓ i n i t = 1 \bar{C}^{\mathrm{init}}_{D,k} = \bar{C}^{\mathrm{init}}_{D,\ell} = 1 C ˉ D , k init = C ˉ D , ℓ init = 1 and r D , ℓ ≠ r D , k r_{D,\ell} \ne r_{D,k} r D , ℓ = r D , k it equals r D , ℓ − r D , k ≠ 0 r_{D,\ell} - r_{D,k} \ne0 r D , ℓ − r D , k = 0 . A polynomial that is not identically zero vanishes only on a closed set of measure zero, including within the open set where all four coefficients are nonzero. Away from that set (A.1) fails, so no rank-one representation exists. The same evaluation shows that distinct decay rates across modes already break rank one within a single channel. This witness also respects the diffusion-kernel restriction: distinct eigenvalues and κ D > 0 \kappa_D > 0 κ D > 0 imply distinct slow rates, and sufficiently small nonzero fast loadings preserve the nonzero determinant. More generally the exceptional set is restricted to the admissible parameter domain, not declared negligible on an arbitrary lower-dimensional restriction. Conversely, if every two-horizon, two-mode minor vanishes, all nonzero rows are proportional. Choosing any one as ( s f i x ) ′ (\mathbf{s}^{\mathrm{fix}})' ( s fix ) ′ constructs the representation
(the all-zero case is trivial). The augmented two-field state remains a standard linear system. Neither matrix representations nor ordinal centrality are ruled out.
Corollary A.1 (Competing graphs). Suppose the social slow field on L \mathcal{L} L and the production slow field on L P \mathcal{L}^{P} L P are both active on non-collinear retained modal supports, with nonzero projected loadings and distinct horizon profiles. Then, generically, no fixed scalar characteristic with a common linear horizon coefficient reproduces pricing relevance across both graphs and all horizons.
Proof. The two fields contribute operator blocks whose decay schedules are governed by different Laplacians. If the retained modal subspaces are not collinear, a rank-one family would require the 2 × 2 2 \times2 2 × 2 determinants formed from one social and one production direction to vanish at all horizon pairs, which again defines a proper algebraic subset of the parameter space. Canonical correlations and projection distances between the two retained subspaces (Appendix § 13 .§ 2 ) measure how close the data are to the collinear case.
Theorem 2
For the slow channel, C D , k ( h ) = C D , k i n i t r D , k h − 1 C_{D,k}(h) = C^{\mathrm{init}}_{D,k}r_{D,k}^{h-1} C D , k ( h ) = C D , k init r D , k h − 1 for h ≥ 1 h \ge1 h ≥ 1 . Taking log magnitudes and substituting r D , k = ρ D e − κ D λ k r_{D,k} = \rho_D e^{-\kappa_D\lambda_k} r D , k = ρ D e − κ D λ k gives (4) . The same steps with C a t t , k ( h ) = C a t t , k i n i t r a t t , k h C_{\mathrm{att},k}(h) = C^{\mathrm{init}}_{\mathrm{att},k}r_{\mathrm{att},k}^{h} C att , k ( h ) = C att , k init r att , k h give (5) . Here ρ a t t = 0 \rho_{\mathrm{att}} = 0 ρ att = 0 is an impact-only response and is excluded. Under two channels, log ∣ C D , k ( h ) + C a t t , k ( h ) ∣ \log\lvert C_{D,k}(h) + C_{\mathrm{att},k}(h)\rvert log ∣ C D , k ( h ) + C att , k ( h )∣ is the log of a sum of two exponentials in h h h with mode-specific rates. Its second difference in h h h is generically nonzero, while every function of the form a k + h c 1 + h λ k c 2 a_k + hc_1 + h\lambda_kc_2 a k + h c 1 + h λ k c 2 has zero second difference in h h h . Figure A.1 illustrates both cases.
Proposition 1
Under one channel, log r m , k = log ρ m − κ m λ k \log r_{m,k} = \log\rho_m - \kappa_m\lambda_k log r m , k = log ρ m − κ m λ k . Horizon variation within a mode identifies log r m , k \log r_{m,k} log r m , k as the slope of log ∣ C m , k ( h ) ∣ \log\lvert C_{m,k}(h)\rvert log ∣ C m , k ( h )∣ in h h h . Differencing across two informative modes with λ k ≠ λ ℓ \lambda_k \ne\lambda_\ell λ k = λ ℓ gives κ m \kappa_m κ m , and then ρ m \rho_m ρ m follows. With only one distinct informative eigenvalue, ( ρ m , κ m ) (\rho_m,\kappa_m) ( ρ m , κ m ) cannot be separated, because every informative mode then has the same log r m , k \log r_{m,k} log r m , k . At λ = 0 \lambda= 0 λ = 0 , only ρ m \rho_m ρ m is identified. Two distinct but close eigenvalues can instead make identification weak. Nearly equal channel rates complicate channel separation, a different problem.
Half-lives. For ρ m > 0 \rho_m > 0 ρ m > 0 , mode k k k of field m m m survives each period at rate r m , k = ρ m e − κ m λ k r_{m,k} = \rho_m e^{-\kappa_m\lambda_k} r m , k = ρ m e − κ m λ k , so its half-life solves r m , k t = 1 / 2 r_{m,k}^{t} = 1/2 r m , k t = 1/2 , which gives
t 1 / 2 ( k ) = ln 2 κ m λ k − ln ρ m periods . (33) t_{1/2}(k) = \frac{\ln2}{\kappa_m\lambda_k - \ln\rho_m}\text{ periods}. \tag*{(33)} t 1/2 ( k ) = κ m λ k − ln ρ m ln 2 periods . (33)
Figure A.1. The modal decay restriction. Each line is the log response of one mode, normalized at horizon 1. Panel A uses r D , k h − 1 r_{D,k}^{h-1} r D , k h − 1 , whereas Panel B uses 0.3 r D , k h − 1 + 0.7 r a t t , k h 0.3r_{D,k}^{h-1}+0.7r_{\mathrm{att},k}^{h} 0.3 r D , k h − 1 + 0.7 r att , k h . Illustrative parameters are ( ρ D , κ D ) = ( 0.97 , 0.4 ) (\rho_D,\kappa_D)=(0.97,0.4) ( ρ D , κ D ) = ( 0.97 , 0.4 ) and ( ρ a t t , κ a t t ) = ( 0.55 , 0.8 ) (\rho_{\mathrm{att}},\kappa_{\mathrm{att}})=(0.55,0.8) ( ρ att , κ att ) = ( 0.55 , 0.8 ) . Single-channel log paths are straight, whereas a mixture generally bends (Theorem 2).
The persistence parameter alone sets the half-life of the stationary mode, ln 2 / ( − ln ρ m ) \ln2/(-\ln\rho_m) ln 2/ ( − ln ρ m ) , and a larger κ m \kappa_m κ m shortens the half-life of rougher modes. When ρ a t t = 0 \rho_{\mathrm{att}}=0 ρ att = 0 , the attention field has no persistence beyond impact.
Proposition 2
In the exact diagonal benchmark, partition the diagonal as diag ( C k ( h ) ) = Ψ h ( K ) ⊕ Ψ h ( t a i l ) \operatorname{diag}(C_k(h))=\Psi_h^{(K)}\oplus\Psi_h^{(\mathrm{tail})} diag ( C k ( h )) = Ψ h ( K ) ⊕ Ψ h ( tail ) . Then O ~ h ( S ) − O ~ h , K ( S ) = M U K c Ψ h ( t a i l ) U K c ′ R \widetilde{O}_h^{(S)}-\widetilde{O}_{h,K}^{(S)}=MU_{K^c}\Psi_h^{(\mathrm{tail})}U_{K^c}^{\prime}R O h ( S ) − O h , K ( S ) = M U K c Ψ h ( tail ) U K c ′ R , and submultiplicativity with ∥ U K c ∥ 2 ≤ 1 \lVert U_{K^c}\rVert_2\le1 ∥ U K c ∥ 2 ≤ 1 gives the first term of (7) , since the spectral norm of a diagonal matrix is its largest absolute entry, with an empty tail contributing zero when K = n C K=n_C K = n C . With off-diagonal loadings, O ~ h S C I \widetilde{O}_h^{SCI} O h S C I has the additional term E h , K \mathcal{E}_{h,K} E h , K , whose aggregate is E h , K ( S ) = M E h , K R \mathcal{E}_{h,K}^{(S)}=M\mathcal{E}_{h,K}R E h , K ( S ) = M E h , K R . The triangle inequality adds ∥ E h , K ( S ) ∥ 2 \lVert\mathcal{E}_{h,K}^{(S)}\rVert_2 ∥ E h , K ( S ) ∥ 2 without a further ∥ M ∥ 2 ∥ R ∥ 2 \lVert M\rVert_2\lVert R\rVert_2 ∥ M ∥ 2 ∥ R ∥ 2 factor, because the aggregation is already inside it. The direct term − γ Φ h -\gamma\Phi_h − γ Φ h is excluded from O ~ h S C I \widetilde{O}_h^{SCI} O h S C I and hence from this bound by construction. Under the directed benchmark of Appendix § 13 .§ 3 , the modal tail carries the extra factor cond 2 ( U ) \operatorname{cond}_2(U) cond 2 ( U ) , the condition number of the eigenbasis.
Lemma A.2 (Retained slow and fast components). Let O ~ m , h ( S , K ) ≡ M U K diag ( C m , k ( h ) ) k ≤ K U K ′ R \widetilde{O}_{m,h}^{(S,K)}\equiv MU_K\operatorname{diag}(C_{m,k}(h))_{k\le K}U_K^{\prime}R O m , h ( S , K ) ≡ M U K diag ( C m , k ( h ) ) k ≤ K U K ′ R for m ∈ { D , a t t } m\in\{D,\mathrm{att}\} m ∈ { D , att } . Then
∥ O ~ a t t , h ( S , K ) ∥ 2 ≤ ∥ M ∥ 2 ∥ R ∥ 2 max k ≤ K { ∣ C a t t , k i n i t ∣ r a t t , k h } , \left\lVert\widetilde{O}_{\mathrm{att},h}^{(S,K)}\right\rVert_2\le\lVert M\rVert_2\lVert R\rVert_2\max_{k\le K}\left\{ \left|C_{\mathrm{att},k}^{\mathrm{init}}\right|r_{\mathrm{att},k}^{h}\right\}, O att , h ( S , K ) 2 ≤ ∥ M ∥ 2 ∥ R ∥ 2 k ≤ K max { C att , k init r att , k h } , ∥ O ~ D , h ( S , K ) ∥ 2 ≤ ∥ M ∥ 2 ∥ R ∥ 2 max k ≤ K { ∣ C D , k i n i t ∣ r D , k h − 1 } , h ≥ 1. (34) \left\lVert\widetilde{O}_{D,h}^{(S,K)}\right\rVert_2\le\lVert M\rVert_2\lVert R\rVert_2\max_{k\le K}\left\{ \left|C_{D,k}^{\mathrm{init}}\right|r_{D,k}^{h-1}\right\},\qquad h\ge1. \tag*{(34)} O D , h ( S , K ) 2 ≤ ∥ M ∥ 2 ∥ R ∥ 2 k ≤ K max { C D , k init r D , k h − 1 } , h ≥ 1. (34)
Proof. Submultiplicativity and orthonormality, as above. The two upper bounds do not show that the slow component dominates. That would also need a lower bound on the aggregated slow term: a nonzero aggregate slow loading, no cancellation among leading modes, and a slow rate above the fast rates. Concentration of fast loadings on rough modes is a separate assumption and does not follow from Assumption 2.
Proposition 3
In the slack region the intermediary’s first-order condition is q = Γ θ / N q = \Gamma\theta/N q = Γ θ / N , so θ = N q / Γ \theta= Nq/\Gamma θ = N q /Γ . Clearing, ζ q + N q / Γ = Q \zeta q + Nq/\Gamma= Q ζ q + N q /Γ = Q , gives q = Γ Q / ( N + ζ Γ ) q = \Gamma Q/(N + \zeta\Gamma) q = Γ Q / ( N + ζ Γ ) and inventory θ = N Q / ( N + ζ Γ ) \theta= NQ/(N + \zeta\Gamma) θ = N Q / ( N + ζ Γ ) , which is below the cap N / χ N/\chi N / χ exactly when Q ≤ ( N + ζ Γ ) / χ Q \le(N + \zeta\Gamma)/\chi Q ≤ ( N + ζ Γ ) / χ . When the cap binds, θ = N / χ \theta= N/\chi θ = N / χ and q = ( Q − N / χ ) / ζ q = (Q - N/\chi)/\zeta q = ( Q − N / χ ) / ζ . The multiplier on the cap is q − Γ / χ q - \Gamma/\chi q − Γ/ χ , which is nonnegative exactly when Q ≥ ( N + ζ Γ ) / χ Q \ge(N + \zeta\Gamma)/\chi Q ≥ ( N + ζ Γ ) / χ . The two branches agree at the boundary. The intermediary’s objective is strictly concave, so its choice is unique, and aggregate demand ζ q + θ ( q ) \zeta q + \theta(q) ζ q + θ ( q ) is strictly increasing in q q q , so clearing is unique at fixed N N N . The derivatives follow by differentiation, using ∂ Q / ∂ S = 1 \partial Q/\partial S = 1 ∂ Q / ∂ S = 1 . Positive claim prices require the additional domain restriction E [ V ] > q E[V] > q E [ V ] > q .
With inherited claims θ − \theta^{-} θ − , capital responds to the price, N = N − − ℓ m t m ( q − q − ) N = N^{-} - \ell_{\mathrm{mtm}}(q - q^{-}) N = N − − ℓ mtm ( q − q − ) with ℓ m t m = θ − / R f ≥ 0 \ell_{\mathrm{mtm}} = \theta^{-}/R_f \ge0 ℓ mtm = θ − / R f ≥ 0 . In the slack region let C N , − = N − + ℓ m t m q − + ζ Γ C_{N,-} = N^{-} + \ell_{\mathrm{mtm}}q^{-} + \zeta\Gamma C N , − = N − + ℓ mtm q − + ζ Γ and Δ − = C N , − 2 − 4 ℓ m t m Γ Q \Delta_{-} = C_{N,-}^{2} - 4\ell_{\mathrm{mtm}}\Gamma Q Δ − = C N , − 2 − 4 ℓ mtm Γ Q . The lower root q − = ( C N , − − Δ − ) / ( 2 ℓ m t m ) q_{-} = (C_{N,-} - \sqrt{\Delta_{-}})/(2\ell_{\mathrm{mtm}}) q − = ( C N , − − Δ − ) / ( 2 ℓ mtm ) is regular for ℓ m t m > 0 \ell_{\mathrm{mtm}} > 0 ℓ mtm > 0 when Δ − > 0 \Delta_{-} > 0 Δ − > 0 , capital and price are positive, and the solution satisfies the slack-regime inequality. For ℓ m t m = 0 \ell_{\mathrm{mtm}} = 0 ℓ mtm = 0 , use the fixed-capital solution instead. The implicit derivative is q S = G S / ( 1 + ℓ m t m G N ) q_{S} = G_{S}/(1+\ell_{\mathrm{mtm}}G_{N}) q S = G S / ( 1 + ℓ mtm G N ) with 1 + ℓ m t m G N = Δ − / ( N + ζ Γ ) > 0 1+\ell_{\mathrm{mtm}}G_{N} = \sqrt{\Delta_{-}}/(N + \zeta\Gamma) > 0 1 + ℓ mtm G N = Δ − / ( N + ζ Γ ) > 0 at that root. In the binding region ( ζ χ − ℓ m t m ) q = χ Q − N − − ℓ m t m q − (\zeta\chi- \ell_{\mathrm{mtm}})q = \chi Q - N^{-} - \ell_{\mathrm{mtm}}q^{-} ( ζ χ − ℓ mtm ) q = χ Q − N − − ℓ mtm q − , and stability requires ζ χ > ℓ m t m \zeta\chi> \ell_{\mathrm{mtm}} ζ χ > ℓ mtm .
Proposition 4
At a reference point strictly inside the slack region, q = Γ ( Q 0 + S ) / ( C N − T ) q = \Gamma(Q_{0} + S)/(C_{N} - T) q = Γ ( Q 0 + S ) / ( C N − T ) with C N = N ‾ + ζ Γ C_{N} = \overline{N} + \zeta\Gamma C N = N + ζ Γ and q 0 = Γ Q 0 / C N q_{0} = \Gamma Q_{0}/C_{N} q 0 = Γ Q 0 / C N . Using the expansion ( C N − T ) − 1 = C N − 1 ( 1 + T / C N + T 2 / C N 2 + ⋯ ) (C_{N} - T)^{-1} = C_{N}^{-1}(1 + T/C_{N} + T^{2}/C_{N}^{2} + \cdots) ( C N − T ) − 1 = C N − 1 ( 1 + T / C N + T 2 / C N 2 + ⋯ ) and collecting terms up to second order gives (15) . The T 2 T^{2} T 2 term is of the same order as S T ST S T and cannot be dropped. For the conditional change, apply E t E_{t} E t to the quadratic at t + 1 t+1 t + 1 and subtract its value at t t t , using E t [ S t + 1 T t + 1 ] = ρ K T t l ′ A D z t + l ′ Cov t ( η t + 1 , e t + 1 ) E_{t}[S_{t+1}T_{t+1}] = \rho_{K}T_{t}l^{\prime}A_{D}z_{t} + l^{\prime}\operatorname{Cov}_{t}(\eta_{t+1},e_{t+1}) E t [ S t + 1 T t + 1 ] = ρ K T t l ′ A D z t + l ′ Cov t ( η t + 1 , e t + 1 ) and E t [ T t + 1 2 ] = ρ K 2 T t 2 + σ e , t 2 E_{t}[T_{t+1}^{2}] = \rho_{K}^{2}T_{t}^{2} + \sigma_{e,t}^{2} E t [ T t + 1 2 ] = ρ K 2 T t 2 + σ e , t 2 . This gives (16) without imposing scalar closure. The identity is exact for the quadratic approximation, not for the piecewise equilibrium, and it holds only within one smooth regime: innovations that cross the margin cap require the piecewise conditional mean. Let σ S e , t ≡ l ′ Cov t ( η t + 1 , e t + 1 ) \sigma_{Se,t} \equiv l^{\prime}\operatorname{Cov}_{t}(\eta_{t+1},e_{t+1}) σ S e , t ≡ l ′ Cov t ( η t + 1 , e t + 1 ) . If additionally l ′ A D = r l ′ l^{\prime}A_{D} = rl^{\prime} l ′ A D = r l ′ , the scalar expression follows: A S ( r − 1 ) S t + A T ( ρ K − 1 ) T t + A S T [ ( ρ K − 1 ) S t T t + σ S e , t ] + A T T [ ( ρ K 2 − 1 ) T t 2 + σ e , t 2 ] A_{S}(r-1)S_{t} + A_{T}(\rho_{K}-1)T_{t} + A_{ST}[(\rho_{K}-1)S_{t}T_{t} + \sigma_{Se,t}] + A_{TT}[(\rho_{K}^{2}-1)T_{t}^{2} + \sigma_{e,t}^{2}] A S ( r − 1 ) S t + A T ( ρ K − 1 ) T t + A S T [( ρ K − 1 ) S t T t + σ S e , t ] + A T T [( ρ K 2 − 1 ) T t 2 + σ e , t 2 ] . This additional invariant-direction restriction generally fails.
Own persistence. If, separately justified, the premium follows q t + 1 = ρ q q t + ξ 0 z t + ξ 1 T t z t + u t + 1 q_{t+1} = \rho_q q_t + \xi_0 z_t + \xi_1 T_t z_t + u_{t+1} q t + 1 = ρ q q t + ξ 0 z t + ξ 1 T t z t + u t + 1 with a scalar field z t + 1 = r z t + u t + 1 z z_{t+1} = r z_t + u^z_{t+1} z t + 1 = r z t + u t + 1 z and persistence ρ K \rho_K ρ K of T \mathcal{T} T , then one date earlier
E t − 1 Δ q t + 1 = ρ q ( ρ q − 1 ) q t − 1 + ξ 0 ( ρ q + r − 1 ) z t − 1 + ξ 1 ( ρ q + r ρ K − 1 ) T t − 1 z t − 1 + ξ 1 Cov t − 1 ( u t z , e t ) , (35) E_{t-1}\Delta q_{t+1} = \rho_q(\rho_q - 1)q_{t-1} + \xi_0(\rho_q + r - 1)z_{t-1} + \xi_1(\rho_q + r\rho_K - 1)T_{t-1}z_{t-1} + \xi_1\operatorname{Cov}_{t-1}(u^z_t,e_t), \tag*{(35)} E t − 1 Δ q t + 1 = ρ q ( ρ q − 1 ) q t − 1 + ξ 0 ( ρ q + r − 1 ) z t − 1 + ξ 1 ( ρ q + r ρ K − 1 ) T t − 1 z t − 1 + ξ 1 Cov t − 1 ( u t z , e t ) , (35)
so, with ξ 1 > 0 \xi_1 > 0 ξ 1 > 0 and the conditional covariance term held fixed, the coefficient on the lagged interaction is positive when ρ q + r ρ K > 1 \rho_q + r\rho_K > 1 ρ q + r ρ K > 1 , negative when ρ q + r ρ K < 1 \rho_q + r\rho_K < 1 ρ q + r ρ K < 1 and zero at equality. None of these cases is a sign theorem for a regression that, like the monthly specification, also conditions on two lagged outcome changes and standardized regressors (Appendix § 15 .§ 2 ).
Level sign. The diffusion kernel, conventionally called the graph heat kernel, is entrywise nonnegative: with A ≡ D − 1 / 2 A D − 1 / 2 ≥ 0 A \equiv D^{-1/2}AD^{-1/2} \ge0 A ≡ D − 1/2 A D − 1/2 ≥ 0 , e − κ D L = e − κ D e κ D A = e − κ D ∑ n ≥ 0 ( κ D A ) n / n ! ≥ 0 e^{-\kappa_D\mathcal{L}} = e^{-\kappa_D}e^{\kappa_D A} = e^{-\kappa_D}\sum_{n\ge0}(\kappa_D A)^n/n! \ge0 e − κ D L = e − κ D e κ D A = e − κ D ∑ n ≥ 0 ( κ D A ) n / n ! ≥ 0 . If l ≥ 0 l \ge0 l ≥ 0 and B D ε t ≥ 0 B_D\varepsilon_t \ge0 B D ε t ≥ 0 (an adverse storm raises the adverse field in hit counties), every entry of the pricing kernel ((14) ) applied to the storm is nonnegative, so the level response of the EBP is nonnegative at every horizon, holding capital fixed.
Stationary and nonstationary financial exposure
Because L = U Λ U ′ \mathcal{L} = U\Lambda U' L = U Λ U ′ , A D = ρ D e − κ D L = U diag ( r D , k ) U ′ A_D = \rho_D e^{-\kappa_D\mathcal{L}} = U\operatorname{diag}(r_{D,k})U' A D = ρ D e − κ D L = U diag ( r D , k ) U ′ and A D h − 1 = U diag ( r D , k h − 1 ) U ′ A_D^{h-1} = U\operatorname{diag}(r_{D,k}^{h-1})U' A D h − 1 = U diag ( r D , k h − 1 ) U ′ , where Λ \Lambda Λ holds the eigenvalues λ k \lambda_k λ k on its diagonal. Hence ( π H ) ′ A D h − 1 B D = ∑ k ( ( π H ) ′ u k ) r D , k h − 1 u k ′ B D (\pi^H)'A_D^{h-1}B_D = \sum_k((\pi^H)'u_k)r_{D,k}^{h-1}u_k'B_D ( π H ) ′ A D h − 1 B D = ∑ k (( π H ) ′ u k ) r D , k h − 1 u k ′ B D , and separating k = 1 k = 1 k = 1 gives ((17) ). The stationary mode is u 1 ∝ D 1 / 2 1 u_1 \propto D^{1/2}\mathbf{1} u 1 ∝ D 1/2 1 , because L D 1 / 2 1 = D 1 / 2 1 − D − 1 / 2 A 1 = D 1 / 2 1 − D − 1 / 2 D 1 = 0 \mathcal{L}D^{1/2}\mathbf{1} = D^{1/2}\mathbf{1} - D^{-1/2}A\mathbf{1} = D^{1/2}\mathbf{1} - D^{-1/2}D\mathbf{1} = 0 L D 1/2 1 = D 1/2 1 − D − 1/2 A 1 = D 1/2 1 − D − 1/2 D 1 = 0 , so λ 1 = 0 \lambda_1 = 0 λ 1 = 0 and r D , 1 = ρ D r_{D,1} = \rho_D r D , 1 = ρ D . The first term therefore weights the shock by square-root degree. The nonstationary term vanishes when π H ∝ u 1 \pi^H \propto u_1 π H ∝ u 1 , that is, when π i H ∝ D i i 1 / 2 \pi_i^H \propto D_{ii}^{1/2} π i H ∝ D ii 1/2 . If w ′ A D = r D , k w ′ w'A_D = r_{D,k}w' w ′ A D = r D , k w ′ , then w ′ A D h − 1 = r D , k h − 1 w ′ w'A_D^{h-1} = r_{D,k}^{h-1}w' w ′ A D h − 1 = r D , k h − 1 w ′ by induction, which gives the collapse statement in the main text. For symmetric A D A_D A D , such w w w lie in a single eigenspace. As τ = κ D ( h − 1 ) → ∞ \tau= \kappa_D(h - 1) \to\infty τ = κ D ( h − 1 ) → ∞ , e − τ L → u 1 u 1 ′ e^{-\tau\mathcal{L}} \to u_1u_1' e − τ L → u 1 u 1 ′ , so any weighted aggregate converges to its incidence term, which is the limit visible in Figure 3 . Every identity in this subsection holds with l l l in place of π H \pi^H π H , which covers loadings of either sign and δ = 0 \delta= 0 δ = 0 .
Signed responses are invariant to the sign of each eigenvector, because u k u_k u k appears twice in each term. An unsigned norm can also be invariant to eigenvector signs, but discards the economic direction of a response. It is therefore not the signed pricing projection in (16). Main-text equation (21) follows by using the two orthonormal eigenvectors ( 1 , 1 ) ′ / 2 (1,1)'/\sqrt{2} ( 1 , 1 ) ′ / 2 and ( 1 , − 1 ) ′ / 2 (1,-1)'/\sqrt{2} ( 1 , − 1 ) ′ / 2 , with eigenvalues 0 0 0 and 2 2 2 .
The common response and Proposition 5
The identities (20) follow from T 1 = 0 T1=0 T 1 = 0 and w ′ 1 = 1 w'1=1 w ′ 1 = 1 . They are a missing-direction result rather than a restriction on the common response’s sign or magnitude. An independent aggregate anchor can supply that response. A zero response under one set of weights also need not mean zero pressure under another. With two counties, w = ( 1 / 2 , 1 / 2 ) ′ w=(1/2,1/2)' w = ( 1/2 , 1/2 ) ′ , z = ( 1 , − 1 ) ′ z=(1,-1)' z = ( 1 , − 1 ) ′ and l = ( 1 , 0 ) ′ l=(1,0)' l = ( 1 , 0 ) ′ give w ′ z = 0 w'z=0 w ′ z = 0 but l ′ z = 1 l'z=1 l ′ z = 1 .
Proof of Proposition 5. If l = C Y ′ a l=C_Y'a l = C Y ′ a then l ′ z = a ′ C Y z l'z=a'C_Yz l ′ z = a ′ C Y z . Otherwise the orthogonal decomposition of l l l with respect to row ( C Y ) \operatorname{row}(C_Y) row ( C Y ) supplies v ∈ ker ( C Y ) v\in\ker(C_Y) v ∈ ker ( C Y ) with l ′ v ≠ 0 l'v\ne0 l ′ v = 0 , and z z z and z + v z+v z + v have the same measurement but different targets.
Proposition A.3 (Observability with a history ). With known A D A_D A D , known inputs and observations at dates 0 , … , H o b s 0,\ldots,H_{\mathrm{obs}} 0 , … , H obs , let Obs H o b s ≡ ( C Y ′ , ( C Y A D ) ′ , … , ( C Y A D H o b s ) ′ ) \operatorname{Obs}_{H_{\mathrm{obs}}}\equiv(C_Y',(C_YA_D)',\ldots,(C_YA_D^{H_{\mathrm{obs}}})') Obs H obs ≡ ( C Y ′ , ( C Y A D ) ′ , … , ( C Y A D H obs ) ′ ) . The initial-state functional l ′ z 0 l'z_0 l ′ z 0 is identified if and only if l ∈ row ( Obs H o b s ) l\in\operatorname{row}(\operatorname{Obs}_{H_{\mathrm{obs}}}) l ∈ row ( Obs H obs ) . Similarly, the final-date functional l ′ z H o b s l'z_{H_{\mathrm{obs}}} l ′ z H obs is identified if and only if ( A D H o b s ) ′ l ∈ row ( Obs H o b s ) (A_D^{H_{\mathrm{obs}}})'l\in\operatorname{row}(\operatorname{Obs}_{H_{\mathrm{obs}}}) ( A D H obs ) ′ l ∈ row ( Obs H obs ) , after known input contributions are removed. If states are restricted to the reachable subspace R = span ( B D , A D B D , … ) \mathcal{R}=\operatorname{span}(B_D,A_DB_D,\ldots) R = span ( B D , A D B D , … ) with basis R R R_{\mathcal{R}} R R , both conditions apply to Obs H o b s R R \operatorname{Obs}_{H_{\mathrm{obs}}}R_{\mathcal{R}} Obs H obs R R and to the target multiplied by R R R_{\mathcal{R}} R R .
Proof. Stacking the observations gives Obs H o b s z 0 \operatorname{Obs}_{H_{\mathrm{obs}}}z_0 Obs H obs z 0 plus known inputs, and z H o b s = A D H o b s z 0 z_{H_{\mathrm{obs}}}=A_D^{H_{\mathrm{obs}}}z_0 z H obs = A D H obs z 0 plus known inputs. Apply the null-space argument above to Obs H o b s \operatorname{Obs}_{H_{\mathrm{obs}}} Obs H obs , with target l l l or ( A D H o b s ) ′ l (A_D^{H_{\mathrm{obs}}})'l ( A D H obs ) ′ l .
This identifies a functional of a known linear system. It does not separately identify A D A_D A D , B D B_D B D and C Y C_Y C Y from a fitted outcome composite. A residual-maker T T T replaces C Y C_Y C Y by T C Y TC_Y T C Y and can remove observability.
Recovery and estimation of the pricing projection
Proof of Proposition 6. Let z = R R x z=R_{\mathcal{R}}x z = R R x and C R = C ∗ R R C_{\mathcal{R}}=C_*R_{\mathcal{R}} C R = C ∗ R R . If R R ′ l = C R ′ a R_{\mathcal{R}}'l=C_{\mathcal{R}}'a R R ′ l = C R ′ a , then l ′ z = a ′ C ∗ z l'z=a'C_*z l ′ z = a ′ C ∗ z for every x x x . Otherwise there exists an x 0 ∈ ker C R x_0\in\ker C_{\mathcal{R}} x 0 ∈ ker C R with l ′ R R x 0 ≠ 0 l'R_{\mathcal{R}}x_0\ne0 l ′ R R x 0 = 0 . States that differ by R R x 0 R_{\mathcal{R}}x_0 R R x 0 have identical measured responses and different financial targets. Applying the identity to z = R h ε z=R_h\varepsilon z = R h ε proves (25) .
The response subspace is part of the maintained specification. One must not infer its dimension or orientation from the same financial outcomes used to test the projection. Observationally equivalent latent systems can yield the same H h H_h H h . The proposition concerns their common pricing functional only when the financial representation is independently justified. It is not a system-identification theorem for unrestricted ( A D , B D , C Y , l ) (A_D,B_D,C_Y,l) ( A D , B D , C Y , l ) .
With full-state spanning, C ∗ ′ a = l C_*'a=l C ∗ ′ a = l , the representation holds for any propagation path. Restricting R \mathcal{R} R can permit recovery with fewer measurements, but the representation then inherits those restrictions. Responses outside R \mathcal{R} R require a separate check or bound.
Alternative household measurements. The map C ∗ C_* C ∗ describes an observation of the household state, rather than a particular data source. If C c C_c C c maps that state into consumption responses and C p C_p C p into fixed-price portfolio demand, jointly observing them gives C ∗ = ( C c ′ , C p ′ ) ′ C_* = (C'_c, C'_p)' C ∗ = ( C c ′ , C p ′ ) ′ . Proposition 6 then applies to this stacked map with the same proof. If a measured demand row is l ′ l' l ′ itself, it directly reveals the financial target. Otherwise, the spanning condition determines which combination of observations suffices. Actual holdings require valuation and price-feedback adjustments before they represent fixed-price demand. Similarly, expenditure and consumption have different maps when purchases provide services over time. Such known temporal filters can be written as measurements of an augmented state. The empirical application and recovery simulations use spending measurements. The conditional holdings problem in § 3 provides the local portfolio-demand schedule, taking household circumstances as given. Appendix § 16.4 .§ 1 derives local joint consumption–portfolio responses under stated conditions. Extending those decision-date restrictions across horizons requires additional dynamics.
More than one financially loaded channel. The slow-only credit-pricing benchmark sets the financial loading on attention to zero. More generally, let the fixed-price demand shift be S t = l D ′ z t + l a t t ′ b t S_t = l'_D z_t + l'_{\mathrm{att}} b_t S t = l D ′ z t + l att ′ b t , allowing signed l a t t l_{\mathrm{att}} l att , and put B a t t = η a t t B B_{\mathrm{att}} = \eta_{\mathrm{att}}B B att = η att B . For a one-time shock and h ≥ 1 h \ge1 h ≥ 1 , the conditional pricing response is
K h = c E [ l D ′ A D h − 1 B D + l a t t ′ A a t t h B a t t ] . K_h = c_E\left[l'_D A_D^{h-1}B_D + l'_{\mathrm{att}}A_{\mathrm{att}}^hB_{\mathrm{att}}\right]. K h = c E [ l D ′ A D h − 1 B D + l att ′ A att h B att ] .
At impact only the fast term can respond through these fields. Stacking the two fields and their measurement maps gives the same recovery condition on their joint response subspace. In the exact diagonal benchmark each mode again has a two-exponential profile, now with coefficients c E ( l m ′ u k ) ( u k ′ B m u k ) c_E(l'_m u_k)(u'_kB_m u_k) c E ( l m ′ u k ) ( u k ′ B m u k ) , m ∈ { D , a t t } m \in\{D,\mathrm{att}\} m ∈ { D , att } . The rank-one-minor criterion of Theorem 1 applies unchanged, and when its nondegeneracy conditions hold for these coefficients, rank one generically fails. The channel paths here are the maintained laws in § 2 . Measuring an additional outcome changes what can be observed, while assigning it an additional propagation law is a separate modeling choice. The reported simulations implement the slow-channel benchmark.
Growth responses versus the level target. For the same outcome, shock and units, let H h L H_h^L H h L denote its level-loss response and H h Δ H_h^\Delta H h Δ its first-difference response. Under no anticipation and a known pre-shock baseline,
H h Δ = H h L − H h − 1 L , H h L = ∑ j = 0 h H j Δ , H − 1 L = 0. (36) H_h^\Delta= H_h^L - H_{h-1}^L,\qquad H_h^L = \sum_{j=0}^{h} H_j^\Delta,\qquad H_{-1}^L = 0. \tag*{(36)} H h Δ = H h L − H h − 1 L , H h L = j = 0 ∑ h H j Δ , H − 1 L = 0. (36)
The accumulated estimator’s covariance contains all cross-horizon covariances, not just the sum of horizon-specific variances. The modal-response estimates have the sign of spending growth, so a loss convention reverses their sign. They also use clipped growth and fixed-effect residualization. The identity therefore does not turn those transformed estimates into unclipped national expenditure or restore omitted event means. Recovering the desired H h H_h H h requires a consistent measurement map, outcome definition, temporal aggregation and common-response anchor. A mode’s growth response cannot simply be inserted as its level-demand loading in the pricing kernel.
