Main paper
Introduction
An institution that must complete a portfolio adjustment can use a public forecast to prepare its position. Other traders can use the same forecast to acquire inventory for the event. Their commitment is different: they participate only when expected liquidation proceeds cover trading and activation costs. Better information therefore changes both the institution’s preparation and the inventory available to it. The resulting execution bill can rise even as the target becomes easier to predict.
We study this interaction in a simultaneous Bayesian quantity game. After a common public signal, the institution chooses its pre-event quantity and other traders decide whether to activate an inventory allocation and how much to trade. All quantities follow a prescribed common schedule with transient impact. At a uniform-price terminal auction, the institution completes its target and suppliers liquidate. Fixing the schedule permits an exact equilibrium calculation with private activation costs and explicit terminal cash payments. The institution chooses before observing realized entry, and its target does not depend on affected execution prices.
The paper makes three contributions.
C1. A classification of institutional information value. In the scalar model, every refinement benefits an institution when the participation roster is committed before the signal. Optional activation changes the comparison. At entry, a more precise forecast can attract inventory whose price impact persists into the auction. Near full participation, it can withdraw helpful inventory asymmetrically, even without pre-event impact. In a smooth interior, harm occurs when the curvature of inventory supply interacts sufficiently strongly with its marginal effect on institutional cost. A full-support lognormal construction proves that a participation boundary is not necessary. We also identify favorable scalar comparisons: any centered target benefits from a nonzero posterior mean relative to no information, and every precision gain in the centered Gaussian family is beneficial. An exact cash decomposition explains why helpful inventory and helpful information are different claims.
C2. An exact finite-size qualification. We solve the symmetric Bayesian equilibrium for every population size N ≥ 2 N \ge2 N ≥ 2 . Each trader’s own impact raises the entry threshold. As a result, the same refinement under the same target law can be harmful in the continuum and beneficial in a finite game. A joint limit identifies N N N times posterior-mean displacement as the relevant scale near entry. Distributional stability and diffuse-forecast expansions show which comparisons survive perturbations. Fixed-prior precision paths and an outside completion facility distinguish the effects of the target law from those of execution technology.
C3. The direction of portfolio information. § 2.7 and Appendix § 7 extend equilibrium to basket activation with cross-market impact. Participation still solves one scalar equation, but the favorable scalar information rankings need not survive. We give an exact matrix criterion for the fixed-roster ranking, a sufficient commuting class, and an admissible counterexample verified from the original cash payments. Even with isotropic execution, shared activation makes the direction of information matter. Under one unchanged target law, two feasible signals reduce ordinary and execution-weighted target-forecast error by exactly the same amount. Yet information about basket magnitude raises institutional cost, while information about composition lowers it. Forecast accuracy alone therefore does not rank the execution value of portfolio information.
Daily-reset leveraged funds motivate the committed-target problem. Announced index adjustments provide another setting in which traders prepare inventory for an institutional event. The empirical application studies target-mean forecasts, conditional uncertainty, and close-related prices. Related-market, volume, and factor analyses assess the benchmark approximation. Impact and activation parameters remain illustrative because historical fund balance sheets and institutional orders are unobserved. § 1.2 develops the motivation and related literature. § 2 –§ 2.5 establish the scalar results, § 2.6 gives numerical implications, § 2.7 presents the portfolio comparison, and § 3 reports price evidence. The internet appendix contains supporting scalar theory, portfolio information, and empirical methods.
Institutional motivation and related literature
A target at the daily valuation time
Daily leveraged and inverse ETFs state an objective relative to one benchmark calculation day, generally measured from one NAV determination to the next. Changes in benchmark value and fund wealth require the manager to restore the stated exposure. This creates a committed portfolio target, but does not require an exact cash-equity order at the closing print. Derivatives, investor flows, tracking tolerances, and prior trades determine the residual implementation. We model exact completion in fixed benchmark units as an idealization of that commitment.
Under a no-flow, no-fee one-period balance sheet, starting assets A 0 A_0 A 0 and exposure β A 0 \beta A_0 β A 0 become wealth A 0 ( 1 + β R ) A_0(1+\beta R) A 0 ( 1 + β R ) and pre-reset exposure β A 0 ( 1 + R ) \beta A_0(1+R) β A 0 ( 1 + R ) . Restoring leverage therefore requires
Δ E = A 0 β ( β − 1 ) R . (1) \Delta E = A_0\beta(\beta-1)R. \tag*{(1)} Δ E = A 0 β ( β − 1 ) R . (1)
The coefficients are 12, 6, 2, 0, 2, 6 for leverage objectives − 3 -3 − 3 , − 2 -2 − 2 , − 1 -1 − 1 , 1, 2, 3, so a one-percent benchmark rise implies these numbers of cents of exposure adjustment per initial dollar. Carrying cost C C C and net flow Φ \Phi Φ under a terminal marking convention add − β C + β Φ -\beta C+\beta\Phi − β C + β Φ , while earlier flows need their own marking convention. The flow evidence in Ivanov and Lenkey [23 ] makes that distinction material. A current AUM snapshot cannot reconstruct those historical quantities.
The target in our game is external to affected execution prices. Converting (1) from dollar exposure into fixed benchmark units also requires a reference price and the appropriate marking terms, so the accounting identity motivates target uncertainty without identifying the institution’s residual order. The reset clock matters for returns as well: an ideal daily reset has log residual
log ( 1 + β R d ) − β log ( 1 + R d ) = − 1 2 β ( β − 1 ) R d 2 + O ( R d 3 ) , (2) \log(1+\beta R_d)-\beta\log(1+R_d)=-\frac{1}{2}\beta(\beta-1)R_d^2+O(R_d^3), \tag*{(2)} log ( 1 + β R d ) − β log ( 1 + R d ) = − 2 1 β ( β − 1 ) R d 2 + O ( R d 3 ) , (2)
whereas resetting at every intraday mark gives the sum of these residuals over intraday returns. The continuous-leverage identity of Avellaneda and Zhang [2 ] concerns the latter clock in a continuous-time limit. Appendix § 8 illustrates the distinction on matched paths. Current issuer documents support the daily mandate interpretation [31 ] (ProShares 2026).
Information, participation, and execution
Public information can change participation and worsen financial outcomes. In Crego [10 ] , news encourages informed risk-averse investors to trade and worsens adverse selection. In Diamond and Verrecchia [12 ] , disclosure can induce market-maker exit and reduce risk-bearing capacity. Admati and Pfleiderer [1 ] study how preannounced liquidity trades affect information asymmetry and coordination. Our participants are risk neutral and share target information. Private activation costs make inventory supply optional, while the committed institution adjusts its pretrade and pays the resulting execution bill.
The comparison uses Blackwell [5 ] and the posterior-mean geometry of Gentzkow and Kamenica [20 ] . Persuasion provides a related framework [24 ] Dworczak and Martini, 2019), but the institution has no disclosure instrument here. Ignoring a public forecast would not stop other traders from responding. Quadratic public-information objectives also admit matrix and spectral characterizations, as in Miyashita and Ui [28 ] . We derive the relevant matrix from institutional execution payments. Optional basket activation then creates the nonlinear comparison between information about magnitude and composition. The contribution is this financial classification and its finite-size qualification, rather than a new general information-design method.
Execution models allow substantially richer decisions than our prescribed schedule. Stochastic targets and terminal constraints appear in Dai et al. [11 ] ; Bank and Voß [3 ] , and endogenous timing and dropout in Fu et al. [18 , 17 ] . Major-minor interaction and anticipation are studied by Chen et al. [8 ] ; Cheng et al. [9 ] ; Neuman and Voß [32 ] . Institutional trades reveal information in Moallemi et al. [29 ] , while Campbell and Nutz [7 ] ; Strehle [37 ] analyze execution equilibria with rate, boundary, or transient-impact costs. The common signal and fixed schedule here isolate information value when activation remains avoidable and terminal settlement is uniform-price.
Table 1 records the closest economic distinctions. Information and entry already interact in quantity competition (Do and Kettering, [14 ] ), and prepared inventory is central to index execution in Pegoraro et al. [34 ] . Zhao [38 ] studies an affected-price leveraged-fund reset with capacity entry, auction mobilization, and finite speculators. In that paper’s October 15 version, capacity in Appendix § 5.4 .1.2 is committed before the open, while the fringe in Appendix § 5.3 .§ 2.7 enters within the auction. Here activation follows the common target forecast, and the target is external to affected execution prices. Entry and finite players alone are not the novelty claims.
The closest empirical comparison is Barbon et al. [4 ] . Using historical assets and constituent weights, they connect leveraged-ETF and option rebalancing to end-of-day liquidity provision and subsequent dissipation. Brøgger [6 ] finds smaller impact coefficients for larger and more predictable leveraged-volatility flows. These findings motivate favorable-information benchmarks alongside the adverse regions. Our price panel characterizes target uncertainty without identifying historical reset orders or institutional costs. Intraday predictability also has competing explanations (Gao et al., [19 ] ). The model ranks the institutional payments induced by information. Directional ETF returns alone cannot identify that comparison.