An aggregate anchor and the nuisance common response. For illustration, let C Y = I d n C C_Y = I d_{n_C} C Y = I d n C and let T T T have rank n C − 1 n_C - 1 n C − 1 with T 1 = 0 T\mathbf{1} = 0 T 1 = 0 . Then stacking T T T and a row w ′ w^{\prime} w ′ with w ′ 1 = 1 w^{\prime}\mathbf{1} = 1 w ′ 1 = 1 gives a full-column-rank measurement map, so the missing mean is measured. If the observed aggregate also contains unrestricted a e a_e a e , however, its value is w ′ R h ε e + a e w^{\prime}R_h\varepsilon_e + a_e w ′ R h ε e + a e , not w ′ R h ε e w^{\prime}R_h\varepsilon_e w ′ R h ε e . The transformation R h ↦ R h + 1 α ′ R_h \mapsto R_h + \mathbf{1}\alpha^{\prime} R h ↦ R h + 1 α ′ and a e ↦ a e − α ′ ε e a_e \mapsto a_e - \alpha^{\prime}\varepsilon_e a e ↦ a e − α ′ ε e preserves the observations. Thus an aggregate anchor needs an exclusion, a separately measured competing response or a bound on a e a_e a e . If only a e ∈ [ a ‾ e , a ‾ e ] a_e \in[\underline{a}_e,\overline{a}_e] a e ∈ [ a e , a e ] is justified, the recovered aggregate household response is an interval of the same width, not a point estimate.
Error and partial identification. With r l = l − C ∗ ′ a r_l = l - C_*^{\prime}a r l = l − C ∗ ′ a , subtraction gives
p ^ h − K h ε = c E a ′ ( H ^ h − H h ) ε − c E r l ′ R h ε . (37) \widehat{p}_h - K_h\varepsilon= c_Ea^{\prime}(\widehat{H}_h - H_h)\varepsilon- c_Er_l^{\prime}R_h\varepsilon. \tag*{(37)} p h − K h ε = c E a ′ ( H h − H h ) ε − c E r l ′ R h ε . (37)
The triangle and Cauchy–Schwarz inequalities give (26) . If P R P_R P R is the orthogonal projector onto the response subspace, replace r l r_l r l by P R r l P_Rr_l P R r l . Consequently, independent bounds ∥ P R r l ∥ ≤ r ‾ \lVert P_Rr_l\rVert\le\overline{r} ∥ P R r l ∥ ≤ r and ∥ R h ε ∥ ≤ z ‾ h \lVert R_h\varepsilon\rVert\le\overline{z}_h ∥ R h ε ∥ ≤ z h give a conditional identification interval of half-width c E r ‾ z ‾ h c_E\overline{r}\overline{z}_h c E r z h around c E a ′ H h ε c_Ea^{\prime}H_h\varepsilon c E a ′ H h ε , before sampling uncertainty. Without such bounds an unobserved direction cannot be made economically small by better measurement of the observed directions. The norm bound allows the least favorable alignment of errors. Independently justified error bounds make it a quantitative, potentially conservative bound.
When a a a or c E c_E c E is estimated, the plug-in error also contains ( c E a ^ − c E a ) ′ H ^ h ε (\widehat{c_Ea} - c_Ea)^{\prime}\widehat{H}_h\varepsilon ( c E a − c E a ) ′ H h ε . Their uncertainty and any common training-data dependence must be propagated jointly. Cross-fitting the spending operator alone does not remove errors-in-variables bias in the financial regression.
Choosing a representation without using the financial outcome. The representation a a a need not be unique. Given C R C_R C R , b = R R ′ l b = R^{\prime}_Rl b = R R ′ l and a positive-definite covariance V g V_g V g for an estimated measurement response, a minimum-variance choice solves
min a a ′ V g a subject to C R ′ a = b , a ∗ = V g − 1 C R ( C R ′ V g − 1 C R ) − 1 b . (38) \min_a\ a^{\prime}V_ga \quad\text{subject to } C_R^{\prime}a = b,\qquad a^* = V_g^{-1}C_R(C_R^{\prime}V_g^{-1}C_R)^{-1}b. \tag*{(38)} a min a ′ V g a subject to C R ′ a = b , a ∗ = V g − 1 C R ( C R ′ V g − 1 C R ) − 1 b . (38)
The dagger is the Moore–Penrose inverse. Feasibility requires b ∈ row ( C R ) b \in\operatorname{row}(C_{\mathsf{R}}) b ∈ row ( C R ) , precisely the spanning condition. The formula follows from the first-order condition V g a = C R μ r e c V_g a = C_{\mathsf{R}}\mu_{\mathrm{rec}} V g a = C R μ rec and the linear constraint. This is a standard constrained quadratic calculation, used here to target the household-demand projection. It does not estimate an unknown l l l or authorize selecting a a a to fit the EBP. Estimate the covariance and any tuning choices in training data. Use nonredundant measurement rows, or a justified covariance regularization, in (29) , since V − 1 V^{-1} V − 1 requires an invertible covariance.
For fixed maps, the covariance criterion selects precision among valid representations, rather than minimizing coefficient size. Adding valid measurement rows cannot increase the optimal population variance when the existing measurements and their covariance block remain unchanged, because the old weights, extended by zeros, remain feasible. Strict improvement requires added rows that load on the response subspace or whose errors correlate with existing measurement errors. Large weights may be unavoidable when a financially important state barely moves the measured outcome. Under approximate spanning, weights must instead balance sampling uncertainty against the projected residual financial direction in (26) .
Which modes require precision? Conditionally on a a a and c E c_E c E , if Σ h \Sigma_h Σ h is the full covariance of ( H ^ h − H h ) ε (\widehat{H}_h-H_h)\varepsilon ( H h − H h ) ε , the variance of the observable projection error is c E 2 a ′ Σ h a c_E^2a^{\prime}\Sigma_h a c E 2 a ′ Σ h a . In the illustrative full-field case C ∗ = Id n C C_*=\operatorname{Id}_{n_C} C ∗ = Id n C and a = l a=l a = l , let that error be U e h U Ue_h^U U e h U in the orthonormal eigenbasis and put l U = U ′ l l^U=U^{\prime}l l U = U ′ l . Then
Var ( p ^ h − p h ) = c E 2 ( l U ) ′ Var ( e h U ) l U = c E 2 ∑ k , j l k U l j U Cov ( e k , h U , e j , h U ) . (39) \operatorname{Var}(\widehat{p}_h-p_h)=c_E^2(l^U)^{\prime}\operatorname{Var}(e_h^U)l^U=c_E^2\sum_{k,j}l_k^U l_j^U\operatorname{Cov}(e_{k,h}^U,e_{j,h}^U). \tag*{(39)} Var ( p h − p h ) = c E 2 ( l U ) ′ Var ( e h U ) l U = c E 2 k , j ∑ l k U l j U Cov ( e k , h U , e j , h U ) . (39)
A sum of individual mode variances is insufficient unless cross-mode covariances vanish. A mode with small unweighted measurement variance can matter financially when ∣ l k U ∣ |l_k^U| ∣ l k U ∣ is large.
Differentiating r D , k h − 1 r_{D,k}^{h-1} r D , k h − 1 while holding ρ D \rho_D ρ D , B D B_D B D , l l l and c E c_E c E fixed gives (30) . Since λ 1 = 0 \lambda_1=0 λ 1 = 0 , the stationary term has zero derivative. The nonstationary products ( l ′ u k ) ( u k ′ B D ε ) (l^{\prime}u_k)(u_k^{\prime}B_D\varepsilon) ( l ′ u k ) ( u k ′ B D ε ) can have either sign, so an increase in diffusion need not increase the financial response even with nonnegative county inputs and loadings.
Observable dynamics and cardinal nonseparability
Timing and continued forcing. The direct consumption responses { Φ h } \{\Phi_h\} { Φ h } of § 2.1 are assumed summable, ∑ h ≥ 0 ∥ Φ h ∥ < ∞ \sum_{h\geq0}\|\Phi_h\|<\infty ∑ h ≥ 0 ∥ Φ h ∥ < ∞ . In the benchmark sequence, the shock at t t t moves the attention field at t t t and the slow demand field at t + 1 t+1 t + 1 , after which propagation is autonomous. A continued-input alternative is
d t + h = η D ∑ j = 0 h − 1 A D h − 1 − j Φ j ε t , h ≥ 1. (40) d_{t+h}=\eta_D\sum_{j=0}^{h-1}A_D^{h-1-j}\Phi_j\varepsilon_t,\qquad h\geq1. \tag*{(40)} d t + h = η D j = 0 ∑ h − 1 A D h − 1 − j Φ j ε t , h ≥ 1. (40)
This follows by substitution into d t + 1 = A D d t + η D ν t d_{t+1}=A_Dd_t+\eta_D\nu_t d t + 1 = A D d t + η D ν t with ν t + j = Φ j ε t \nu_{t+j}=\Phi_j\varepsilon_t ν t + j = Φ j ε t . If one mode has transition r r r but is repeatedly forced at rate r i n r_{\mathrm{in}} r in , its response is proportional to
∑ j = 0 h − 1 r h − 1 − j r i n j = r h − r i n h r − r i n , r ≠ r i n , \sum_{j=0}^{h-1}r^{h-1-j}r_{\mathrm{in}}^j=\frac{r^h-r_{\mathrm{in}}^h}{r-r_{\mathrm{in}}},\qquad r\ne r_{\mathrm{in}}, j = 0 ∑ h − 1 r h − 1 − j r in j = r − r in r h − r in h , r = r in ,
and equals h r h − 1 hr^{h-1} h r h − 1 when r = r i n r=r_{\mathrm{in}} r = r in . Thus a two-exponential profile need not mean two fields. Discrimination requires observed inputs, an input-duration exclusion, or a restricted forcing family tested against the same outcomes. Controls do not make an unrestricted Φ j \Phi_j Φ j disappear. Different transmission graphs supply another competing explanation. For stationary repeated shocks, stability must be combined with suitable stationary inputs and finite innovation moments, because impulse summability alone is not a stochastic existence proof.
Lemma A.4 (Observation restriction). Suppose A D = U diag ( r D , k ) U ′ A_D=U\operatorname{diag}(r_{D,k})U' A D = U diag ( r D , k ) U ′ , C Y = U diag ( c k ) U ′ C_Y=U\operatorname{diag}(c_k)U' C Y = U diag ( c k ) U ′ and B D = U diag ( b k D ) U ′ B_D=U\operatorname{diag}(b^D_k)U' B D = U diag ( b k D ) U ′ *on all excited and observed directions. Let a known linear operator F e , h \mathcal{F}_{e,h} F e , h apply the actual observation mask, temporal filter and residualization to the response, and apply exactly the same operator to every modal design column. Suppose omitted terms are zero or included in the nuisance span, the residualized design has full column rank, and its columns are orthogonal in conditional expectation to the observation error. Then joint least squares identifies the coefficients of the filtered observable model. Without temporal filtering these are c k b k D r D , k h − 1 c_kb^D_kr_{D,k}^{h-1} c k b k D r D , k h − 1 for the spending-loss level at h ≥ 1 h\ge1 h ≥ 1 .
Proof. Expand C Y A D h − 1 B D = ∑ k c k b k D r D , k h − 1 u k u k ′ C_YA_D^{h-1}B_D=\sum_k c_kb^D_kr_{D,k}^{h-1}u_ku_k' C Y A D h − 1 B D = ∑ k c k b k D r D , k h − 1 u k u k ′ . Linearity gives the same expansion after applying F e , h \mathcal{F}_{e,h} F e , h . The rank and orthogonality conditions identify its coefficients by the normal equations. Missing counties need not commute with the graph if the same mask is applied to the correct design. Selecting outcomes or omitting relevant modes can violate orthogonality.
This is a sufficient condition, not a claim that the empirical regressions satisfy it. With general measurement and input loadings,
( U ′ H h L U ) k k = ∑ j ( U ′ C Y U ) k j r D , j h − 1 ( U ′ B D U ) j k . (U'H_h^LU)_{kk}=\sum_j(U'C_YU)_{kj}r_{D,j}^{h-1}(U'B_DU)_{jk}. ( U ′ H h L U ) k k = j ∑ ( U ′ C Y U ) k j r D , j h − 1 ( U ′ B D U ) j k .
Even one field can mix several rates in an observed diagonal entry. Exact diagonal input loading alone simplifies that entry but does not remove all off-diagonal outcome responses from a masked regression. For county-specific direct impacts Φ 0 = diag ( ϕ i ) \Phi_0=\operatorname{diag}(\phi_i) Φ 0 = diag ( ϕ i ) , commuting globally with a connected Laplacian requires ( ϕ i − ϕ j ) L i j = 0 (\phi_i-\phi_j)L_{ij}=0 ( ϕ i − ϕ j ) L ij = 0 along every edge and hence homogeneous ϕ i \phi_i ϕ i . Retained-subspace approximate diagonality is weaker, but needs an explicit residual check.
Levels, growth and the observation filter. Take the adverse field to load on the spending-loss level, H h L = C Y A D h − 1 B D H_h^L = C_Y A_D^{h-1}B_D H h L = C Y A D h − 1 B D for h ≥ 1 h \ge1 h ≥ 1 , with H 0 L = 0 H_0^L = 0 H 0 L = 0 . Then
H 1 Δ = C Y B D , H h Δ = C Y ( A D − I d n C ) A D h − 2 B D ( h ≥ 2 ) . H_1^\Delta= C_YB_D,\qquad H_h^\Delta= C_Y(A_D-\mathrm{Id}_{n_C})A_D^{h-2}B_D \qquad(h \ge2). H 1 Δ = C Y B D , H h Δ = C Y ( A D − Id n C ) A D h − 2 B D ( h ≥ 2 ) .
A scalar decaying level therefore has an initial jump C C C followed by C ( r − 1 ) r h − 2 C(r-1)r^{h-2} C ( r − 1 ) r h − 2 , of the opposite sign when 0 < r < 1 0<r<1 0 < r < 1 . Conversely, a growth response C r h − 1 Cr^{h-1} C r h − 1 accumulates to C ( 1 − r h ) / ( 1 − r ) C(1-r^h)/(1-r) C ( 1 − r h ) / ( 1 − r ) and leaves a permanent displacement. Selecting a same-sign growth tail cannot decide between these level models. If published levels satisfy y ˉ t = ∑ j = 0 J g j f i l t e r y t − j \bar{y}_t=\sum_{j=0}^{J}g_j^{\mathrm{filter}}y_{t-j} y ˉ t = ∑ j = 0 J g j filter y t − j , apply this filter before differencing model predictions. Away from onset a geometric response has multiplier ∑ j g j f i l t e r r − j \sum_j g_j^{\mathrm{filter}}r^{-j} ∑ j g j filter r − j , but onset and a change in g j f i l t e r g_j^{\mathrm{filter}} g j filter require the actual filter. Nonlinear clipping is not a linear observation map. The clipped-growth surface is therefore retained as an estimator diagnostic, not a level-demand kernel.
How much does rank one miss? Let K s t a c k K_{\mathrm{stack}} K stack stack the pricing rows w h K h Σ ε 1 / 2 \sqrt{w_h}K_h\Sigma_\varepsilon^{1/2} w h K h Σ ε 1/2 over fixed horizons, with prespecified nonnegative horizon weights and shock second-moment matrix. For ∥ K s t a c k ∥ F > 0 \lVert K_{\mathrm{stack}}\rVert_F>0 ∥ K stack ∥ F > 0 , the relative best rank-one error is
E 1 2 = min rank ( G ) ≤ 1 ∥ K s t a c k − G ∥ F 2 ∥ K s t a c k ∥ F 2 = 1 − σ 1 ( K s t a c k ) 2 ∑ j σ j ( K s t a c k ) 2 . \mathcal{E}_1^2=\frac{\min_{\operatorname{rank}(G)\le1}\lVert K_{\mathrm{stack}}-G\rVert_F^2}{\lVert K_{\mathrm{stack}}\rVert_F^2}=1-\frac{\sigma_1(K_{\mathrm{stack}})^2}{\sum_j\sigma_j(K_{\mathrm{stack}})^2}. E 1 2 = ∥ K stack ∥ F 2 min rank ( G ) ≤ 1 ∥ K stack − G ∥ F 2 = 1 − ∑ j σ j ( K stack ) 2 σ 1 ( K stack ) 2 .
This is the standard singular-value approximation identity. Nonseparability says the error is generically nonzero, not that it is large, precisely estimated, or improves prediction. It says nothing about ordinal reversals. For example, ( 2 , 1 ) (2,1) ( 2 , 1 ) and ( 3 , 1 ) (3,1) ( 3 , 1 ) are nonproportional but retain the same ordering. Estimated response kernels require held-out loss comparisons and joint uncertainty. With repeated eigenvalues, interpret projected eigenspaces, because individual eigenvectors within them have no unique orientation.
Aggregation allocation and omitted-node memory
Let O h = O ~ h S C I O_h=\widetilde{O}_h^{SCI} O h = O h S C I and let O h , K O_{h,K} O h , K be its approximation. For an actual county shock, the exact identity is
M O h ε − M O h , K R M ε = M ( O h − O h , K ) R M ε + M O h ( I d n C − R M ) ε . (41) MO_h\varepsilon-MO_{h,K}RM\varepsilon=M(O_h-O_{h,K})RM\varepsilon+MO_h(\mathrm{Id}_{n_C}-RM)\varepsilon. \tag*{(41)} M O h ε − M O h , K R M ε = M ( O h − O h , K ) R M ε + M O h ( Id n C − R M ) ε . (41)
The first term concerns approximation of the specified lifted state shock. The second concerns its wrong county allocation. Its norm is at most ∥ M O h ( I d n C − R M ) ∥ 2 ∥ ε ∥ 2 \lVert MO_h(\mathrm{Id}_{n_C}-RM)\rVert_2\lVert\varepsilon\rVert_2 ∥ M O h ( Id n C − R M ) ∥ 2 ∥ ε ∥ 2 , but without bounds on allocation this need not be small. The identity M R = I d n S MR=\mathrm{Id}_{n_S} M R = Id n S imposes no such bound. Similarly,
M A D 2 R − ( M A D R ) 2 = M A D ( I d n C − R M ) A D R . MA_D^2R-(MA_DR)^2=MA_D(\mathrm{Id}_{n_C}-RM)A_DR. M A D 2 R − ( M A D R ) 2 = M A D ( Id n C − R M ) A D R .
Closure requires, for example, A D R = R A ˉ D A_D R = R\bar{A}_D A D R = R A ˉ D with A ˉ D = M A D R \bar{A}_D = M A_D R A ˉ D = M A D R . A right inverse alone does not make state aggregates Markov. The remainder term in Proposition 2 must also be bounded before that proposition provides a quantitative guarantee.
Partition a national field into observed V o b s \mathcal{V}_{\mathrm{obs}} V obs and omitted V o m i t \mathcal{V}_{\mathrm{omit}} V omit counties. Use transition blocks ( A D ) o b s , o b s (A_D)_{\mathrm{obs},\mathrm{obs}} ( A D ) obs , obs , ( A D ) o b s , o m i t (A_D)_{\mathrm{obs},\mathrm{omit}} ( A D ) obs , omit , ( A D ) o m i t , o b s (A_D)_{\mathrm{omit},\mathrm{obs}} ( A D ) omit , obs , ( A D ) o m i t , o m i t (A_D)_{\mathrm{omit},\mathrm{omit}} ( A D ) omit , omit and input blocks ( B D ) o b s (B_D)_{\mathrm{obs}} ( B D ) obs , ( B D ) o m i t (B_D)_{\mathrm{omit}} ( B D ) omit . Eliminating the omitted state gives
z o b s , t + 1 = ( A D ) o b s , o b s z o b s , t + ( B D ) o b s ε t + ( A D ) o b s , o m i t ( A D ) o m i t , o m i t t z o m i t , 0 + ∑ j = 0 t − 1 ( A D ) o b s , o m i t ( A D ) o m i t , o m i t j ( ( A D ) o m i t , o b s z o b s , t − 1 − j + ( B D ) o m i t ε t − 1 − j ) . z_{\mathrm{obs},t+1} = (A_D)_{\mathrm{obs},\mathrm{obs}}z_{\mathrm{obs},t} + (B_D)_{\mathrm{obs}}\varepsilon_t + (A_D)_{\mathrm{obs},\mathrm{omit}}(A_D)_{\mathrm{omit},\mathrm{omit}}^t z_{\mathrm{omit},0} + \sum_{j=0}^{t-1}(A_D)_{\mathrm{obs},\mathrm{omit}}(A_D)_{\mathrm{omit},\mathrm{omit}}^j\left((A_D)_{\mathrm{omit},\mathrm{obs}}z_{\mathrm{obs},t-1-j} + (B_D)_{\mathrm{omit}}\varepsilon_{t-1-j}\right). z obs , t + 1 = ( A D ) obs , obs z obs , t + ( B D ) obs ε t + ( A D ) obs , omit ( A D ) omit , omit t z omit , 0 + j = 0 ∑ t − 1 ( A D ) obs , omit ( A D ) omit , omit j ( ( A D ) omit , obs z obs , t − 1 − j + ( B D ) omit ε t − 1 − j ) .
This follows by solving the omitted recursion and substituting it into the observed one. Even a single first-order national field gives the observed subsystem distributed memory. Inducing and renormalizing a graph on observed outcomes is not equivalent to selecting observations after propagation. Source completion helps when it restores measured inputs and paths, but does not certify that a retained graph is the national economic system.
No-diffusion comparison, sharp common-direction bounds and relative demand
Equation (18) follows by subtracting the fixed-input counterfactual A D 0 = ρ D I d n C A_D^0 = \rho_D I_{d n_C} A D 0 = ρ D I d n C . In graph coordinates it is
Δ K h d i f f = c E ρ D h − 1 ∑ k ≥ 2 ( e − κ D ( h − 1 ) λ k − 1 ) ( l ′ u k ) u k ′ B D . \Delta K_h^{\mathit{diff}} = c_E \rho_D^{h-1}\sum_{k\ge2}\left(e^{-\kappa_D(h-1)\lambda_k}-1\right)(l^\prime u_k)u_k^\prime B_D. Δ K h diff = c E ρ D h − 1 k ≥ 2 ∑ ( e − κ D ( h − 1 ) λ k − 1 ) ( l ′ u k ) u k ′ B D .
The stationary term is zero, but a nonstationary exposure is not itself this difference. At h = 1 h=1 h = 1 the entire difference is zero even with unequal initial financial exposures. The economic contrast holds inputs, financial loadings and price impact fixed. Thus, if B D B_D B D already contains nonlocal transmission, it removes subsequent diffusion only. Changing topology is a separate intervention and needs a matched degree/input benchmark.
Proposition A.5 (A missing common direction). Suppose C Y = I d n C C_Y = I_{d n_C} C Y = I d n C , ker T = span { 1 } \ker T = \operatorname{span}\{1\} ker T = span { 1 } and the contrast g = T z g = Tz g = T z is known. Choose the unique z ⊥ z_\perp z ⊥ with T z ⊥ = g Tz_\perp= g T z ⊥ = g and 1 ′ z ⊥ = 0 1^\prime z_\perp= 0 1 ′ z ⊥ = 0 . Then every observationally equivalent field is z = z ⊥ + α 1 z = z_\perp+ \alpha1 z = z ⊥ + α 1 . Given known l l l with 1 ′ l > 0 1^\prime l > 0 1 ′ l > 0 and an independently justified interval α ∈ [ α ‾ , α ‾ ] \alpha\in[\underline{\alpha},\overline{\alpha}] α ∈ [ α , α ] , the sharp identified interval for demand pressure is
[ l ′ z ⊥ + ( 1 ′ l ) α ‾ , l ′ z ⊥ + ( 1 ′ l ) α ‾ ] . [l^\prime z_\perp+ (1^\prime l)\underline{\alpha}, l^\prime z_\perp+ (1^\prime l)\overline{\alpha}]. [ l ′ z ⊥ + ( 1 ′ l ) α , l ′ z ⊥ + ( 1 ′ l ) α ] . Without a bound on α \alpha α , that functional is unbounded even with error-free contrasts. If 1 ′ l < 0 1^\prime l < 0 1 ′ l < 0 the two endpoints change places. If 1 ′ l = 0 1^\prime l = 0 1 ′ l = 0 the functional equals l ′ z ⊥ l^\prime z_\perp l ′ z ⊥ and is identified without any bound.
Proof. The null space characterizes every solution. Substitution gives an affine function of α \alpha α with slope 1 ′ l \mathbf{1}^{\prime}l 1 ′ l . It is strictly monotone when the slope is nonzero, so both endpoints and every point between them are attained, and it is constant when the slope is zero. Sharpness is relative to these restrictions, without additional county bounds. □
The interval separates a missing direction from sampling error. It is not an empirical bound until the common-response bounds and l l l are independently supported. Observed national spending with an unrestricted common nuisance cannot supply those bounds. With additional restrictions such as county-level sign or magnitude limits, intersect the feasible set before taking extrema. Estimation error in contrasts, loadings and price impact requires simultaneous confidence sets rather than plugging in point estimates.
For two assets with known claim-demand loadings l a l_a l a , l b l_b l b and 1 ′ l b ≠ 0 \mathbf{1}^{\prime}l_b \ne0 1 ′ l b = 0 , let
r a b l o a d = 1 ′ l a 1 ′ l b , Δ l a b = l a − r a b l o a d l b . r_{ab}^{\mathrm{load}} = \frac{\mathbf{1}^{\prime}l_a}{\mathbf{1}^{\prime}l_b}, \qquad\Delta l_{ab} = l_a-r_{ab}^{\mathrm{load}}l_b. r ab load = 1 ′ l b 1 ′ l a , Δ l ab = l a − r ab load l b .
Then 1 ′ Δ l a b = 0 \mathbf{1}^{\prime}\Delta l_{ab}=0 1 ′ Δ l ab = 0 , so the contrast ( Δ l a b ) ′ z (\Delta l_{ab})^{\prime}z ( Δ l ab ) ′ z is identified by T z Tz T z on this unrestricted state space even when the two aggregate pressures are not. This is a relative-demand result, not automatically a relative-price result. Let c a c l a i m = P a c a c_a^{\mathrm{claim}}=P_ac_a c a claim = P a c a and c b c l a i m = P b c b c_b^{\mathrm{claim}}=P_bc_b c b claim = P b c b be the claim-unit market maps corresponding to the dollar-unit maps in § 7 . If these maps are separately identified and competing capital and payoff components are removed,
Δ p a c a c l a i m − r a b l o a d Δ p b c b c l a i m = ( Δ l a b ) ′ z . \frac{\Delta p_a}{c_a^{\mathrm{claim}}}-r_{ab}^{\mathrm{load}}\frac{\Delta p_b}{c_b^{\mathrm{claim}}}=(\Delta l_{ab})^{\prime}z. c a claim Δ p a − r ab load c b claim Δ p b = ( Δ l ab ) ′ z .
A multivariate absorption matrix requires the corresponding known linear transformation. Neither the present municipal pilot nor dividend income supplies the required holdings map, demand sensitivities or absorption. This extension does not rescue national EBP.
Assumptions, units and uncertainty
The conditional equilibrium is restricted to positive capital and claim prices, interior household demand, fixed payoff moments and a specified smooth regime. Its uniqueness at fixed capital is not global uniqueness with endogenous capital. The surplus-wealth illustration is first order, since expanding x i = ( s ˉ i − m i z i ) q / ( γ i σ i 2 ) x_i=(\bar{s}_i-m_i z_i)q/(\gamma_i\sigma_i^2) x i = ( s ˉ i − m i z i ) q / ( γ i σ i 2 ) also produces − m i z i ( q − q ˉ ) / ( γ i σ i 2 ) -m_i z_i(q-\bar{q})/(\gamma_i\sigma_i^2) − m i z i ( q − q ˉ ) / ( γ i σ i 2 ) . The second-order coefficients in Proposition 4 belong to the primitive linear-demand schedule, not a second-order derivation of that illustration. A stationary capital law further requires ∣ ρ K ∣ < 1 |\rho_K|<1 ∣ ρ K ∣ < 1 , suitable innovations and a flow/payout rule consistent with its accounts.
Joint response fitting should minimize a quadratic loss on the outcome coefficients without
Table A.1. Notation needed to connect the empirical response to financial demand
sign- or size-based selection, with their full cross-mode and cross-horizon covariance. Near-zero coefficients are not logged or discarded by sign. Covariance regularization is selected using training data, and stability constraints are maintained restrictions, not evidence for diffusion. Whole-calendar blocks preserve common shocks across recipients and overlapping event windows. However, conditional loss resampling does not cover re-estimation and model selection. A full procedure must repeat those steps inside every inference or signal-injection replication. Oracle and feasible power must use the same observation masks, continued-input alternatives and economic signal scale. An uninformative held-out comparison does not reject propagation.
Model details
Retained-mode diagonality
Let U K U_K U K collect the first K K K eigenvectors and U K c U_{K^c} U K c the rest. Assumption 2 states that U K ′ Φ 0 U K = diag ( ϕ 0 , k ) k ≤ K + E Φ , K U_K^\prime\Phi_0U_K = \operatorname{diag}(\phi_{0,k})_{k\leq K} + E_{\Phi,K} U K ′ Φ 0 U K = diag ( ϕ 0 , k ) k ≤ K + E Φ , K and U K ′ B U K = diag ( b k a t t ) k ≤ K + E B , K U_K^\prime BU_K = \operatorname{diag}(b_k^{\mathrm{att}})_{k\leq K} + E_{B,K} U K ′ B U K = diag ( b k att ) k ≤ K + E B , K , with ∥ E Φ , K ∥ 2 \lVert E_{\Phi,K}\rVert_2 ∥ E Φ , K ∥ 2 and ∥ E B , K ∥ 2 \lVert E_{B,K}\rVert_2 ∥ E B , K ∥ 2 small relative to the diagonal entries, and that the cross-block terms U K c ′ Φ 0 U K U_{K^c}^\prime\Phi_0U_K U K c ′ Φ 0 U K , U K ′ Φ 0 U K c U_K^\prime\Phi_0U_{K^c} U K ′ Φ 0 U K c , U K c ′ B U K U_{K^c}^\prime BU_K U K c ′ B U K and U K ′ B U K c U_K^\prime BU_{K^c} U K ′ B U K c are small relative to the retained block. The exact diagonal benchmark sets all off-diagonal terms to zero. The assumption restricts retained and cross-block terms, not the entire discarded block. Spending-side closure diagnostics are proxies for these magnitudes rather than identifying every structural entry. The same conditions are assumed for the production basis.
The production field
The competing slow field runs on a symmetric production affinity A P A^P A P with positive degrees D P D^P D P :
d t + 1 P = A D P d t P + η D P ν t , A D P ≡ ρ D P e − κ D P L P , L P ≡ I d n C − ( D P ) − 1 / 2 A P ( D P ) − 1 / 2 , (42) d^P_{t+1}=A_D^P d^P_t+\eta_D^P\nu_t,\qquad A_D^P\equiv\rho_D^P e^{-\kappa_D^P\mathcal{L}^P},\qquad\mathcal{L}^P\equiv\mathrm{Id}_{n_C}-(D^P)^{-1/2}A^P(D^P)^{-1/2}, \tag*{(42)} d t + 1 P = A D P d t P + η D P ν t , A D P ≡ ρ D P e − κ D P L P , L P ≡ Id n C − ( D P ) − 1/2 A P ( D P ) − 1/2 , (42)
with ρ D P ∈ ( 0 , 1 ) \rho_D^P\in(0,1) ρ D P ∈ ( 0 , 1 ) , κ D P > 0 \kappa_D^P>0 κ D P > 0 , η D P ≥ 0 \eta_D^P\geq0 η D P ≥ 0 . The empirical exposures are the directed operators W u p W^{up} W u p and W d o w n W^{down} W d o w n of § 4 , while the spectral analysis uses their symmetrized affinity. Excluding production from the household-demand equation is a restriction, not evidence that production cannot affect capital, payoff risk or credit premia. Social specificity therefore requires a matched incremental comparison on common shocks, counties, targets and information sets. Separate identification weakens when the leading modal spaces are nearly collinear. The canonical correlations between the retained SCI and production modal spaces, and the projection distance ∥ U K U K ′ − U K u p ( U K u p ) ′ ∥ F \lVert U_KU_K^\prime-U_K^{up}(U_K^{up})^\prime\rVert_F ∥ U K U K ′ − U K u p ( U K u p ) ′ ∥ F with its downstream analogue, measure this.
The directed benchmark
The operator formulation does not need a symmetric graph. The directed benchmark replaces L \mathcal{L} L by Id n C − W \operatorname{Id}_{n_C}-W Id n C − W . Its eigenbasis is generally bi-orthogonal, and if W W W is non-normal its eigenvectors can be ill-conditioned and transient amplification can matter. The diagnostic δ N ( W ) ≡ ∥ W W ′ − W ′ W ∥ F / ∥ W ∥ F 2 \delta_N(W)\equiv\lVert WW^{\prime}-W^{\prime}W\rVert_F/\lVert W\rVert_F^2 δ N ( W ) ≡ ∥ W W ′ − W ′ W ∥ F / ∥ W ∥ F 2 , together with the condition number of the retained basis, diagnoses that eigenvector representation, not whether the directed transition is usable. A matrix exponential or Schur calculation avoids inverting an unstable eigenbasis.
Symmetrization imposes an economic restriction on the direction of transmission and changes the transition matrix. If A = ( W + W ′ ) / 2 A=(W+W^{\prime})/2 A = ( W + W ′ ) /2 , the reversible walk is W r e v = D − 1 A W^{\mathrm{rev}}=D^{-1}A W rev = D − 1 A , and D − 1 / 2 A D − 1 / 2 = D 1 / 2 W r e v D − 1 / 2 D^{-1/2}AD^{-1/2}=D^{1/2}W^{\mathrm{rev}}D^{-1/2} D − 1/2 A D − 1/2 = D 1/2 W rev D − 1/2 . Thus the diffusion kernel’s first-order neighbor term is a degree-rescaled version of W r e v W^{\mathrm{rev}} W rev , not the pruned, row-normalized empirical W W W . A one-hop spending coefficient is not an estimate of κ D \kappa_D κ D .
Path weights and behavioral strength. If W W W is a neighbor-averaging rule, a two-step response uses ( W 2 ) i j = ∑ k W i k W k j (W^2)_{ij}=\sum_k W_{ik}W_{kj} ( W 2 ) ij = ∑ k W ik W k j , so link weights multiply along each path and parallel paths add. For nonnegative row-stochastic W W W , W 2 1 = 1 W^2\mathbf{1}=\mathbf{1} W 2 1 = 1 . Hence a small product on one path does not imply loss of total row weight. These are weighted walks, including repeated visits, rather than counts of distinct friends. The SCI becomes such a transition only under a behavioral interpretation and a specified normalization.
For example, with z t + 1 = W h o p z t z_{t+1}=W_{\mathrm{hop}}z_t z t + 1 = W hop z t and W h o p = ( 1 − α ) Id n C + α W W_{\mathrm{hop}}=(1-\alpha)\operatorname{Id}_{n_C}+\alpha W W hop = ( 1 − α ) Id n C + α W , 0 < α < 1 0<\alpha<1 0 < α < 1 , the binomial expansion gives
W h o p h = ∑ j = 0 h ω j h o p ( h ) W j , ω j h o p ( h ) = ( h j ) α j ( 1 − α ) h − j , ω 2 h o p ( h ) ω 1 h o p ( h ) = ( h − 1 ) α 2 ( 1 − α ) , h ≥ 2. (43) W_{\mathrm{hop}}^h=\sum_{j=0}^{h}\omega_j^{\mathrm{hop}}(h)W^j,\qquad\omega_j^{\mathrm{hop}}(h)=\binom{h}{j}\alpha^j(1-\alpha)^{h-j},\qquad\frac{\omega_2^{\mathrm{hop}}(h)}{\omega_1^{\mathrm{hop}}(h)}=\frac{(h-1)\alpha}{2(1-\alpha)},\qquad h\ge2. \tag*{(43)} W hop h = j = 0 ∑ h ω j hop ( h ) W j , ω j hop ( h ) = ( j h ) α j ( 1 − α ) h − j , ω 1 hop ( h ) ω 2 hop ( h ) = 2 ( 1 − α ) ( h − 1 ) α , h ≥ 2. (43)
Thus relative hop contributions depend on the response speed and horizon, not an arbitrary discount on a second connection. Multiplying the whole transition by a persistence factor leaves this ratio unchanged. Halving a regressor with an unrestricted coefficient merely doubles that coefficient, while fitted values and its t t t -statistic are unchanged. Distinguishing transmission laws therefore requires restrictions across horizons or measured behavior, rather than rescaling an exposure.
Stable-sign support
Definition B.1 (Stable-sign support). Under a single channel the outcome-side modal response is θ k Y ( h ) = C k Y r k h − h 0 \theta_k^Y(h)=C_k^Y r_k^{h-h_0} θ k Y ( h ) = C k Y r k h − h 0 with r k = ρ e − κ λ k ∈ ( 0 , 1 ) r_k=\rho e^{-\kappa\lambda_k}\in(0,1) r k = ρ e − κ λ k ∈ ( 0 , 1 ) on horizons where one channel dominates and entry horizon h 0 h_0 h 0 . Within mode k k k it has the constant sign of C k Y C_k^Y C k Y and decays geometrically. The stable-sign support is, for each retained
:::`
mode, the longest contiguous run of horizons with the same sign, subject to a minimum run length and a signal threshold c s i g c_{\mathrm{sig}} c sig fixed in advance.
Proposition B.2 (Identification on the support). If θ k Y ( h ) = C k Y ρ k h − h 0 \theta_k^Y(h) = C_k^Y \rho_k^{h-h_0} θ k Y ( h ) = C k Y ρ k h − h 0 holds on nonzero cells, then on a fixed support log ∣ θ k Y ( h ) ∣ = a k + h log ρ − κ h λ k \log\lvert\theta_k^Y(h)\rvert= a_k + h \log\rho- \kappa h \lambda_k log ∣ θ k Y ( h )∣ = a k + h log ρ − κh λ k . The slopes ( log ρ , κ ) (\log\rho,\kappa) ( log ρ , κ ) are identified if the columns h h h and h λ k h\lambda_k h λ k have rank two after mode fixed effects. Two distinct eigenvalues with within-mode horizon variation suffice, whereas constant signs alone do not.
Proof. Take logs and absorb the timing shift into a k a_k a k . After removing mode intercepts, full column rank determines the two slopes uniquely.
In population, incompatibility with the admissible law rejects that law, not all propagation. An inadmissible estimated slope is not by itself a size-controlled rejection, and a support chosen from noisy outcomes must be re-selected within training and inside the inference procedure. The two-step estimator reported in Table D.4 fits each mode’s slope on its support by weighted least squares and then regresses the slopes on λ k \lambda_k λ k . With three cells per mode, a per-mode standard error can be close to zero by chance, which is how one mode came to carry almost all of the weight (Appendix § 14.2 .§ 1 ).
Capital dynamics
Intermediary capital obeys the accounting identity
N t + 1 = R f N t + θ t ( V t + 1 − R f P t ) + F t + 1 K , (44) N_{t+1} = R_f N_t + \theta_t(V_{t+1} - R_f P_t) + F^K_{t+1}, \tag*{(44)} N t + 1 = R f N t + θ t ( V t + 1 − R f P t ) + F t + 1 K , (44)
where F K F^K F K collects net external flows. At fixed current capital, expected trading earnings θ t q t \theta_t q_t θ t q t rise with pressure in both regions, with ∂ S ( θ q ) = θ G S + q θ S > 0 \partial_S(\theta q) = \theta G_S + q\theta_S > 0 ∂ S ( θ q ) = θ G S + q θ S > 0 , θ S = N / ( N + ζ Γ ) \theta_S = N/(N + \zeta\Gamma) θ S = N / ( N + ζ Γ ) in the slack region and 0 0 0 when the cap binds. Inherited mark-to-market losses can still reduce capital along an endogenous path. These are different derivatives, and neither identifies a capital law.
Within one smooth regime, let k t k_t k t be the deviation of capital from its reference. Its law and the premium relation are
k t + 1 = ρ K k t + c K ′ z t + D K ε t + u t + 1 , q ^ t = g q ′ z t − b K k t , g q = G S l , b K = − G N > 0. (45) k_{t+1} = \rho_K k_t + c'_K z_t + D_K \varepsilon_t + u_{t+1}, \qquad\widehat{q}_t = g'_q z_t - b_K k_t, \qquad g_q = G_S l, \ b_K = -G_N > 0. \tag*{(45)} k t + 1 = ρ K k t + c K ′ z t + D K ε t + u t + 1 , q t = g q ′ z t − b K k t , g q = G S l , b K = − G N > 0. (45)
The vector c K c_K c K depends on retention, payouts, inherited exposures and capital flows. With
A a u g = ( A D 0 c K ′ ρ K ) , B a u g = ( B D D K ) , C a u g E = b E ( g q ′ , − b K ) , (46) A_{\mathrm{aug}} =
\begin{pmatrix}
A_D & 0 \\
c'_K & \rho_K
\end{pmatrix},
\qquad
B_{\mathrm{aug}} =
\begin{pmatrix}
B_D \\
D_K
\end{pmatrix},
\qquad
C^E_{\mathrm{aug}} = b_E(g'_q,-b_K), \tag*{(46)} A aug = ( A D c K ′ 0 ρ K ) , B aug = ( B D D K ) , C aug E = b E ( g q ′ , − b K ) , (46)
the EBP impulse response is I R h E B P = C a u g E A a u g h − 1 B a u g IR_h^{EBP}=C_{\mathrm{aug}}^{E}A_{\mathrm{aug}}^{h-1}B_{\mathrm{aug}} I R h E B P = C aug E A aug h − 1 B aug . With D K = 0 D_K=0 D K = 0 , block-matrix powers give
I R h E B P = b E [ g q ′ A D h − 1 B D − b K ∑ j = 0 h − 2 ρ K h − 2 − j c K ′ A D j B D ] , (47) IR_h^{EBP}=b_E\left[g_q^{\prime}A_D^{h-1}B_D-b_K\sum_{j=0}^{h-2}\rho_K^{h-2-j}c_K^{\prime}A_D^jB_D\right], \tag*{(47)} I R h E B P = b E [ g q ′ A D h − 1 B D − b K j = 0 ∑ h − 2 ρ K h − 2 − j c K ′ A D j B D ] , (47)
where the sum is empty at h = 1 h=1 h = 1 . Capital can offset or reinforce the fixed-capital kernel. A scalar example shows why autocorrelations do not sign the dynamics. Let q t = a z z t − a k k t q_t=a_z z_t-a_k k_t q t = a z z t − a k k t with a z , a k > 0 a_z,a_k>0 a z , a k > 0 , z t + 1 = r z t + u t + 1 z z_{t+1}=r z_t+u_{t+1}^{z} z t + 1 = r z t + u t + 1 z and k t + 1 = ρ K k t − c z z t + u t + 1 K k_{t+1}=\rho_K k_t-c_z z_t+u_{t+1}^{K} k t + 1 = ρ K k t − c z z t + u t + 1 K . Then
q t + 1 = ρ K q t + [ a z ( r − ρ K ) + a k c z ] z t + a z u t + 1 z − a k u t + 1 K . (48) q_{t+1}=\rho_K q_t+\left[a_z(r-\rho_K)+a_kc_z\right]z_t+a_z u_{t+1}^{z}-a_k u_{t+1}^{K}. \tag*{(48)} q t + 1 = ρ K q t + [ a z ( r − ρ K ) + a k c z ] z t + a z u t + 1 z − a k u t + 1 K . (48)
If r < ρ K r<\rho_K r < ρ K , the lagged-field coefficient is negative at c z = 0 c_z=0 c z = 0 and becomes positive only if a k c z > a z ( ρ K − r ) a_kc_z>a_z(\rho_K-r) a k c z > a z ( ρ K − r ) . Neither coefficient alone signs the total change q t + 1 − q t q_{t+1}-q_t q t + 1 − q t . The autocorrelations of the EBP and of an estimated field do not identify ρ K \rho_K ρ K or c z c_z c z .
The units bridge
Holdings and S S S are quantities of the risky claim, while P ˉ S \bar{P}S P ˉ S is the corresponding fixed-price dollar demand shift. It is not executed dollar selling, because the price response induces offsetting household demand as in main-text equation (13) . Holding E [ V ] E[V] E [ V ] and R f R_f R f fixed in q = E [ V ] − R f P q=E[V]-R_fP q = E [ V ] − R f P gives Δ q = − R f Δ P \Delta q=-R_f\Delta P Δ q = − R f Δ P . Modified duration implies Δ y i e l d ≃ − Δ P / ( P D m o d ) \Delta\mathrm{yield}\simeq-\Delta P/(P D_{\mathrm{mod}}) Δ yield ≃ − Δ P / ( P D mod ) . Appendix § 16.4 .§ 3 combines this relation with residual absorption to convert a fixed-price dollar demand shift into basis points, (H.9), and defines its symbols. For EBP measured in basis points, the local loading in the pricing kernel is b E = 10 4 φ E / ( R f P D ˉ m o d ) b_E=10^4\varphi_E/(R_fP\bar{D}_{\mathrm{mod}}) b E = 1 0 4 φ E / ( R f P D ˉ mod ) under these fixed-payoff assumptions. It is not a unit-free equality between EBP and q q q . Market coverage, issuer mix and market-value conventions must match, and a separate fire-sale multiplier would count feedback already in ξ P \xi_P ξ P . Economy-wide net trades sum to zero, so total trading volume is not the demand shift.
In the surplus-wealth illustration (10) , if observed spending loss is c i Y z i c_i^Y z_i c i Y z i with c i Y > 0 c_i^Y>0 c i Y > 0 , the fixed-price credit-demand shift per dollar of spending is λ i C = P ˉ q ˉ m i / ( γ i σ i 2 C i Y ) \lambda_i^C=\bar{P}\bar{q}m_i/(\gamma_i\sigma_i^2 C_i^Y) λ i C = P ˉ q ˉ m i / ( γ i σ i 2 C i Y ) , which requires evidence on wealth loadings, spending loadings, participation and portfolio risk that is independent of the EBP. For independently justified nonnegative bounds on the parameters of (H.9), the smallest response uses the lower bounds on φ E \varphi_E φ E , λ C \lambda_C λ C , C l o s s C_{\mathrm{loss}} C loss and the upper bounds on A m k t A_{\mathrm{mkt}} A mkt , ξ P \xi_P ξ P , D m o d D_{\mathrm{mod}} D mod , and the largest reverses them. If λ C \lambda_C λ C or φ E \varphi_E φ E can be zero, the lower bound is zero.
Power planning
In the Gaussian linear model with n o b s n_{\mathrm{obs}} n obs stacked observations, y = X s i g θ s i g + Z c t r l ϕ c t r l + v p l a n y = X^{\mathrm{sig}}\theta^{\mathrm{sig}} + Z^{\mathrm{ctrl}}\phi^{\mathrm{ctrl}} + v^{\mathrm{plan}} y = X sig θ sig + Z ctrl ϕ ctrl + v plan with known Σ = Var ( v p l a n ) > 0 \Sigma= \operatorname{Var}(v^{\mathrm{plan}}) > 0 Σ = Var ( v plan ) > 0 , let X w s i g = Σ − 1 / 2 X s i g X_w^{\mathrm{sig}} = \Sigma^{-1/2}X^{\mathrm{sig}} X w sig = Σ − 1/2 X sig and Z w c t r l = Σ − 1 / 2 Z c t r l Z_w^{\mathrm{ctrl}} = \Sigma^{-1/2}Z^{\mathrm{ctrl}} Z w ctrl = Σ − 1/2 Z ctrl . Profiling out the nuisance parameters, I = ( X w s i g ) ′ ( I n o b s − P Z w c t r l ) X w s i g \mathcal{I} = (X_w^{\mathrm{sig}})'(I_{n_{\mathrm{obs}}} - P_{Z_w^{\mathrm{ctrl}}})X_w^{\mathrm{sig}} I = ( X w sig ) ′ ( I n obs − P Z w ctrl ) X w sig , and an unbiased estimator satisfies Var ( a ′ θ ^ s i g ) ≥ a ′ I + a \operatorname{Var}(a'\hat{\theta}^{\mathrm{sig}}) \ge a'\mathcal{I}^{+}a Var ( a ′ θ ^ sig ) ≥ a ′ I + a for a ∈ range ( I ) a \in\operatorname{range}(\mathcal{I}) a ∈ range ( I ) . Noise, weak excitation of the relevant modes and collinearity with nuisance regressors are separate sources of lost information.
If the estimated state is x ^ = x + u \hat{x} = x + u x ^ = x + u with independent classical error, unconditional reliability Rel \operatorname{Rel} Rel , and controls that explain a share R x 2 R_x^2 R x 2 of the true state’s variance, then
Rel c o n d = 1 − R x 2 1 − R x 2 + ( 1 − Rel ) / Rel , plim β ^ n a i v e = β Rel c o n d . (49) \operatorname{Rel}_{\mathrm{cond}} = \frac{1-R_x^2}{1-R_x^2+(1-\operatorname{Rel})/\operatorname{Rel}}, \qquad\operatorname{plim}\hat{\beta}_{\mathrm{naive}} = \beta\operatorname{Rel}_{\mathrm{cond}}. \tag*{(49)} Rel cond = 1 − R x 2 + ( 1 − Rel ) / Rel 1 − R x 2 , plim β ^ naive = β Rel cond . (49)
Regression calibration corrects the attenuation but leaves the zero-slope statistic unchanged.
For a perfectly measured state, effect B B B , outcome noise σ \sigma σ and equivalence band ± c e q B \pm c_{\mathrm{eq}}B ± c eq B , a normal planning approximation is
n ≳ 1 1 − R x 2 [ ( z 1 − α / 2 + z 0.90 ) σ c e q B ] 2 . (50) n \gtrsim\frac{1}{1-R_x^2}\left[\frac{(z_{1-\alpha/2}+z_{0.90})\sigma}{c_{\mathrm{eq}}B}\right]^2. \tag*{(50)} n ≳ 1 − R x 2 1 [ c eq B ( z 1 − α /2 + z 0.90 ) σ ] 2 . (50)
At α = 0.025 \alpha= 0.025 α = 0.025 , c e q = 0.5 c_{\mathrm{eq}} = 0.5 c eq = 0.5 and B = σ = 10 B = \sigma= 10 B = σ = 10 basis points this gives about 50 independent episodes with R x 2 = 0 R_x^2 = 0 R x 2 = 0 and 249 with R x 2 = 0.8 R_x^2 = 0.8 R x 2 = 0.8 . In a simulation with an imposed 10-basis-point effect, 10 basis points of noise and 60 independent episodes, joint detection-and-calibration power is 84.6 percent with a perfect state and orthogonal controls, 0.2 percent when controls explain 80 percent of the state’s variance, and zero when the state’s reliability is also 0.03. These are sensitivities of the stated model, not estimates of this paper’s power.
The common-response identity in main-text equation (20) has a geometric counterpart. For a propagated field z e ( τ ) = e − τ L b e z_e(\tau) = e^{-\tau L}b_e z e ( τ ) = e − τ L b e and recipient set R e R_e R e , an event intercept removes the fraction
f e m e a n ( τ ) = ∣ R e ∣ z ˉ e , R e ( τ ) 2 ∑ i ∈ R e z e i ( τ ) 2 . (51) f_e^{\mathrm{mean}}(\tau) = \frac{\lvert R_e\rvert\bar{z}_{e,R_e}(\tau)^2}{\sum_{i\in R_e}z_{ei}(\tau)^2}. \tag*{(51)} f e mean ( τ ) = ∑ i ∈ R e z e i ( τ ) 2 ∣ R e ∣ z ˉ e , R e ( τ ) 2 . (51)
of its squared norm. This is a projection onto a constant within one event, not the variation removed jointly by all regression controls and fixed effects. For a connected symmetric normalized Laplacian with positive degrees d i d e g d_i^{\mathrm{deg}} d i deg and a nonzero nonnegative seed, the limiting field is proportional to d i d e g \sqrt{d_i^{\mathrm{deg}}} d i deg .
Hence
f e m e a n ( ∞ ) = ( ∣ R e ∣ − 1 ∑ i ∈ R e d i d e g ) 2 ∣ R e ∣ − 1 ∑ i ∈ R e d i d e g . (52) f_e^{\mathrm{mean}}(\infty)=\frac{\left(|R_e|^{-1}\sum_{i\in R_e}\sqrt{d_i^{\mathrm{deg}}}\right)^2}{|R_e|^{-1}\sum_{i\in R_e}d_i^{\mathrm{deg}}}. \tag*{(52)} f e mean ( ∞ ) = ∣ R e ∣ − 1 ∑ i ∈ R e d i deg ( ∣ R e ∣ − 1 ∑ i ∈ R e d i deg ) 2 . (52)
The limit need not equal one, because the stationary field of this symmetric normalization is not constant.
Table B.1 illustrates this calculation on the 1,590-county graph, using onset-week recorded-damage shares as seeds and the primary expanded-footprint recipient screen. Of the paper’s 48 declarations, 38 have positive recorded-damage seed mass on this graph. The other ten have their positive recorded damage outside the retained nodes and are excluded from this illustration, not from the spending regression. The shares describe the retained recorded-damage construction, not complete physical shock support.
Table B.1. Event-intercept removal of a diffusion-kernel field. Notes: Percentages of squared field norm, across 38 declarations. All columns are in percent. The operator is e − τ L e^{-\tau L} e − τ L throughout, with A = ( W + W ′ ) / 2 A=(W+W')/2 A = ( W + W ′ ) /2 and L = I d n C − D − 1 / 2 A D − 1 / 2 L=I d_{n_C}-D^{-1/2}AD^{-1/2} L = I d n C − D − 1/2 A D − 1/2 . The limit is analytic. Diffusion time is not a calendar horizon, and no decay speed or spending response is fitted here.
The rising median illustrates how mixing can weaken within-event contrasts as a field approaches its stationary shape. It complements the aggregation identity, which shows that weighted aggregates and within-event contrasts retain different directions of the same field. It is neither a monotonicity result for every seed nor evidence that the observed spending response follows diffusion.
Data and local evidence
Sample
The local spending comparison holds the recipient sample fixed across recorded-loss constructions. It contains 74,779 county–declaration observations, 1,574 recipient counties and 48 declarations. Recipients lie outside the expanded recorded storm footprint and have all required spending levels from thirteen weeks before to twelve weeks after onset. The index is used without standard-deviation normalization. Its index-point unit is defined in Appendix § 16.3 .§ 1 . Table 2 reports the matched four-week regressions. Declarations sharing a physical episode are grouped in the episode-bootstrap sensitivity, rather than treated as separate shock realizations.
Table C.1. Exposure support and recorded damage across source constructions
Notes: Panel A uses the same recipient–declaration observations in all rows. Zero-exposure shares are 87.35, 86.25, 86.13 and 86.13 percent, respectively. Onset uses original onset-week county reports. The other constructions use eligible reports within the declared incident assigned through week 4. Mapped zones link uniquely to eligible counties, and the last row assigns multiday reports at their end rather than start. Panel B weights each declaration once and gives recorded property-damage subtotals. A zero recorded subtotal for one declaration is not an assessed zero loss, so unknown amounts remain flagged. Currency conversion uses each report month. Original onset loss totals are omitted because their conversion was event-level rather than record-month-level.
Persistence
The event-time comparison in Appendix § 16.3 .§ 7 uses the same source construction and recipient controls as the four-week regression. It holds the four-week source geography fixed across horizons, while the separate dated-input analysis in Appendix § 16.3 .§ 4 distinguishes inputs at onset from later destruction. The four-week spatial gradient does not by itself determine a decay rate or establish the persistence of a single impulse.
Identification diagnostics
Measured own-weather controls retain a broad recipient sample while accounting for observed rain, temperature, gusts, warnings and reports. Appendix § 16.3 .§ 2 describes these variables. The pre-event, leave-one-declaration-out and alternative-baseline comparisons are reported in Appendix § 16.3 .§ 7 . These diagnostics assess timing and influential episodes. In contrast, the source-construction agreement in the main table addresses a different measurement question. Together they support a spatial spending association, rather than a fully isolated causal social channel.
Economic Impact Payments
The county-month panel uses the 2020–2021 Economic Impact Payment rounds as remote demand inputs. For this comparison I fix the spatial weighting at J E I P ( W ) = ∑ ℓ = 1 6 0.65 ℓ − 1 W ℓ J_{\mathrm{EIP}}(W)=\sum_{\ell=1}^{6}0.65^{\ell-1}W^{\ell} J EIP ( W ) = ∑ ℓ = 1 6 0.6 5 ℓ − 1 W ℓ and normalize each row before aggregating weights to source states. I then exclude the recipient’s state and states whose nearest county centroid is within 300 kilometers of the recipient, renormalizing over the remaining states before weighting payments per resident. The same construction applies to production and distance weights. These fixed weights define the transfer comparison rather than estimate its weekly propagation law. Within each round, the SCI shift-share is residualized against the production and distance exposures and direct-home payouts. All regressions have county and month fixed effects and two-way clustered standard errors. The design needs source-shock and exclusion assumptions, including a treatment of correlated pandemic and policy exposure, and it says nothing about aggregate pricing.
Table C.2. Summary statistics for the remote-payment comparison
Notes: The sample is the impact regression’s January 2020–June 2021 county–month sample, including non-payment months. Spending is the monthly spending index in percentage points. Social exposure is standardized over this sample after residualizing remote source-state payments against production, distance and home-state payments within each payment round. Negative standardized values indicate below-average residual exposure.
Table C.3. Remote-source Economic Impact Payment spillovers
Notes: Coefficients and standard errors are in spending-index percentage points per standard deviation of exposure. The outcome is 100 times the monthly mean of OI’s seasonally adjusted spending-change index, observed h h h months relative to the payment month. Exposures are standardized within each horizon’s estimation sample. Remote-source exposures exclude source states within 300 km, and the SCI exposure is residualized against production, distance and home-state payments within each round. County and month fixed effects are included. Standard errors clustered by county and calendar month appear in parentheses. The difference row is the within-regression SCI–production contrast. * p < 0.10 p < 0.10 p < 0.10 , ** p < 0.05 p < 0.05 p < 0.05 , *** p < 0.01 p < 0.01 p < 0.01 .
Figure C.1. Remote-source Economic Impact Payment lead-lag profile, with 95 percent confidence intervals, in spending-index percentage points per standard deviation of exposure. The SCI residual is the social shift-share after excluding nearby source states and partialling out production, distance and home payments within each round. Exposures are standardized separately within each horizon’s estimation sample.
Table C.4. Leave-one-round-out stability of the remote EIP design
Notes: Each row re-estimates the county-month specification after removing one payment round. Coefficients and standard errors are in spending-index percentage points per standard deviation of exposure. The outcome is 100 times OI’s monthly mean spending-change index, and exposures are standardized within each estimation sample. Standard errors clustered by county and calendar month appear in parentheses. * p < 0.10 p < 0.10 p < 0.10 , ** p < 0.05 p < 0.05 p < 0.05 , *** p < 0.01 p < 0.01 p < 0.01 .
Source responses and aggregation
The local regression measures differences across observed recipients. A source county’s recorded property loss is an input to that comparison, not an estimate of its consumption loss. Insurance, transfers, borrowing and rebuilding can change the mapping from destruction to spending or portfolio demand. Appendix § 16.3 .§ 3 therefore compares monetary source measures separately. Recovering total household pressure additionally requires independently measured financial sensitivities and the common-response information in Appendix § 12.8 .10. An unsigned sum of county coefficients is not an aggregate multiplier.
Pre-2020 labor outcomes
The pre-2020 exercise reruns the storm design at a quarterly clock on QCEW wages and employment for 658 non-winter FEMA events from 1996 to 2019, using the modern SCI under a topology-stability assumption. County migration shares correlate 0.784 with log SCI over 2004–2010 (Table E.6), which is consistent with that assumption but does not establish it. Table C.6 adds production exposures and a year-over-year outcome.
Table C.5. Historical SCI transmission in quarterly county wages, 1996–2019
Notes: Cumulative log county QCEW wage change over h h h quarters. The sample covers non-winter FEMA storms in 1996–2019 and excludes directly hit counties, with county/event effects and two-way clustered errors. The modern SCI imposes a topology-stability assumption. Migration overlap in 2004–2010 (correlation 0.784) does not validate that assumption for every earlier year. A wage response is not a spending or portfolio-demand response. * p < 0.10 p < 0.10 p < 0.10 , ** p < 0.05 p < 0.05 p < 0.05 , *** p < 0.01 p < 0.01 p < 0.01 .
Modal surfaces and transport
These exercises distinguish the pooled exposure coefficient from recovery of the dynamic response surface. The modal fits do not pin down a stable decay law, and held-out predictive performance for residual state contrasts is close to the zero-response benchmark. The tables document that conclusion and the observation map to which it applies. The target-specific simulation in Appendix § 16.5 then evaluates what additional measurements and restrictions contribute to financial prediction. The
Table C.6. Historical wage-response robustness: production controls and seasonality. Notes: All panels use the pre-2020 storm non-winter event panel (658 events, 1996–2019) with county and event fixed effects and two-way clustered standard errors, hit counties excluded, gravity included throughout. Panel A adds upstream and downstream production exposures built from 2019 CBP county sector shares and the BEA input–output matrix, applied to the historical events under a disclosed share-stability assumption. Panels B and C replace the quarterly-diff cumulative outcome with year-over-year log wage growth at t + h t+h t + h , which nets out county-quarter seasonality mechanically. * p < 0.10 p<0.10 p < 0.10 , ** p < 0.05 p<0.05 p < 0.05 , *** p < 0.01 p<0.01 p < 0.01 .
modal and transport exercises use weekly growth and onset-week source shares, rather than the level changes and through-week-4 source shares in the local-spending comparison.
Noise in the modal surfaces and the per-mode decay fits
The modal response surfaces use the 34 lowest nonstationary eigenvectors, excluding the stationary mode. At each horizon the county-level regression jointly includes all 34 exposure columns E j k e = u j k ( u k ′ b r a w ) E_{jke} = u_{jk}(u'_{k}b^{raw}) E j k e = u j k ( u k ′ b r a w ) , county, event and target-week fixed effects, and the controls in main-text equation (23) , so the estimates are jointly fitted exposure coefficients, not separate regressions of a projected outcome on a projected shock. Controls include distance bins, a same-hit-state indicator, COVID-19 cases and deaths per capita, state stringency, economic-support and income-support indices, one-week lags of spending growth and log spending, and mean spending growth over the previous four weeks. The outcome is weekly log spending growth, clipped at the first and ninety-ninth percentiles, rather than the unclipped level changes in Table 2 . Storm inputs are county shares of onset-week recorded damage restricted to the retained counties. They are not rescaled to sum to one within that support. Standard errors are clustered by county and event.
Table D.1 summarizes the K = 34 K = 34 K = 34 modal surfaces behind § 5 . Each surface has 34 modes and 17 weekly horizons. The squared-estimate/variance ratio is ∑ θ ^ 2 / ∑ s e ^ 2 \sum\hat{\theta}^{2} / \sum\widehat{\mathrm{se}}^{2} ∑ θ ^ 2 / ∑ se 2 over the cells in question. It is about one for a surface of pure noise. Cells are not independent, so the shares and ratios are descriptive rather than a joint test. Figure D.1 plots the per-mode log-decay slopes on the SCI support and the two second-step fits.
Table D.1. Cell-level signal in the modal response surfaces (K = 34 K = 34 K = 34 ). Notes: 578 cells per surface. Smooth and rough refer to each geometry’s own Laplacian spectrum. The retained production modes have eigenvalues near one, the SCI modes between 0.08 and 0.32. Responses are estimated jointly at each horizon as in main equation (25).
Figure D.1. Per-mode log-decay slopes b k d e c a y b_k^{\mathrm{decay}} b k decay on the SCI stable-sign support (K = 34 K = 34 K = 34 , c s i g = 0.5 c_{\mathrm{sig}} = 0.5 c sig = 0.5 ), with 95 percent intervals, and the second-step regressions of b k d e c a y b_k^{\mathrm{decay}} b k decay on λ k \lambda_k λ k under inverse-variance and equal weights. κ ^ \hat{\kappa} κ ^ is minus the slope. The highlighted mode receives 99.6 percent of the inverse-variance weight.
The decay fit across retained-mode counts
Table Table D.2 repeats the inverse-variance second step for each evaluated choice of K K K at c s i g = 0.5 c_{\mathrm{sig}} = 0.5 c sig = 0.5 . The estimate of κ \kappa κ changes sign across K K K , ρ ^ \hat{\rho} ρ ^ exceeds one for K = 8 K = 8 K = 8 and K = 16 K = 16 K = 16 , and the largest single-mode weight share is above one half in seven of the ten cases.
Table D.2. Inverse-variance decay fit by number of retained SCI modes. Notes: Largest weight share is the share of total second-step weight on one mode (0.996 at K = 34 K = 34 K = 34 ). Each row uses the same two-step decay estimator at a different mode count.
Out-of-fold transport
The transport experiment uses 64 storms, 45 onset dates, 1,590 counties and 34 modes, with five purged outer folds. Overlapping response windows leave 13 calendar-overlap groups, 12 with complete cumulative paths. Penalties and aggregation weights are tuned within training folds, while the fixed-decay benchmark retains the decay shape estimated from the full-sample modal fit. Held-out county, event and week effects use held-out outcomes while nuisance coefficients use training data, and population-weighted state means are taken over observed non-hit counties, so the target is not total consumption.
Table D.3. Out-of-fold aggregate transport with calendar-group resampling intervals
Notes: R 2 = 1 − ∑ ( y − y ^ ) 2 / ∑ y 2 R^2 = 1 - \sum(y - \hat{y})^2 / \sum y^2 R 2 = 1 − ∑ ( y − y ^ ) 2 / ∑ y 2 , relative to a zero conditional response. Intervals resample calendar-overlap groups with fitted predictions fixed, omitting training and selection uncertainty. There are 25,145 weekly and 2,632 cumulative cells, far more than independent shocks. The nested rule selects zero in two of five folds. The identity R 2 = ( 2 y ′ y ^ − y ^ ′ y ^ ) / y ′ y R^2 = (2y^\prime\hat{y} - \hat{y}^\prime\hat{y})/y^\prime y R 2 = ( 2 y ′ y ^ − y ^ ′ y ^ ) / y ′ y shows that small predictions give R 2 R^2 R 2 near zero regardless of the true signal.
Sensitivity to weighting the mode-level decay slopes
The representative inverse-variance fit places 99.6 percent of its weight on one mode. Equal weighting, omitting that mode and pooling the support cells instead give negative, imprecise diffusion-rate estimates. The result is a sensitivity of the estimated decay law, rather than an independently replicated structural parameter.
Table D.4. Weighting sensitivity of the single-channel decay fit
Notes: SCI stable-sign support, K = 34 K = 34 K = 34 , c s i g = 0.5 c_{\mathrm{sig}} = 0.5 c sig = 0.5 . The two-step fit regresses each support mode’s log-decay slope on graph frequency, so κ ^ D \hat{\kappa}_D κ ^ D is minus the slope and ρ ^ D \hat{\rho}_D ρ ^ D the exponential of the intercept. Largest weight is the largest share of total regression weight on one mode. The pooled row fits main equation (4) to support cells with mode intercepts. Inference conditions on the estimated, selected surface. Table D.3 reports the corresponding held-out transport comparisons.
Financial evidence and measurement
Aggregate credit-premium and return associations
Table E.1 reports the aggregate regressions discussed in the main text. The SCI amplitude is the monthly mean of the absolute slow coordinate from a VAR(2) fitted to two principal factors of the weekly state-response surface projected onto the SCI basis. The SCI state and the production-family state (the first principal component of the upstream and downstream amplitudes) enter jointly with macro and COVID controls. Panel B reports a twelve-month intermediary-return regression whose autoregressive columns used overlapping future-return lags. They are not valid predictive evidence and are shown only for completeness.
Table E.1. Aggregate macro-finance associations
Notes: HAC standard errors in parentheses. Panel A outcomes are one-month EBP changes in percentage points. Panel B uses the sum of twelve forward monthly He–Kelly–Manela (HKM) value-weighted intermediary investment returns, in decimal units. Tightness is the negative of the standardized He–Kelly–Manela intermediary capital ratio. Stars are conventional and do not account for the specification search described in Appendix § 15 .§ 2 . * p < 0.10 p < 0.10 p < 0.10 , ** p < 0.05 p < 0.05 p < 0.05 , *** p < 0.01 p < 0.01 p < 0.01 .
Specification selection and inference
The state construction, the tightness interaction and the autoregressive specification were chosen after several alternatives had been examined, and the pricing model of § 3 was written after both the positive association and the later negative observer association were known. Neither sign is therefore a test of that model. The recursive and block bootstrap p p p -values in Table E.11 were also computed after the results were known.
In the AR(2) specification the regressor y t − 1 = E t B P − E t − 1 B P y_{t-1}=E_t^{BP}-E_{t-1}^{BP} y t − 1 = E t B P − E t − 1 B P uses information through month t t t , and regressors were standardized on the full sample, so the regression is not a real-time forecast. Standardization preserves the sign of the interaction but not its magnitude or structural interpretation. The storm count explains 62 percent of the monthly amplitude variance. Separately, the SCI interaction is 62 percent spanned by lagged macro and COVID controls (72 percent for the count interaction). Twelve-month return regressions restricted to information available at the forecast date, estimated for two alternative states, give negative coefficients, with 40 observations and only four non-overlapping annual windows.
Intermediary presence
The intermediary-presence regression uses county-level 2019 assets from Compustat headquarters records mapped from ZIP codes to counties. Banks load negatively on both the SCI and production legs, and fund-trust and real-estate intensity is dominated by administrator addresses rather than local presence, so the non-bank composite (broker-dealers and insurers) is the headline object. The conditional model does not uniquely select it. Broker-dealers alone have the same sign but only 14 counties with positive assets, the largest holding 91 percent (Table E.4).
The specification displayed in Table E.2 omits some lower-order interaction terms. A test of a triple interaction requires the corresponding lower-order terms, including exposure-by-local-intensity and tightness-by-local-intensity, unless absorbed by the fixed effects. The reported coefficients are therefore descriptive rather than a test of the fully hierarchical interaction model.