A game with pre-event trading and terminal completion
Information, timing, and admissible positions
The pre-event interval is [ 0 , 1 ] [0,1] [ 0 , 1 ] , with one public signal Y Y Y observed at time zero and the terminal target D D D revealed at time one. Conditional moments are
μ = E [ D ∣ Y ] , ν = E [ ( D − μ ) 2 ∣ Y ] , \mu= \mathbb{E}[D \mid Y], \qquad\nu= \mathbb{E}[(D-\mu)^2 \mid Y], μ = E [ D ∣ Y ] , ν = E [( D − μ ) 2 ∣ Y ] ,
where D ∈ L 2 D \in L^2 D ∈ L 2 and the signal experiment is independent of trading decisions. The institution and all prospective anticipators observe Y Y Y . Each anticipator also observes a private nonnegative entry cost f f f , drawn independently of the target and the entire family of compared public experiments. Its distribution has continuous CDF F F F and F ( 0 ) = 0 F(0)=0 F ( 0 ) = 0 . The limiting population has total mass one, with its private types distributed according to F F F in every public state. There is no additional information before terminal revelation on which strategies can be conditioned.
The charge f f f is avoidable after observing the signal, for example a fixed cost of opening an event-specific inventory program or allocating an operational slot. Infrastructure or staffing already paid for before the signal belongs instead to the fixed-roster comparison. Price and volume observations do not identify these activation costs.
The institution chooses a signed pre-event quantity z ∈ R z \in\mathbb{R} z ∈ R . A type f f f either remains inactive or pays f f f and chooses a signed inventory q f ∈ R q_f \in\mathbb{R} q f ∈ R . Pre-event trades use a common deterministic schedule ϕ ∈ L 2 [ 0 , 1 ] \phi\in L^2[0,1] ϕ ∈ L 2 [ 0 , 1 ] , with ϕ ≥ 0 \phi\ge0 ϕ ≥ 0 and ∫ 0 1 ϕ t d t = 1 \int_0^1 \phi_t\,dt = 1 ∫ 0 1 ϕ t d t = 1 . Thus the institution trades at rate z ϕ t z\phi_t z ϕ t , and an entrant at rate q f ϕ t q_f\phi_t q f ϕ t . Aggregate anticipator inventory is m = ∫ q f 1 { f e n t e r s } d F ( f ) m = \int q_f\mathbf{1}_{\{f\ \mathrm{enters}\}}\,dF(f) m = ∫ q f 1 { f enters } d F ( f ) , so aggregate pre-event rate is ( z + m ) ϕ t (z+m)\phi_t ( z + m ) ϕ t . All positions are chosen simultaneously after the signal, in the Nash sense that the institution holds the anticipator strategy fixed when it deviates. The institution is not a Stackelberg leader that commits first and internalizes subsequent re-entry decisions.
The common schedule may represent a prescribed execution window or an operational schedule whose size can be changed at a decision time. The game optimizes quantities along that schedule. Information comparisons hold the decision time and execution opportunity fixed. Waiting would shorten the horizon and change the schedule or integrated costs, requiring a different optimization problem. Once a trader enters, its entry status remains fixed until the event. Inventory follows q f ∫ 0 t ϕ s d s q_f \int_0^t \phi_s \, \mathrm{d}s q f ∫ 0 t ϕ s d s and becomes zero at terminal liquidation. Even in an extension allowing intermediate trading, flattening the position would not undo the entry decision.
Transient impact and the terminal price
Take the unaffected benchmark price as the numeraire and initially constant, so only execution-price deviations need be recorded. The pre-event impact process satisfies
I ˙ t = − ρ I t + κ ( z + m ) ϕ t , I 0 = 0 , (3) \dot{I}_t=-\rho I_t+\kappa(z+m)\phi_t,\qquad I_0=0, \tag*{(3)} I ˙ t = − ρ I t + κ ( z + m ) ϕ t , I 0 = 0 , (3)
with ρ ≥ 0 \rho\ge0 ρ ≥ 0 and κ ≥ 0 \kappa\ge0 κ ≥ 0 . Define the average execution loading and the terminal carryover loading by
K = κ ∫ 0 1 ∫ 0 t ϕ t e − ρ ( t − s ) ϕ s d s d t , (4) K=\kappa\int_0^1\int_0^t \phi_t e^{-\rho(t-s)}\phi_s\,\mathrm{d}s\,\mathrm{d}t, \tag*{(4)} K = κ ∫ 0 1 ∫ 0 t ϕ t e − ρ ( t − s ) ϕ s d s d t , (4)
H = κ ∫ 0 1 e − ρ ( 1 − s ) ϕ s d s . (5) H=\kappa\int_0^1 e^{-\rho(1-s)}\phi_s\,\mathrm{d}s. \tag*{(5)} H = κ ∫ 0 1 e − ρ ( 1 − s ) ϕ s d s . (5)
The common pre-event price component paid per unit is K ( z + m ) K(z+m) K ( z + m ) . The displacement still present just before the auction is H ( z + m ) H(z+m) H ( z + m ) . The institution also pays a temporary execution charge η 0 z 2 / 2 \eta_0 z^2/2 η 0 z 2 /2 , and each entrant pays η q f 2 / 2 \eta q_f^2/2 η q f 2 /2 , where η 0 , η > 0 \eta_0,\eta>0 η 0 , η > 0 . These coefficients include the schedule’s squared-rate integral and use the same normalized cash units as price impact.
At the event, the institutional order is J = D − z J=D-z J = D − z , the aggregate anticipator order is − m -m − m , and net outside demand is n = D − z − m n=D-z-m n = D − z − m . An outside liquidity supplier has the linear inverse supply rule
p A = H ( z + m ) + λ n , λ > 0 , (6) p_A=H(z+m)+\lambda n,\qquad\lambda>0, \tag*{(6)} p A = H ( z + m ) + λn , λ > 0 , (6)
and every event trade executes at this uniform price relative to the unaffected benchmark. The institution pays ( D − z ) p A (D-z)p_A ( D − z ) p A , anticipators collectively receive m p A mp_A m p A , and the outside supplier receives n p A np_A n p A . The institution finishes with z + J = D z+J=D z + J = D , each anticipator finishes flat, and the supplier absorbs the residual inventory. Any remaining supplier inventory is valued at the unaffected benchmark immediately after the accounting horizon. This convention completes the cash account without introducing a second market-clearing problem.
Equation (6) specifies an elastic supply schedule. It is not derived from a strategic dealer equilibrium. Net demand can have either sign, and the model imposes no auction capacity constraint or additional post-auction manipulation opportunity. The institution internalizes its price effect because its order is large. An individual trader in the continuum takes the uniform auction price as given. The finite-player version below restores each minor trader’s own price effect at its population scale.
Assumption 1. The execution coefficients satisfy λ > 0 \lambda> 0 λ > 0 , 0 ≤ H < λ 0 \le H < \lambda 0 ≤ H < λ , K ≥ 0 K \ge0 K ≥ 0 , and η 0 , η > 0 \eta_0, \eta> 0 η 0 , η > 0 . Private costs have a continuous CDF F F F on [ 0 , ∞ ) [0,\infty) [ 0 , ∞ ) with F ( 0 ) = 0 F(0) = 0 F ( 0 ) = 0 , are independent of the target and the entire family of compared public signals taken jointly, and the target has finite second moment.
The restriction on H H H is sufficient rather than necessary, and compares inherited displacement per pre-event unit with the auction price response per unit of net terminal demand. Here λ \lambda λ is inverse depth, so a larger value means less terminal liquidity for a given price change. It is checked for every reported numerical scenario. The associated coefficient b = λ + K − H b = \lambda+ K - H b = λ + K − H is positive, which also excludes profitable zero-target round trips within the stipulated strategy class.