Table E.2. Intermediary presence and SCI amplification in the EIP panel
Notes: Dependent variable: one-month-ahead county log spending. SCI triple and Production triple denote residual social or production exposure interacted with lagged tightness and the indicated mechanism measure. Each completed interaction is standardized within its estimation sample. County and month fixed effects, direct-home payout interactions and remote-source pandemic controls. Two-way clustered standard errors in parentheses. The county-bootstrap specifications in Table E.3 use different scaling and inference. Their p p p -values are not interchangeable p p p -values for these coefficients. The non-bank composite rests on 68 counties, the largest holding 64 percent of assets. The insurance residual is insurance intensity orthogonalized with respect to the other intermediary types. {sup}* p < 0.10 p < 0.10 p < 0.10 , {sup}** p < 0.05 p < 0.05 p < 0.05 , {sup}*** $p < 0.01.
Table E.3. Restricted-null wild cluster bootstrap for the firm-tier mechanism
Notes: County-cluster bootstrap specification, with its own coefficient scaling and standard errors, rather than the two-way covariance fit in the preceding table. Restricted-null bootstrap, B = 1999 B = 1999 B = 1999 valid draws. The non-bank row has zero exceedances, so < 0.001 < 0.001 < 0.001 describes Monte Carlo resolution, not an exact zero population p p p -value. The final column is a null t ∗ t^* t ∗ reference interval, not a confidence interval for the coefficient. These non-hierarchical interactions are descriptive. * p < 0.10 p < 0.10 p < 0.10 , ** p < 0.05 p < 0.05 p < 0.05 , *** p < 0.01 p < 0.01 p < 0.01 .
Table E.4. County-intensity concentration in the local-home intermediary measures
Notes: Shares are computed from the raw 2019 asset aggregates underlying the log-intensity measures.
Table E.5. Drop-top stress test for the local-home broker-dealer mechanism
Notes: With only 14 counties having any listed broker-dealer assets in 2019, the broker-dealer result is most vulnerable to dropping the largest broker-dealer headquarters counties. The table re-estimates the headline broker-dealer specification after dropping the top-N N N broker-dealer-home counties. Nominal two-way-clustered t t t statistics in parentheses, not independent replications. Dropping ten of these 14 counties leaves four. The resulting t = 11.07 t = 11.07 t = 11.07 is not persuasive evidence of robustness with so few exposed clusters. These non-hierarchical interactions do not identify household demand or EBP transmission. * p < 0.10 p < 0.10 p < 0.10 , ** p < 0.05 p < 0.05 p < 0.05 , *** p < 0.01 p < 0.01 p < 0.01 .
A long-sample migration proxy
County and state migration networks are available before 2020 and track the SCI closely (Table E.6), and the monthly propagated-shock series from the two networks correlate 0.949 over 2005–2024. Table E.7, in which T i g h t t − 1 \mathrm{Tight}_{t-1} Tight t − 1 denotes lagged tightness, reports monthly regressions over that period. Its return outcomes sum monthly HKM value-weighted intermediary investment returns over the stated horizon, in decimal units. The EBP interaction is null for both networks. The return interactions for the migration proxy, +0.023 at six months and +0.021 at twelve, are exploratory and have not been checked for release timing or specification search.
Table E.6. Migration network as a historical SCI proxy
Notes: State ACS overlap uses 19 available years in 2005–2024, since 2020 is absent. County overlap uses seven years. High cross-network correlation is not evidence of historical topology stability or of shock transmission.
Table E.7. Long-run macro-finance regressions with the migration proxy
Notes: HAC standard errors in parentheses. * p < 0.10 p < 0.10 p < 0.10 , ** p < 0.05 p < 0.05 p < 0.05 , *** p < 0.01 p < 0.01 p < 0.01 .
Stationary and nonstationary exposure: computation details
The calculation behind Figure 3 uses the retained SCI graph on the 1,590 support counties, not a source-complete national graph. The graph is symmetrized as A = ( W + W ′ ) / 2 A = (W + W^{\prime})/2 A = ( W + W ′ ) /2 with a zero diagonal, and the calculation uses its full eigendecomposition. The first eigenvalue is zero to machine precision, the first eigenvector has correlation one with square-root degree, and the next 34 eigenvalues agree with the modal-regression basis to 10 − 15 10^{-15} 1 0 − 15 . The largest eigenvalue is 1.60. Of 85 candidate storms, 18 have no damage inside the support and drop out. Let D e \mathcal{D}_e D e be event e e e ’s total recorded damage and b e r a w \mathbf{b}^{raw}_e b e r a w the retained counties’ shares of it. The unit footprint is then b e = b e r a w / m e \mathbf{b}_e = \mathbf{b}^{raw}_e/m_e b e = b e r a w / m e with m e = 1 ′ b e r a w m_e = \mathbf{1}^{\prime}\mathbf{b}^{raw}_e m e = 1 ′ b e r a w .
Here 48 of the 67 retained events have m e < 1 m_e < 1 m e < 1 . Damage-weighted comparisons multiply this unit footprint by in-support damage m e D e m_e\mathcal{D}_e m e D e , giving D e b e r a w \mathcal{D}_e\mathbf{b}^{raw}_e D e b e r a w . Thus the in-support share enters once. The normalization changes cross-event weighting but cancels in each event’s pressure-to-incidence ratio. The dividend-based weights put 18.3 percent of their squared norm on the stationary mode, population weights 32.9 percent, and weights confined to the top 5 percent of dividend counties 10.8 percent. Table E.8 reports selected values.
Table E.8. Proxy-weighted pressure relative to incidence, by diffusion time. Notes: 67 storms. τ = κ D ( h − 1 ) \tau= \kappa_D(h - 1) τ = κ D ( h − 1 ) . Weighted pressure is w ′ e − τ L b e w'e^{-\tau L}b_e w ′ e − τ L b e , while incidence is ( w ′ u 1 ) ( u 1 ′ b e ) (w'u_1)(u_1'b_e) ( w ′ u 1 ) ( u 1 ′ b e ) . Square-root-degree weights give a ratio of exactly one at every τ \tau τ . Computed from the full eigendecomposition and normalized in-support event footprints.
Horizon-specific source importance
The county comparison in Figure 2 evaluates the same symmetric diffusion operator source by source. Let e j e_j e j denote a unit initial slow-field input in county j j j . With homogeneous input loading, the pricing response is proportional to
I j ( τ ) = w ′ e − τ L e j = [ e − τ L w ] j , τ = κ D ( h − 1 ) . (53) I_j(\tau) = w'e^{-\tau L}e_j = [e^{-\tau L}w]_j,\qquad\tau= \kappa_D(h - 1). \tag*{(53)} I j ( τ ) = w ′ e − τ L e j = [ e − τ L w ] j , τ = κ D ( h − 1 ) . (53)
The common price-impact, input-loading and persistence factors cancel in within-horizon county ratios. The unit is an initial field input in the model’s symmetric coordinates, so converting a dollar shock into that field requires a separately specified input map. The calculation uses the same retained graph and dividend-share weights as Figure 3 , with all 1,590 modes. Here τ = 0 \tau= 0 τ = 0 denotes the first slow-field response, h = 1 h = 1 h = 1 . A calendar interpretation additionally requires κ D \kappa_D κ D .
Three relations explain why a fixed score tracks different parts of the response:
I ( 0 ) = w , I ( τ ) = ( 1 − τ ) w + τ D − 1 / 2 A D − 1 / 2 w + O ( τ 2 ) , lim τ → ∞ I ( τ ) = u 1 ( u 1 ′ w ) . (54) I(0) = w,\qquad I(\tau) = (1-\tau)w + \tau D^{-1/2}AD^{-1/2}w + O(\tau^2),\qquad\lim_{\tau\to\infty} I(\tau) = u_1(u_1'w). \tag*{(54)} I ( 0 ) = w , I ( τ ) = ( 1 − τ ) w + τ D − 1/2 A D − 1/2 w + O ( τ 2 ) , τ → ∞ lim I ( τ ) = u 1 ( u 1 ′ w ) . (54)
Initially the loading selects a county’s own weight. Its first change combines that weight with its neighbors’ weights. Over longer diffusion times the stationary direction, u 1 ∝ d e g r e e u_1 \propto\sqrt{\mathrm{degree}} u 1 ∝ degree , governs the limit. Panel A of Figure E.1 follows these transitions. Own dividend shares have rank correlation one initially, neighbors’ shares reach 0.95 at τ = 1 \tau= 1 τ = 1 , and connectedness reaches 0.997 at τ = 32 \tau= 32 τ = 32 . These benchmark scores contain different information about the horizon-specific loading.
Figure E.1. Source importance over the full diffusion-time range. Panel A extends main-text Figure 2 ’s comparisons on a zero-based vertical scale and a numerical diffusion-time axis through 32. Friends’ dividends are D − 1 / 2 A D − 1 / 2 w D^{-1/2}AD^{-1/2}w D − 1/2 A D − 1/2 w . Panel B indexes each within-pair ratio to 100 initially. The unindexed Loudon/Washington ratio has stationary limit 0.999. Model-implied for equal unit initial slow-field inputs in symmetric coordinates, with observed SCI and dividend-proxy weights.
County paths and the pair comparison
Loudon County, Tennessee, and Washington County, New York, have populations of 55,129 and 61,091 in the county frame and total SCI degrees differing by 0.277 percent. Their dividend weight ratio is 2.52. Their model-implied importance ratio is 0.64 at τ = 1 \tau= 1 τ = 1 and 0.50 at τ = 4 \tau= 4 τ = 4 , approaching their square-root-degree ratio, 0.999, in the limit. Both counties gain importance relative to the county mean over τ = 0 \tau= 0 τ = 0 to 4, but Washington gains more (Figure E.2, Panel A). The reversal concerns their relative source loadings.
The broader comparison includes all 27,169 county pairs satisfying ∣ log d i − log d j ∣ ≤ 0.01 |\log d_i - \log d_j| \le0.01 ∣ log d i − log d j ∣ ≤ 0.01 . Of these, 24.5 percent reverse their initial importance ordering by τ = 4 \tau= 4 τ = 4 . The displayed example is selected from the 3,162 pairs in that set with both populations at least 50,000 and an initial importance ratio of at least two, oriented so the initially larger county is in the numerator. The selection rule chooses the smallest ratio at τ = 4 \tau= 4 τ = 4 , making this a clear illustration rather than a typical pair. Panel B reports the entire eligible family’s median and interquartile range, and 17.0 percent of that family reverse by τ = 4 \tau= 4 τ = 4 .
Wake and Mecklenburg counties, North Carolina, provide the comparatively stable case in Figure 2 . Under the same graph, diffusion law and dividend-proxy weights, their ratio is 1.065 initially, 1.023 at τ = 4 \tau= 4 τ = 4 and 1.001 at τ = 30 \tau= 30 τ = 30 . Its largest relative departure from the initial value is about 6 percent over the plotted range. Their total SCI degrees differ by 0.0224 percent. This pair is chosen illustratively after examining the calculated paths, for its comparable connectedness and population and nearly stable relative loadings. It illustrates a useful approximate fixed-score description for these sources, rather than exact proportionality of the full operator. The contrasting pairs use the same inputs and propagation law throughout.
To compare changes across pairs with different initial ratios, Panel B of Figure E.1 displays
J a / b ( τ ) = 100 I a ( τ ) / I b ( τ ) I a ( 0 ) / I b ( 0 ) . (55) J_{a/b}(\tau)=100\frac{I_a(\tau)/I_b(\tau)}{I_a(0)/I_b(0)}. \tag*{(55)} J a / b ( τ ) = 100 I a ( 0 ) / I b ( 0 ) I a ( τ ) / I b ( τ ) . (55)
Both indices start at 100. Loudon/Washington reaches 19.87 at τ = 4 \tau= 4 τ = 4 and 38.59 at τ = 30 \tau= 30 τ = 30 , while Wake/Mecklenburg is 96.04 and 94.01. An index of 100 means the pair’s relative importance remains at its initial value. Equality of the two counties instead means the unindexed ratio is one. The index measures changes in relative source loadings, separately from the common persistence and price-impact factors. The horizontal scale measures accumulated network spreading, τ = κ D ( h − 1 ) \tau= \kappa_D(h - 1) τ = κ D ( h − 1 ) . Translating it into calendar weeks requires the weekly spreading rate κ D \kappa_D κ D .
Figure E.2. County paths and the full example family. Panel A divides each county’s importance by the mean across the retained counties at that diffusion time. Panel B compares the Loudon/Washington ratio with the median and middle 50 percent of the 3,162 eligible pairs, and its vertical axis is logarithmic. The horizontal axes measure numerical diffusion time. All paths use the model diffusion law and dividend-proxy weights.
Financial geography and alternative weights
Figure E.3 locates these changes on the observed graph. The color is the change from τ = 0 \tau= 0 τ = 0 to 4 in log importance relative to the county mean. A county can gain relative importance through ties to financially weighted counties even when its own initial weight is small. The map makes this geography visible, while the pricing response remains the model projection in (E.1).
Figure E.3. Geography of the change in relative source importance. Markers locate retained continental U.S. counties, with size proportional to population subject to display limits. Blue indicates a rise and red a fall in importance relative to the county mean from τ = 0 \tau= 0 τ = 0 to 4 4 4 . The color scale caps values at − 2 -2 − 2 and 2 2 2 . Model-implied using observed SCI and dividend-proxy weights.
Table E.9 repeats the calculation with alternative proxy weights. The changes in relative loadings persist across these choices. The financial geography determines their magnitude. If w ∝ u 1 w \propto u_1 w ∝ u 1 , for example, then I ( τ ) = w I(\tau) = w I ( τ ) = w for every τ \tau τ and the stationary direction alone suffices. Thus the comparison illustrates the economic content of Theorem 1, namely that horizon-specific financial relevance depends jointly on transmission and the weights selecting household responses.
Table E.9. Horizon dependence under alternative proxy weights
Notes: The rank-one error is the best relative Frobenius error after normalizing each horizon’s loading vector to unit Euclidean norm, on τ ∈ { 0 , 1 , 2 , 4 , 8 } \tau\in\{0, 1, 2, 4, 8\} τ ∈ { 0 , 1 , 2 , 4 , 8 } , and it measures cardinal approximation error. Participation uses the 2019–22 average number of IRS returns reporting ordinary dividends. Top 5% retains 80 counties with positive weight, so its pair share uses only the 34 near-connectedness pairs with both weights positive (11 reversals). Its rank correlation among positive-weight counties alone is 0.69. The other pair shares use the common 27,169-pair comparison. Proxy weights describe scenarios for financial geography rather than measured marginal bond-demand sensitivities.
Municipal fund-flow pilot
The pilot asks whether remote network exposure predicts fund cash flows beyond direct weather and common storm timing. It is exploratory. Its measurement rules and specifications were set after preliminary flow analyses and before the results below were estimated. It is not an estimate of household selling weights or a corporate-credit test.
The background database combines CRSP, including dead and merged funds, with SEC N-PORT. Its single-state open-end bond subset contains 906 fund lineages across 48 states in 1991–2024. The broader municipal database has 2,482 lineages. Implied net flow is T N A t − T N A t − 1 ( 1 + r t ) \mathrm{TNA}_{t} - \mathrm{TNA}_{t-1}(1 + r_{t}) TNA t − TNA t − 1 ( 1 + r t ) , aggregated over a state’s funds and divided by lagged assets. Where reported flows cover most assets, the two measures correlate at 0.965. The national descriptive series in Figure E.4 has a monthly standard deviation of 0.83 percent of assets. It provides historical context, not an exposure-response estimate or investor-location evidence.
Figure E.4. Implied net flows into single-state municipal bond funds, summed over all states, as a percent of lagged assets covered by clean flow records, January 1991 to December 2024. Labels mark selected episodes. Sources: CRSP Survivor-Bias-Free Mutual Fund Database and SEC N-PORT.
Fund histories and flow measurement. Reported net subscriptions are new sales minus redemptions, divided by the same portfolios’ lagged assets. Reinvested distributions are separate. Unique portfolio-to-lineage links must be supported by dated header-history or portfolio-map intervals at the observed monthly date. A later unique portfolio mapping is not backcast. Header histories establish 9,182 portfolio-month links, 4,950 before 2008. Five unsupported links are excluded and each accepted report maps to one lineage. These are links, not additional independent shocks. Current and lagged raw assets must be present for all retained share-class records, with their sums agreeing with the portfolio totals. Carried-forward or unresolved amounts are excluded. Coverage is measured against the full observed eligible asset universe, including funds excluded for incomplete flow records. Five missing-asset observations concern four share classes. Franklin’s New Jersey, Pennsylvania and Arizona class-B records have February 2013 end dates and liquidation flags, but their last monthly balances are not observed. The Nuveen Ohio class-I gap is longer. Its last observed balance is in December 1996, not a valid substitute for March–April 2009 assets. These gaps remain excluded rather than filled. CRSP’s header end date denotes the latest NAV observation. It bounds recorded support, not necessarily the date an economic portfolio ceased to exist. Completeness here means completeness of retained recorded classes, not proof that no class is absent from the source or that the historical lineage key is infallible.
Fixed design. The primary outcome is the first full calendar month after onset. Onset-month and preceding-month flows are separate diagnostics. The source-complete declarations are collapsed to one state–onset-month row. The primary weather screen covers onset through the end of the outcome month, excludes focal states and FEMA-affected recipients, and requires known NOAA damage below $1 million. Unknown damage is not zero. State aggregates require at least three eligible funds and 50 percent reporting coverage. Controls are state and onset-month effects, gravity, upstream, downstream and nearby exposure, adjacent-month social exposure, local damage, and pre-window flows and returns. State weights use retrospective 2022 tax-exempt-interest income. Neither investor residence nor historical marginal portfolio demand is observed.
The primary design has 15 states, 42 onset months and 27 calendar quarters, with 19 residual degrees of freedom. Onsets span November 2006–October 2017. Contemporaneous fund-history and complete-asset requirements change sample membership. Of the initial 103 observations, 72 qualify and 31 are excluded, while 14 additional observations qualify. Controls absorb 95.30 percent of exposure variation. October 2017 supplies 31.71 percent of the residual exposure sum of squares. Concentration corresponds to 4.90 equally informative states and 5.64 onset months. Median coverage is 70.46 percent of the observed eligible asset universe, not all municipal wealth.
Detection and economic magnitude. A conditional Gaussian simulation uses the fixed design, shrunk cross-state flow covariance and a common AR(1) coefficient of 0.356. The covariance requires positive-semidefinite projection. No pair covariance is missing. The three-endpoint maximum-statistic cutoff is 2.381. Across 10,000 independent null draws, familywise rejection is 4.95 percent. The primary 80-percent detection threshold is 9.73 basis points on the fixed scale. For a 10-basis-point outflow, joint detection and lead equivalence within half the effect have 81.79 percent power, falling to 10.99 percent when noise standard deviations double. This is oracle power conditional on measured exposure. It does not incorporate exposure measurement error or re-estimation, nor establish the clustered test’s size.
Before estimating this specification family, the protocol set an illustrative economic benchmark
Table E.10. Municipal fund flows: primary specification and sensitivity analyses
Notes: Basis points of lagged assets per fixed reference residual-exposure SD, 9.101529 × 10 − 5 9.101529 \times10^{-5} 9.101529 × 1 0 − 5 , computed on the initial sample rather than re-estimated in each specification. Positive net flow is inflow, while positive gross redemptions are withdrawals. Intervals and nominal p p p -values use state and calendar-quarter two-way clustering with the smaller cluster count minus one as inference degrees of freedom. They are not simultaneous intervals. The three timing tests have Bonferroni-adjusted p = 1 p = 1 p = 1 . The matched sample has nuisance rank 55 and four residual degrees of freedom after exposure, below the fixed minimum of ten. Different timing samples do not constitute an impulse-response path. Adding current returns is descriptive, not causal mediation. Before applying the complete fund-history requirements, the estimate is −3.10, interval [−17.31, 11.11].
of one basis point in a matched municipal yield. If outflow is a fraction f f f of the tested funds’ assets, s s s is their share of the matched market, χ f \chi_f χ f the fraction passed into security sales, ξ P \xi_P ξ P the residual demand elasticity and D m o d D_{mod} D m o d duration, then
Δ y [ b p ] = χ f s ξ P D m o d f b p . (56) \Delta y\ [\mathrm{bp}] = \frac{\chi_f s}{\xi_P D_{mod}} f_{\mathrm{bp}}. \tag*{(56)} Δ y [ bp ] = ξ P D m o d χ f s f bp . (56)
At χ f = 0.5 \chi_f = 0.5 χ f = 0.5 , s = 0.1 s = 0.1 s = 0.1 , ξ P = 1 \xi_P = 1 ξ P = 1 and D m o d = 5 D_{mod} = 5 D m o d = 5 , the hurdle requires a 100-basis-point outflow, or $17.84 million at the sample’s median covered assets of $1.784 billion. The primary point estimate instead corresponds to a $0.091 million inflow, with nominal interval [−$1.621, $1.802] million. Both the point estimate and its interval are far from that assumed central hurdle. The protocol also reports a fixed sensitivity grid over market share, pass-through, elasticity and duration, none of which is estimated here. A hypothetical five-basis-point EBP effect with unit cross-market loading would require a 500-basis-point outflow under the same assumptions. Zero household response or zero cross-market transmission remains admissible, so there is no empirically established positive lower bound that this mechanism must exceed.
Interpretation. Contemporary fund histories and complete recorded assets improve flow measurement. Identifying the household-demand channel additionally requires investor residence and causal exclusion. These estimates therefore describe municipal fund cash flows rather than the household demand shift required for the corporate-credit EBP channel.
Alternative aggregate-state constructions
Table E.11 compares the SCI amplitude with incidence, seasonal, signed-observer and production-network states under the monthly design.
Table E.11. Monthly EBP interactions and alternative state constructions
Notes: Monthly regressions of the one-month EBP change on the listed state, lagged tightness, their interaction, macro and COVID controls and two outcome lags. HAC t t t -statistics are in square brackets. State construction and standardization differ across columns. Column (3) includes both interactions jointly. Its components cannot be added to recover column (1). “SCI” is the recovered amplitude, and “Count” counts 85 active storm events. The count explains 62 percent of the monthly variance of the amplitude. The observer is a signed filter of 34 SCI modes imposing the modal-fit decay rates. Bootstrap entries use recursive AR(2) null paths with six-month innovation blocks and 4,999 draws, holding states fixed. All columns are descriptive. Appendix § 15 .§ 2 discusses selection and timing.
Corporate-bond fund flows and the pricing margin
The exploratory pricing exercise uses SEC N-PORT fund flows and corporate-security holdings. The fund-level universe permits open-end funds and funds reporting a fund-of-funds designation, while excluding funds ever classified as ETFs, money-market, closed-end, interval, insurance-underlying or master-feeder vehicles within the reporting window. Captive and placeholder series, central-fund names and funds predominantly held by other reporting funds are excluded. The fund-of-funds treatment and normalized instrument were considered after the initial construction. This is an exploratory specification. Fund-months require lagged net assets of at least $10 million. Flow rates are winsorized at the first and ninety-ninth percentiles within month.
Let f j t f_{jt} f j t be fund j j j ’s net inflow divided by lagged net assets and let s j , t − 1 s_{j,t-1} s j , t − 1 be its lagged corporate-security holdings divided by the market denominator. The aggregate flow proxy, in percentage units, is F t = 100 ∑ j s j , t − 1 f j t F_t = 100 \sum_j s_{j,t-1} f_{jt} F t = 100 ∑ j s j , t − 1 f j t . The denominator is the Federal Reserve’s nonfinancial corporate-business debt-securities series (NCDBDIQ027S), interpolated between quarterly endpoints. Corporate portfolios can include financial and foreign issuers, so the denominator and numerator therefore have different scopes. This normalization is a reporting scale, not an estimated economy-wide demand elasticity.
Principal-component loadings use a balanced July 2020–June 2024 block of 318 funds. The two factor scores in other months are projections onto those loadings. Regressing each eligible fund’s flow rate on an intercept and these factors over its available sample gives residuals u j t u_{jt} u j t , requiring at least 12 months. For the set J t J_t J t with residuals, the granular instrument is
Z t G I V = 100 { ∑ j ∈ J t s j , t − 1 u j t − ( ∑ j ∈ J t s j , t − 1 ) 1 ∣ J t ∣ ∑ j ∈ J t u j t } . (57) Z_t^{GIV}=100\left\{\sum_{j\in J_t}s_{j,t-1}u_{jt}-\left(\sum_{j\in J_t}s_{j,t-1}\right)\frac{1}{|J_t|}\sum_{j\in J_t}u_{jt}\right\}. \tag*{(57)} Z t G I V = 100 ⎩ ⎨ ⎧ j ∈ J t ∑ s j , t − 1 u j t − j ∈ J t ∑ s j , t − 1 ∣ J t ∣ 1 j ∈ J t ∑ u j t ⎭ ⎬ ⎫ . (57)
This is the size-minus-equal-weighted construction in market units. The aggregate inflow proxy includes all eligible fund-month flows, including those without enough observations to estimate a loading.
Factors, fund selection and loadings use the full reporting window. The exercise relates same-month flows to EBP changes rather than forecasting in real time. The regression includes an intercept, the two factors and two lagged EBP changes. HAC calculations use actual calendar-month gaps after observations are removed. They condition on estimated factors and do not propagate fund-selection or generated-instrument uncertainty. A strong first stage does not establish exclusion, because residual flows can still reflect credit news, return chasing or omitted common exposures.
Table E.12 reports the fixed measurement and timing comparisons. Removing reinvested distributions produces the external-cash measure, sales less redemptions. It leaves the sign and full-sample magnitude close to the original measure. Using only predated holdings and non-backcast asset estimates has little effect. It does not make full-sample factor estimates predetermined. The crisis exclusions yield wider intervals. Omitting one year at a time from the external-cash specification gives coefficients from −1.48 to −5.76 basis points, with respective standard errors 1.66 and 1.25. These are endpoints of a sensitivity range, not a confidence interval.
From executed flows to the operator. In the conditional equilibrium X = X ‾ + ζ q − S X=\overline{X}+\zeta q-S X = X + ζ q − S , q = G ( S , N ) q=G(S,N) q = G ( S , N ) , fix capital, payoff moments and demand parameters strictly inside one regime. Dollar executed selling at reference price P ‾ > 0 \overline{P}>0 P > 0 obeys
d F s e l l = P ‾ ( − d X ) = P ‾ ( 1 − ζ G S ) d S , d E B P = b E G S d S . (58) dF^{\mathrm{sell}}=\overline{P}(-dX)=\overline{P}(1-\zeta G_S)dS,\qquad dE^{\mathrm{BP}}=b_EG_SdS. \tag*{(58)} d F sell = P ( − d X ) = P ( 1 − ζ G S ) d S , d E BP = b E G S d S . (58)
If 1 − ζ G S > 0 1-\zeta G_S>0 1 − ζ G S > 0 , division yields
ϑ F ≡ d E B P d F s e l l = b E G S P ‾ ( 1 − ζ G S ) , d E B P d ( P ‾ S ) = b E G S P ‾ . (59) \vartheta_F\equiv\frac{dE^{\mathrm{BP}}}{dF^{\mathrm{sell}}}=\frac{b_EG_S}{\overline{P}(1-\zeta G_S)},\qquad\frac{dE^{\mathrm{BP}}}{d(\overline{P}S)}=\frac{b_EG_S}{\overline{P}}. \tag*{(59)} ϑ F ≡ d F sell d E BP = P ( 1 − ζ G S ) b E G S , d ( P S ) d E BP = P b E G S . (59)
Table E.12. Fund-flow pricing: measurement and sample comparisons. Notes: Monthly EBP change in basis points per inflow of 0.01% of the nonfinancial corporate debt securities denominator. Three calendar-month Bartlett HAC standard errors in parentheses. The instrument is the holdings-weighted residual flow net of its equally weighted counterpart, scaled as above. Every column includes two common flow factors and two lagged EBP changes. Column (3) also excludes backcast asset estimates. First-stage statistics are squared HAC t t t -statistics, not tests of exclusion. Stars use the HAC standard errors and normal critical values: ∗ ^{*} ∗ p < 0.10 p < 0.10 p < 0.10 , ∗ ∗ ^{**} ∗∗ p < 0.05 p < 0.05 p < 0.05 , ∗ ∗ ∗ ^{***} ∗∗∗ p < 0.01 p < 0.01 p < 0.01 .
Substitution into K h = b E G S l ′ A D h − 1 B D K_h = b_E G_S l' A_D^{h-1} B_D K h = b E G S l ′ A D h − 1 B D gives (24) . In the slack regime G S = Γ / ( N + ζ Γ ) G_S = \Gamma/(N + \zeta\Gamma) G S = Γ/ ( N + ζ Γ ) and hence ϑ F = b E Γ / ( P ˉ N ) \vartheta_F = b_E\Gamma/(\bar{P}N) ϑ F = b E Γ/ ( P ˉ N ) . Positive inflows have the opposite orientation to selling. At the binding cap, where 1 − ζ G S = 0 1 - \zeta G_S = 0 1 − ζ G S = 0 , the price can change without a first-order quantity change, and the flow-based inversion is unavailable. One-sided limits must be used at a regime boundary. Subscriptions multiplied by corporate portfolio shares are a proxy for executed corporate-claim transactions, not a measurement of the desired shift S S S . Applying the estimated pricing coefficient to the household operator would additionally require matching asset scope, measuring the absorption share and identifying the household loading. Appendix § 16.5 .§ 7 implements estimation of the two market terms in simulation and reports the limits on identifying them in the actual 57-month sample.
Subscriptions and changes in corporate holdings. A separate quarterly holdings check supplies an observed quantity margin. For consecutive three-month reports, approximate purchases equal end-of-quarter corporate holdings less beginning holdings compounded at the fund’s total return. The corporate sleeve’s own return and security-level executions are not observed. Both purchases and subscriptions are divided by initial net assets. The regression includes quarter effects, uses initial-net-asset weights and clusters standard errors by fund.
The subscription slope divided by mean corporate allocation is 0.864, or 0.874 using external cash. This is a proportional-allocation diagnostic, not an estimate of the equilibrium absorption share m F m_F m F , and no uncertainty interval for the ratio is estimated. It connects subscriptions to corporate-portfolio adjustment without identifying the household-to-fund ownership map or removing the pricing regression’s exclusion concerns.
Table E.13. Fund subscriptions and approximate corporate-bond purchases
Notes: Consecutive reports exactly three months apart, all three monthly flow reports present, initial net assets over $10 million, positive initial corporate holdings. The sample jointly trims the outer one percent of purchase, subscription, external-cash and net-asset roll-forward ratios. The portfolio share is weighted by beginning net assets. Parentheses give fund-clustered standard errors. Purchases are a return-adjusted holdings proxy, not observed trades. * p < 0.10 p < 0.10 p < 0.10 , ** p < 0.05 p < 0.05 p < 0.05 , *** p < 0.01 p < 0.01 p < 0.01 .
Conditional power and economic scale. An illustrative magnitude calculation assumes modified duration 6.5 and a unit mapping from the yield change to the EBP. A price multiplier m p r i c e m_{\mathrm{price}} m price then implies an absolute slope 100 m p r i c e / 6.5 100m_{\mathrm{price}}/6.5 100 m price /6.5 basis points per inflow of 1% of the market denominator, which is 15.4 at m p r i c e = 1 m_{\mathrm{price}} = 1 m price = 1 and 76.9 at m p r i c e = 5 m_{\mathrm{price}} = 5 m price = 5 , or 0.154 and 0.769 basis points per 0.01%. These are assumed scenarios, not independently estimated multipliers or demand elasticities.
The simulation fixes the original-flow design in column (1), including 57 monthly observations, its flows, instruments and factors, and adds independent Gaussian errors at the standard deviation of EBP changes residualized on controls alone. At h = 0 h = 0 h = 0 that scale is 22.54 basis points. Each cell in Table E.14 uses 2,000 draws and the same 2SLS procedure with three-lag HAC standard errors. The decision rule is ∣ t ∣ > 1.96 |t| > 1.96 ∣ t ∣ > 1.96 , but its finite-sample null size has not been calibrated. Thus the table reports conditional rejection frequencies under a nominal 5% rule, rather than a validated power guarantee. It does not propagate uncertainty in factors, fund selection or instrument exclusion. The simulations use the absolute slope, since changing its sign leaves two-sided power unchanged under the symmetric error model.
Table E.14. Conditional sensitivity of the original fund-flow design
Notes: Simulated rejection frequencies. The outcome is E B P t + h − E B P t − 1 EBP_{t+h} - EBP_{t-1} E B P t + h − E B P t − 1 , and h = 0 h = 0 h = 0 is the same-month specification in Table 5 . Errors, duration and the mapping to EBP are assumed as described in the text. These simulations use the original-flow measure, including reinvested distributions, not the external-cash measure.
Across the broader set of four fund/instrument constructions and three endpoints, rejection frequencies at m p r i c e = 5 m_{\mathrm{price}} = 5 m price = 5 range from 10.3% to 17.4%. A separate 4,000-draw calculation for the original-flow, same-month design at m p r i c e = 5 m_{\mathrm{price}} = 5 m price = 5 gives a mean absolute estimate among significant draws 4.12 times the imposed slope. This describes selection on significance under the assumed data-generating process, but it provides no correction factor for the observed coefficient. As a post-result sample sensitivity, excluding March–May 2020 from the original-flow specification gives −4.03 (3.08) basis points per 0.01% inflow over 54 months. The external-cash comparison and leave-one-year range are reported above. Together these calculations motivate separating a provisional pricing association from a calibrated market-wide response.
A longer spending sample: feasibility and power
This planning exercise evaluates a longer spending design and a joint detection-and-calibration gate. Appendix § 16.5 separately evaluates prediction accuracy for a financially weighted target with hypothetical aggregate and financial measurements. The two exercises answer different questions and use different data-generating assumptions and success criteria.
As main § 7.4 notes, the NielsenIQ Consumer Panel (from 2004) and Retail Scanner data (from 2006) would add pre-pandemic storms. They measure different objects, so the spending map must be chosen with care. Store revenue and resident purchases require different maps C ∗ C_{*} C ∗ , because evacuation can relocate sales without an equivalent change in household demand. Coverage, product scope and household precision must be assessed before choosing a map. Neither panel supplies the financial loadings l l l , and hence a a a , or the market map c E c_{E} c E .
Public inputs
The feasibility of a 2006–2019 event design can be assessed using public inputs, without observing licensed spending outcomes. Table F.1 reports the result. The eligibility rules are research choices (contiguous-U.S. county storm declarations, non-winter, complete dates, duration of at most 14 days and complete support from four weeks before to twelve weeks after onset), not official definitions of unexpected shocks, and inputs are equal-weighted affected-county incidence within each declaration rather than damage dollars. NOAA links are candidates, not verified physical identity, and warning dates and anticipation remain unresolved for the historical events.
Synthetic estimator experiments
A synthetic design with 18 counties, three states, unequal expenditure weights, serial correlation 0.6, spatially correlated noise and dose-dependent heteroskedasticity tests the nested cross-fitting estimator. Three outer folds each retune rank, decay, ridge penalty and aggregation weight within training, and no financial outcome selects an operator. Both a shape oracle, which knows the response pattern but must estimate its amplitude, and the feasible estimator must beat incidence, gravity and production benchmarks at two endpoints and satisfy slope and intercept equivalence.
Table F.1. Public-data feasibility of the 2006–2019 event design. Notes: Components are transitive overlap groups, not verified independent shocks. Training is 2006–2013 and validation 2014–2019, with a 12-week purge on each side of the split. The 102 columns use 34 nonstationary SCI modes and three temporal terms over lags 4–12, projected nationally with population-proxy weights.
With calibration evaluated at zero, the strong-signal case passes in 5.25 (oracle) and 9.75 percent (feasible) of 400 draws, and every other scenario, including the incidence and production nulls, passes in none. An alternative calibration at the training-mean prediction was chosen after inspecting the first five scenarios and evaluated on 200 new draws. It raises the strong case to 69.5 percent for both methods, still below an 80 percent target. A linked experiment that adds a known 10-basis-point financial bridge finds no draw that passes both the spending and the financial stage. In its first strong case, incidence explains 96.7 percent of the true aggregate state’s training variance, the collapse in (17) appearing in simulation. These experiments diagnose the design rather than measure the power of licensed data.
Recorded losses, spending measurement and source timing
The source ledger and the exposure unit
The matched spending comparison uses 48 declarations and the same 74,779 recipient–declaration observations in every column. FEMA supplies incident dates and designated counties, while NOAA supplies retrospective weather reports and property-damage amounts. The loss ledger retains 7,014 distinct NOAA report identifiers and the original 684 eligible source county–declaration combinations. It distinguishes positive recorded amounts, zero-coded reports, blank amounts and absence of a report. Sums of recorded contributions are not estimates of complete county losses.
Let Δ D r e , s \Delta D_{re,s} Δ D r e , s be the eligible recorded amount in source county r r r assigned to response week s s s , and let D r e ( h ) = ∑ s ≤ h Δ D r e , s D_{re}(h) = \sum_{s \le h} \Delta D_{re,s} D r e ( h ) = ∑ s ≤ h Δ D r e , s and D e ( h ) = ∑ r D r e ( h ) D_e(h) = \sum_r D_{re}(h) D e ( h ) = ∑ r D r e ( h ) . The source shares and the index, in index points, are
s r e ( h ) = D r e ( h ) D e ( h ) , E j e S C I ( h ) = 100 ∑ r W j r l o c s r e ( h ) , (60) s_{re}^{(h)} = \frac{D_{re}(h)}{D_e(h)}, \qquad E_{je}^{SCI}(h) = 100 \sum_r W_{jr}^{loc}s_{re}^{(h)}, \tag*{(60)} s r e ( h ) = D e ( h ) D r e ( h ) , E j e S C I ( h ) = 100 r ∑ W j r l oc s r e ( h ) , (60)
and (21) is the case h = 4 h = 4 h = 4 . When D e ( h ) = 0 D_e(h) = 0 D e ( h ) = 0 , set the recorded source vector and E j e S C I ( h ) E_{je}^{SCI}(h) E j e S C I ( h ) to zero and retain the loss-coverage flags. This does not assert zero actual destruction. As in § 4 , the local weights W j r l o c W_{jr}^{loc} W j r l oc retain each recipient’s 100 largest out-of-county SCI-times-source-population links, normalized by row. An index point, one unit of E S C I E^{SCI} E S C I , is one percentage point of the damage-share-weighted link index. It is not a percentage point of friends affected, since source weights sum to one across damaged counties, rather than assigning weight one to each affected county. Multiplying every recorded loss in a declaration by the same number leaves E S C I E^{SCI} E S C I unchanged. The index therefore describes the geography of recorded destruction, while the monetary alternatives below retain its scale.
In (22) , the outcome Y j e h Y_{jeh} Y j e h is 100 times the log spending change, in log points, and exposure is in index points. Scaling both by 100 leaves the regression coefficient unchanged, so it equals the coefficient of the unscaled log change on the unscaled share index. Thus the primary −0.352585 is reported as approximately −0.353 percent spending per index point. This change of reporting units leaves the fitted responses and test statistics unchanged. No exposure cap is applied. The incident-aligned and onset-week constructions also use different loss-record windows, so their comparison is not a unit-only rescaling.
The primary input uses county-coded reports beginning within the declared incident. A coverage comparison adds NOAA zones mapping uniquely to one eligible source county. Ambiguous multi-county zones remain unallocated. The inherited zone crosswalk is not a validation of all historical zone boundaries, so direct county assignment remains primary. The ledger preserves potentially overlapping reports rather than treating matching amounts or narratives as conclusive duplicates.
The same source definition is applied to the gravity and two production exposures.
Table G.1. Four-week spending gradients across recorded-loss constructions
Notes: The outcome is week-4 log spending relative to week −1, multiplied by 100. Each source construction rebuilds the social, gravity and production fields on identical recipients. Three one-way, null-imposed Webb-weight bootstraps use 9,999 draws, with 48 declaration, 39 episode and 13 sixteen-week onset-calendar groups. Stars use declaration bootstrap p p p -values: * p < 0.10 p < 0.10 p < 0.10 , ** p < 0.05 p < 0.05 p < 0.05 , *** p < 0.01 p < 0.01 p < 0.01 . The related comparisons were developed after the original outcomes were examined.
The negative four-week gradient is similar across all eight constructions. The reviewed subtotal retains NOAA property-loss entries and documented county-specific public amounts, with source links and scope notes recorded in the county–declaration ledger. It is a partial recorded-property subtotal, not an independent complete assessment. Here 435 of 684 county–declaration amounts are blank, and two declarations have no populated revised amount. Taking the larger recorded value changes only one county relative to the original input. The Sumner comparison replaces its NOAA county total by a reported $130 million property assessment, never adds the two. The source is a news quotation of the county official rather than a verified primary assessment. These alternatives test sensitivity to recorded inputs, while the agreement supports the spatial spending gradient.
Recipient weather and regression inference
The regressions retain the recipient sample and control for its measured weather instead of screening on any report over a long future window. The weather block contains county precipitation and temperature, heat and cold indicators, station gust measurements, warning exposure and damaging-weather reports over the aligned event window. Coverage and missing-value indicators distinguish an unavailable measurement from measured low weather. The remaining controls are distance bins, same-hit-state exposure, gravity, upstream and downstream production fields, and predated spending level and growth. County and declaration fixed effects remove persistent recipient differences and the response common to a declaration.
Ordinary least squares does not impose a Gaussian regression-error distribution. Preferred tests impose the tested coefficient’s null, multiply restricted residuals by a common six-point Webb weight within each resampling group, reapply the nuisance projections, and compare studentized statistics over 9,999 draws. The weights are ± 1 / 2 \pm\sqrt{1/2} ± 1/2 , ± 1 \pm1 ± 1 and ± 3 / 2 \pm\sqrt{3/2} ± 3/2 , each with probability 1 / 6 1/6 1/6 . This procedure preserves the fitted design’s heteroskedastic residual pattern rather than generating Gaussian observation errors.
Declaration and episode resampling give closely aligned inference for the primary coefficient. Calendar-block inference is weaker, with p = 0.055 p = 0.055 p = 0.055 for the incident-aligned county input and 0.112 0.112 0.112 –0.134 0.134 0.134 for the mapped-zone timing comparisons. The dependence models answer different sampling questions. None accommodates arbitrary dependence or resolves concentration by construction. Appendix § 16.3 .§ 7 reports the corresponding pre-event and influence checks.
Monetary severity and the household source response
A monetary input asks how much response follows the destruction, rather than where recorded destruction occurs. In millions of 2024 dollars, let E j e $ ( h ) = 10 − 6 ∑ r W j r l o c D r e 2024 ( h ) E^{\$}_{je}(h) = 10^{-6}\sum_r W^{\mathrm{loc}}_{jr}D^{2024}_{re}(h) E j e $ ( h ) = 1 0 − 6 ∑ r W j r loc D r e 2024 ( h ) . The dollar exposure retains total recorded event loss. Before any nonlinear transformation, E j e $ ( h ) = 10 − 6 D e 2024 ( h ) E j e S C I ( h ) / 100 E^{\$}_{je}(h) = 10^{-6}D^{2024}_{e}(h)E^{\mathrm{SCI}}_{je}(h)/100 E j e $ ( h ) = 1 0 − 6 D e 2024 ( h ) E j e SCI ( h ) /100 when the same deflator is used. The income-scaled alternative is
E j e I ( h ) = 100 ∑ r W j r l o c D r e 2024 ( h ) I r , 2019 2024 , (61) E^{I}_{je}(h) = 100\sum_r W^{\mathrm{loc}}_{jr}\frac{D^{2024}_{re}(h)}{I^{2024}_{r,2019}}, \tag*{(61)} E j e I ( h ) = 100 r ∑ W j r loc I r , 2019 2024 D r e 2024 ( h ) , (61)
where I r , 2019 I_{r,2019} I r , 2019 is observed annual county personal income, converted to the same prices. All 684 source county–declaration rows match the BEA denominator. A percentage point of E I E^{I} E I is damage relative to annual source income, not an observed income loss or a percentage point of friends affected. The dollar and income-scaled coefficients therefore cannot be compared by numerical size alone.
The negative spatial gradient is more precisely estimated than a proportional monetary source loading. A diminishing-response dollar map accommodates a larger response at high observed exposure. For example, the county-input log model predicts about 1.87 percent lower spending at the positive-exposure 90th percentile, compared with 0.026 percent at the positive median. Its declaration-bootstrap p = 0.118 p = 0.118 p = 0.118 leaves that response law imprecise. This contrast shows that locating a recipient response and calibrating dollars of destruction into a household disturbance require different information. The estimates do not select attention over resource disruption, nor establish that damage has no role in response magnitude.
Table G.2. Monetary source inputs on the matched recipient sample
Notes: Each row is a separate matched regression using incident-aligned reports assigned through week 4. Every competing exposure uses the same source input as its social counterpart. Dollars are in 2024 prices. The log-dose model is a different response law, not a change of coefficient units. Stars and p p p -values use the 9,999-draw declaration wild bootstrap. No recipient reaches one percentage point of the damage/income input. * p < 0.10 p < 0.10 p < 0.10 , ** p < 0.05 p < 0.05 p < 0.05 , *** p < 0.01 p < 0.01 p < 0.01 .
For the model, let a dated household injection be g C ( Δ D r e , s ) g_C(\Delta D_{re,s}) g C ( Δ D r e , s ) for resource disruption and g R ( Δ D r e , s ) g_R(\Delta D_{re,s}) g R ( Δ D r e , s ) for a risk-related or attention response. Linear dollars and concave maps are alternative source specifications. Conditional on the chosen map, the existing propagation operator describes transmission across counties. Changing source amplitude under a fixed linear operator changes response size, but not the normalized decay rate of that operator.
Testing a severity restriction. Event fixed effects remove a common event response, not the cross-recipient variation in D e E j e S C I D_eE_{je}^{SCI} D e E j e S C I . A regression containing both E S C I E^{SCI} E S C I and D e E S C I D_eE^{SCI} D e E S C I consequently tests the incremental linear severity gradient conditional on location. It is narrower than a general test of whether damage matters. Grouped restrictions must also account for the nuisance projection. If Z g r p Z^{grp} Z g r p contains grouped exposures, M Z c t r l M_{Z^{ctrl}} M Z c t r l removes the other controls and v v v is the proportional-dollar exposure, the implied grouped coefficient vector under Y = b v + Z c t r l ϕ c t r l + u Y = bv + Z^{ctrl}\phi^{ctrl} + u Y = b v + Z c t r l ϕ c t r l + u is
E [ β ^ ∣ Z g r p , Z c t r l , v ] = b [ ( Z g r p ) ′ M Z c t r l Z g r p ] − 1 ( Z g r p ) ′ M Z c t r l v . \mathbb{E}[\hat{\beta}\mid Z^{grp}, Z^{ctrl}, v] = b\left[(Z^{grp})^\prime M_{Z^{ctrl}}Z^{grp}\right]^{-1}(Z^{grp})^\prime M_{Z^{ctrl}}v. E [ β ^ ∣ Z g r p , Z c t r l , v ] = b [ ( Z g r p ) ′ M Z c t r l Z g r p ] − 1 ( Z g r p ) ′ M Z c t r l v .
Squared residual-exposure weights alone do not generally give this restriction. On the onset-week recorded input, a conditional two-group proportionality test gives declaration-bootstrap p = 0.077 p = 0.077 p = 0.077 , while on the reviewed subtotal it gives p = 0.018 p = 0.018 p = 0.018 . A more permissive location-plus-dollar model does not reject either the zero-location restriction (p = 0.206 p = 0.206 p = 0.206 ) or the zero-incremental-dollar restriction (p = 0.869 p = 0.869 p = 0.869 ). The nuisance models differ, and the reviewed input has unknown county losses. These tests therefore do not deliver a specification-independent verdict on proportional damage scaling.
Continuing destruction and response age
The dated ledger assigns each reported amount to its local beginning date and, in a separate timing comparison, to its end date. These are weather-report dates, not observations of within-report loss accrual or real-time publication of damage estimates. Incident membership is likewise not proof of physical storm identity. Later reports enter only endpoints on or after their assigned date, so a whole-incident total is never placed mechanically at onset.
For a dated monetary input Δ D r e , s \Delta D_{re,s} Δ D r e , s and a response-age bin B b B_b B b , define
E j e h m , b = ∑ s ≤ h h − s ∈ B b ∑ r W j r m Δ D r e , s , B b ∈ { [ 0 , 1 ] , [ 2 , 4 ] , [ 5 , 12 ] } , E_{jeh}^{m,b} = \sum_{\substack{s \le h \\ h-s \in B_b}} \sum_r W_{jr}^{m} \Delta D_{re,s}, \qquad B_b \in\{[0,1],[2,4],[5,12]\}, E j e h m , b = s ≤ h h − s ∈ B b ∑ r ∑ W j r m Δ D r e , s , B b ∈ {[ 0 , 1 ] , [ 2 , 4 ] , [ 5 , 12 ]} ,
where m m m is social, gravity, upstream or downstream. Stacking horizons h ∈ { 1 , 2 , 4 , 8 , 12 } h \in\{1,2,4,8,12\} h ∈ { 1 , 2 , 4 , 8 , 12 } gives
Y j e h = α j , h + μ e , h + ∑ b β b E j e h S C I , b + ∑ m ≠ S C I ∑ b ϕ m , b c o m p E j e h m , b + Z j e ′ ϕ h + u j e h . (62) Y_{jeh} = \alpha_{j,h} + \mu_{e,h} + \sum_b \beta_b E_{jeh}^{SCI,b} + \sum_{m \ne SCI} \sum_b \phi_{m,b}^{comp} E_{jeh}^{m,b} + \mathbf{Z}'_{je} \phi_h + u_{jeh}. \tag*{(62)} Y j e h = α j , h + μ e , h + b ∑ β b E j e h S C I , b + m = S C I ∑ b ∑ ϕ m , b co m p E j e h m , b + Z j e ′ ϕ h + u j e h . (62)
The horizon-specific nuisance terms are residualized within each response block. Inference groups repeated horizons together. The 373,895 stacked rows are five measurements of the same 74,779 recipient–declaration pairs, not five times as many independent shocks. Dollar and income-scaled inputs are estimated separately.
Table G.3. Dated-dollar spending responses by age of destruction
Notes: Spending levels are relative to week −1, with 2024-dollar inputs dated at report beginning. The three age bins partition the assigned input at each endpoint. Each coefficient measures a spending-level response at that response age, not expenditure cumulated across weeks. All tests use 9,999 null-imposed declaration-cluster Webb draws. The joint test re-estimates and studentizes all three social coefficients together.
The dated construction prevents continuing destruction from being mistaken for slow decay of an initial impulse. The late-age dollar coefficients are negative, but their bootstrap uncertainty does not establish a persistent severity loading. For the income-scaled model, joint declaration-bootstrap p p p -values are 0.184 for county reports and 0.122 with mapped zones. These distributed responses inform source timing, but they are not a new estimate of the graph’s decay law.
Calendar-pattern measurement and operator comparisons
The spending tracker uses a fixed 2019 comparison year. Recurring county-specific fluctuations can therefore contain a comparison-year component as well as genuine seasonality. In the schematic observation model L j t o b s = ℓ j t − b j , w ( t ) + η j t L_{jt}^{\mathrm{obs}} = \ell_{jt} - b_{j,w(t)} + \eta_{jt} L j t obs = ℓ j t − b j , w ( t ) + η j t , temporally independent η \eta η produces adjacent differenced-noise covariance − σ j 2 -\sigma_j^2 − σ j 2 . A shared noisy baseline also correlates response errors across horizons. These observations motivate measurement-aware comparisons rather than treating every short-run change as noise.
The recovery-design noise calibration uses a retrospective calendar adjustment. Let H P 13 ( L ) HP_{13}(L) H P 13 ( L ) be the residual from a centered thirteen-week moving average, computed with at least seven observed weeks. For declaration e e e , donor dates are t + 52 n t + 52n t + 52 n , n ∈ { − 4 , … , 4 } ∖ { 0 } n \in\{-4,\ldots,4\} \setminus\{0\} n ∈ { − 4 , … , 4 } ∖ { 0 } , on or after September 7, 2020, excluding the declaration’s weeks [ − 19 , 18 ] [-19,18] [ − 19 , 18 ] . With m m m eligible donors,
L j e t ∗ = L j t o b s − m ρ c a l m ρ c a l + 1 − ρ c a l 1 m ∑ n ∈ N j e t H P 13 ( L j , t + 52 n o b s ) (63) L_{jet}^{*} = L_{jt}^{\mathrm{obs}} - \frac{m\rho_{\mathrm{cal}}}{m\rho_{\mathrm{cal}} + 1 - \rho_{\mathrm{cal}}}\frac{1}{m}\sum_{n\in\mathcal{N}_{jet}} HP_{13}\left(L_{j,t+52n}^{\mathrm{obs}}\right) \tag*{(63)} L j e t ∗ = L j t obs − m ρ cal + 1 − ρ cal m ρ cal m 1 n ∈ N j e t ∑ H P 13 ( L j , t + 52 n obs ) (63)
with no subtraction if donors are unavailable. Here ρ c a l \rho_{\mathrm{cal}} ρ cal is the mean uncentered cross-cycle correlation at 52, 104 and 156 weeks. Donors can come from later years. The adjustment is a measurement sensitivity, not real-time information or an independently observed household response. The local comparison remains on unadjusted spending levels.
For a fixed linear preprocessing map F e c a l F_e^{\mathrm{cal}} F e cal , a candidate operator must be passed through the same map as its observed response:
F e c a l y e and F e c a l H ( ϑ ) ε e + F e c a l Z e ϕ , V e F = F e c a l V e ( F e c a l ) ′ . (64) F_e^{\mathrm{cal}}y_e \quad\text{and} \quad F_e^{\mathrm{cal}}\mathcal{H}(\vartheta)\varepsilon_e + F_e^{\mathrm{cal}}Z_e\phi,\qquad V_e^F = F_e^{\mathrm{cal}}V_e\left(F_e^{\mathrm{cal}}\right)^{\prime}. \tag*{(64)} F e cal y e and F e cal H ( ϑ ) ε e + F e cal Z e ϕ , V e F = F e cal V e ( F e cal ) ′ . (64)
The prediction must cover donor and baseline dates and overlapping inputs. Filtering can remove observable response directions, so Proposition 6 applies to the measurement map actually used. More precise measurements and a valid financial projection play distinct roles in the recovery exercise.
Cross-disaster scope and supporting household finance
A separate exploratory extension examined 67 additional declarations, mostly floods, hurricanes, winter storms and fires, with a rebuilt exposure and conditioning design. It did not establish a common negative coefficient across those disaster types. Source allocations and controls differed from the primary sample, including equal county allocations when recorded damage was unavailable. The result confines the demonstrated local gradient to the primary storm setting, and the extension is not a matched validation of the incident-aligned construction. Unknown damage allocation alone does not explain the extension’s result, since its recorded-damage subset also lacked a negative association. A cross-disaster transport test requires comparable source measurement before interpreting differences as household or operator heterogeneity.
An exploratory monthly mortgage comparison supplies complementary evidence of household financial strain. Across 433 covered recipient counties and 20,565 county–declaration pairs, the onset-week recorded-loss index predicts a 0.0208-percentage-point increase per index point in mortgages 30–89 days delinquent, averaged over months 1–3 relative to month −1. Declaration- and episode-bootstrap p p p -values are 0.038 and 0.040, while sixteen-week calendar-block inference is weaker (approximately 0.20–0.23). The response is positive under all four compared loss constructions. The mortgage sample and its weather controls differ from the spending design, and exposure information is concentrated in a few declarations. The result is supporting financial-stress evidence, not a measure of fixed-price portfolio demand or its credit-market price impact.
Local-response diagnostics
The incident-aligned index preserves the ordering of highly exposed recipients. The four-week gradient remains negative in all 48 leave-one-declaration-out samples and when the acute pandemic onsets are excluded. Table G.4 separates stability of the point estimate from its precision under alternative dependence structures.
Reported standard errors are the sample standard deviation of 9,999 unrestricted-residual wild-bootstrap coefficient estimates, using the Webb six-point weights [32 ] . The original finite-sample variance correction is applied once, as its square root multiplying the residuals. Stars use the corresponding restricted-null, studentized bootstrap tests. The same dependence groups and multiplier family are used for variance estimation and testing. Calendar-panel tests additionally re-estimate Bartlett-HAC studentization in each draw. The reported coefficient standard errors are nevertheless bootstrap standard deviations, not HAC standard errors. For the event-time comparison, the four-week damage geography is held fixed at every horizon. This retrospective construction makes the pre-event coefficients and later spending differences comparable. It is not an information set available before week 4 or an estimated impulse response. The shared week −9 baseline in the figure also displays the week immediately preceding onset rather than normalizing it to zero. Its coefficient is −0.020 per index point (declaration-bootstrap p = 0.909 p = 0.909 p = 0.909 ). The four-week estimate remains negative under either baseline. Table G.5 reports the week −1 baseline used in the main spending table.
Table G.4. Four-week spending gradient: sample and weather comparisons. Notes: Coefficients are interpreted as the change in spending from week − 1 -1 − 1 to week 4, in log percentage points, associated with a 1-percentage-point increase in the damage-share-weighted link index. Columns use declaration, physical-episode and sixteen-week onset-block wild-bootstrap standard errors, respectively (9,999 Webb draws). The first two rows use 74,779 observations and 48 declarations, while the last uses 63,889 observations and 41 declarations. All rows include the same non-weather controls and consistently rebuilt competing exposures. The full sample contains 48 declaration, 39 episode and 13 calendar clusters. ∗ {}^{*} ∗ p < 0.10 p < 0.10 p < 0.10 , ∗ ∗ {}^{**} ∗∗ p < 0.05 p < 0.05 p < 0.05 , ∗ ∗ ∗ {}^{***} ∗∗∗ p < 0.01 p < 0.01 p < 0.01 .
Figure G.1. Event-time spending gradient for the incident-aligned county exposure. The same four-week loss geography and 74,779 observations are used at every date, with spending relative to week − 9 -9 − 9 . This is a retrospective event-study comparison, not a known onset-pulse response. Error bars give 95 percent pointwise approximate studentized declaration-bootstrap intervals. Activity controls predate the leads and measured weather covers the path. The main spending table instead subtracts week − 1 -1 − 1 .
Table G.5. Spending differences with a fixed four-week damage geography.
Notes: Coefficients are changes in spending, in log percentage points, per point of the damage-share-weighted link index. The post-onset rows subtract week − 1 -1 − 1 and use the same 74,779 observations, fixed effects and controls as Table G.4. Columns use declaration, physical-episode and sixteen-week onset-block wild-bootstrap standard errors, respectively (9,999 Webb draws). Joint pre-event tests instead use week − 9 -9 − 9 and common bootstrap multipliers across horizons, preserving their covariance. Their reference is the empirical joint Wald distribution, not a normal or chi-squared approximation. All draws are retained. Calendar inference has 13 clusters. The joint tests do not establish parallel trends. The figure uses approximate studentized null-bootstrap-t t t intervals, not exact confidence-set inversions. ∗ p < 0.10 ^{*}\ p < 0.10 ∗ p < 0.10 , ∗ ∗ p < 0.05 ^{**}\ p < 0.05 ∗∗ p < 0.05 , ∗ ∗ ∗ p < 0.01 ^{***}\ p < 0.01 ∗∗∗ p < 0.01 .
The negative gradient is clearest at four weeks. Later point estimates are smaller and their uncertainty is wide, so the event profile supplies evidence about the location and timing of the spending response rather than a measured relaxation rate. Dated destruction entering after onset is treated separately in the source-input analysis, keeping the duration of the disturbance distinct from transmission after a given input.
Table G.6. Leave-one-declaration-out stability of the spending gradient
Notes: All exposures and controls are re-estimated for each deletion. Significance uses conventional county/declaration two-way clustered inference, while the bootstrap results above concern the full and restricted samples.
Dated repeated inputs and fixed-origin recipients
The county-calendar comparison in Table 3 separates damage arriving at different dates and retains counties after later direct hits. The full panel has 1,574 counties and 211 common weekly origins, February 17, 2020 through February 26, 2024. Actual spending levels from five weeks before each origin through twelve weeks after it, and complete seven-day rain and temperature measurements, give 332,114 county-origin observations. The fixed-origin cohort excludes a county only if it has an own county-coded NOAA weather report in the starting week, giving 300,357 observations. Both samples follow later hits. This is an observational cohort comparison, not randomized intention to treat.
The source universe, the 2,802 counties that supply the local network inputs, remains separate from spending coverage. The original 48 incident windows contain 2,617 distinct county-coded NOAA reports, with recorded property damage of $1.592 billion nominal, or $1.665 billion in 2024 dollars. The entire amount of each report enters once in its start week. Within-report accrual is unobserved. Blank amounts remain flagged. Other county-coded NOAA reports supply background forcing controls, although omitted zone reports and counties outside the source graph limit coverage. No input is capped or interpolated.
Let Δ D r e , t \Delta D_{re,t} Δ D r e , t be new recorded damage for source r r r in week t t t and D e f u l l = ∑ r , t ∈ e Δ D r e , t D_{e}^{\mathrm{full}} = \sum_{r,t\in e} \Delta D_{re,t} D e full = ∑ r , t ∈ e Δ D r e , t its declaration’s fixed full-incident total. Let
s r t = ∑ e Δ D r e , t D e f u l l , x j t S C I = 100 ∑ r W j r l o c s r t . s_{rt} = \sum_e \frac{\Delta D_{re,t}}{D_e^{\mathrm{full}}}, \qquad x_{jt}^{SCI} = 100 \sum_r W_{jr}^{\mathrm{loc}}s_{rt}. s r t = e ∑ D e full Δ D r e , t , x j t S C I = 100 r ∑ W j r loc s r t .
For a zero recorded total, define the corresponding Δ D r e , t / D e f u l l \Delta D_{re,t}/D^{full}_e Δ D r e , t / D e f u l l contribution as zero and retain its coverage flag. This is not an assessed zero actual loss. This input dates each report separately and uses a full-incident denominator, whereas (21) sums reports through week 4. Thus the calendar comparison extends the design rather than merely changing the original coefficient’s units.
For every h = 0 , … , 12 h = 0, \ldots, 12 h = 0 , … , 12 , the outcome is 100 ( L j , t + h − L j , t − 1 ) 100(L_{j,t+h} - L_{j,t-1}) 100 ( L j , t + h − L j , t − 1 ) . Current direct and social inputs enter jointly where direct input varies, alongside current gravity/upstream/downstream exposure and four lags of the direct and all mapped inputs. The regressions also include the pre-origin spending level and four growth lags. Additionally, they include background damage with report-presence and unknown-amount flags, and county and calendar-origin effects. Both specifications adjust actual own weather before and after origin. The future-conditioned model additionally includes the next twelve weeks of direct and all mapped focal/background damage and report flags. Its complete control set is held fixed at every horizon. As a result, projecting the summed weekly outcomes gives exactly the sum of endpoint coefficients.
The comparison asks how the conditional path changes when subsequent recorded arrivals are held fixed. Interpreting it as a causal pulse requires strict conditional exogeneity and separability of the direct, social and competing inputs. Future monetary losses need not satisfy these conditions simply because meteorology is external. Neither the all-county comparison nor the origin cohort conditions on surviving unhit until the response horizon. The earlier event sample was also fixed across horizons. This is a separate repeated-input design.
Inference uses common Webb calendar-block multipliers. Standard errors follow the unrestricted-residual coefficient bootstrap described in Appendix § 16.3 .§ 7 . Null tests use restricted residuals and Bartlett time-HAC studentization. Each of 9,999 draws imposes the zero-network-response null, reprojects the complete nuisance space and re-estimates HAC studentization. The same multipliers apply across counties and horizons. Thirteen- and twenty-six-week blocks give 17 and 9 calendar blocks, respectively. The origin cohort uses its exact unbalanced fixed-effect projection. All draws are valid. Independent full-row refits reproduce the compressed calculations. These are approximate dependence checks, not exact randomization inference. Calendar blocks need not coincide with independent physical storms.
The fixed-origin four-week coefficient changes by 3.5 percent when later recorded arrivals are controlled. The negative timing pattern is retained without excluding later-hit counties. In the all-county panel, the corresponding change is 5.5 percent, with more block-length-sensitive inference. This supports the timing of the local spending gradient while preserving the distinction between spatial contrasts and a full propagation law.
The integrated response remains imprecise. A plateau need not reverse accumulated losses, and a confidence interval including zero does not establish disappearance. An economic-equivalence claim would require an independently justified tolerance and simultaneous endpoint uncertainty over a specified tail. No exact recovery date is inferred. Similarly, smaller later spending estimates do not determine the duration of portfolio adjustment, because the separate financial direction in § 3.5 may be weakly measured by spending.
Table G.7. Weekly spending response in the fixed-origin cohort. Notes: The 300,357 observations have no own county-coded weather report at origin. Later hits are retained. Coefficients are log percentage points per point of the damage-share-weighted link index. Adjacent columns repeat each specification using 13/26-week calendar-block wild-bootstrap standard errors in parentheses (9,999 Webb draws). Outcomes, controls and source timing follow the text. The whole endpoint path has max-t t t p p p -values 0.026/0.023 without and 0.038/0.026 with future damage controls, for 13/26-week blocks. ∗ ^{*} ∗ p < 0.10 p < 0.10 p < 0.10 , ∗ ∗ ^{**} ∗∗ p < 0.05 p < 0.05 p < 0.05 , ∗ ∗ ∗ ^{***} ∗∗∗ p < 0.01 p < 0.01 p < 0.01 .
The dollar point estimates remain negative with later forcing controls, but the calendar-block tests and concentrated support do not provide a precise monetary loading. The income-scaled input has a different sign because it reweights source locations, not because its reporting units change. Missing recorded losses and unobserved within-report accrual also prevent these tests from establishing that true damage size is irrelevant. The location-gradient result and the dollar-dose question remain distinct.
Table G.8. Accumulated response in the fixed-origin cohort Notes: Sum of weekly log-level responses, in log-percentage-point-weeks per point of the damage-share-weighted link index, rather than a percentage change in total-window expenditure. Adjacent columns repeat each specification using 13/26-week calendar-block wild-bootstrap standard errors in parentheses (9,999 Webb draws). Averaging the response requires division by the number of weeks. Additionally, expenditure-dollar conversion requires spending bases and weights. ∗ p < 0.10 ^{*}p < 0.10 ∗ p < 0.10 , ∗ ∗ p < 0.05 ^{**}p < 0.05 ∗∗ p < 0.05 , ∗ ∗ ∗ p < 0.01 ^{***}p < 0.01 ∗∗∗ p < 0.01 .
Table G.9. Monetary input alternatives in the calendar panel Notes: Same 332,114 all-county observations and competing-input definitions. Each row repeats the specification using 13/26-week calendar-block wild-bootstrap standard errors in parentheses (9,999 Webb draws). The damage-share coefficient is per point of the index, while dollar inputs use millions of 2024 dollars. Damage/income is 100 times recorded loss divided by source counties’ actual 2019 annual personal income, in common dollars. Normalizations reweight economic inputs rather than simply rescale them. The largest residualized exposure week supplies 84.5 percent of dollar-dose information, 34.7 percent for damage shares and 42.2 percent for income-scaled inputs. Effective exposure weeks are 1.38, 6.83 and 3.43. These are concentration summaries, not independent-storm counts. The original unrestricted-score bootstrap is not used as the main inference in this highly leveraged comparison. ∗ p < 0.10 ^{*}p < 0.10 ∗ p < 0.10 , ∗ ∗ p < 0.05 ^{**}p < 0.05 ∗∗ p < 0.05 , ∗ ∗ ∗ p < 0.01 ^{***}p < 0.01 ∗∗∗ p < 0.01 .
Financial measurement and a physical-weather design
Household demand: a measurable derivative
Main equation (27) is a first-order accounting decomposition of a fixed-price desired holding. It does not equate spending losses, cash needs and executed bond sales. For an asset a a a , let county desired dollar holdings at the reference price be V i a d = s i a W i V_{ia}^{d}=s_{ia}W_i V ia d = s ia W i . Differentiation gives
− d V i a d = ( s i a χ i W + W i ψ i a p o r t ) d z i , χ i W = − ∂ W i ∂ z i , ψ i a p o r t = − ∂ s i a ∂ z i . -dV_{ia}^{d}=\left(s_{ia}\chi_i^{W}+W_i\psi_{ia}^{\mathrm{port}}\right)dz_i,\qquad\chi_i^{W}=-\frac{\partial W_i}{\partial z_i},\qquad\psi_{ia}^{\mathrm{port}}=-\frac{\partial s_{ia}}{\partial z_i}. − d V ia d = ( s ia χ i W + W i ψ ia port ) d z i , χ i W = − ∂ z i ∂ W i , ψ ia port = − ∂ z i ∂ s ia .
For heterogeneous households this is summed over household-level derivatives. Products of unweighted county averages generally do not give that sum. Price-adjusted holdings and account flows measure adjustment after removing valuation changes. Transfers, insurance, borrowing and cash balances determine whether a spending disturbance depletes investable resources. A fall in W W W with stable shares is the cash-resource margin, whereas a fall in the credit share with stable W W W is the rebalancing margin. With a complete asset menu, ∑ a s i a = 1 \sum_a s_{ia}=1 ∑ a s ia = 1 implies ∑ a ψ i a p o r t = 0 \sum_a\psi_{ia}^{\mathrm{port}}=0 ∑ a ψ ia port = 0 , so a lower credit share can mean a higher cash or Treasury share without net withdrawals from the investment account.
Neither χ i W \chi_i^{W} χ i W nor ψ i a p o r t \psi_{ia}^{\mathrm{port}} ψ ia port need be positive. The sign of the adverse field is a normalization rather than an assumption about a household’s financial reaction. If direct payoff losses also move wealth, they must be separated before attributing χ W \chi^{W} χ W to the social channel. If the measured field is rescaled to z ~ = b z \tilde{z}=bz z ~ = b z , its demand derivative rescales to ℓ ~ $ = ℓ $ / b \tilde{\ell}^{\$}=\ell^{\$}/b ℓ ~ $ = ℓ $ / b , preserving predicted dollar demand. Independently estimated spending and holdings models must share that normalization.
An executed quantity equation illustrates the measurement problem:
Δ x i a e = − ℓ i a z i e + ζ i a Δ q a e + payoff and other demand news i a e + u i a e . (65) \Delta x_{iae}=-\ell_{ia}z_{ie}+\zeta_{ia}\Delta q_{ae}+\text{payoff and other demand news}_{iae}+u_{iae}. \tag*{(65)} Δ x ia e = − ℓ ia z i e + ζ ia Δ q a e + payoff and other demand news ia e + u ia e . (65)
Here ζ i a \zeta_{ia} ζ ia is the premium sensitivity, not part of the autonomous demand shift. Household–asset effects remove persistent ownership differences. Asset–event effects can absorb common prices and issuer news only when their household loadings are common or modeled. Otherwise, heterogeneous demand slopes leave differential price responses. Instruments or an independently identified demand system are needed when the remaining price term is endogenous. Within-asset contrasts still remove a common demand shift. Account-level budget measurements, an independent aggregate anchor or justified bounds are needed for absolute national pressure. Return-adjusted trades alone do not recover a fixed-price demand curve, and delegated managers need not change earnings forecasts for either household margin to operate.