Cash objectives and their curvature
Conditional expected institutional expenditure relative to the unaffected target value is
C 0 ( z , m ; μ , v ) = K z ( z + m ) + η 0 2 z 2 + E [ ( D − z ) { H ( z + m ) + λ ( D − z − m ) } ∣ Y ] = A 2 z 2 + b z m − d μ z − ( λ − H ) μ m + λ ( μ 2 + v ) , (7) \begin{aligned}
C_0(z,m;\mu,v) &= Kz(z+m) + \frac{\eta_0}{2}z^2 + \mathbb{E}[(D-z)\{H(z+m)+\lambda(D-z-m)\}\mid Y] \\
&= \frac{A}{2}z^2 + bzm - d\mu z - (\lambda-H)\mu m + \lambda(\mu^2+v), \tag*{(7)}
\end{aligned} C 0 ( z , m ; μ , v ) = K z ( z + m ) + 2 η 0 z 2 + E [( D − z ) { H ( z + m ) + λ ( D − z − m )} ∣ Y ] = 2 A z 2 + b z m − d μ z − ( λ − H ) μ m + λ ( μ 2 + v ) , (7)
where
b = λ + K − H , A = η 0 + 2 b , d = 2 λ − H . (8) b = \lambda+ K - H,\qquad A = \eta_0 + 2b,\qquad d = 2\lambda- H. \tag*{(8)} b = λ + K − H , A = η 0 + 2 b , d = 2 λ − H . (8)
An entrant’s conditional expected profit before its fixed cost is
π ( q ; z , m , μ ) = q { λ μ − b ( z + m ) } − η 2 q 2 . (9) \pi(q;z,m,\mu) = q\{\lambda\mu-b(z+m)\} - \frac{\eta}{2}q^2. \tag*{(9)} π ( q ; z , m , μ ) = q { λ μ − b ( z + m )} − 2 η q 2 . (9)
These expressions follow by subtracting pre-event purchases and temporary charges from terminal liquidation receipts. The variance term λ v \lambda v λ v is retained in (7) , even though it does not affect the pre-event best response under risk neutrality, because it changes when information changes.
Both optimization problems are globally well posed in their own controls. The major’s objective has curvature A > 0 A > 0 A > 0 , and the entrant’s profit has curvature − η < 0 -\eta< 0 − η < 0 . These conditions do not require a positive joint signal-inventory state matrix. Completing the square in the actual control bounds the economically necessary linear signal-profit term. Every equilibrium position is bounded by a constant times ∣ μ ∣ |\mu| ∣ μ ∣ , so finite second moments suffice for ex ante integrability.
An unaffected square-integrable price martingale S S S can be restored in a filtration ( F t ) (\mathcal{F}_t) ( F t ) with σ ( Y ) ⊆ F 0 \sigma(Y) \subseteq\mathcal{F}_0 σ ( Y ) ⊆ F 0 , provided it remains a martingale in that enlarged filtration and the positions retain the stated deterministic schedule. Its conditional expected contribution z ∫ ϕ t ( S t − S 1 ) d t z\int\phi_t(S_t-S_1)\,\mathrm{d}t z ∫ ϕ t ( S t − S 1 ) d t to institutional shortfall is zero, and the corresponding inventory profit has the opposite zero expectation. The target may be correlated with the terminal martingale innovation without changing that calculation, since the institution’s terminal fundamental payment cancels against the value of its mandated final inventory. The present analysis therefore concerns the execution component of the target rather than a fundamental-return alpha.
Equilibrium and cost incidence
Define
s = λ − b d A = λ η 0 + b H A > 0 , c = b ( 1 − b A ) = b ( η 0 + b ) A > 0. (10) s = \lambda-\frac{bd}{A}=\frac{\lambda\eta_0+bH}{A}>0,\qquad c=b\left(1-\frac{b}{A}\right)=\frac{b(\eta_0+b)}{A}>0. \tag*{(10)} s = λ − A b d = A λ η 0 + b H > 0 , c = b ( 1 − A b ) = A b ( η 0 + b ) > 0. (10)
The institutional best response to m m m is ( d μ − b m ) / A (d\mu-bm)/A ( d μ − bm ) / A . Substituting it into the entrant’s first-order condition gives η q = s μ − c m \eta q=s\mu-cm η q = s μ − c m . This is an algebraic solution of simultaneous best responses. It does not give the institution an earlier commitment opportunity.
Theorem 1 (Unique participation equilibrium). Under Assumption 1, for every conditional mean μ \mu μ there is a unique equilibrium entered mass a ∈ [ 0 , 1 ] a \in[0,1] a ∈ [ 0 , 1 ] , aggregate inventory m m m , and institutional quantity z z z , up to entry decisions by indifferent zero-measure types. The common entrant inventory q q q denotes a hypothetical entrant’s optimum when a = 0 a=0 a = 0 . They satisfy
a = F ( η s 2 μ 2 2 ( η + c a ) 2 ) , (11) a=F\left(\frac{\eta s^2\mu^2}{2(\eta+ca)^2}\right), \tag*{(11)} a = F ( 2 ( η + c a ) 2 η s 2 μ 2 ) , (11) q = s μ η + c a , m = a q , z = d μ − b m A . (12) q=\frac{s\mu}{\eta+ca},\qquad m=aq,\qquad z=\frac{d\mu-bm}{A}. \tag*{(12)} q = η + c a s μ , m = a q , z = A d μ − bm . (12) A type enters exactly when f ≤ η q 2 / 2 f\leq\eta q^2/2 f ≤ η q 2 /2 . The solution is continuous in μ \mu μ , its entered mass is nondecreasing in ∣ μ ∣ |\mu| ∣ μ ∣ , and its positions have finite second moments. The equilibrium is independent of v v v , although its institutional cost increases by λ v \lambda v λ v .
Proof. Strict own-control curvature gives unique conditional best responses. Every entrant has the same trading technology and hence the same optimal q q q , with maximized gross profit η q 2 / 2 \eta q^2/2 η q 2 /2 . Aggregation gives m = a q m=aq m = a q and therefore (12) , while the entry rule gives (11) . Let h ( a ) = a − F ( η s 2 μ 2 / [ 2 ( η + c a ) 2 ] ) h(a)=a-F(\eta s^2\mu^2/[2(\eta+ca)^2]) h ( a ) = a − F ( η s 2 μ 2 / [ 2 ( η + c a ) 2 ]) . Continuity of F F F implies continuity of h h h , its endpoint values satisfy h ( 0 ) ≤ 0 h(0)\leq0 h ( 0 ) ≤ 0 and h ( 1 ) ≥ 0 h(1)\geq0 h ( 1 ) ≥ 0 , and for a 2 > a 1 a_2>a_1 a 2 > a 1 the nonincreasing argument of F F F gives h ( a 2 ) − h ( a 1 ) ≥ a 2 − a 1 > 0 h(a_2)-h(a_1)\geq a_2-a_1>0 h ( a 2 ) − h ( a 1 ) ≥ a 2 − a 1 > 0 . There is therefore exactly one zero, and compactness of [ 0 , 1 ] [0,1] [ 0 , 1 ] , continuity of the equation, and uniqueness imply continuous dependence on μ \mu μ , while the same ordering establishes monotonicity in ∣ μ ∣ |\mu| ∣ μ ∣ . Finally, ∣ q ∣ ≤ s ∣ μ ∣ / η |q|\leq s|\mu|/\eta ∣ q ∣ ≤ s ∣ μ ∣/ η and ∣ m ∣ ≤ ∣ q ∣ |m|\leq|q| ∣ m ∣ ≤ ∣ q ∣ , with an analogous bound on z z z , establish the moment claim and complete verification of both agents’ global best responses.
If F F F has density f F f_F f F at the entry cutoff and participation is interior, implicit differentiation for u = ∣ μ ∣ > 0 u=|\mu|>0 u = ∣ μ ∣ > 0 gives
d a d u = f F ( g ) η s 2 u / ( η + c a ) 2 1 + f F ( g ) η s 2 u 2 c / ( η + c a ) 3 , g = η s 2 u 2 2 ( η + c a ) 2 . (13) \frac{da}{du}=\frac{f_F(g)\eta s^2u/(\eta+ca)^2}{1+f_F(g)\eta s^2u^2c/(\eta+ca)^3},\qquad g=\frac{\eta s^2u^2}{2(\eta+ca)^2}. \tag*{(13)} d u d a = 1 + f F ( g ) η s 2 u 2 c / ( η + c a ) 3 f F ( g ) η s 2 u / ( η + c a ) 2 , g = 2 ( η + c a ) 2 η s 2 u 2 . (13)
This derivative is nonnegative, and it is strictly positive only where the density is positive and the participation constraint does not bind. Equation (11) is an implicit characterization for a general distribution, rather than a closed-form solution for a a a . Its monotonicity supports globally reliable bisection without requiring a contraction estimate for the entry response.
When inventory provision raises institutional cost
Let z 0 = d μ / A z_0=d\mu/A z 0 = d μ / A be the institution’s optimal pre-event position in the absence of anticipators, and let C 0 ∅ = λ ( μ 2 + v ) − d 2 μ 2 / ( 2 A ) C_0^\varnothing=\lambda(\mu^2+v)-d^2\mu^2/(2A) C 0 ∅ = λ ( μ 2 + v ) − d 2 μ 2 / ( 2 A ) be the corresponding cost. Define
e = d b A − ( λ − H ) = H b − ( λ − H ) η 0 A , k = b 2 2 A . (14) e=\frac{db}{A}-(\lambda-H)=\frac{Hb-(\lambda-H)\eta_0}{A},\qquad k=\frac{b^2}{2A}. \tag*{(14)} e = A d b − ( λ − H ) = A H b − ( λ − H ) η 0 , k = 2 A b 2 . (14)
Proposition 2 (Institutional cost decomposition). For any signed anticipator inventory m m m , the institution’s optimized cost satisfies
min z C 0 ( z , m ; μ , v ) − C 0 ∅ = e μ m − k m 2 . (15) \min_z C_0(z,m;\mu,v)-C_0^\varnothing=e\mu m-km^2. \tag*{(15)} z min C 0 ( z , m ; μ , v ) − C 0 ∅ = e μ m − k m 2 . (15)
If its pre-event quantity is frozen at z 0 z_0 z 0 , the cost change is e μ m e\mu m e μ m , and adapting that quantity saves k m 2 km^2 k m 2 . At equilibrium, μ m ≥ 0 \mu m \ge0 μ m ≥ 0 , so entry weakly lowers institutional cost whenever e ≤ 0 e \le0 e ≤ 0 . If e > 0 e > 0 e > 0 and μ ≠ 0 \mu\ne0 μ = 0 , entry raises cost exactly when 0 < m / μ < e / k 0 < m/\mu< e/k 0 < m / μ < e / k .
Proof. Completing the square in (7) gives the minimum
λ ( μ 2 + v ) − ( λ − H ) μ m − ( d μ − b m ) 2 2 A . \lambda(\mu^2+v)-(\lambda-H)\mu m-\frac{(d\mu-bm)^2}{2A}. λ ( μ 2 + v ) − ( λ − H ) μ m − 2 A ( d μ − bm ) 2 .
Subtracting C ∞ C^\infty C ∞ proves (15) . Evaluation at z 0 z_0 z 0 gives the frozen-position expression, and subtraction yields b 2 m 2 / ( 2 A ) b^2m^2/(2A) b 2 m 2 / ( 2 A ) . The equilibrium sign follows from (12) , after which the stated conditions are the signs of e ( m / μ ) − k ( m / μ ) 2 e(m/\mu)-k(m/\mu)^2 e ( m / μ ) − k ( m / μ ) 2 .
The coefficient e e e measures the cost effect of the first unit of anticipator inventory at the institution’s unperturbed optimum. When H b > ( λ − H ) η 0 Hb>(\lambda-H)\eta_0 H b > ( λ − H ) η 0 , the price displacement carried into the auction exceeds the initial benefit of additional auction supply. Adjusting its own pretrade gives the institution a quadratic saving. At sufficiently small participation, however, that saving cannot offset a positive first-order cost effect. When H = 0 H=0 H = 0 , e = − λ η 0 / A < 0 e=-\lambda\eta_0/A<0 e = − λ η 0 / A < 0 , and anticipator inventory is beneficial throughout this benchmark. This sign determines the onset mechanism. Section 5 shows that information can also be harmful when all anticipator inventory is beneficial.
Information and optional inventory supply
The cost function of a posterior mean
Let m ( μ ) m(\mu) m ( μ ) denote equilibrium inventory from Theorem 1, and define
ψ ( μ ) = − d 2 2 A μ 2 + e μ m ( μ ) − k m ( μ ) 2 . (16) \psi(\mu)=-\frac{d^2}{2A}\mu^2+e\mu m(\mu)-km(\mu)^2. \tag*{(16)} ψ ( μ ) = − 2 A d 2 μ 2 + e μ m ( μ ) − k m ( μ ) 2 . (16)
Throughout, implementation costs are independent of the target and all public signals in the comparison, considered jointly. Their conditional distribution therefore remains F F F . For any such public information H \mathcal{H} H , the institution’s ex ante optimized expenditure is
C ( H ) = λ E [ D 2 ] + E [ ψ ( E [ D ∣ H ] ) ] . (17) C(\mathcal{H})=\lambda\mathbb{E}[D^2]+\mathbb{E}\left[\psi\left(\mathbb{E}[D\mid\mathcal{H}]\right)\right]. \tag*{(17)} C ( H ) = λ E [ D 2 ] + E [ ψ ( E [ D ∣ H ] ) ] . (17)
The first term is unchanged across experiments under the same target law. The second combines the institution’s direct forecasting benefit with the cost effect of the inventory supplied in response. The information comparison therefore reduces to the shape of one equilibrium cost function.
Theorem 3 (Information characterization). Under Assumption 1 , (17) holds for every square-integrable target. On an interval I I I , every feasible Blackwell refinement with posterior means in I I I weakly lowers cost if ψ \psi ψ is concave on I I I . Conversely, if ψ \psi ψ is not concave on I I I , there exists a two-point target law supported in I I I for which revelation raises cost relative to no information. Thus concavity is equivalent to a beneficial information ranking across all target laws supported in I I I .
Where m m m is twice differentiable,
ψ ′ ′ = − d 2 A + 2 e m ′ − 2 k ( m ′ ) 2 + ( e μ − 2 k m ) m ′ ′ . (18) \psi''=-\frac{d^2}{A}+2em'-2k(m')^2+(e\mu-2km)m''. \tag*{(18)} ψ ′′ = − A d 2 + 2 e m ′ − 2 k ( m ′ ) 2 + ( e μ − 2 k m ) m ′′ . (18)
At a point τ \tau τ where m m m is continuous with one-sided derivatives, the derivative jump is
J τ : = ψ + ′ ( τ ) − ψ − ′ ( τ ) = [ e τ − 2 k m ( τ ) ] [ m + ′ ( τ ) − m − ′ ( τ ) ] . (19) J_{\tau} := \psi_{+}^{\prime}(\tau) - \psi_{-}^{\prime}(\tau) = [e\tau- 2km(\tau)] [m_{+}^{\prime}(\tau) - m_{-}^{\prime}(\tau)]. \tag*{(19)} J τ := ψ + ′ ( τ ) − ψ − ′ ( τ ) = [ e τ − 2 k m ( τ )] [ m + ′ ( τ ) − m − ′ ( τ )] . (19)
For a positive-signal interior region with continuously differentiable cost density f F f_F f F , put g = η q 2 / 2 g = \eta q^2/2 g = η q 2 /2 , B F = F ( g ) + 2 g f F ( g ) B_F = F(g) + 2g f_F(g) B F = F ( g ) + 2 g f F ( g ) , and R F = η + c B F R_F = \eta+ cB_F R F = η + c B F . Then
m ′ = s B F R F , m ′ ′ = s 2 η q { 3 f F ( g ) + 2 g f F ′ ( g ) } R F 3 . (20) m^{\prime} = \frac{sB_F}{R_F}, \qquad m^{\prime\prime} = \frac{s^2\eta q\{3f_F(g) + 2g f_F^{\prime}(g)\}}{R_F^3}. \tag*{(20)} m ′ = R F s B F , m ′′ = R F 3 s 2 η q { 3 f F ( g ) + 2 g f F ′ ( g )} . (20)
Proof. Substitute Proposition 2 and the conditional variance identity into expected cost to obtain (17) . A Blackwell refinement admits posterior means M 0 M_0 M 0 , M 1 M_1 M 1 satisfying E [ M 1 ∣ H 0 ] = M 0 \mathbb{E}[M_1 \mid\mathcal{H}_0] = M_0 E [ M 1 ∣ H 0 ] = M 0 , so conditional Jensen proves the concavity implication. If concavity fails, choose x , y ∈ I x,y \in I x , y ∈ I and θ ∈ ( 0 , 1 ) \theta\in(0,1) θ ∈ ( 0 , 1 ) such that θ ψ ( x ) + ( 1 − θ ) ψ ( y ) > ψ ( θ x + ( 1 − θ ) y ) \theta\psi(x) + (1-\theta)\psi(y) > \psi(\theta x + (1-\theta)y) θ ψ ( x ) + ( 1 − θ ) ψ ( y ) > ψ ( θ x + ( 1 − θ ) y ) . A target equal to x x x or y y y with these probabilities and its full revelation implement precisely that strict inequality, with the same target law in both experiments. The moment bounds in Theorem 1 justify all expectations. Differentiating (16) gives (18) and (19) . Finally, m = q F ( g ) m = qF(g) m = q F ( g ) and μ = q [ η + c F ( g ) ] / s \mu= q[\eta+ cF(g)]/s μ = q [ η + c F ( g )] / s imply d m / d q = B F dm/dq = B_F d m / d q = B F and d μ / d q = R F / s > 0 d\mu/dq = R_F/s > 0 d μ / d q = R F / s > 0 . A second differentiation, using B F ′ = η q ( 3 f F + 2 g f F ′ ) B_F^{\prime} = \eta q(3f_F + 2gf_F^{\prime}) B F ′ = η q ( 3 f F + 2 g f F ′ ) , proves (20) .