A household problem in surplus. The two financial margins can be derived without making the spending sign a restriction on credit demand. Suppress the household index and let c 0 s u r = c 0 − c ‾ 0 c_0^{\mathrm{sur}}=c_0-\underline{c}_0 c 0 sur = c 0 − c 0 denote current consumption above a predetermined commitment. Let T 0 T_0 T 0 be assistance paid now and T 1 T_1 T 1 a firm future assistance obligation. Net resources are w 0 n e t = w 0 − c ‾ 0 − T 0 w_0^{\mathrm{net}}=w_0-\underline{c}_0-T_0 w 0 net = w 0 − c 0 − T 0 and Y 1 n e t = Y 1 − c ‾ 1 − T 1 Y_1^{\mathrm{net}}=Y_1-\underline{c}_1-T_1 Y 1 net = Y 1 − c 1 − T 1 . The household chooses current surplus consumption, safe investment a s a f e ≥ 0 a^{\mathrm{safe}}\geq0 a safe ≥ 0 and credit holdings x ≥ 0 x\geq0 x ≥ 0 to solve
max u ( c 0 s u r ) + ϖ E [ u ( c 1 s u r ) ] , u ( c ) = c 1 − γ 1 − γ , γ > 0 , ϖ > 0 , \max\quad u(c_0^{\mathrm{sur}})+\varpi E[u(c_1^{\mathrm{sur}})],\qquad u(c)=\frac{c^{1-\gamma}}{1-\gamma},\qquad\gamma>0,\quad\varpi>0, max u ( c 0 sur ) + ϖ E [ u ( c 1 sur )] , u ( c ) = 1 − γ c 1 − γ , γ > 0 , ϖ > 0 ,
subject to
c 0 s u r + a s a f e + P x = w 0 n e t , c 1 s u r = R f a s a f e + V x + Y 1 n e t > 0. (66) c_0^{\mathrm{sur}}+a^{\mathrm{safe}}+Px=w_0^{\mathrm{net}},\qquad c_1^{\mathrm{sur}}=R_f a^{\mathrm{safe}}+Vx+Y_1^{\mathrm{net}}>0. \tag*{(66)} c 0 sur + a safe + P x = w 0 net , c 1 sur = R f a safe + V x + Y 1 net > 0. (66)
At γ = 1 \gamma=1 γ = 1 , u ( c ) = log c u(c)=\log c u ( c ) = log c . Prices and the distribution of the claim payoff are fixed. The payoff has two possible values, V L < R f P < V H V_L<R_fP<V_H V L < R f P < V H , each with positive probability. Background income Y 1 Y_1 Y 1 is independent of V V V and can already be risky. Current resources and both commitments are known at the decision date. Thus future assistance reduces resources available for own consumption, rather than being assumed to change tastes. Uncertainty about assistance belongs in net background income instead of being counted again as a firm obligation.
The result below concerns a strictly interior optimum, with positive safe and credit positions that can both adjust. Surplus consumption stays positive in every state under the local perturbations, and differentiation can be passed through expectations. An increase in background risk means adding ν ξ \sqrt{\nu}\,\xi ν ξ to Y 1 n e t Y_1^{\mathrm{net}} Y 1 net , where E ξ = 0 E\xi=0 E ξ = 0 , E ξ 2 = 1 E\xi^2=1 E ξ 2 = 1 and ξ \xi ξ is bounded and independent of both the existing income risk and V V V . The derivative is evaluated at ν = 0 \nu=0 ν = 0 . This gives a local addition to existing risk, not a comparison of arbitrary finite changes in risk or of risks correlated with the credit payoff.
Proposition H.1 (Local transfer and precautionary responses). Under these conditions, a marginal increase in either T 0 T_0 T 0 or T 1 T_1 T 1 lowers desired credit holdings. The marginal propensity to consume current net resources is strictly between zero and one. A small independent addition to background income risk lowers current consumption and credit holdings, while increasing safe investment. These are fixed-price responses at the decision date.
Proof. Define effective future utility U h h ( k ) = E Y [ u ( k + Y 1 n e t ) ] U^{\mathrm{hh}}(k)=E_Y[u(k+Y_1^{\mathrm{net}})] U hh ( k ) = E Y [ u ( k + Y 1 net )] and let M j h h ( k ) = E Y [ ( k + Y 1 n e t ) − γ − j ] M_j^{\mathrm{hh}}(k)=E_Y[(k+Y_1^{\mathrm{net}})^{-\gamma-j}] M j hh ( k ) = E Y [( k + Y 1 net ) − γ − j ] . Its absolute risk aversion and absolute prudence are
A h h ( k ) = γ M 1 h h M 0 h h , P h h ( k ) = ( γ + 1 ) M 2 h h M 1 h h . \mathcal{A}^{\mathrm{hh}}(k)=\gamma\frac{M_1^{\mathrm{hh}}}{M_0^{\mathrm{hh}}},\qquad\mathcal{P}^{\mathrm{hh}}(k)=(\gamma+1)\frac{M_2^{\mathrm{hh}}}{M_1^{\mathrm{hh}}}. A hh ( k ) = γ M 0 hh M 1 hh , P hh ( k ) = ( γ + 1 ) M 1 hh M 2 hh .
Both are strictly decreasing. For example,
( A h h ) ′ = γ − ( γ + 1 ) M 2 h h M 0 h h + γ ( M 1 h h ) 2 ( M 0 h h ) 2 < 0 , (67) \left(\mathcal{A}^{\mathrm{hh}}\right)'=\gamma\frac{-(\gamma+1)M_2^{\mathrm{hh}}M_0^{\mathrm{hh}}+\gamma\left(M_1^{\mathrm{hh}}\right)^2}{\left(M_0^{\mathrm{hh}}\right)^2}<0, \tag*{(67)} ( A hh ) ′ = γ ( M 0 hh ) 2 − ( γ + 1 ) M 2 hh M 0 hh + γ ( M 1 hh ) 2 < 0 , (67)
because ( M 1 h h ) 2 ≤ M 0 h h M 2 h h (M_1^{\mathrm{hh}})^2 \le M_0^{\mathrm{hh}}M_2^{\mathrm{hh}} ( M 1 hh ) 2 ≤ M 0 hh M 2 hh . The same argument, using ( M 2 h h ) 2 ≤ M 1 h h M 3 h h (M_2^{\mathrm{hh}})^2 \le M_1^{\mathrm{hh}}M_3^{\mathrm{hh}} ( M 2 hh ) 2 ≤ M 1 hh M 3 hh , gives ( P h h ) ′ < 0 (\mathcal{P}^{\mathrm{hh}})' < 0 ( P hh ) ′ < 0 . Also ( U h h ) ′ ′ ′ > 0 (U^{\mathrm{hh}})''' > 0 ( U hh ) ′′′ > 0 . This is the decreasing-risk-aversion and decreasing-prudence structure underlying standard risk aversion [23 ] . The calculation here also allows consumption and saving to adjust.
Let e j V = V j − R f P e_j^V = V_j - R_fP e j V = V j − R f P and k j f i n = R f ( w 0 n e t − c 0 s u r ) + e j V x k_j^{\mathrm{fin}} = R_f(w_0^{\mathrm{net}} - c_0^{\mathrm{sur}}) + e_j^V x k j fin = R f ( w 0 net − c 0 sur ) + e j V x . The two first-order conditions are
u ′ ( c 0 s u r ) = ϖ R f E V [ ( U h h ) ′ ( k j f i n ) ] , E V [ e V ( U h h ) ′ ( k j f i n ) ] = 0. (68) u'(c_0^{\mathrm{sur}}) = \varpi R_f E_V[(U^{\mathrm{hh}})'(k_j^{\mathrm{fin}})], \qquad E_V[e^V(U^{\mathrm{hh}})'(k_j^{\mathrm{fin}})] = 0. \tag*{(68)} u ′ ( c 0 sur ) = ϖ R f E V [( U hh ) ′ ( k j fin )] , E V [ e V ( U hh ) ′ ( k j fin )] = 0. (68)
All moments in the rest of this proof are evaluated at this optimum. To keep them distinct from the propagation matrices, give them the subscript hh and let
a h h = − u ′ ′ ( c 0 s u r ) , b h h = − E V [ ( U h h ) ′ ′ ] , k h h = E V [ e V ( U h h ) ′ ′ ] , a_{\mathrm{hh}} = -u''(c_0^{\mathrm{sur}}), \qquad b_{\mathrm{hh}} = -E_V[(U^{\mathrm{hh}})''], \qquad k_{\mathrm{hh}} = E_V[e^V(U^{\mathrm{hh}})''], a hh = − u ′′ ( c 0 sur ) , b hh = − E V [( U hh ) ′′ ] , k hh = E V [ e V ( U hh ) ′′ ] ,
d h h = − E V [ ( e V ) 2 ( U h h ) ′ ′ ] , u h h = E V [ ( U h h ) ′ ′ ] , m h h = E V [ e V ( U h h ) ′ ′ ] . d_{\mathrm{hh}} = -E_V[(e^V)^2(U^{\mathrm{hh}})''], \qquad u_{\mathrm{hh}} = E_V[(U^{\mathrm{hh}})''], \qquad m_{\mathrm{hh}} = E_V[e^V(U^{\mathrm{hh}})'']. d hh = − E V [( e V ) 2 ( U hh ) ′′ ] , u hh = E V [( U hh ) ′′ ] , m hh = E V [ e V ( U hh ) ′′ ] .
The determinant of the negative Hessian, after dividing its credit row by ϖ \varpi ϖ , is
D h h = ( a h h + ϖ R f 2 b h h ) d h h − ϖ R f 2 k h h 2 > 0. \mathcal{D}_{\mathrm{hh}} = (a_{\mathrm{hh}} + \varpi R_f^2 b_{\mathrm{hh}})d_{\mathrm{hh}} - \varpi R_f^2 k_{\mathrm{hh}}^2 > 0. D hh = ( a hh + ϖ R f 2 b hh ) d hh − ϖ R f 2 k hh 2 > 0.
Strict concavity gives the inequality. In particular, b h h d h h − k h h 2 > 0 b_{\mathrm{hh}}d_{\mathrm{hh}} - k_{\mathrm{hh}}^2 > 0 b hh d hh − k hh 2 > 0 by the weighted Cauchy–Schwarz inequality and the two distinct excess payoffs.
For the signs, let μ ˉ h h = E V [ ( U h h ) ′ ] \bar{\mu}_{\mathrm{hh}} = E_V[(U^{\mathrm{hh}})'] μ ˉ hh = E V [( U hh ) ′ ] and reweight each payoff-state probability by its marginal utility divided by μ ˉ h h \bar{\mu}_{\mathrm{hh}} μ ˉ hh . The second condition in (H.3) makes the reweighted mean excess payoff zero. Let
j ˙ r e t = − e L V e H V e H V − e L V > 0 , A j h h = A h h ( k j f i n ) , P j h h = P h h ( k j f i n ) . \dot{j}_{\mathrm{ret}} = \frac{-e_L^V e_H^V}{e_H^V-e_L^V} > 0, \qquad A_j^{\mathrm{hh}} = A^{\mathrm{hh}}(k_j^{\mathrm{fin}}), \qquad\mathcal{P}_j^{\mathrm{hh}} = \mathcal{P}^{\mathrm{hh}}(k_j^{\mathrm{fin}}). j ˙ ret = e H V − e L V − e L V e H V > 0 , A j hh = A hh ( k j fin ) , P j hh = P hh ( k j fin ) .
Since x > 0 x > 0 x > 0 , k H f i n > k L f i n k_H^{\mathrm{fin}} > k_L^{\mathrm{fin}} k H fin > k L fin , and both risk aversion and prudence are larger in the low-payoff state. Consequently,
k h h = μ ˉ h h j ˙ r e t ( A L h h − A H h h ) > 0 , k_{\mathrm{hh}} = \bar{\mu}_{\mathrm{hh}}\dot{j}_{\mathrm{ret}}(A_L^{\mathrm{hh}} - A_H^{\mathrm{hh}}) > 0, k hh = μ ˉ hh j ˙ ret ( A L hh − A H hh ) > 0 ,
m h h = − μ ˉ h h j ˙ r e t ( A L h h P L h h − A H h h P H h h ) < 0. m_{\mathrm{hh}} = -\bar{\mu}_{\mathrm{hh}}\dot{j}_{\mathrm{ret}}(A_L^{\mathrm{hh}}\mathcal{P}_L^{\mathrm{hh}} - A_H^{\mathrm{hh}}\mathcal{P}_H^{\mathrm{hh}}) < 0. m hh = − μ ˉ hh j ˙ ret ( A L hh P L hh − A H hh P H hh ) < 0.
Differentiating (H.3) with respect to net resources gives
∂ x ∂ w 0 n e t = R f a h h k h h D h h > 0 , ∂ c 0 s u r ∂ w 0 n e t = ϖ R f 2 ( b h h d h h − k h h 2 ) D h h ∈ ( 0 , 1 ) . (69) \frac{\partial x}{\partial w_0^{\mathrm{net}}} = \frac{R_f a_{\mathrm{hh}}k_{\mathrm{hh}}}{\mathcal{D}_{\mathrm{hh}}} > 0, \qquad\frac{\partial c_0^{\mathrm{sur}}}{\partial w_0^{\mathrm{net}}} = \frac{\varpi R_f^2(b_{\mathrm{hh}}d_{\mathrm{hh}} - k_{\mathrm{hh}}^2)}{\mathcal{D}_{\mathrm{hh}}} \in(0,1). \tag*{(69)} ∂ w 0 net ∂ x = D hh R f a hh k hh > 0 , ∂ w 0 net ∂ c 0 sur = D hh ϖ R f 2 ( b hh d hh − k hh 2 ) ∈ ( 0 , 1 ) . (69)
Current assistance subtracts from w 0 n e t w_0^{\mathrm{net}} w 0 net , so its credit-demand derivative is the negative of the first expression. A future assistance obligation instead gives
∂ x ∂ T 1 = − a h h k h h D h h < 0. \frac{\partial x}{\partial T_1}=-\frac{a_{\mathrm{hh}}k_{\mathrm{hh}}}{\mathcal{D}_{\mathrm{hh}}}<0. ∂ T 1 ∂ x = − D hh a hh k hh < 0.
Thus current consumption does not absorb the whole current resource loss, and risky investment is normal at this interior optimum. No fixed portfolio share is assumed.
For the small additional risk, ∂ ν ( U h h ) ′ = 1 2 ( U h h ) ′ ′ ′ \partial_\nu(\mathcal{U}^{\mathrm{hh}})'=\frac{1}{2}(\mathcal{U}^{\mathrm{hh}})''' ∂ ν ( U hh ) ′ = 2 1 ( U hh ) ′′′ at ν = 0 \nu=0 ν = 0 . Implicit differentiation then gives
x ν = ϖ R f 2 k h h u h h + ( a h h + ϖ R f 2 b h h ) m h h 2 D h h , (70) x_\nu=\frac{\varpi R_f^2k_{\mathrm{hh}}u_{\mathrm{hh}}+(a_{\mathrm{hh}}+\varpi R_f^2b_{\mathrm{hh}})m_{\mathrm{hh}}}{2\mathcal{D}_{\mathrm{hh}}}, \tag*{(70)} x ν = 2 D hh ϖ R f 2 k hh u hh + ( a hh + ϖ R f 2 b hh ) m hh , (70)
( c 0 s u r ) ν = − ϖ R f ( d h h u h h + k h h m h h ) 2 D h h . (71) (c_0^{\mathrm{sur}})_\nu=-\frac{\varpi R_f(d_{\mathrm{hh}}u_{\mathrm{hh}}+k_{\mathrm{hh}}m_{\mathrm{hh}})}{2\mathcal{D}_{\mathrm{hh}}}. \tag*{(71)} ( c 0 sur ) ν = − 2 D hh ϖ R f ( d hh u hh + k hh m hh ) . (71)
The two payoff states imply the identities
b h h ( − m h h ) − k h h u h h = μ ˉ h h 2 j r e t A L h h A H h h ( P L h h − P H h h ) > 0 , b_{\mathrm{hh}}(-m_{\mathrm{hh}})-k_{\mathrm{hh}}u_{\mathrm{hh}}=\bar{\mu}_{\mathrm{hh}}^2j_{\mathrm{ret}}\mathcal{A}^{\mathrm{hh}}_L\mathcal{A}^{\mathrm{hh}}_H(\mathcal{P}^{\mathrm{hh}}_L-\mathcal{P}^{\mathrm{hh}}_H)>0, b hh ( − m hh ) − k hh u hh = μ ˉ hh 2 j ret A L hh A H hh ( P L hh − P H hh ) > 0 ,
d h h u h h + k h h m h h = μ ˉ h h 2 j r e t A L h h A H h h ( e H V P L h h − e L V P H h h ) > 0. d_{\mathrm{hh}}u_{\mathrm{hh}}+k_{\mathrm{hh}}m_{\mathrm{hh}}=\bar{\mu}_{\mathrm{hh}}^2j_{\mathrm{ret}}\mathcal{A}^{\mathrm{hh}}_L\mathcal{A}^{\mathrm{hh}}_H(e^{V}_{H}\mathcal{P}^{\mathrm{hh}}_L-e^{V}_{L}\mathcal{P}^{\mathrm{hh}}_H)>0. d hh u hh + k hh m hh = μ ˉ hh 2 j ret A L hh A H hh ( e H V P L hh − e L V P H hh ) > 0.
The numerator of (H.5) can therefore be written as − a h h ( − m h h ) − ϖ R f 2 [ b h h ( − m h h ) − k h h u h h ] < 0 -a_{\mathrm{hh}}(-m_{\mathrm{hh}})-\varpi R_f^2[b_{\mathrm{hh}}(-m_{\mathrm{hh}})-k_{\mathrm{hh}}u_{\mathrm{hh}}]<0 − a hh ( − m hh ) − ϖ R f 2 [ b hh ( − m hh ) − k hh u hh ] < 0 . Both consumption and credit holdings fall. Since current resources have not changed, ( a s a f e ) ν = − ( c 0 s u r ) ν − P x ν > 0 (a^{\mathrm{safe}})_\nu=-(c^{\mathrm{sur}}_0)_\nu-Px_\nu>0 ( a safe ) ν = − ( c 0 sur ) ν − P x ν > 0 .
Interpretation and the boundary. Precaution increases total saving in this problem, but the increase is held safely and is accompanied by a reduction in credit. Saving and portfolio reallocation are chosen together, not placed in a sequence. No separate or earmarked saving account is needed. Similarly, the current-transfer result follows from normal risky demand and a consumption response below one, not from assuming that every transferred dollar was previously invested in bonds. With CARA utility and independent background risk, the unconstrained risky position has no wealth effect, so this particular transfer argument vanishes.
Lower current commitments provide another interpretation of a spending decline. Since c 0 = c 0 + c 0 s u r c_0=c_0+c_0^{\mathrm{sur}} c 0 = c 0 + c 0 sur , (H.4) implies, with other inputs fixed,
∂ c 0 ∂ c ˉ 0 = 1 − ∂ c 0 s u r ∂ w 0 n e t > 0 , ∂ x ∂ c ˉ 0 = − ∂ x ∂ w 0 n e t < 0. \frac{\partial c_0}{\partial\bar{c}_0}=1-\frac{\partial c_0^{\mathrm{sur}}}{\partial w_0^{\mathrm{net}}}>0,\qquad\frac{\partial x}{\partial\bar{c}_0}=-\frac{\partial x}{\partial w_0^{\mathrm{net}}}<0. ∂ c ˉ 0 ∂ c 0 = 1 − ∂ w 0 net ∂ c 0 sur > 0 , ∂ c ˉ 0 ∂ x = − ∂ w 0 net ∂ x < 0.
Thus lower current commitments reduce observed consumption while raising surplus resources and credit holdings at this interior optimum.
Interiority is an economic scope condition. For example, let current resources be 100, future income initially 50, P = R f = ϖ = 1 P=R_f=\varpi=1 P = R f = ϖ = 1 , γ = 2 \gamma=2 γ = 2 , and V V V be 0.8 or 1.4 with equal probabilities. With no borrowing, the optimum has consumption 74.92, credit 25.08 and zero safe balances. Replacing certain income by independent income of 10 or 90 with equal probabilities lowers consumption to 63.05 but raises credit to 31.00 and safe balances to 5.95. This finite-change example starts at a binding portfolio boundary and is not covered by Proposition H.1 . Observing positive balances in both accounts after the shock would not reveal that initial constraint.
Equity and the direction of precaution. Corporate bonds can be safer than equities without being safe assets. A cautious household could therefore move into credit as well as out of it. To study this margin, add equity holdings x E ≥ 0 x^{E} \ge0 x E ≥ 0 with price P E P^{E} P E and payoff V E V^{E} V E to (H.2), so that
c 0 s u r + a s a f e + P x + P E x E = w 0 n e t , c 1 s u r = R f a s a f e + V x + V E x E + Y 1 n e t > 0. c_{0}^{\mathrm{sur}} + a^{\mathrm{safe}} + Px + P^{E}x^{E} = w_{0}^{\mathrm{net}}, \qquad c_{1}^{\mathrm{sur}} = R_{f}a^{\mathrm{safe}} + Vx + V^{E}x^{E} + Y_{1}^{\mathrm{net}} > 0. c 0 sur + a safe + P x + P E x E = w 0 net , c 1 sur = R f a safe + V x + V E x E + Y 1 net > 0.
The payoffs take three values across ordered financial states, a worst state (L L L ), a downturn (M M M ) and a boom (H H H ), with probabilities p j > 0 p_j > 0 p j > 0 . Both payoffs are nonnegative and weakly increasing across these states. The two claims need not be issued by the same firm, and the worst state need not be an issuer’s default. Let Δ 1 V = V M − V L \Delta_{1}V = V_{M} - V_{L} Δ 1 V = V M − V L and Δ 2 V = V H − V M \Delta_{2}V = V_{H} - V_{M} Δ 2 V = V H − V M , define Δ 1 V E \Delta_{1}V^{E} Δ 1 V E and Δ 2 V E \Delta_{2}V^{E} Δ 2 V E in the same way, and let
s h h = Δ 1 V Δ 2 V E − Δ 2 V Δ 1 V E ≠ 0 , \mathfrak{s}_{hh} = \Delta_{1}V\Delta_{2}V^{E} - \Delta_{2}V\Delta_{1}V^{E} \ne0, s hh = Δ 1 V Δ 2 V E − Δ 2 V Δ 1 V E = 0 ,
so that the three assets span the three states. When s h h > 0 \mathfrak{s}_{hh} > 0 s hh > 0 , credit loses a larger part of its payoff range than equity between the downturn and the worst state, so its losses are concentrated in the worst state. Credit with a capped payoff, which pays the same amount outside the worst state (Δ 2 V = 0 < Δ 1 V \Delta_{2}V = 0 < \Delta_{1}V Δ 2 V = 0 < Δ 1 V ), is the leading case. When s h h < 0 \mathfrak{s}_{hh} < 0 s hh < 0 , credit loses a smaller part of its payoff range there than equity does. Net future income Y 1 n e t Y_{1}^{\mathrm{net}} Y 1 net is certain at the reference point. The optimum is strictly interior, with positive safe, credit and equity positions that can all adjust. Let c j s u r c_{j}^{\mathrm{sur}} c j sur denote future surplus in state j j j at this optimum, and let the added income risk be ν ξ \sqrt{\nu}\,\xi ν ξ as before.
Proposition H.2 (Precaution with credit and equity). Under these conditions, a small independent addition to income risk lowers current consumption and raises future surplus in the worst state. Let g 1 = 1 / ( c L s u r c M s u r ) + τ h h \mathfrak{g}_{1} = 1/(c_{L}^{\mathrm{sur}}c_{M}^{\mathrm{sur}}) + \tau_{hh} g 1 = 1/ ( c L sur c M sur ) + τ hh and g 2 = 1 / ( c M s u r c H s u r ) + τ h h \mathfrak{g}_{2} = 1/(c_{M}^{\mathrm{sur}}c_{H}^{\mathrm{sur}}) + \tau_{hh} g 2 = 1/ ( c M sur c H sur ) + τ hh , with τ h h > 0 \tau_{hh} > 0 τ hh > 0 given in the proof, so that g 1 > g 2 > 0 \mathfrak{g}_{1} > \mathfrak{g}_{2} > 0 g 1 > g 2 > 0 . The two risky holdings respond as
x ν = − γ + 1 2 [ x g 2 + Δ 2 V E ( c M s u r − c L s u r ) ( g 1 − g 2 ) s h h ] , x_{\nu} = -\frac{\gamma+1}{2}\left[x\mathfrak{g}_{2} + \frac{\Delta_{2}V^{E}(c_{M}^{\mathrm{sur}}-c_{L}^{\mathrm{sur}})(\mathfrak{g}_{1}-\mathfrak{g}_{2})}{\mathfrak{s}_{hh}}\right], x ν = − 2 γ + 1 [ x g 2 + s hh Δ 2 V E ( c M sur − c L sur ) ( g 1 − g 2 ) ] , x ν E = − γ + 1 2 [ x E g 2 − Δ 2 V ( c M s u r − c L s u r ) ( g 1 − g 2 ) s h h ] . (72) x_{\nu}^{E} = -\frac{\gamma+1}{2}\left[x^{E}\mathfrak{g}_{2} - \frac{\Delta_{2}V(c_{M}^{\mathrm{sur}}-c_{L}^{\mathrm{sur}})(\mathfrak{g}_{1}-\mathfrak{g}_{2})}{\mathfrak{s}_{hh}}\right]. \tag*{(72)} x ν E = − 2 γ + 1 [ x E g 2 − s hh Δ 2 V ( c M sur − c L sur ) ( g 1 − g 2 ) ] . (72)
*The first terms lower both holdings at a common proportional rate. The second terms tilt the portfolio away from the claim whose losses are concentrated in the worst state, relative to the other claim. (i) If s h h > 0 \mathfrak{s}_{hh} > 0 s hh > 0 , credit holdings fall, and by a larger proportion than equity holdings. With a capped credit payoff, equity holdings also fall and safe investment rises. (ii) If s h h < 0 s_{\mathrm{hh}} < 0 s hh < 0 , equity holdings fall, and credit holdings fall by a smaller proportion or rise. A marginal increase in T 0 T_0 T 0 or T 1 T_1 T 1 lowers credit and equity holdings in the same proportion. These are fixed-price responses at the decision date.
Proof. The three assets span the payoff states and Y 1 n e t Y_1^{\mathrm{net}} Y 1 net is certain. The household therefore chooses c 0 s u r c_0^{\mathrm{sur}} c 0 sur and the three values c j s u r c_j^{\mathrm{sur}} c j sur subject to c 0 s u r + ∑ j s j c j s u r = w 0 n e t + Y 1 n e t / R f c_0^{\mathrm{sur}} + \sum_j s_j c_j^{\mathrm{sur}} = w_0^{\mathrm{net}} + Y_1^{\mathrm{net}}/R_f c 0 sur + ∑ j s j c j sur = w 0 net + Y 1 net / R f , where the state prices s j > 0 s_j > 0 s j > 0 are fixed by ( R f , P , P E ) (R_f, P, P^E) ( R f , P , P E ) and equal ϖ p j ( c 0 s u r / c j s u r ) γ \varpi p_j(c_0^{\mathrm{sur}}/c_j^{\mathrm{sur}})^\gamma ϖ p j ( c 0 sur / c j sur ) γ at the optimum. The added income risk is not spanned. With it, the condition for state j j j is ϖ p j E ξ [ u ′ ( c j s u r + ν ξ ) ] = s j u ′ ( c 0 s u r ) \varpi p_j E_\xi[u'(c_j^{\mathrm{sur}} + \sqrt{\nu}\xi)] = s_j u'(c_0^{\mathrm{sur}}) ϖ p j E ξ [ u ′ ( c j sur + ν ξ )] = s j u ′ ( c 0 sur ) , and ∂ ν E ξ [ u ′ ( c + ν ξ ) ] = u ′ ′ ′ ( c ) / 2 \partial_\nu E_\xi[u'(c + \sqrt{\nu}\xi)] = u'''(c)/2 ∂ ν E ξ [ u ′ ( c + ν ξ )] = u ′′′ ( c ) /2 at ν = 0 \nu= 0 ν = 0 . Differentiating these conditions and the budget gives
( c 0 s u r ) ν = − γ + 1 2 t h h c 0 s u r , ( c j s u r ) ν = γ + 1 2 ( 1 c j s u r − t h h c j s u r ) , t h h = ∑ j s j / c j s u r c 0 s u r + ∑ j s j c j s u r . (c_0^{\mathrm{sur}})_\nu= -\frac{\gamma+1}{2}t_{\mathrm{hh}}c_0^{\mathrm{sur}}, \qquad(c_j^{\mathrm{sur}})_\nu= \frac{\gamma+1}{2}\left(\frac{1}{c_j^{\mathrm{sur}}}-t_{\mathrm{hh}}c_j^{\mathrm{sur}}\right), \qquad t_{\mathrm{hh}} = \frac{\sum_j s_j/c_j^{\mathrm{sur}}}{c_0^{\mathrm{sur}}+\sum_j s_jc_j^{\mathrm{sur}}}. ( c 0 sur ) ν = − 2 γ + 1 t hh c 0 sur , ( c j sur ) ν = 2 γ + 1 ( c j sur 1 − t hh c j sur ) , t hh = c 0 sur + ∑ j s j c j sur ∑ j s j / c j sur .
Since t h h < 1 / ( c L s u r ) 2 t_{\mathrm{hh}} < 1/(c_L^{\mathrm{sur}})^2 t hh < 1/ ( c L sur ) 2 , surplus in the worst state rises. For any two states, ( c j s u r − c i s u r ) ν = − γ + 1 2 ( c j s u r − c i s u r ) [ 1 / ( c i s u r c j s u r ) + t h h ] (c_j^{\mathrm{sur}}-c_i^{\mathrm{sur}})_\nu= -\frac{\gamma+1}{2}(c_j^{\mathrm{sur}}-c_i^{\mathrm{sur}})[1/(c_i^{\mathrm{sur}}c_j^{\mathrm{sur}})+t_{\mathrm{hh}}] ( c j sur − c i sur ) ν = − 2 γ + 1 ( c j sur − c i sur ) [ 1/ ( c i sur c j sur ) + t hh ] . Precaution thus narrows every gap in surplus between a better and a worse state, and it narrows the gap to the worst state by the larger proportion. The positions determine the two gaps through
c M s u r − c L s u r = Δ 1 V x + Δ 1 V E x E , c H s u r − c M s u r = Δ 2 V x + Δ 2 V E x E . c_M^{\mathrm{sur}}-c_L^{\mathrm{sur}} = \Delta_1 Vx+\Delta_1 V^E x^E, \qquad c_H^{\mathrm{sur}}-c_M^{\mathrm{sur}} = \Delta_2 Vx+\Delta_2 V^E x^E. c M sur − c L sur = Δ 1 V x + Δ 1 V E x E , c H sur − c M sur = Δ 2 V x + Δ 2 V E x E .
both of which are positive. Differentiating these identities and solving for the two responses gives (H.7). In case (i) the bracket for credit exceeds x g 2 x\,g_2 x g 2 , and the bracket for equity is at most x E g 2 x^E g_2 x E g 2 , with equality for a capped credit payoff. In case (ii) the two roles are reversed. Safe investment satisfies R f a ν s a f e = ( c L s u r ) ν − V L x ν − V L E x ν E R_f a^{\mathrm{safe}}_\nu= (c_L^{\mathrm{sur}})_\nu- V_Lx_\nu- V_L^E x_\nu^E R f a ν safe = ( c L sur ) ν − V L x ν − V L E x ν E , which is positive when both holdings fall. Finally, at fixed prices CRRA utility makes c 0 s u r c_0^{\mathrm{sur}} c 0 sur and every c j s u r c_j^{\mathrm{sur}} c j sur proportional to surplus wealth w 0 n e t + Y 1 n e t / R f w_0^{\mathrm{net}} + Y_1^{\mathrm{net}}/R_f w 0 net + Y 1 net / R f . Both risky positions are linear in the two gaps, so ∂ x / ∂ w 0 n e t = x / ( w 0 n e t + Y 1 n e t / R f ) \partial x/\partial w_0^{\mathrm{net}} = x/(w_0^{\mathrm{net}} + Y_1^{\mathrm{net}}/R_f) ∂ x / ∂ w 0 net = x / ( w 0 net + Y 1 net / R f ) , and the same expression holds for x E x^E x E . Current assistance lowers w 0 n e t w_0^{\mathrm{net}} w 0 net , and a future obligation lowers Y 1 n e t Y_1^{\mathrm{net}} Y 1 net .
Existing income risk. When Y 1 n e t Y_1^{\mathrm{net}} Y 1 net is already risky, the same steps apply to effective utility U h h U^{\mathrm{hh}} U hh and to financial wealth k j f i n = R f a s a f e + V j x + V j E x E k_j^{\mathrm{fin}} = R_fa^{\mathrm{safe}} + V_jx + V_j^E x^E k j fin = R f a safe + V j x + V j E x E . They give ( k j f i n ) ν = 1 2 P j h h − Λ h h / A j h h (k_j^{\mathrm{fin}})_\nu= \frac{1}{2}\mathcal{P}_j^{\mathrm{hh}} - \Lambda_{\mathrm{hh}}/A_j^{\mathrm{hh}} ( k j fin ) ν = 2 1 P j hh − Λ hh / A j hh , with Λ h h = − γ ( c 0 s u r ) ν / c 0 s u r > 0 \Lambda_{\mathrm{hh}} = -\gamma(c_0^{\mathrm{sur}})_\nu/c_0^{\mathrm{sur}} > 0 Λ hh = − γ ( c 0 sur ) ν / c 0 sur > 0 . Risk aversion and prudence are both decreasing, so this response is larger in worse states and both gaps still narrow,
Δ 1 V x ν + Δ 1 V E x ν E < 0 , Δ 2 V x ν + Δ 2 V E x ν E < 0. \Delta_1 Vx_\nu+\Delta_1 V^E x_\nu^E < 0, \qquad\Delta_2 Vx_\nu+\Delta_2 V^E x_\nu^E < 0. Δ 1 V x ν + Δ 1 V E x ν E < 0 , Δ 2 V x ν + Δ 2 V E x ν E < 0.
The household therefore lowers its exposure to the worst state and cannot raise both holdings. With a capped credit payoff, equity holdings fall. The division of the reduction between credit and equity is pinned down by (H.7) only when net future income is certain. In particular, with three assets credit demand need not rise with resources once income is already risky, so the transfer statement of the proposition is a benchmark under certain income.