The economic content of (18) is clearest after separating marginal inventory incidence from the direct forecasting gain. For μ > 0 \mu> 0 μ > 0 , define
G ( μ ) = e μ − 2 k m ( μ ) , Q ( h ) = d 2 A − 2 e h + 2 k h 2 . (21) G(\mu) = e\mu- 2km(\mu), \qquad Q(h) = \frac{d^2}{A} - 2eh + 2kh^2. \tag*{(21)} G ( μ ) = e μ − 2 k m ( μ ) , Q ( h ) = A d 2 − 2 e h + 2 k h 2 . (21)
Here G G G is the marginal effect of supplied inventory on optimized institutional cost, holding the target mean fixed. For every h ≥ 0 h \ge0 h ≥ 0 , Q ( h ) > 0 Q(h) > 0 Q ( h ) > 0 . If e ≤ 0 e \le0 e ≤ 0 , each term in Q Q Q is nonnegative and the first is positive. If e > 0 e > 0 e > 0 , the identity e = d b / A − ( λ − H ) < d b / A e = db/A - (\lambda- H) < db/A e = d b / A − ( λ − H ) < d b / A and 2 k = b 2 / A 2k = b^2/A 2 k = b 2 / A give Q ( h ) ≥ d 2 / A − e 2 / ( 2 k ) > 0 Q(h) \ge d^2/A - e^2/(2k) > 0 Q ( h ) ≥ d 2 / A − e 2 / ( 2 k ) > 0 . Consequently,
ψ ′ ′ ( μ ) = G ( μ ) m ′ ′ ( μ ) − Q ( m ′ ( μ ) ) . \psi^{\prime\prime}(\mu) = G(\mu)m^{\prime\prime}(\mu) - Q(m^{\prime}(\mu)). ψ ′′ ( μ ) = G ( μ ) m ′′ ( μ ) − Q ( m ′ ( μ )) .
Smooth local harm occurs precisely when G m ′ ′ > Q ( m ′ ) Gm^{\prime\prime} > Q(m^{\prime}) G m ′′ > Q ( m ′ ) . Marginally costly inventory requires sufficiently accelerating supply. Marginally helpful inventory requires sufficiently decelerating supply. Opposite signs of G G G and m ′ ′ m^{\prime\prime} m ′′ imply strict concavity. At a slope discontinuity the corresponding condition is J τ = G ( τ ) Δ m ′ ( τ ) J_{\tau} = G(\tau)\Delta m^{\prime}(\tau) J τ = G ( τ ) Δ m ′ ( τ ) . The total inventory effect, e μ m − k m 2 e\mu m - km^2 e μ m − k m 2 , has a different threshold: for e > 0 e > 0 e > 0 and m > 0 m > 0 m > 0 , its sign changes at m / μ = e / k m/\mu= e/k m / μ = e / k , whereas marginal incidence changes at e / ( 2 k ) e/(2k) e / ( 2 k ) . Reflection gives the negative-mean case.
The converse allows the target law to vary across candidate counterexamples. It does not allow arbitrary posterior-mean distributions under a fixed prior. For a given nonbinary prior, feasible posterior means satisfy the additional restrictions characterized by Gentzkow and Kamenica [20 ] . The ordering argument uses standard information geometry. The financial content lies in the coefficients and inventory derivatives that determine the shape of ψ \psi ψ .
If a roster of mass a a a is fixed before information arrives, with implementation costs already sunk, inventory is m = t ( a ) μ m=t(a)\mu m = t ( a ) μ , where t ( a ) = a s / ( η + c a ) t(a)=as/(\eta+ca) t ( a ) = a s / ( η + c a ) . In this case
ψ a ( μ ) = [ − ( λ − H ) t ( a ) − { d − b t ( a ) } 2 2 A ] μ 2 . (22) \psi_a(\mu)=\left[-(\lambda-H)t(a)-\frac{\{d-bt(a)\}^{2}}{2A}\right]\mu^{2}. \tag*{(22)} ψ a ( μ ) = [ − ( λ − H ) t ( a ) − 2 A { d − b t ( a ) } 2 ] μ 2 . (22)
The bracket is strictly negative. Every Blackwell refinement therefore lowers cost for each fixed roster, strictly whenever it increases the second moment of the posterior mean. The institution reoptimizes its quantity whether information arrives before or after activation. Only information arriving before avoidable activation can also change who participates.
An exact cash decomposition. The same mechanism has a finite-refinement interpretation. For nested posterior means M 0 , M 1 M_0,M_1 M 0 , M 1 , write m i = m ( M i ) m_i=m(M_i) m i = m ( M i ) , Δ M = M 1 − M 0 \Delta M=M_1-M_0 Δ M = M 1 − M 0 , and Δ m = m 1 − m 0 \Delta m=m_1-m_0 Δ m = m 1 − m 0 . Equation (16) gives
Δ C = − d 2 2 A E [ ( Δ M ) 2 ] + E [ ( e M 1 − 2 k m 0 ) Δ m ] − k E [ ( Δ m ) 2 ] . (23) \Delta C=-\frac{d^{2}}{2A}\mathbb{E}[(\Delta M)^2]+\mathbb{E}[(eM_1-2km_0)\Delta m]-k\mathbb{E}[(\Delta m)^2]. \tag*{(23)} Δ C = − 2 A d 2 E [( Δ M ) 2 ] + E [( e M 1 − 2 k m 0 ) Δ m ] − k E [( Δ m ) 2 ] . (23)
Indeed, E [ M 1 2 − M 0 2 ] = E [ ( Δ M ) 2 ] \mathbb{E}[M_1^2-M_0^2]=\mathbb{E}[(\Delta M)^2] E [ M 1 2 − M 0 2 ] = E [( Δ M ) 2 ] and E [ m 0 Δ M ] = 0 \mathbb{E}[m_0\Delta M]=0 E [ m 0 Δ M ] = 0 . Expanding M 1 m 1 − M 0 m 0 = M 1 Δ m + m 0 Δ M M_1m_1-M_0m_0=M_1\Delta m+m_0\Delta M M 1 m 1 − M 0 m 0 = M 1 Δ m + m 0 Δ M and m 1 2 − m 0 2 = 2 m 0 Δ m + ( Δ m ) 2 m_1^2-m_0^2=2m_0\Delta m+(\Delta m)^2 m 1 2 − m 0 2 = 2 m 0 Δ m + ( Δ m ) 2 proves the identity, including refinements that cross participation boundaries.
The three terms can be read directly from the original cash objective. First hold suppliers’ inventory at m 0 m_0 m 0 and let the institution adjust its pretrade to M 1 M_1 M 1 . The first term is the direct forecasting saving. Next change inventory to m 1 m_1 m 1 , keeping that pretrade fixed. The middle term is the resulting change in cash expenditure. Finally let the institution reoptimize against m 1 m_1 m 1 . The final cost change is − k E [ ( Δ m ) 2 ] -k\mathbb{E}[(\Delta m)^2] − k E [( Δ m ) 2 ] , since k = b 2 / ( 2 A ) k=b^2/(2A) k = b 2 / ( 2 A ) . Information is harmful exactly when the middle cost increase exceeds the two savings together. These intermediate steps are accounting counterfactuals, not separate asymmetric-information equilibria. Appendix § 5 .§ 2 evaluates them in the smooth example.
Information conditional on a trading environment
A public range or activity measure may reveal trading conditions as well as the target. To distinguish these channels, let Z Z Z be observed before both target experiments and let execution coefficients and the activation-cost law be measurable functions of Z Z Z . Suppose the primitive assumptions hold conditional on Z Z Z , and costs are conditionally independent of the target and the entire compared signal family given Z Z Z . For the finite game, private costs must also be conditionally independent across allocations. With H 0 ⊆ H 1 \mathcal{H}_0\subseteq\mathcal{H}_1 H 0 ⊆ H 1 both containing σ ( Z ) \sigma(Z) σ ( Z ) , the same calculation gives
C ( H i ) = E [ λ Z D 2 ] + E [ ψ Z ( E [ D ∣ H i ] ) ] . (24) C(\mathcal{H}_i)=\mathbb{E}[\lambda_ZD^2]+\mathbb{E}[\psi_Z(\mathbb{E}[D\mid\mathcal{H}_i])]. \tag*{(24)} C ( H i ) = E [ λ Z D 2 ] + E [ ψ Z ( E [ D ∣ H i ])] . (24)
For example, square-integrable D D D and execution coefficients in a compact admissible set bounded away from zero denominators ensure integrability. Equilibrium and Jensen arguments then apply conditional on Z Z Z and can be integrated. This is a conditional formulation of the existing game, not identification of F Z F_Z F Z from trading volume. Newly revealing Z Z Z , changing the conditional cost law, or delaying execution changes more than target information within a fixed environment.
Forecast accuracy and institutional information value are different comparisons. True nested conditional means M i = E [ D ∣ H i ] M_i = \mathbb{E}[D \mid\mathcal{H}_i] M i = E [ D ∣ H i ] satisfy
E [ ( D − M 0 ) 2 ] − E [ ( D − M 1 ) 2 ] = E [ ( M 1 − M 0 ) 2 ] ≥ 0. (25) \mathbb{E}[(D-M_0)^2]-\mathbb{E}[(D-M_1)^2]=\mathbb{E}[(M_1-M_0)^2]\ge0. \tag*{(25)} E [( D − M 0 ) 2 ] − E [( D − M 1 ) 2 ] = E [( M 1 − M 0 ) 2 ] ≥ 0. (25)
The adverse effect in Theorem 3 can therefore coexist with improved target prediction. Worse performance by a fitted regression reflects estimation or specification error, rather than harmful information in this sense. Appendix § 5 .§ 2.4 bounds the error in a model-implied cost comparison when posterior means are estimated.
Entry onset and saturation
Assume a cost distribution supported on [ f 0 , f 1 ] [f_0,f_1] [ f 0 , f 1 ] , where 0 < f 0 < f 1 < ∞ 0<f_0<f_1<\infty 0 < f 0 < f 1 < ∞ , with endpoint expansions
F ( f 0 + x ) = ℓ 0 x + O ( x 2 ) , 1 − F ( f 1 − x ) = ℓ 1 x + O ( x 2 ) , ℓ 0 , ℓ 1 > 0. (26) F(f_0+x)=\ell_0x+O(x^2), \qquad1-F(f_1-x)=\ell_1x+O(x^2), \qquad\ell_0,\ell_1>0. \tag*{(26)} F ( f 0 + x ) = ℓ 0 x + O ( x 2 ) , 1 − F ( f 1 − x ) = ℓ 1 x + O ( x 2 ) , ℓ 0 , ℓ 1 > 0. (26)
Define the onset and full-participation thresholds and the saturated inventory slope by
μ 0 = 2 η f 0 s , μ 1 = η + c s 2 f 1 η , t 1 = s η + c . (27) \mu_0=\frac{\sqrt{2\eta f_0}}{s}, \qquad\mu_1=\frac{\eta+c}{s}\sqrt{\frac{2f_1}{\eta}}, \qquad t_1=\frac{s}{\eta+c}. \tag*{(27)} μ 0 = s 2 η f 0 , μ 1 = s η + c η 2 f 1 , t 1 = η + c s . (27)
Entry starts at μ 0 \mu_0 μ 0 , while the entire pool participates above μ 1 \mu_1 μ 1 . Full participation exhausts the entry margin, not the inventory response: each active quantity remains unrestricted and aggregate inventory continues to grow as m = t 1 μ m=t_1\mu m = t 1 μ . There is no hard position or balance-sheet cap.
Theorem 4 (Two sources of harmful information). Under Assumption 1 and (26), set
L 0 = 2 s f 0 ℓ 0 η + 2 c f 0 ℓ 0 , B 1 = 1 + 2 f 1 ℓ 1 , L 1 = s B 1 η + c B 1 > t 1 . L_0=\frac{2s f_0\ell_0}{\eta+2c f_0\ell_0}, \qquad B_1=1+2f_1\ell_1, \qquad L_1=\frac{sB_1}{\eta+cB_1}>t_1. L 0 = η + 2 c f 0 ℓ 0 2 s f 0 ℓ 0 , B 1 = 1 + 2 f 1 ℓ 1 , L 1 = η + c B 1 s B 1 > t 1 . The jumps of ψ \psi ψ at the positive boundaries are
J 0 = e μ 0 L 0 , J 1 = μ 1 ( e − 2 k t 1 ) ( t 1 − L 1 ) . (28) J_0=e\mu_0L_0, \qquad J_1=\mu_1(e-2kt_1)(t_1-L_1). \tag*{(28)} J 0 = e μ 0 L 0 , J 1 = μ 1 ( e − 2 k t 1 ) ( t 1 − L 1 ) . (28) Consequently, onset has an upward jump if e > 0 e>0 e > 0 , and saturation has an upward jump if e < 2 k t 1 e<2kt_1 e < 2 k t 1 , including every e ≤ 0 e\le0 e ≤ 0 . At any boundary τ \tau τ with J τ > 0 J_\tau>0 J τ > 0 , fix δ > 0 \delta>0 δ > 0 and the target D = τ + δ B D=\tau+\delta B D = τ + δ B , where B B B is a symmetric sign. A noisy observation of B B B with accuracy ( 1 + ϵ / δ ) / 2 (1+\epsilon/\delta)/2 ( 1 + ϵ / δ ) /2 , for 0 < ϵ < δ 0<\epsilon<\delta 0 < ϵ < δ , refines no information and changes institutional cost by
Δ C = J τ 2 ϵ + O ( ϵ 2 ) > 0 for sufficiently small ϵ . (29) \Delta C=\frac{J_\tau}{2}\epsilon+O(\epsilon^2)>0 \quad\text{for sufficiently small }\epsilon. \tag*{(29)} Δ C = 2 J τ ϵ + O ( ϵ 2 ) > 0 for sufficiently small ϵ . (29) Every admissible execution environment with (26) therefore admits a harmful refinement for a suitably located prior, as well as beneficial refinements in regions with constant participation.
Proof. At onset, m = 0 m=0 m = 0 below the boundary and expansion of the equilibrium equation gives
m ( μ 0 + x ) = L 0 x + O ( x 2 ) . m(\mu_0+x)=L_0x+O(x^2). m ( μ 0 + x ) = L 0 x + O ( x 2 ) .
At saturation, m ( μ 1 ) = t 1 μ 1 m(\mu_1)=t_1\mu_1 m ( μ 1 ) = t 1 μ 1 , the left slope is L 1 L_1 L 1 by the endpoint expansion, and the right slope is t 1 t_1 t 1 . The endpoint expansions imply quadratic remainders for these inventory expansions, as shown in Appendix § 5.5 . Formula (19) gives the stated jumps. Bayes’ rule gives posterior means τ ± ϵ \tau\pm\epsilon τ ± ϵ with equal probability, so the cost change is the symmetric difference [ ψ ( τ + ϵ ) + ψ ( τ − ϵ ) ] / 2 − ψ ( τ ) [\psi(\tau+\epsilon)+\psi(\tau-\epsilon)]/2-\psi(\tau) [ ψ ( τ + ϵ ) + ψ ( τ − ϵ )] /2 − ψ ( τ ) . Its linear term is J τ ϵ / 2 J_\tau\epsilon/2 J τ ϵ /2 . If e > 0 e>0 e > 0 the onset jump is positive, while if e ≤ 0 e\leq0 e ≤ 0 the saturation jump is positive because k , t 1 > 0 k,t_1>0 k , t 1 > 0 and L 1 > t 1 L_1>t_1 L 1 > t 1 . A binary refinement confined to the nonempty no-entry region has the strictly negative change − d 2 ϵ 2 / ( 2 A ) -d^2\epsilon^2/(2A) − d 2 ϵ 2 / ( 2 A ) , which proves the final assertion.
At onset, the first entrants create more inherited price displacement than they save in the institution’s auction bill. Saturation has a different mechanism when e ≤ 0 e\leq0 e ≤ 0 . Inventory lowers institutional cost at every signal, but a downward forecast revision withdraws helpful supply. An upward revision cannot attract more participants from the exhausted pool, although incumbents still increase their quantities. This asymmetry explains why helpful participation need not make information helpful. Persistent pre-event impact is necessary for harmful onset in this model. It is not necessary for harmful information near saturation.
For an explicit example without pre-event impact, take λ = η 0 = η = 1 \lambda=\eta_0=\eta=1 λ = η 0 = η = 1 , K = H = 0 K=H=0 K = H = 0 , and uniform costs on [ 0.02 , 0.04 ] [0.02,0.04] [ 0.02 , 0.04 ] . Then A = 3 A=3 A = 3 , b = 1 b=1 b = 1 , d = 2 d=2 d = 2 , s = 1 / 3 s=1/3 s = 1/3 , c = 2 / 3 c=2/3 c = 2/3 , e = − 1 / 3 e=-1/3 e = − 1/3 , k = 1 / 6 k=1/6 k = 1/6 , and μ 1 = 2 \mu_1=\sqrt{2} μ 1 = 2 . The inventory slopes are 5 / 13 5/13 5/13 from the left and 1 / 5 1/5 1/5 from the right, yielding
J 1 = 24 2 325 , Δ C = 12 2 325 ϵ + O ( ϵ 2 ) . (30) J_1=\frac{24\sqrt{2}}{325},\qquad\Delta C=\frac{12\sqrt{2}}{325}\epsilon+O(\epsilon^2). \tag*{(30)} J 1 = 325 24 2 , Δ C = 325 12 2 ϵ + O ( ϵ 2 ) . (30)
With participation fixed at one, the same experiment instead changes cost by − 0.74 ϵ 2 -0.74\epsilon^2 − 0.74 ϵ 2 .