Three examples. Let P = P E = R f = ϖ = 1 P = P^{E} = R_f = \varpi= 1 P = P E = R f = ϖ = 1 , γ = 2 \gamma= 2 γ = 2 and current resources be 100, and let the worst state, the downturn and the boom have probabilities 0.2, 0.3 and 0.5. First, let credit have a capped payoff, paying 0.65 in the worst state and 1.15 otherwise, and let equity pay 0.30, 0.90 and 1.60. With certain future income of 10, the optimum has consumption 55.22, safe balances 26.38, credit 9.09 and equity 9.31. Replacing certain income by independent income of 2 or 18 with equal probabilities lowers consumption to 54.35, credit to 8.55 and equity to 8.93, and raises safe balances to 28.17. Credit falls by 5.9 percent and equity by 4.1 percent. Second, let credit lose no additional value in the worst state, paying 0.95, 0.95 and 1.10, while equity pays 0.30, 0.90 and 1.80. The optimum has consumption 55.48, safe balances 22.90, credit 5.52 and equity 16.10. The same change in income lowers consumption to 54.59 and equity to 15.25 but raises credit to 6.46 and safe balances to 23.69. This household moves from stocks into bonds although its safe balance can adjust.
Third, return to the capped credit payoff and let future income be 60. With no borrowing, the optimum has consumption 80.22, credit 3.32, equity 16.47 and zero safe balances. Replacing certain income by 50 or 70 lowers consumption to 79.31 and equity to 15.26 but raises credit to 5.43, with safe balances still at zero. Credit is then the safest claim that can adjust, and precaution moves the household into it even though its losses are concentrated in the worst state. This last case starts at a binding boundary and is not covered by Proposition H.2 . In all three examples a loss of one unit of current resources lowers credit holdings.
The household problem holds all asset prices fixed. It therefore delivers a conditional shift in credit demand, which the pricing map of § 3 takes as given. It is not a joint equilibrium of the credit and equity markets.
Equation (10) remains a local surplus-wealth illustration, not the exact global demand function of (H.2). Surplus consumption in utility and wealth above the present value of commitments are distinct quantities. The optimization supplies conditions for the sign of the local fixed-price derivative. Under their respective conditions, the two propositions give assistance needs that rise with z i z_i z i a nonnegative term in l i l_i l i at the reference state. If independent background risk also rises with z i z_i z i , its term is nonnegative in the two-asset problem and in case (i) of Proposition H.2 , which assumes certain net income. It can be negative in case (ii) or at a portfolio boundary. The net loading l i l_i l i is the sum of these terms and can therefore have either sign. The propositions supply neither an empirical value for l i l_i l i nor a duration for which that derivative stays fixed. Correlated income risk, more payoff states than assets and later balance-sheet accumulation require separate restrictions.
Signed loadings in the pricing results. Proposition 6 and its proof in Appendix § 12.8 .10 place no restriction on the sign of l l l . When households raise desired credit holdings as conditions worsen, the corresponding entries of l l l are negative. The target remains l ′ z l^{\prime}z l ′ z , and the same recovery condition applies to these signed loadings. The recovery weights change with the loadings. They are computed from independently measured signed loadings and are not chosen to fit the EBP. Signed loadings require three qualifications. First, the normalization π H = l / δ \pi^{H}=l/\delta π H = l / δ requires δ ≠ 0 \delta\ne0 δ = 0 , and the kernel can always be written with l l l directly. Loadings that sum to zero still respond to a geographically uneven field. Second, the level-sign result of Appendix § 12.8 .§ 7 requires l ≥ 0 l \ge0 l ≥ 0 . Without it the kernel can change sign across horizons, because diffusion moves the response between counties whose loadings differ in sign. In the two-county example of (21), loadings ( l 1 , l 2 ) (l_1,l_2) ( l 1 , l 2 ) and a unit input in county 1 give
K 1 ( h ) = c E ρ D h − 1 2 [ ( l 1 + l 2 ) + ( l 1 − l 2 ) e − 2 κ D ( h − 1 ) ] . (73) K_1(h)=\frac{c_E\rho_D^{h-1}}{2}\left[(l_1+l_2)+(l_1-l_2)e^{-2\kappa_D(h-1)}\right]. \tag*{(73)} K 1 ( h ) = 2 c E ρ D h − 1 [ ( l 1 + l 2 ) + ( l 1 − l 2 ) e − 2 κ D ( h − 1 ) ] . (73)
For κ D > 0 \kappa_D>0 κ D > 0 and 0 < ρ D < 1 0<\rho_D<1 0 < ρ D < 1 , this response switches sign once over the full horizon path if l 1 ( l 1 + l 2 ) < 0 l_1(l_1+l_2)<0 l 1 ( l 1 + l 2 ) < 0 , and does not switch otherwise. The switch need not fall inside the horizons studied, and without diffusion there is none. Third, with C Y = Id n C C_Y=\operatorname{Id}_{n_C} C Y = Id n C and ker T = span { 1 } \ker T=\operatorname{span}\{1\} ker T = span { 1 } , row ( T ) \operatorname{row}(T) row ( T ) is the set of vectors orthogonal to 1 1 1 . For unrestricted county responses, within-event contrasts therefore recover the target exactly when l ′ 1 = 0 l^{\prime}1=0 l ′ 1 = 0 , that is, when a response common to all counties leaves credit demand unchanged. On a restricted response space, Proposition 6 gives the corresponding condition. Theorem 1 uses the weights only through the modal coefficients ω k \omega_k ω k , whose signs are unrestricted, and applies unchanged. When household responses offset at a horizon, l ′ z = 0 l^{\prime}z=0 l ′ z = 0 and the household-demand component of the premium does not move at that horizon, holding the other channels fixed. This is a zero demand shift. It differs from the attenuated response that arises when intermediaries absorb part of a nonzero shift.
Table H.1. Liquid balances among holders of non-retirement credit proxies
Notes: The 2022 Survey of Consumer Finances, 4,595 public-use families and five implicates (184–196 holders of these categories per implicate). Survey-weighted ratios are computed in each implicate and averaged. Direct bonds are corporate or other bonds (X7639). Bond funds are other bond mutual funds and ETFs (X3828). Both exclude separately recorded retirement accounts, but neither category isolates corporate credit. The combined column sums the two categories. Transaction balances sum checking, savings, money-market accounts and funds, brokerage call accounts and prepaid accounts. Months divide these balances by one-twelfth of annual normal income, in common dollar units. Asset columns use all dollars in the named category as the denominator. The last column uses all families holding either category. Nonpositive normal income does not satisfy a months threshold. These are descriptive point estimates, not estimates of portfolio interiority or adjustment capacity. Source: Federal Reserve Board, 2022 SCF.
Netting donors and recipients. Transfers redistribute resources. Let g i c r e d = P ∂ x i ∂ w i 0 n e t g_i^{\mathrm{cred}} = P \frac{\partial x_i}{\partial w_{i0}^{\mathrm{net}}} g i cred = P ∂ w i 0 net ∂ x i be the household’s fixed-price marginal credit demand per dollar of current net resources. A transfer of t d r ≥ 0 t_{dr} \ge0 t d r ≥ 0 from donor d d d to recipient r r r lowers donor desired dollar holdings by g d c r e d t d r g_d^{\mathrm{cred}}t_{dr} g d cred t d r and raises recipient desired dollar holdings by g r c r e d t d r g_r^{\mathrm{cred}}t_{dr} g r cred t d r . With all other inputs fixed, its net selling pressure is
S t r $ = ∑ d , r ( g d c r e d − g r c r e d ) t d r . (74) S_{tr}^{\$} = \sum_{d,r}\left(g_d^{\mathrm{cred}} - g_r^{\mathrm{cred}}\right)t_{dr}. \tag*{(74)} S t r $ = d , r ∑ ( g d cred − g r cred ) t d r . (74)
Thus transfers generate positive aggregate selling pressure only when the transfer-weighted donor response exceeds the recipient response. Recipient spending or low credit participation can make this condition plausible, but do not establish it. Recipients remain in the financial aggregation even when directly hit counties are excluded from a spending comparison. The pricing kernel applies to the net fixed-price demand shift, not gross donor sales. This is a static accounting restriction and does not require a new propagation law.
Observable scope. Table H.1 describes liquid balances among households holding two non-retirement credit proxies in the 2022 SCF. The categories are held by about 1.35 percent of families and do not cover all household credit exposure. They are used because the public survey separates them from retirement accounts, not because they isolate the relevant corporate claims perfectly.
These observations make liquidity and account access measurable dimensions of the sign condition. They do not classify an optimum as interior. Transaction needs, committed balances and trading restrictions can bind even when both balances are positive. Moreover, withdrawal access matters for financing a transfer, whereas precautionary rebalancing can occur within a retirement account without a withdrawal. Measuring those restrictions alongside the derivatives in main equation (27) is what separates the two routes empirically.
Relative identification and the role of the operator
Let S a ϕ = ( ℓ a ϕ ) ′ z S_a^{\phi} = (\ell_a^{\phi})'z S a ϕ = ( ℓ a ϕ ) ′ z be fixed-price dollar demand reduction. At a specified reference state, suppose Δ p a = c a S a ϕ + n a \Delta p_a = c_a S_a^{\phi} + n_a Δ p a = c a S a ϕ + n a , where n a n_a n a collects other premium movements. Equation (28) applies after n a , n b n_a,n_b n a , n b are independently removed and c a , c b c_a,c_b c a , c b are known positive scalars, with both totals δ a ϕ \delta_a^{\phi} δ a ϕ and δ b ϕ \delta_b^{\phi} δ b ϕ nonzero. If cross-asset absorption is a matrix C m k t C^{\mathrm{mkt}} C mkt , the corresponding identified linear transformation of Δ p − n \Delta p-n Δ p − n must be used. Separate division by two diagonal entries is generally invalid. Identical investor loadings give zero relative signal, however much county spending moves.
Cancellation is not sufficient when the spending map is incomplete. With observed T C Y z TC_Yz T C Y z and Δ π a b = π a − π b \Delta\pi_{ab} = \pi_a - \pi_b Δ π ab = π a − π b , the contrast is recoverable on R \mathcal{R} R only if ( Δ π a b ) ′ R R = a ′ T C Y R R (\Delta\pi_{ab})'R_{\mathcal{R}} = a'TC_YR_{\mathcal{R}} ( Δ π ab ) ′ R R = a ′ T C Y R R for some independently justified a a a . When C Y = Id n C C_Y = \operatorname{Id}_{n_C} C Y = Id n C and ker T = span { 1 } \ker T = \operatorname{span}\{1\} ker T = span { 1 } this follows from ( Δ π a b ) ′ 1 = 0 (\Delta\pi_{ab})'1 = 0 ( Δ π ab ) ′ 1 = 0 . Otherwise Proposition 6 imposes an additional measurement condition.
For the two-county example, let L = ( 1 − 1 − 1 1 ) \mathcal{L} = \begin{pmatrix} 1 & -1 \\ -1 & 1 \end{pmatrix} L = ( 1 − 1 − 1 1 ) , B D = I d 2 B_D = \mathrm{Id}_2 B D = Id 2 , π a = ( 0.8 , 0.2 ) ′ \pi_a = (0.8, 0.2)^{\prime} π a = ( 0.8 , 0.2 ) ′ , π b = ( 0.2 , 0.8 ) ′ \pi_b = (0.2, 0.8)^{\prime} π b = ( 0.2 , 0.8 ) ′ , and the shock be e 1 e_1 e 1 . At h ≥ 1 h \ge1 h ≥ 1 , let n = h − 1 n = h - 1 n = h − 1 denote elapsed periods, distinct from diffusion time τ = κ D n \tau= \kappa_D n τ = κ D n .
z h = ρ D n 2 ( 1 + e − 2 κ D n , 1 − e − 2 κ D n ) ′ , z_h = \frac{\rho_D^n}{2}(1 + e^{-2\kappa_D n}, 1 - e^{-2\kappa_D n})^{\prime}, z h = 2 ρ D n ( 1 + e − 2 κ D n , 1 − e − 2 κ D n ) ′ ,
( π a − π b ) ′ z h = 0.6 ρ D n e − 2 κ D n , (\pi_a - \pi_b)^{\prime}z_h = 0.6\rho_D^n e^{-2\kappa_D n}, ( π a − π b ) ′ z h = 0.6 ρ D n e − 2 κ D n ,
relative diffusion increment = 0.6 ρ D n ( e − 2 κ D n − 1 ) . \text{relative diffusion increment} = 0.6\rho_D^n(e^{-2\kappa_D n} - 1). relative diffusion increment = 0.6 ρ D n ( e − 2 κ D n − 1 ) .
An arbitrary additive common field cancels in the difference. The increment is zero at the first slow-field response, h = 1 h = 1 h = 1 , and at κ D = 0 \kappa_D = 0 κ D = 0 , and negative thereafter for this origin and loading configuration. Reversing the shock origin reverses the relative sign. Those restrictions are declared by the model before observing asset returns. They do not predict a universal positive diffusion effect or reverse the counties’ ranking. Unknown ρ D \rho_D ρ D , changing loadings or differential capital news could mimic the path, so the spending and financial measurements must distinguish them.
This example shows what success would mean qualitatively. Holding initial dose fixed, remote adverse responses should occur where the spending operator predicts, with stronger demand adjustment where independently measured l s l^{s} l s is larger. The relative premium should then follow the resulting horizon profile after the heterogeneous no-diffusion response and competing news are controlled. In the scalar slack-capital benchmark, the same positive fixed-price demand reduction raises the level premium more when capital is low, since G S N < 0 G_{SN} < 0 G S N < 0 . At the cap this derivative changes, and endogenous capital can alter the total path. There is still no sign theorem for the old lagged-state EBP-change regression.
The absolute EBP route requires the measured dollar pressure to satisfy the units bridge (H.9) below. An observed zero net withdrawal is compatible with portfolio rotation, whereas zero fixed-price credit-demand adjustment rules out that margin of the proposed mechanism. A null relative-price response is informative only when demand, absorption and the design’s power are known. The municipal fund-flow pilot measures none of these complete links. Kilts can improve spending measurement, not automatically the investor map or absorption elasticity.
Conditional financial visibility
The two household margins need not coincide. Ignoring valuation, financial accumulation satisfies the event-change budget identity
Δ F = Δ Y + Δ T + Δ B − Δ C − Δ I n o n f i n a n c i a l , \Delta F = \Delta Y + \Delta T + \Delta B - \Delta C - \Delta I^{nonfinancial}, Δ F = Δ Y + Δ T + Δ B − Δ C − Δ I n o n f inan c ia l ,
where transfers and net borrowing are included, all terms cover the same interval, and internal financial-account transfers cancel. If the other components are unchanged, lower consumption increases financial accumulation. Thus a remote spending decline does not establish a cash need. Portfolio rotation remains a distinct testable margin.
For specified counterfactual expenditures C i , t + h 0 C^{0}_{i,t+h} C i , t + h 0 , the dollar shortfall is
C l o s s ( H ) = − ∑ h = 0 H ∑ i C i , t + h 0 ( e g i , h − 1 ) , C_{loss}(H)=-\sum_{h=0}^{H}\sum_i C^{0}_{i,t+h}\left(e^{g_{i,h}}-1\right), C l oss ( H ) = − h = 0 ∑ H i ∑ C i , t + h 0 ( e g i , h − 1 ) ,
where g i , h g_{i,h} g i , h is a log-level response. Accumulated growth alone is not this dollar sum. Let λ C \lambda_C λ C be the fixed-price credit-demand reduction per dollar of that shortfall, A m k t A_{mkt} A mk t the matched market value, ξ P \xi_P ξ P the residual absorption elasticity, D m o d D_{mod} D m o d modified duration and ϕ E \phi_E ϕ E the loading of the EBP on yields. Residual holders absorb the fixed-price demand shift λ C C l o s s \lambda_C C_{loss} λ C C l oss with − Δ P / P ≃ λ C C l o s s / ( A m k t ξ P ) -\Delta P/P \simeq\lambda_C C_{loss}/(A_{mkt}\xi_P) − Δ P / P ≃ λ C C l oss / ( A mk t ξ P ) . The duration relation of Appendix § 13 .§ 6 , Δ yield ≃ − Δ P / ( P D m o d ) \Delta\text{yield} \simeq-\Delta P/(P D_{mod}) Δ yield ≃ − Δ P / ( P D m o d ) , converts this price change into a yield change, and ϕ E \phi_E ϕ E maps the yield change into the EBP:
Δ E B P [ bp ] = 10 4 ϕ E λ C C l o s s A m k t ξ P D m o d . (75) \Delta EBP\ [\text{bp}]=10^{4}\frac{\phi_E\lambda_C C_{loss}}{A_{mkt}\xi_P D_{mod}}. \tag*{(75)} Δ E B P [ bp ] = 1 0 4 A mk t ξ P D m o d ϕ E λ C C l oss . (75)
Table H.2 makes the economic hurdle explicit. Every input is illustrative, including a $1 trillion matched credit market, duration five and unit yield-to-EBP loading. Positive pressure denotes a fixed-price reduction in desired credit holdings, not executed turnover. The resource and share components are first-order changes, whereas their product is second order.
Table H.2. Conditional household demand and EBP magnitudes
Notes: Equation (H.9) with assumed inputs, not estimated responses or achieved power. Wealth must belong to the exposed households, and absorption cannot be chosen to fit EBP. The last row cancels at first order. Large movements require checking the local-linear price approximation. “pp” denotes percentage points.
For illustration, a five-basis-point increase in a $1 trillion market requires $2.5 billion of demand reduction at unit absorption, $2.50 per dollar of a $1 billion shortfall, or $0.5 billion at ξ P = 0.2 \xi_P = 0.2 ξ P = 0.2 . At the 57-month mean of the debt-securities denominator of Appendix § 15 .§ 8 , about $7,970 billion, the same increase would require about $20 per dollar. A lower EBP loading raises those requirements inversely. The demand object is a change in desired holdings at the endpoint, not the sum of the same persistent adjustment each week. Relating it to cumulative spending requires matched timing and population. A statistically detectable household response may still be too small to explain national EBP. Additionally, strong absorption can mute prices.
For a single unbiased normal estimate of aggregate demand pressure, the approximate 80-percent-power, two-sided 5-percent-test requirement is
SE ( S a ^ ) ≤ S a ∗ z 0.975 + z 0.80 \operatorname{SE}(\widehat{S_a}) \le\frac{S_a^*}{z_{0.975}+z_{0.80}} SE ( S a ) ≤ z 0.975 + z 0.80 S a ∗
The respective illustrative SE ceilings are $0.892 billion and $0.178 billion. These do not establish power for the joint validation criteria. Financial sampling, episode dependence, operator estimation, price endogeneity and uncertain scaling require full-design simulation. The appropriate first financial observations are therefore resource changes and valuation-adjusted credit-share rotation, followed by the independently weighted operator contrast in Appendix § 16.4 .§ 2 . No such response has been established in the available panel.
Physical-weather coverage and recipient measurement
The measured-weather adjustment in Appendix § 16.3 .§ 2 accounts for observed county precipitation and temperature, station gusts, warnings and reports. A stronger physical exclusion would require high-resolution rainfall and wind coverage, relevant flood indicators, and reconciliation of FEMA declarations with NOAA physical episodes. Warning geometry measures threatened areas, rather than realized damage, and station coverage alone cannot certify the absence of local destructive weather.
Sources include NOAA MRMS precipitation, IEM archives of NWS warning polygons, NOAA ISD/ASOS gust observations, HRRR wind fields and USGS stage or discharge observations. Product versions and archived coverage matter when comparing early 2020 with later events. A proposed low-hazard recipient design would retain explicit coverage flags and align each hazard window to the spending endpoint. Absence of a report would not be converted into measured zero exposure. Sources and recipients remain separate, so a revised footprint must be passed through the social, gravity, production and direct fields before selecting observed recipients.6
Warning geography and the cost of long exclusion windows
An outcome-blind warning-coverage exercise reconciles 23,544 selected warning keys to archived geometries, with no missing key or null geometry. Keys include issuing office, year, warning number, phenomenon, zone and issue/expiration time. The 4,638 distinct geometries are intersected with 2020 Census county boundaries in an equal-area projection. The IEM exporter uses the warning’s archived geography identifier rather than today’s zone code. This resolves the geography of the selected warnings, not complete physical-hazard coverage.
Table H.3 reports sample sizes only. These are intersections of recorded-weather, county-warning and archived zone-warning screens on the earlier 75,237-row candidate sample, not additional spending regressions or the current 74,779-row measured-weather sample. These sample counts describe warning coverage, not identifying variation in the spending regression.
Table H.3. Recipient coverage under alternative warning-exclusion windows. Notes: All windows exclude expanded storm support, county-coded acute NOAA reports, selected county warnings, positive archived zone-warning area contact and unknown warning timing. These outcome-blind sample counts do not estimate a spending response, its identifying variation or achieved power.
The comparison explains the use of measured own-weather controls in the main application, because a long future-weather screen sharply changes the recipient population and can exclude counties for conditions occurring after the outcome endpoint. Conservative warning times also use a plus/minus-one-day UTC envelope rather than county-local conversion. Remaining hazard gaps belong in the interpretation of the local association. Financial recovery still requires the distinct demand and price measurements derived in Appendices § 16.4 .§ 1 –§ 16.4 .§ 3 .
Simulated recovery on the sample design
This appendix documents the implementation summarized in § 7.4 and Table 6 . It is a simulation on the observed network and storm design with hypothetical financial measurements, so it does not recover the actual mechanism. Code, frozen protocols, saved replications and the full run record are kept with the replication files.
Design
The simulation uses the 2,793-county source-complete SCI graph, in which 1,574 counties are observed recipients, and all 85 candidate storm declarations (79 episodes), not only the 48 declarations of the empirical sample, on one weekly calendar from January 2020 to June 2024. Sixty-nine storms are held out across three calendar folds. Training weeks exclude every held-out storm’s 52-week response support, and footprint counties are masked for 52 weeks after onset. The household field follows the slow operator of § 2 with ρ D = 0.85 \rho_D = 0.85 ρ D = 0.85 and κ D = 1 \kappa_D = 1 κ D = 1 per week, simulated in coordinates in which county log spending is the state, an added measurement assumption. The input is local damage relative to one week of baseline spending.
The target is the financially weighted response of observed recipient households for each held-out storm at horizons one to twelve. Loadings are proportional to county dividend and interest income, measured in five groups, and the market map is 10 9 c E $ = 0.48 10^9 c_E^{\$} = 0.48 1 0 9 c E $ = 0.48 basis points per billion dollars of desired selling, where c E $ = c E / P ˉ c_E^{\$} = c_E/\bar{P} c E $ = c E / P ˉ is the map per dollar at reference price P ˉ \bar{P} P ˉ . In the anchor route , a common household response proportional to damage doubles the average recipient response. A hypothetical event-level aggregate anchor measures each storm’s recipient-mean response path, propagated field plus common response, with a reported standard deviation equal to half these paths’ root mean square over all storms and horizons. In the structure route , the data-generating process has no common household response outside the propagation law, and no anchor is used. That restriction cannot be tested from spending contrasts.
Estimator, comparisons, inference and the planning diagnostic
The structural estimator is the diffusion-kernel operator under criterion (29) in its simplest case, with ω = 0 \omega= 0 ω = 0 and λ P = 0 \lambda_{\mathrm{P}} = 0 λ P = 0 . Its persistence is continuous on [ 0.50 , 0.985 ] [0.50, 0.985] [ 0.50 , 0.985 ] , and its diffusion rate is profiled on a 13-point grid that excludes the scenario value and is then refined. Its amplitude is estimated jointly with a geographic nuisance and, when an anchor exists, the common response. Held-out anchors are not used. Instead, held-out common responses are predicted from damage with the per-dollar profile estimated on training anchors, a form correct in this scenario. Spending contrasts are weighted by inverse county noise variance estimated in training, and anchors by their reported precision. The comparisons use the same criterion with graph-polynomial features (ridge penalty chosen by inner blocked cross-validation), static one-hop features and a shape oracle that knows the true field shape. An earlier two-step regression and an anchor-only prediction are also reported.
The financial target is projected with county-resolution loadings, each equal to the group scale, known or estimated from hypothetical auxiliary measurements with assumed errors, times the county income shape within the group. Uncertainty comes from a parametric bootstrap under a declared measurement-error model, county AR(1) noise plus a state AR(1) factor, fitted to training residuals by the method of moments with an indirect-inference correction for the fixed-effect projection. Anchors use their reported standard deviation and the financial and market maps their reported errors. Each of 199 bootstrap samples re-estimates the operator with the training weights held fixed. The bands are pointwise Wald intervals and sup-t t t bands over each storm’s twelve-week path, both centred at the estimate.
For a held-out storm e e e , the estimation error decomposes as k ^ e − k e = b e + s e + G e m \hat{k}_e-k_e=b_e+s_e+G_e m k ^ e − k e = b e + s e + G e m , the sum of specification bias, sampling error with covariance Σ e \Sigma_e Σ e , and the first-order effect of the error m m m in the financial and market maps, with sensitivity G e G_e G e and covariance V m V_m V m . If s e s_e s e and m m m have mean zero and the three components are uncorrelated, then to first order
E ∥ k ^ e − k e ∥ 2 = ∥ b e ∥ 2 + tr Σ e + tr ( G e V m G e ′ ) . (76) \mathbb{E}\lVert\hat{k}_e-k_e\rVert^2=\lVert b_e\rVert^2+\operatorname{tr}\Sigma_e+\operatorname{tr}\left(G_eV_mG_e^{\prime}\right). \tag*{(76)} E ∥ k ^ e − k e ∥ 2 = ∥ b e ∥ 2 + tr Σ e + tr ( G e V m G e ′ ) . (76)
Under a normal approximation to the estimation error, a calibrated pointwise interval at horizon h h h covers with probability Φ ( z − β e h ) − Φ ( − z − β e h ) \Phi(z-\beta_{eh})-\Phi(-z-\beta_{eh}) Φ ( z − β e h ) − Φ ( − z − β e h ) , where z z z is the normal critical value, β e h = b e h / σ e h \beta_{eh}=b_{eh}/\sigma_{eh} β e h = b e h / σ e h , and σ e h \sigma_{eh} σ e h is the total standard error at that horizon, including map error. A bound on the bias gives a bound on the mean squared error, not its value. A path-level bound ∥ b e ∥ ≤ ϱ ∥ k e ∥ \lVert b_e\rVert\leq\varrho\lVert k_e\rVert ∥ b e ∥ ≤ ϱ ∥ k e ∥ implies only ∣ β e h ∣ ≤ ϱ ∥ k e ∥ / σ e h \lvert\beta_{eh}\rvert\leq\varrho\lVert k_e\rVert/\sigma_{eh} ∣ β e h ∣ ≤ ϱ ∥ k e ∥ / σ e h , which can be uninformative, because a horizon can have a small standard error even when the path’s total uncertainty is large. Coverage predictions therefore use β e h \beta_{eh} β e h horizon by horizon. Heterogeneous loadings enter through their covariance and the target’s sensitivity to them, G e V m G e ′ G_eV_mG_e^{\prime} G e V m G e ′ .
Equation (I.1) supplies a planning diagnostic. For each design, the sampling term is the spread of the estimate across 40 noise draws, and the bias term is the error of a noiseless fit. Both use the scenario’s assumed signal magnitudes and specification errors. Pooling over held-out storms and horizons, with SNR the target’s root mean square over the root-mean-square standard error and ϱ \varrho ϱ the relative bias, the predicted normalized error is ( S N R − 2 + ϱ 2 + σ a 2 + σ c 2 ) 1 / 2 (\mathrm{SNR}^{-2}+\varrho^2+\sigma_a^2+\sigma_c^2)^{1/2} ( SNR − 2 + ϱ 2 + σ a 2 + σ c 2 ) 1/2 , where the map terms σ a = 0.25 / 5 \sigma_a=0.25/\sqrt{5} σ a = 0.25/ 5 and σ c = 0.38 \sigma_c=0.38 σ c = 0.38 enter only when the maps carry assumed errors. They are this study’s planning approximation for five equally precise group scales and a log-normal market map, not the general term tr ( G e V m G e ′ ) \operatorname{tr}(G_eV_mG_e^{\prime}) tr ( G e V m G e ′ ) for arbitrary loadings. The diagnostic is scenario-based. It plans measurement, but it does not certify that a given data set will recover the response. Across the nineteen non-null designs of Table I.Table 1 , Panel A, predicted and achieved log errors correlate at 0.998, and the prediction falls on the correct side of the 0.5 threshold in eighteen.
Measurement-noise models
Panel A uses a design-based noise model, estimated from raw residuals of the county panel after county and week effects, that combines county AR(1) noise with coefficient 0.26 and a state factor taken as the state-week mean residual. Panel B uses noise calibrated more carefully to the panel after September 7, 2020, with storm-affected cells excluded. Estimated by within-state cross-county covariances, the regional factor carries about 6 percent of the noise variance, whereas the raw state-mean factor of Panel A carries 27 percent, inflated by averaging over a few counties in small states.
The residual autocorrelation decays slowly, 0.15 at four weeks and 0.09 at thirteen in the raw panel and 0.22 and 0.12 after denoising. A two-component model fits it, with white noise plus a persistent county drift whose autoregressive coefficient is 0.99. The drift carries about 45 percent of the variance in the denoised panel. Denoising subtracts the 2019-base calendar pattern (Appendix § 16.3 .§ 5 ) and lowers the white component, but it leaves the drift. County piecewise-linear trends with knots every 26 weeks, removed identically from the data and from every model column, take out the drift’s low-frequency part. Without an anchor at the 85 declarations, they lower the predicted error from 2.91 to 1.79. Their spacing was chosen by the planning diagnostic, before the Panel B evaluation draws.
Results
Table I.Table 1 reports every design, and Figures I.Figure 1 and I.Figure 2 illustrate them.
With an anchor at the assumed precision, the operator recovers the response at current spending noise (normalized error 0.170), under the panel-calibrated noise (0.168), and with loading scales and market map that carry assumed errors (0.400). It narrowly misses when the anchor’s error is three times that precision (0.502, Monte Carlo s.e. 0.010) and fails under an aggregate-innovation calibration (14.7), in which the anchor’s standard deviation equals the innovation standard deviation of an AR(1) benchmark for the constructed spending aggregate, 87.5 times the assumed precision (128 times over all weeks). That innovation includes genuine aggregate movements, so it is not a floor on the error of every event-attribution procedure. Precise spending and more storms substitute for a weaker anchor (0.206 and 0.385). Under the same draws and final inference (Table I.Table 2 ), an anchor-only prediction has normalized error 0.30 with known maps and 0.51 with maps carrying assumed errors, and its pointwise bands cover 24 and 48 percent of true values, against 24 and 49 percent under the earlier predeclared inference. The operator’s advantage is therefore not an artifact of the inference method. The operator also estimates the cross-county composition, about 4 percent of the target.
Without an anchor , recovery requires a correctly specified common response together with a quarter of the design-based noise and twice the storms (0.320), or half the noise and four times the storms (0.478). With panel-calibrated noise, denoising and county trends, the error is 1.63 at the 85 declarations and 0.69 with twice the storms, although the bands remain honest (pointwise coverage 0.95 and 0.89).
Comparisons (final inference for every method, Table I.Table 2 ). Without an anchor, at a quarter of the noise and twice the storms, graph-polynomial features reach 0.71 and the one-hop response 0.92, against 0.32 for the structural operator, and their bands cover 14 and 8 percent of true values. With an anchor the criterion-based estimators are closer (0.17 to 0.22 in the first
Table I.1. Simulated recovery on the sample design: all environments (structural estimator)
Notes: The estimator is the parsimonious diffusion-kernel operator under criterion (29) , fitted on training storms and evaluated on 69 held-out storms over horizons 1–12. Normalized error is the root-mean-squared error over the root mean square of the true response. Coverage is for nominal 90 percent bands. Final inference: county-resolution loadings and a parametric bootstrap that re-estimates the operator (Appendix § 16.5 .§ 2 ). Predeclared: the first inference, a Gauss–Newton sandwich with group-mean loadings (Appendix § 16.5 .§ 5 ). Labels apply the study’s diagnostic criteria to the final inference, under which a design counts as recovered if normalized error is at most 0.50, pointwise coverage at least 0.85, path coverage at least 0.80 and failures at most 2 percent of replications, for nominal 90 percent bands. Predicted errors come from the planning diagnostic of Appendix § 16.5 .§ 2 , evaluated at the scenario’s signal magnitudes. Under the nulls, with no diffusion (κ D = 0 \kappa_D = 0 κ D = 0 ) the composition band excludes zero for 0.7 percent of storms, and with zero financial loading it excludes zero for 0.0 percent. Panel A uses 100 replications in the leading environments and 40 in the others, and Panel B uses 100 (anchor) and 40 (no anchor).
row’s design), but only the structural operator’s bands are calibrated. Under the predeclared inference the two-step regression had error 1.00 with an anchor and 0.47 without, the latter with bands covering 52 percent.
Composition. With an anchor and panel-calibrated noise, the cross-county composition component has normalized error 0.38, against 1 for predicting zero, with coverage 0.88 pointwise and 0.85 by path.
Failures and nulls. In the joint two-timescale and geographic-misspecification scenarios, where the true geographic nuisance uses the squared kernel at three times the size while every estimator fits the first-order kernel, the tested feasible estimators (structural, graph-polynomial, one-hop, two-step and anchor-only) fail, with errors of 0.6 or more with an anchor and 1.1 or more without, as the specification error predicts. These scenarios do not isolate the number of temporal channels. Without diffusion, the composition band excludes zero for 0.7 percent of storms, and with zero financial loading, it never excludes zero.
Inference sensitivity. When the noise is more persistent than the bootstrap family allows (A R ( 2 ) AR(2) A R ( 2 ) dynamics with Student-t t t innovations), pointwise coverage is 0.860, close to the loss implied by a 22–27 percent understatement of the long-run variance. The shape oracle’s bands undercover slightly without an anchor (0.85 in the leading design and 0.82 under panel-calibrated noise with twice the storms), while the structural estimator’s remain calibrated.
Table I.2. Matched comparison: every estimator under the final inference on the same draws
Notes: Normalized error / pointwise coverage of nominal 90 percent bands for the recipients’ financially weighted response, horizons 1–12, 69 held-out storms. Columns (1), (3) and (4) are the designs of main Table 6 , rows 1–3, and column (2) is row 1’s design with design-based noise. Each column uses the 100 production replications of Table I.Table 1 , the same data for every estimator. Every estimator uses county-resolution loadings and the parametric bootstrap of Appendix § 16.5 .§ 2 (B = 199) with its own fitted noise model, while the graph-polynomial and one-hop fits hold their training-fold ridge penalty fixed in the bootstrap. The structural operator has the lower error in 100 percent of paired replications against anchor only and 100 percent against one-hop in column (1), and in 97 percent against the graph polynomial in column (4). Monte Carlo standard errors are at most 0.035 for errors and 0.03 for coverage.
Figure I.1. True and recovered financially weighted response paths for the three designs of main Table 6 , for the held-out storm with the largest true response (the 2023 Illinois declaration) in one simulated replication. Black shows the true response of observed recipient households, in basis points, red the structural estimate and shading the 90 percent path band. Panel A: anchors, noise calibrated to the denoised panel, known maps (row 1). Panel B: anchors, design-based noise, loading scales and market map with assumed errors (row 2). Panel C: no anchors, a quarter of the design-based noise and twice the storms (row 3). Panels A and C repeat main Figure 4 . Simulation with hypothetical measurements for one storm, not average performance.
Figure I.2. Predicted versus achieved normalized error (left) and achieved pointwise coverage (right) across the designs of Table I.Table 1 , Panel A, for the structural estimator and the shape oracle. Dashed lines mark the recovery thresholds.
Development history
The procedure was developed in stages. Each stage was evaluated on fresh simulation draws under a protocol frozen beforehand. Development used separate draws, and every earlier result is retained.
Under the predeclared inference, the two-step regression gives normalized error 1.00 in the baseline anchored design. Without an anchor, at a quarter of the spending noise and twice the storms, its error is 0.47 and its nominal 90 percent pointwise bands cover 52 percent of true responses.