A primitive classification between the boundaries
Proposition 5 (Uniform implementation costs). For F F F uniform on [ f 0 , f 1 ] [f_0,f_1] [ f 0 , f 1 ] , inventory has m ′ > 0 m'>0 m ′ > 0 and m ′ ′ > 0 m''>0 m ′′ > 0 throughout the positive-signal interior. If e ≤ 0 e\leq0 e ≤ 0 , ψ \psi ψ is strictly concave in every smooth participation region, its onset jumps are nonpositive, and its only failures of concavity occur at saturation. If e > 0 e>0 e > 0 , positive smooth curvature can occur only where
m ( μ ) μ < e 2 k , μ > 0. (31) \frac{m(\mu)}{\mu}<\frac{e}{2k},\qquad\mu>0. \tag*{(31)} μ m ( μ ) < 2 k e , μ > 0. (31) Negative signals obey the reflected classification.
Proof. The uniform density gives 3 f F + 2 g f F ′ = 3 / ( f 1 − f 0 ) > 0 3f_F+2gf'_F=3/(f_1-f_0)>0 3 f F + 2 g f F ′ = 3/ ( f 1 − f 0 ) > 0 in (20) . For e ≤ 0 e\leq0 e ≤ 0 , every term in (18) is nonpositive and its first term is strictly negative. Onset jumps have the sign of e e e . On constant-participation regions (22) gives strict concavity. For e > 0 e>0 e > 0 , completing the square in the first three terms of (18) bounds their sum above by − d 2 / A + e 2 / ( 2 k ) < 0 -d^2/A+e^2/(2k)<0 − d 2 / A + e 2 / ( 2 k ) < 0 , since 0 < e < d b / A 0<e<db/A 0 < e < d b / A and 2 k = b 2 / A 2k=b^2/A 2 k = b 2 / A . Positive curvature therefore requires e μ − 2 k m > 0 e\mu-2km>0 e μ − 2 k m > 0 . Finally, the equilibrium has odd inventory m m m and even cost ψ \psi ψ , which gives reflection across zero.
Condition (31) is stricter than m / μ < e / k m/\mu<e/k m / μ < e / k , the condition under which participation raises the level of cost. Participation can therefore be costly even where additional information is locally beneficial. The institution’s direct forecasting gain can outweigh the additional supply response before inventory provision becomes beneficial in levels.
When information remains beneficial
The information ranking also has a favorable benchmark that does not fix participation. The alternative representation
ψ ( μ ) = − ( λ − H ) μ m ( μ ) − { d μ − b m ( μ ) } 2 2 A (32) \psi(\mu) = -(\lambda- H)\mu m(\mu) - \frac{\{d\mu- bm(\mu)\}^{2}}{2A} \tag*{(32)} ψ ( μ ) = − ( λ − H ) μ m ( μ ) − 2 A { d μ − bm ( μ ) } 2 (32)
makes that benchmark transparent.
Proposition 6 (Centered targets and Gaussian precision). For every cost law in Assumption 1, ψ \psi ψ is even, satisfies ψ ( 0 ) = 0 \psi(0) = 0 ψ ( 0 ) = 0 , and is strictly decreasing on [ 0 , ∞ ) [0,\infty) [ 0 , ∞ ) . Hence, for any square-integrable target with E D = 0 \mathbb{E}D = 0 E D = 0 , a public signal lowers institutional cost relative to no information, strictly if its posterior mean is nonzero with positive probability. If D ∼ N ( 0 , σ D 2 ) D \sim N(0,\sigma_D^2) D ∼ N ( 0 , σ D 2 ) with σ D > 0 \sigma_D > 0 σ D > 0 , every increase in the precision of an independent Gaussian signal strictly lowers cost, including with optional participation.
Proof. For μ ≥ 0 \mu\ge0 μ ≥ 0 , parameterize the equilibrium by q ≥ 0 q \ge0 q ≥ 0 . Both q q q and m = q F ( η q 2 / 2 ) m = qF(\eta q^2/2) m = q F ( η q 2 /2 ) are nondecreasing in μ \mu μ , and q q q is strictly increasing because η q + c m = s μ \eta q + cm = s\mu η q + c m = s μ . For μ 2 > μ 1 ≥ 0 \mu_2 > \mu_1 \ge0 μ 2 > μ 1 ≥ 0 , this identity gives
0 ≤ m ( μ 2 ) − m ( μ 1 ) ≤ s c ( μ 2 − μ 1 ) , d − b s c = A ( λ − H ) η 0 + b > 0. 0 \le m(\mu_2) - m(\mu_1) \le\frac{s}{c}(\mu_2-\mu_1), \qquad d-\frac{bs}{c}=\frac{A(\lambda-H)}{\eta_0+b}>0. 0 ≤ m ( μ 2 ) − m ( μ 1 ) ≤ c s ( μ 2 − μ 1 ) , d − c b s = η 0 + b A ( λ − H ) > 0.
Thus d μ − b m d\mu-bm d μ − bm is positive and strictly increasing for μ > 0 \mu>0 μ > 0 , while μ m \mu m μ m is nonnegative and nondecreasing. Equation (32) proves the first assertion, without differentiability of F F F . Relative to a zero posterior mean, every nonzero posterior mean strictly reduces ψ \psi ψ . For a centered Gaussian prior, the posterior mean has distribution σ D r Z \sigma_D\sqrt{r}Z σ D r Z , where Z ∼ N ( 0 , 1 ) Z\sim N(0,1) Z ∼ N ( 0 , 1 ) and r ∈ [ 0 , 1 ] r\in[0,1] r ∈ [ 0 , 1 ] is the fraction of target variance explained. Coupling these distributions with the same ∣ Z ∣ |Z| ∣ Z ∣ gives strict decrease as r r r increases. The quadratic moment bound justifies taking expectations.
The no-information benchmark differs from a refinement after the direction is known. Let S , B S,B S , B be independent symmetric signs, fix 0 < ϵ < δ < τ 0<\epsilon<\delta<\tau 0 < ϵ < δ < τ , and put D = S ( τ + δ B ) D=S(\tau+\delta B) D = S ( τ + δ B ) . This target is centered. Coarse information reveals S S S . A symmetric channel independent of S S S reports T ∈ { − 1 , 1 } T\in\{-1,1\} T ∈ { − 1 , 1 } with Pr ( T = B ) = ( 1 + ϵ / δ ) / 2 \Pr(T=B)=(1+\epsilon/\delta)/2 Pr ( T = B ) = ( 1 + ϵ / δ ) /2 . Fine information reveals ( S , T ) (S,T) ( S , T ) , giving
M 0 = S τ , M 1 = S ( τ + ϵ T ) , Δ C = ψ ( τ − ϵ ) + ψ ( τ + ϵ ) 2 − ψ ( τ ) . M_0=S\tau,\qquad M_1=S(\tau+\epsilon T),\qquad\Delta C=\frac{\psi(\tau-\epsilon)+\psi(\tau+\epsilon)}{2}-\psi(\tau). M 0 = S τ , M 1 = S ( τ + ϵ T ) , Δ C = 2 ψ ( τ − ϵ ) + ψ ( τ + ϵ ) − ψ ( τ ) .
Evenness of ψ \psi ψ gives the last equality. Choosing τ \tau τ in an adverse region and sufficiently small ϵ \epsilon ϵ makes this refinement harmful, although learning S S S initially is beneficial. This four-point prior is distinct from the centered Gaussian precision family. A signal that reveals magnitude while keeping the posterior mean zero gives equality under risk neutrality. Appendix § 5 .§ 2.2 gives a complementary global example: with no pre-event impact and exponential activation costs, ψ \psi ψ is strictly concave and every nontrivial posterior-mean refinement is beneficial.
Harmful information in a smooth interior
Derivative jumps are sufficient for harmful information but are not essential. Equation (21) also permits smooth supply curvature to outweigh the institution’s direct forecasting benefit. The following construction uses accelerating supply with positive marginal incidence.