The target-oriented and structural estimators under (29) , with a predeclared Gauss–Newton sandwich inference and group-mean loadings, achieved the predicted accuracy, but pointwise coverage was 0.70 to 0.83 in most designs.
A diagnosis on separate development draws found four sources of the undercoverage. First, group-mean loadings leave a deterministic resolution error of 2.6 to 5.3 percent of the target. Second, a few high-damage storms dominate the common-response information. Third, serial dependence crosses the two-week clusters. Finally, percentile intervals are miscentred through the nonlinearity in persistence and diffusion.
The final inference, county-resolution loadings with the parametric bootstrap, was then frozen and evaluated on fresh draws (Panel A), and later under the panel-calibrated noise (Panel B).
The final inference was chosen after the predeclared inference failed, so only its evaluation is fresh.
A diffusion-specific EBP restriction
Apply one dated linear contrast T e t i m e T_e^{\mathrm{time}} T e time , the mean over three full months after onset minus the mean over three full months before, to both the observed EBP and the predicted path. Construct the basis-point path p t = ∑ e ′ K t − e ′ ε e ′ p_t = \sum_{e'} K_{t-e'}\varepsilon_{e'} p t = ∑ e ′ K t − e ′ ε e ′ , including overlapping shocks, and set p e = T e t i m e { p t } p_e = T_e^{\mathrm{time}}\{p_t\} p e = T e time { p t } . Under no anticipation, K h = 0 K_h = 0 K h = 0 for h < 1 h < 1 h < 1 . Equation (18) gives the decomposition p e = p e 0 + p e d i f f p_e = p_e^0 + p_e^{\mathit{diff}} p e = p e 0 + p e diff . The proposed validation is
Y e = α + β p e d i f f + γ 0 p e 0 + Z e ′ γ + u e , β = 1 under the complete, correctly measured bridge. (77) Y_e = \alpha+ \beta p_e^{\mathit{diff}} + \gamma_0 p_e^0 + Z_e'\gamma+ u_e,\qquad\beta= 1 \quad\text{under the complete, correctly measured bridge.} \tag*{(77)} Y e = α + β p e diff + γ 0 p e 0 + Z e ′ γ + u e , β = 1 under the complete, correctly measured bridge. (77)
The complete model also implies γ 0 = 1 \gamma_0 = 1 γ 0 = 1 . The nuisance span includes an intercept, p 0 p^0 p 0 , and predetermined production, geography, seasonality, macro and insurer channels. If M [ 1 , p 0 , Z ] M_{[1,p^0,Z]} M [ 1 , p 0 , Z ] projects off that span, identifying diffusion requires ( p d i f f ) ′ M [ 1 , p 0 , Z ] p d i f f > 0 (p^{\mathit{diff}})'M_{[1,p^0,Z]}p^{\mathit{diff}} > 0 ( p diff ) ′ M [ 1 , p 0 , Z ] p diff > 0 , so stationary incidence alone is insufficient. The lead contrast tests anticipation and timing. Parameter uncertainty in H ^ h \hat{H}_h H ^ h , a a a and c E c_E c E must enter inference, since an unbiased unit coefficient is not guaranteed by plugging noisy estimates into a regression. Capital or other omitted responses can invalidate the restriction even when spending transport succeeds, and the monthly change regression does not test it. Appendix § 13 .§ 7 illustrates how noise, weak excitation of the relevant modes and collinearity with nuisance controls determine the power of this proposed test.
Recovering the credit-premium response with an estimated market map
This section extends the first design of Table 6 , which treats both financial maps as known, by estimating them from simulated data. In contrast, the second design draws them with assumed errors. Household loading scales are estimated from hypothetical county account flows, and the market map from an independently simulated panel of fund observations whose instrument is built from the simulated fund flows, not supplied as the true demand innovation. The section states the fund demand–price system and its mapping to the model, the conditions under which fund-level shocks identify its two market terms, the simulated inputs, the market-only and bridge results, and, separately, what the actual 57-month sample supports.
The fund demand–price system. Let f j t f_{jt} f j t be fund j j j ’s net subscriptions over lagged net assets, w j t h o l d w_{jt}^{\mathrm{hold}} w j t hold its lagged corporate-security holdings as a share of the market denominator of Appendix § 15 .§ 8 , F t = 100 ∑ j w j t h o l d f j t F_t = 100 \sum_j w_{jt}^{\mathrm{hold}} f_{jt} F t = 100 ∑ j w j t hold f j t the holdings-weighted subscription proxy for aggregate purchases, in percent of that denominator, and Δ E t B P \Delta E_t^{BP} Δ E t B P the monthly EBP change in basis points. With common flow factors f t c o m f_t^{\mathrm{com}} f t com and controls Z t m k t Z_t^{\mathrm{mkt}} Z t mkt (an intercept and two lagged EBP changes),
f j t = α j F + ζ j F Δ E t B P + ( L j F ) ′ f t c o m + u j t , f_{jt} = \alpha_j^F + \zeta_j^F \Delta E_t^{BP} + \left(L_j^F\right)' f_t^{\mathrm{com}} + u_{jt}, f j t = α j F + ζ j F Δ E t B P + ( L j F ) ′ f t com + u j t ,
Δ E t B P = ϕ F t + ( Z t m k t ) ′ b + g ′ f t c o m + u t m k t . \Delta E_t^{BP} = \phi F_t + \left(Z_t^{\mathrm{mkt}}\right)'b + g' f_t^{\mathrm{com}} + u_t^{\mathrm{mkt}}. Δ E t B P = ϕ F t + ( Z t mkt ) ′ b + g ′ f t com + u t mkt .
With ζ a g g F = 100 ∑ j w j t h o l d ζ j F \zeta_{\mathrm{agg}}^F = 100 \sum_j w_{jt}^{\mathrm{hold}}\zeta_j^F ζ agg F = 100 ∑ j w j t hold ζ j F , aggregate purchases are F = ζ a g g F Δ E B P + U p c t F = \zeta_{\mathrm{agg}}^F \Delta E^{BP} + U^{\mathrm{pct}} F = ζ agg F Δ E B P + U pct , where U p c t U^{\mathrm{pct}} U pct denotes desired purchases at fixed prices, so
Δ E B P = ϕ m F U p c t + m F ( ( Z m k t ) ′ b + g ′ f c o m + u m k t ) , m F = ( 1 − ζ a g g F ϕ ) − 1 . \Delta E^{BP} = \phi m_F U^{\mathrm{pct}} + m_F\left(\left(Z^{\mathrm{mkt}}\right)'b + g' f^{\mathrm{com}} + u^{\mathrm{mkt}}\right), \qquad m_F = \left(1-\zeta_{\mathrm{agg}}^F\phi\right)^{-1}. Δ E B P = ϕ m F U pct + m F ( ( Z mkt ) ′ b + g ′ f com + u mkt ) , m F = ( 1 − ζ agg F ϕ ) − 1 .
For desired selling S p c t = − U p c t S^{\mathrm{pct}}=-U^{\mathrm{pct}} S pct = − U pct the premium loading is c E s e l l = − ϕ m F c_E^{\mathrm{sell}}=-\phi m_F c E sell = − ϕ m F .
The market-percentage units give S p c t = 100 P ˉ S / A m k t S^{\mathrm{pct}}=100\bar{P}S/A_{\mathrm{mkt}} S pct = 100 P ˉ S / A mkt , where S S S is the model’s claim-quantity shift and A m k t A_{\mathrm{mkt}} A mkt is dollar market value. Thus ϕ = − ( A m k t / 100 ) ϑ F \phi=-(A_{\mathrm{mkt}}/100)\vartheta_F ϕ = − ( A mkt /100 ) ϑ F for the dollar-flow slope in Appendix E.8, and m F = 1 − ζ G S m_F = 1 - \zeta_{GS} m F = 1 − ζ GS is the executed share. The maps satisfy
c E $ = 100 A m k t c E s e l l = ϑ F ( 1 − ζ G S ) = b E G S P ‾ , c E = P ‾ c E $ = 100 P ‾ A m k t c E s e l l . c_E^{\$} = \frac{100}{A_{\mathrm{mkt}}}c_E^{\mathrm{sell}} = \vartheta_F(1-\zeta_{GS}) = \frac{b_EG_S}{\overline{P}}, \qquad c_E = \overline{P}c_E^{\$} = \frac{100\overline{P}}{A_{\mathrm{mkt}}}c_E^{\mathrm{sell}}. c E $ = A mkt 100 c E sell = ϑ F ( 1 − ζ GS ) = P b E G S , c E = P c E $ = A mkt 100 P c E sell .
The final expression is the factor in (24) , now with the market-percentage conversion explicit. Matching ζ a g g F = 100 P ‾ ζ / ( A m k t b E ) \zeta_{\mathrm{agg}}^F = 100\overline{P}\zeta/(A_{\mathrm{mkt}}b_E) ζ agg F = 100 P ζ / ( A mkt b E ) assumes that the observed funds are the only price-elastic non-intermediary demand. A granular instrument Z G I V Z^{\mathrm{GIV}} Z GIV has first-stage slope
m F Cov ( Z G I V , U p c t ) / Var ( Z G I V ) , m_F\operatorname{Cov}(Z^{\mathrm{GIV}},U^{\mathrm{pct}})/\operatorname{Var}(Z^{\mathrm{GIV}}), m F Cov ( Z GIV , U pct ) / Var ( Z GIV ) ,
which equals m F m_F m F only under the unknown normalization of Z G I V Z^{\mathrm{GIV}} Z GIV in units of desired demand. An estimated first-stage coefficient is not by itself an estimate of the executed share.
Identification of the two market terms. The construction is the granular instrumental-variables approach of Gabaix and Koijen (2024), applied to the two equations of the system. Let w j p r e c ∝ 1 / σ j 2 w_j^{\mathrm{prec}} \propto1/\sigma_j^2 w j prec ∝ 1/ σ j 2 be precision weights normalized to sum to one, with σ j 2 \sigma_j^2 σ j 2 the idiosyncratic variance of fund j j j , and let w ~ j t p r e c \widetilde{w}_{jt}^{\mathrm{prec}} w j t prec be the corresponding precision-weighted shares, normalized so that ∑ j w ~ j t p r e c = ∑ j w j t h o l d \sum_j \widetilde{w}_{jt}^{\mathrm{prec}} = \sum_j w_{jt}^{\mathrm{hold}} ∑ j w j t prec = ∑ j w j t hold . The instrument is Z t G I V = 100 ∑ j ( w j t h o l d − w ~ j t p r e c ) u ^ j t Z_t^{\mathrm{GIV}} = 100\sum_j(w_{jt}^{\mathrm{hold}}-\widetilde{w}_{jt}^{\mathrm{prec}})\widehat{u}_{jt} Z t GIV = 100 ∑ j ( w j t hold − w j t prec ) u j t , where u ^ j t \widehat{u}_{jt} u j t are fund residuals after removing factors estimated from the cross-sectionally demeaned panel. With a common ζ F \zeta^F ζ F , the demeaned panel contains neither the price term nor any common rate. The demand equation uses the precision-weighted average of fund-demeaned flows, f p r e c , t f_{\mathrm{prec},t} f prec , t , whose idiosyncratic part is u p r e c , t u_{\mathrm{prec},t} u prec , t . Two moments follow. The instrumental-variables regression of Δ E B P \Delta E^{BP} Δ E B P on F F F with instrument Z G I V Z^{\mathrm{GIV}} Z GIV identifies ϕ \phi ϕ when Cov ( Z G I V , u m k t ) = 0 \operatorname{Cov}(Z^{\mathrm{GIV}},u^{\mathrm{mkt}})=0 Cov ( Z GIV , u mkt ) = 0 and Cov ( Z G I V , F ) ≠ 0 \operatorname{Cov}(Z^{\mathrm{GIV}},F)\ne0 Cov ( Z GIV , F ) = 0 . Similarly, the regression of f p r e c f_{\mathrm{prec}} f prec on Δ E B P \Delta E^{BP} Δ E B P with the same instrument identifies ζ F \zeta^F ζ F when Cov ( Z G I V , u p r e c ) = 0 \operatorname{Cov}(Z^{\mathrm{GIV}},u_{\mathrm{prec}})=0 Cov ( Z GIV , u prec ) = 0 and Cov ( Z G I V , Δ E B P ) ≠ 0 \operatorname{Cov}(Z^{\mathrm{GIV}},\Delta E^{BP})\ne0 Cov ( Z GIV , Δ E B P ) = 0 . Both require the common factors to be removed validly.
The second exclusion condition holds in population under the following assumptions. Let the weights be fixed or predetermined relative to current idiosyncratic shocks, conditional on the factor and information structure. Let the conditional covariance of the idiosyncratic shocks across funds be diagonal, with Var ( u j ) = σ j 2 \operatorname{Var}(u_j)=\sigma_j^2 Var ( u j ) = σ j 2 . Also let the normalized population precision weights be w j p r e c = ϖ / σ j 2 w_j^{\mathrm{prec}}=\varpi/\sigma_j^2 w j prec = ϖ / σ j 2 , with ∑ j w ~ j p r e c = ∑ j w j h o l d \sum_j\widetilde{w}_j^{\mathrm{prec}}=\sum_jw_j^{\mathrm{hold}} ∑ j w j prec = ∑ j w j hold . Then
Cov ( Z G I V , u p r e c ∣ i n f o r m a t i o n ) = 100 ∑ j ( w j h o l d − w ~ j p r e c ) w j p r e c σ j 2 = 100 ϖ ( ∑ j w j h o l d − ∑ j w ~ j p r e c ) = 0. \begin{aligned}
\operatorname{Cov}(Z^{\mathrm{GIV}},u_{\mathrm{prec}}\mid\mathrm{information}) = 100\sum_j(w_j^{\mathrm{hold}}-\widetilde{w}_j^{\mathrm{prec}})w_j^{\mathrm{prec}}\sigma_j^2 \\
&= 100\varpi\left(\sum_jw_j^{\mathrm{hold}}-\sum_j\widetilde{w}_j^{\mathrm{prec}}\right)=0.
\end{aligned} Cov ( Z GIV , u prec ∣ information ) = 100 j ∑ ( w j hold − w j prec ) w j prec σ j 2 = 100 ϖ ( j ∑ w j hold − j ∑ w j prec ) = 0.
This is a population statement, not an exact finite-sample property of estimated residual weights.
The cancellation need not hold with residual common or category shocks, misspecified weights, heterogeneous feedback that the factors do not remove, or factors and weights estimated in a finite sample. The simulations and exclusion diagnostics below examine these departures. Equal weights generally violate the condition, and in simulated 228-month panels their median estimate of ζ S F \zeta_{S}^{F} ζ S F is 2.0 times the truth.
Household stocks, fund flows and the premium path. Fix the intermediary regime and payoff news, and express the household responses at monthly measurement dates. Let S m $ = ( a $ ) ′ H m ε S_{m}^{\$} = (a^{\$})^{\prime}H_{m}\varepsilon S m $ = ( a $ ) ′ H m ε be the reduction in desired credit holdings at fixed reference prices, in dollars. a $ a^{\$} a $ contains the separately measured financial sensitivities and any observed-sector pass-through. Let X m $ X_{m}^{\$} X m $ be the deviation in actual holdings net of valuation gains and E m B P , d e v E_{m}^{BP,\mathrm{dev}} E m B P , dev the premium-level deviation in basis points. With a fixed price-responsive demand slope ζ $ = P ˉ ζ / b E \zeta^{\$} = \bar{P}\zeta/b_{E} ζ $ = P ˉ ζ / b E ,
X m $ = ζ $ E m B P , d e v − S m $ , X_{m}^{\$} = \zeta^{\$}E_{m}^{BP,\mathrm{dev}} - S_{m}^{\$}, X m $ = ζ $ E m B P , dev − S m $ ,
F m b u y = Δ X m $ = ζ $ Δ E m B P , d e v − Δ S m $ . F_{m}^{\mathrm{buy}} = \Delta X_{m}^{\$} = \zeta^{\$}\Delta E_{m}^{BP,\mathrm{dev}} - \Delta S_{m}^{\$}. F m buy = Δ X m $ = ζ $ Δ E m B P , dev − Δ S m $ .
Thus monthly purchases respond to changes in desired holdings, not their level. Under static inverse absorption E m B P , d e v = ϕ $ X m $ E_{m}^{BP,\mathrm{dev}} = \phi_{\$}X_{m}^{\$} E m B P , dev = ϕ $ X m $ , with ϕ $ < 0 \phi_{\$} < 0 ϕ $ < 0 and 1 − ζ $ ϕ $ > 0 1-\zeta^{\$}\phi_{\$}>0 1 − ζ $ ϕ $ > 0 ,
m F = ( 1 − ζ $ ϕ $ ) − 1 , F m b u y = − m F Δ S m $ , m_{F} = (1-\zeta^{\$}\phi_{\$})^{-1}, \qquad F_{m}^{\mathrm{buy}} = -m_{F}\Delta S_{m}^{\$}, m F = ( 1 − ζ $ ϕ $ ) − 1 , F m buy = − m F Δ S m $ ,
Δ E m B P , d e v = c E $ Δ S m $ , c E $ = − ϕ $ m F > 0. \Delta E_{m}^{BP,\mathrm{dev}} = c_{E}^{\$}\Delta S_{m}^{\$}, \qquad c_{E}^{\$}=-\phi_{\$}m_{F}>0. Δ E m B P , dev = c E $ Δ S m $ , c E $ = − ϕ $ m F > 0.
Starting from zero deviations, cumulation gives the exact level identity E m B P , d e v = c E $ S m $ E_{m}^{BP,\mathrm{dev}} = c_{E}^{\$}S_{m}^{\$} E m B P , dev = c E $ S m $ . Let R m R_{m} R m denote the field response evaluated at the monthly measurement date, so H m = C ∗ R m H_{m}=C^{*}R_{m} H m = C ∗ R m . In the main text’s claim units, S m $ = P ˉ l ′ R m ε S_{m}^{\$}=\bar{P}l^{\prime}R_{m}\varepsilon S m $ = P ˉ l ′ R m ε , yielding (24) . If purchases are recorded as F m p c t = 100 F m b u y / A m k t F_{m}^{pct}=100F_{m}^{buy}/A_{\mathrm{mkt}} F m p c t = 100 F m b u y / A mkt , for dollar market size A m k t A_{\mathrm{mkt}} A mkt , the coefficient per one percent of the market satisfies ϕ $ = 100 ϕ / A m k t \phi_{\$}=100\phi/A_{\mathrm{mkt}} ϕ $ = 100 ϕ / A mkt and c E $ = − 100 ϕ m F / A m k t c_{E}^{\$}=-100\phi m_{F}/A_{\mathrm{mkt}} c E $ = − 100 ϕ m F / A mkt . Identifying weighted subscriptions with purchases is a separate observation restriction, examined through the holdings diagnostic in Appendix § 15 .§ 8 .
When the lagged price terms represent structural market adjustment, let the cleared response instead satisfy
Δ E m B P , d e v = c E $ Δ S m $ + d 1 Δ E m − 1 B P , d e v + d 2 Δ E m − 2 B P , d e v . \Delta E_{m}^{BP,\mathrm{dev}} = c_{E}^{\$}\Delta S_{m}^{\$} + d_{1}\Delta E_{m-1}^{BP,\mathrm{dev}} + d_{2}\Delta E_{m-2}^{BP,\mathrm{dev}}. Δ E m B P , dev = c E $ Δ S m $ + d 1 Δ E m − 1 B P , dev + d 2 Δ E m − 2 B P , dev .
If the original inverse-absorption equation has lag coefficients b 1 , b 2 b_{1}, b_{2} b 1 , b 2 on price changes, simultaneous clearing gives d j = m F b j d_j = m_F b_j d j = m F b j , not simply b j b_j b j . Integrating with zero initial deviations yields
E m B P , d e v = c E $ S m $ + d 1 E m − 1 B P , d e v + d 2 E m − 2 B P , d e v , E_m^{BP,\mathrm{dev}} = c_E^{\$} S_m^{\$} + d_1 E_{m-1}^{BP,\mathrm{dev}} + d_2 E_{m-2}^{BP,\mathrm{dev}}, E m B P , dev = c E $ S m $ + d 1 E m − 1 B P , dev + d 2 E m − 2 B P , dev ,
K m m a r k e t = c E $ ( a $ ) ∑ j = 0 m ψ j m k t H m − j m k t , K_m^{\mathrm{market}} = c_E^{\$}(a^{\$}) \sum_{j=0}^{m} \psi_j^{\mathrm{mkt}} H_{m-j}^{\mathrm{mkt}}, K m market = c E $ ( a $ ) j = 0 ∑ m ψ j mkt H m − j mkt ,
where ψ 0 m k t = 1 \psi_0^{\mathrm{mkt}} = 1 ψ 0 mkt = 1 , ψ j m k t = d 1 ψ j − 1 m k t + d 2 ψ j − 2 m k t \psi_j^{\mathrm{mkt}} = d_1\psi_{j-1}^{\mathrm{mkt}} + d_2\psi_{j-2}^{\mathrm{mkt}} ψ j mkt = d 1 ψ j − 1 mkt + d 2 ψ j − 2 mkt and negative-index coefficients are zero. The stable permanent-shift multiplier ( 1 − d 1 − d 2 ) − 1 (1-d_1-d_2)^{-1} ( 1 − d 1 − d 2 ) − 1 does not replace this finite-horizon convolution for a transient household response. The scored conditional impact-loading target below is distinct from the dynamically filtered premium-level path.
Weekly measurements must likewise be aligned to monthly flows. To do so, one evaluates the holdings stock at the month’s boundaries, takes its difference, and sums matched premium changes. Summing weekly holdings levels would instead create a stock–flow mismatch. The household measurement and financial-sensitivity conditions remain necessary under either static or dynamic clearing.
Simulated fund observations. The market stage uses simulated fund observations with the sample’s design, namely the 510 funds, their monthly shares and reporting pattern (79 percent of fund-months observed), flow volatilities (median 3.4 percent per month) and fund mean flows, factor loadings and persistence, the price equation’s lag and factor coefficients, and the calm-month EBP-change standard deviation (19.8 bp). The outcomes and structural parameters are simulated. The demand feedback is homogeneous, and ( ζ a g g F , ϕ ) (\zeta_{\mathrm{agg}}^F,\phi) ( ζ agg F , ϕ ) are set to the scenario values c E s e l l = 277.5 c_E^{\mathrm{sell}} = 277.5 c E sell = 277.5 bp per desired selling equal to 1 percent of the market denominator and m F = 0.683 m_F = 0.683 m F = 0.683 , the reduced form with its sign reversed and the first stage of column (1) of Table 5 . They are scenario values, not estimates of either quantity. The executed-purchase coefficient is ϕ ≈ − 406 \phi\approx-406 ϕ ≈ − 406 bp per 1 percent, matching the magnitude and sign of that column’s estimate. Its magnitude corresponds to a price multiplier of about 26 on the conversion of Appendix § 15 .§ 8 , against 1 and 5 in that section’s power calculation. The reported precision is conditional on this price-impact calibration. Performance at smaller impacts requires a separate sensitivity exercise. Idiosyncratic shocks are Gaussian and independent across funds given two factors, household demand passes through the observed funds in proportion to their size, and the simulation identifies the holdings-weighted subscription proxy with executed corporate-security purchases, assuming proportional portfolio allocation. At the 57-month mean market value of $7,970 billion, the scenario’s loading is 3.48 bp per billion dollars of desired selling, against 0.48 in Appendix § 16.5 .§ 1 . Household loading scales are reduced by the same factor, 0.48/3.48, so that the target keeps the first design’s root mean square, 0.72 bp in total and 0.26 bp for the propagated part.
Table I.3. The premium loading estimated from simulated fund observations: market-only simulations
Notes: 200 independent simulated panels per scenario, in basis points per 1% of the market of desired selling. The bias is the mean error. “Oracle” observes the true granular desired-demand innovation and uses the same estimated-factor controls. “Bootstrap”: share of panels whose 90% percentile interval for c E s e l l c_E^{\mathrm{sell}} c E sell , from a bootstrap that regenerates the fund panel and repeats the whole construction, contains the true value. The bootstrap is not robust to weak identification. “AR set”: share of panels whose joint Anderson–Rubin (1949) set for ( ζ F , ϕ ) (\zeta_F, \phi) ( ζ F , ϕ ) , with the χ 2 2 \chi^2_2 χ 2 2 critical value, contains the true values. Its projection onto c E s e l l c_E^{\mathrm{sell}} c E sell is bounded in only 0.48 of valid 57-month panels.
a ^a a One panel has a near-singular 1 − ζ ^ a g g ϕ ^ 1-\hat{\zeta}_{\mathrm{agg}}\hat{\phi} 1 − ζ ^ agg ϕ ^ and c ^ E s e l l = 14,171 \hat{c}_E^{\mathrm{sell}} = 14{,}171 c ^ E sell = 14 , 171 . Without it the RMSE is 173.8. Its bootstrap failed numerically (177 of 199 draws), so its interval is not used. Bootstrap coverage is therefore over the 199 panels with usable inference, and over all 200 attempted panels it lies between 0.875 and 0.880.
Market-only results. Table I.Table 3 reports the market stage alone. In the valid 57-month scenario the estimate from fund data has no detectable bias, and its root-mean-square error is 144.4 against 127.3 for the oracle. Pooled over every independent 57-month sample (800 panels), the fund-data estimate’s root-mean-square error is 9 percent above the oracle’s. On the same pooled panels, the oracle’s bias is −21.0 bp per 1% (Monte Carlo s.e. 4.6), against 1.0 (5.0) for the fund-data estimate, so the oracle’s advantage is in variance. With the true factors as controls its bias is 1.5 (9.0) on 200 of these panels, so the bias comes from the estimated factor controls. When short-duration category flows coincide with credit news in the stress months, the exclusion fails and the estimate is biased upward. The joint Anderson–Rubin set undercovers at this sample length. At 57 months it contains the true values in 0.78 of valid panels at nominal 90 percent, and the shortfall persists when the instrument, factors and weights are replaced by their true values (0.79, with the lag controls and HAC(3) still estimated) and with no HAC lags (0.81). A critical value calibrated by the same bootstrap is a partial, design-specific correction, giving 0.85 at 57 months and 0.90 and 0.91 at 114 and 228.
The bridge. The spending stage is the first design of Table 6 , namely the structural operator on held-out storms, with the same anchor serving the common response in every method. Household loading scales for the five groups are estimated from hypothetical county account flows, with noise of 0.1 percent of holdings for the median county, by regression on the recovered state change in purged training months, and are refitted on every spending-bootstrap draw. The within-group loading shape remains the supplied income shape, so estimating the five scales does not learn an unrestricted county-to-fund ownership map. The market stage estimates c E s e l l c^{\mathrm{sell}}_{E} c E sell from an independently simulated 57-month calibration panel with the sample’s fund geometry. It builds the precision-weighted granular instrument, estimates the factors, ζ F \zeta^{F} ζ F and ϕ \phi ϕ , and forms c E s e l l = − ϕ m F c^{\mathrm{sell}}_{E}=-\phi m_{F} c E sell = − ϕ m F . Every bootstrap draw regenerates that panel and repeats each step, so the uncertainty of the construction is carried into the response. The oracle benchmark instead observes the true granular desired-demand innovation, with the same estimated-factor controls.
The scored target is
K e , h R = c E s e l l S e , h R , p c t , h = 1 , … , 12 , K^{R}_{e,h}=c^{\mathrm{sell}}_{E}S^{R,\mathrm{pct}}_{e,h}, \qquad h=1,\ldots,12, K e , h R = c E sell S e , h R , pct , h = 1 , … , 12 ,
where S e , h R , p c t S^{R,\mathrm{pct}}_{e,h} S e , h R , pct is the desired-selling stock, in percent of the market denominator, of observed recipients h h h weeks after the onset of storm e e e , for observed recipient counties outside the storm footprint. Footprint households and unobserved counties are excluded. It uses the contemporaneous pricing map at fixed capital, the conditional within-period premium response. The calibration panel’s outcome is a monthly EBP change with lag dynamics, and the storms do not enter that independent panel. In this scenario the price equation’s lags imply a long-run multiplier of about 0.81 on a permanent desired-selling stock, so the scored path is not the complete cumulative EBP response that the lag equation would generate. The results are not rescaled by 0.81, which would change the estimand without supplying the horizon-dependent response.
Bridge results. With propagation following the operator’s diffusion kernel, its advantage over one-hop is concentrated in the propagated part of the response (Table I.Table 4 ). Removing information one stage at a time, the normalized error is 0.171 with the true loadings and market map, 0.299 with loading scales estimated from accounts, 0.486 or 0.487 with the market map from the oracle instrument or from fund data, and 0.975 when the demand feedback is ignored. The largest part of the remaining error therefore comes from the market map’s sampling error. Indeed, replacing the constructed instrument by the oracle changes the bridge’s error by −0.001 (Monte Carlo s.e. 0.018), no detectable difference, although not a demonstration of equivalence. Calibration panels that tile the sample’s fund geometry over 114 and 228 months lower the error to 0.388 and 0.355. These are scenario comparisons, not a forecast of what additional real months would identify. Account precision matters as well. With account noise of 0.05 or 0.25 percent of holdings instead of 0.1 percent, the error is 0.451 or 0.754.
On its point estimate the fund-data error, 0.487, meets the 0.50 criterion of Table I.Table 1 . Its Monte Carlo 95 percent interval, 0.44 to 0.53, includes that threshold. Pointwise coverage of nominal 90 percent bands is 0.881, with a Monte Carlo 95 percent interval from 0.850 to 0.913, and path coverage is 0.844. Coverage is near nominal at 57 months partly through compensation. It is 0.890 with the Notes: Mean normalized error of the conditional within-period premium response of observed recipients, weeks 1–12 after onset, held-out storms, 200 simulated replications. Coverage of nominal 90% pointwise bands is in parentheses. Path coverage with the 57-month fund-data market map is 0.84. With the true loadings and market map the operator’s error is 0.171, the counterpart on new simulation draws of 0.168 in Table 6 , row 1. Household loading scales are estimated from hypothetical account flows in every column. “Oracle instrument” observes the true granular demand innovation, with the same estimated-factor controls. Its bands use normal-approximation draws of the market map, so their coverage is not comparable with the other columns. “From fund data” uses only simulated fund flows, shares and EBP changes on an independent calibration panel with the sample’s fund geometry, and its bootstrap regenerates that panel. The 228-month panel tiles that geometry. “Anchor only” is the anchor-only prediction of Table I.Table 2 , the anchor’s common response with no propagated part. Here its loading scales are fitted to that common-response state, are more dispersed than the operator’s, and give a heavy-tailed error (median 0.73). Monte Carlo standard errors of the errors are at most 0.022, except 0.053 and 0.039 for anchor only. The paired one-hop minus operator difference with the 57-month fund-data map is 0.056 (0.022) for the total and 0.204 (0.018) for the propagated part. The simulated propagation law is the operator’s diffusion kernel.
Table I.4. The credit-premium response recovered with estimated financial maps (simulation)
true loadings and market map, 0.847 with estimated loadings and the true market map, and 0.842 with the 228-month market panel. The account stage’s bands are too narrow, and a longer market panel improves prediction without, by itself, calibrating the bands.
With zero financial loading, the 5 percent joint test, a 95 percent simultaneous band over every held-out event and horizon, rejects in 0 of 200 replications (exact 95 percent interval from 0.000 to 0.018). Without diffusion, where the true propagated part is zero, its estimate has root-mean-square error 0.038 bp and its joint test rejects in 0 of 100 replications. These nulls show only that the test does not over-reject. Where the response is not zero, the same test rejects in 8 of 200 replications with a 57-month market map (0 for the propagated part), in 122 of 200 with 228 months, and in all 200 with the true market map. At 57 months the point estimate is uninformative, but the joint bands rarely exclude a zero response.
In a near-noiseless fixture, with no spending noise, accounts with standard deviation 10 − 6 10^{-6} 1 0 − 6 and a 1,140-month calibration panel with 2 bp price news (fund shocks remain), the error is 0.019, against 0.001 with the true maps and 0.011 with the true market map alone. Large valid granular shocks in the stress months do not detectably change accuracy (error 0.481). When category flows coincide with credit news the procedure is biased (error 0.842). Its pointwise bands still cover 0.854 of true values, and the joint Anderson–Rubin set of the market stage contains the true values at least as often as in the valid world, so neither reveals the violation.
Numerical failures are reported separately from point estimation. The baseline bridge panels have none. In the no-diffusion family, one large-shock sensitivity panel (replication 45) has 183 of 199 failed market-bootstrap draws and an inadmissible point estimate, m ^ F = − 16.6 \hat{m}_F=-16.6 m ^ F = − 16.6 . Its point estimate is retained, its inference is treated as numerically failed rather than imputed, and this sensitivity enters no table. In the market-only scenarios, the panel of note a a a in Table I.Table 3 is the only one with failed draws, and three point estimates (one with factors correlated with news, two with category news) have m ^ F ≤ 0 \hat{m}_F\le0 m ^ F ≤ 0 . All are retained in the error statistics.
The historical window cannot serve as its own calibration sample. Purging each fold’s held-out storm windows (52-week support) leaves 24, 24 and 38 of the 57 months when the outcome and both lag months must be clean (26, 26 and 38 if only the outcome month must be), and the first fold loses both stress months. The design also builds in favourable features. The simulated propagation law is the operator’s diffusion kernel, and the simulated fund panel shares the estimator’s assumptions (two factors, homogeneous ζ F \zeta^F ζ F , independent idiosyncratic shocks). Resampling whole real residual cross-sections instead lowers c E s e l l c_E^{\mathrm{sell}} c E sell by about 16 percent, and by about 14 percent with the oracle instrument, so independence across funds matters for its level.
The actual 57-month sample. Applied to the actual fund panel (precision weights, two factors and fund fixed effects in the demand average, including reinvested distributions), the construction gives ζ ^ F = − 2.26 × 10 − 4 \hat{\zeta}^F=-2.26\times10^{-4} ζ ^ F = − 2.26 × 1 0 − 4 per bp (Wald t = − 4.9 t=-4.9 t = − 4.9 ), ϕ ^ = − 299 \hat{\phi}=-299 ϕ ^ = − 299 , m ^ F = 1.99 \hat{m}_F=1.99 m ^ F = 1.99 and c ^ E s e l l = 594 \hat{c}_E^{\mathrm{sell}}=594 c ^ E sell = 594 . Taken at face value, flows chase returns and the system amplifies desired selling. Inference robust to weak identification does not support this reading. The Anderson–Rubin set for the demand feedback is the whole real line, and the projection of the joint set onto c E s e l l c_E^{\mathrm{sell}} c E sell is unbounded. With the χ 2 \chi^2 χ 2 critical value these sets are, if anything, too small at this sample length, as the undercoverage in the simulated 57-month panels shows, and a size-corrected critical value can only enlarge them. March 2020 carries 69 percent of the demand moment and 65 percent of the price moment. Refitting without March 2020 gives c ^ E s e l l = 275 \hat{c}_E^{\mathrm{sell}}=275 c ^ E sell = 275 , and without both stress months 95. Instruments from short-duration funds alone give 1,530, the direction of the upward bias that the category-news simulation produces, although the simulation does not reproduce the category split itself. The external-cash definition gives the same conclusions. The network estimating equation of Chodorow-Reich, Gabaix and Viviano [9 ] is no more informative here. Its iteration reaches no root, minimum distance finds several far-apart roots, one at essentially zero price impact, and its joint set accepts every admissible point. The sample supports the pricing association of Table 5 , not the demand–price decomposition.