Proposition 7 (Smooth-interior harm). Suppose the cost density is continuously differentiable near the equilibrium cutoffs on a neighborhood of μ ∗ > 0 \mu_*>0 μ ∗ > 0 , and ψ ′ ′ ( μ ∗ ) > 0 \psi''(\mu_*)>0 ψ ′′ ( μ ∗ ) > 0 . There exist r ∗ > 0 r_*>0 r ∗ > 0 and an open interval I I I containing μ ∗ \mu_* μ ∗ such that ψ ′ ′ ≥ r ∗ \psi''\ge r_* ψ ′′ ≥ r ∗ on I I I . Let X X X have a continuously differentiable density with compact support inside I I I , let B B B be an independent symmetric sign, and fix D = X + δ B D=X+\delta B D = X + δ B , δ > 0 \delta>0 δ > 0 . Coarse information reveals X X X . A noisy observation of B B B with accuracy ( 1 + ϵ / δ ) / 2 (1+\epsilon/\delta)/2 ( 1 + ϵ / δ ) /2 , where 0 < ϵ < δ 0<\epsilon<\delta 0 < ϵ < δ and supp ( X ) + [ − ϵ , ϵ ] ⊂ I \operatorname{supp}(X)+[-\epsilon,\epsilon]\subset I supp ( X ) + [ − ϵ , ϵ ] ⊂ I , changes cost by
Δ C = E [ ψ ( X + ϵ ) + ψ ( X − ϵ ) 2 − ψ ( X ) ] ≥ r ∗ 2 ϵ 2 > 0. (33) \Delta C=\mathbb{E}\left[\frac{\psi(X+\epsilon)+\psi(X-\epsilon)}{2}-\psi(X)\right]\ge\frac{r_*}{2}\epsilon^2>0. \tag*{(33)} Δ C = E [ 2 ψ ( X + ϵ ) + ψ ( X − ϵ ) − ψ ( X ) ] ≥ 2 r ∗ ϵ 2 > 0. (33) Such an economy exists with a smooth positive density on all positive costs and participation a = 1 / 4 a=1/4 a = 1/4 . The strict inequality persists under sufficiently small admissible perturbations of the execution parameters, for a fixed experiment contained in a smaller interval.
Proof. The strictly increasing map q ↦ q [ η + c F ( η q 2 / 2 ) ] / s q\mapsto q[\eta+cF(\eta q^2/2)]/s q ↦ q [ η + c F ( η q 2 /2 )] / s has positive derivative. Its inverse and (20) therefore give continuous ψ ′ ′ \psi'' ψ ′′ locally. Choose I , r ∗ I,r_* I , r ∗ by continuity. Bayes’ rule gives the two refined posterior means X ± ϵ X\pm\epsilon X ± ϵ . Integrating the bound on ψ ′ ′ \psi'' ψ ′′ twice proves (33) . The unconditional target distribution is continuous, and all compared posterior means stay in the same smooth region.
For existence, use a uniform execution schedule and
( λ , κ , ρ , η 0 , η ) = ( 1 , 0.99 , 0 , 0.01 , 4 ) . (\lambda,\kappa,\rho,\eta_0,\eta)=(1,0.99,0,0.01,4). ( λ , κ , ρ , η 0 , η ) = ( 1 , 0.99 , 0 , 0.01 , 4 ) .
Let costs be lognormal with mean 0.021 0.021 0.021 and variance 0.002 2 / 12 0.002^2/12 0.00 2 2 /12 , so
σ f 2 = log ( 1 + 1 / 1323 ) , ℓ f = log ( 0.021 ) − σ f 2 / 2. \sigma_f^2=\log(1+1/1323),\qquad\ell_f=\log(0.021)-\sigma_f^2/2. σ f 2 = log ( 1 + 1/1323 ) , ℓ f = log ( 0.021 ) − σ f 2 /2.
Put z ∗ = Φ − 1 ( 1 / 4 ) z_*=\Phi^{-1}(1/4) z ∗ = Φ − 1 ( 1/4 ) , g ∗ = exp ( ℓ f + σ f z ∗ ) g_*=\exp(\ell_f+\sigma_f z_*) g ∗ = exp ( ℓ f + σ f z ∗ ) , q ∗ = 2 g ∗ / η q_*=\sqrt{2g_*/\eta} q ∗ = 2 g ∗ / η , and μ ∗ = q ∗ ( η + c / 4 ) / s \mu_*=q_*(\eta+c/4)/s μ ∗ = q ∗ ( η + c /4 ) / s . Then a = 1 / 4 a=1/4 a = 1/4 exactly. Substitution in (20) and (18) , certified in Appendix § 5 , gives
μ ∗ ≈ 0.82506008 , 4.4003 < ψ ′ ′ ( μ ∗ ) < 4.4005. \mu_*\approx0.82506008,\qquad4.4003<\psi''(\mu_*)<4.4005. μ ∗ ≈ 0.82506008 , 4.4003 < ψ ′′ ( μ ∗ ) < 4.4005.
There is neither a positive cost floor nor a finite full-participation threshold. The same local formulas are continuous in admissible execution parameters, so the strict curvature bound survives in a smaller fixed neighborhood, including for sufficiently small positive decay.
The construction combines persistent price displacement, a low institutional temporary charge, a higher supplier charge, and concentrated activation costs. These assumptions produce a rapidly changing participation response. They define a model scenario rather than estimates from ETF prices. The matched gamma law also has positive curvature at one-quarter participation, whereas the lognormal example has negative curvature at median participation. The result identifies an adverse region. It does not give an adverse ranking for every smooth cost law or every forecast.
Distributional robustness and dispersed forecasts
Continuous cost laws and robustness of a fixed experiment
Proposition 8 (Stability in the entry-cost distribution). Let F , G F,G F , G be continuous cost distributions satisfying Assumption 1, with the execution coefficients unchanged, and let Δ F , Δ G \Delta_F,\Delta_G Δ F , Δ G be the cost changes under the same pair of public experiments. With d F G = ∥ F − G ∥ ∞ d_{FG}=\lVert F-G\rVert_{\infty} d F G = ∥ F − G ∥ ∞ ,
∣ a F − a G ∣ ≤ d F G , ∣ m F − m G ∣ ≤ s ∣ μ ∣ η d F G . (34) \lvert a_F-a_G\rvert\le d_{FG}, \qquad\lvert m_F-m_G\rvert\le\frac{s\lvert\mu\rvert}{\eta}d_{FG}. \tag*{(34)} ∣ a F − a G ∣ ≤ d F G , ∣ m F − m G ∣ ≤ η s ∣ μ ∣ d F G . (34) ∣ ψ F ( μ ) − ψ G ( μ ) ∣ ≤ C ∗ μ 2 d F G , C ∗ = s η ( ∣ e ∣ + 2 k s η + c ) . (35) \lvert\psi_F(\mu)-\psi_G(\mu)\rvert\le C_*\mu^2d_{FG}, \qquad C_*=\frac{s}{\eta}\left(\lvert e\rvert+\frac{2ks}{\eta+c}\right). \tag*{(35)} ∣ ψ F ( μ ) − ψ G ( μ )∣ ≤ C ∗ μ 2 d F G , C ∗ = η s ( ∣ e ∣ + η + c 2 k s ) . (35) ∣ Δ F − Δ G ∣ ≤ 2 C ∗ E [ D 2 ] d F G . (36) \lvert\Delta_F-\Delta_G\rvert\le2C_*\mathbb{E}[D^2]d_{FG}. \tag*{(36)} ∣ Δ F − Δ G ∣ ≤ 2 C ∗ E [ D 2 ] d F G . (36) Thus a fixed strictly harmful experiment remains harmful whenever the last bound is smaller than its original cost gap.
Proof. The root map h F ( a ) = a − F ( η s 2 μ 2 2 ( η + c a ) 2 ) h_F(a)=a-F\left(\frac{\eta s^2\mu^2}{2(\eta+ca)^2}\right) h F ( a ) = a − F ( 2 ( η + c a ) 2 η s 2 μ 2 ) increases by at least the change in a a a . At the root for G G G , its absolute value is at most d F G d_{FG} d F G , proving the first bound. The function a s μ / ( η + c a ) as\mu/(\eta+ca) a s μ / ( η + c a ) has derivative bounded in absolute value by s ∣ μ ∣ / η s\lvert\mu\rvert/\eta s ∣ μ ∣ / η , and ∣ m ∣ ≤ s ∣ μ ∣ / ( η + c ) \lvert m\rvert\le s\lvert\mu\rvert/(\eta+c) ∣ m ∣ ≤ s ∣ μ ∣ / ( η + c ) . Factoring the difference of the inventory squares in (16) proves (35) . Applying that bound to both experiments and using E [ E [ D ∣ H ] 2 ] ≤ E [ D 2 ] \mathbb{E}[\mathbb{E}[D\mid\mathcal{H}]^2]\le\mathbb{E}[D^2] E [ E [ D ∣ H ] 2 ] ≤ E [ D 2 ] proves (36) .
Continuity also preserves a strictly positive finite information gap under sufficiently small changes in execution parameters, provided the experiment and its prior remain fixed. One can instead recenter the prior at each new threshold, but that comparison uses a family of different experiments. The numerical results identify those cases explicitly. Neither argument gives a perturbation neighborhood that remains uniform as the information displacement tends to zero.
How quickly entry begins
A positive endpoint density is sufficient for the onset mechanism but is not necessary. Suppose
F ( f 0 + x ) = C x α ( 1 + o ( 1 ) ) , C , α , f 0 > 0. (37) F(f_0+x)=Cx^\alpha(1+o(1)), \qquad C,\alpha,f_0>0. \tag*{(37)} F ( f 0 + x ) = C x α ( 1 + o ( 1 )) , C , α , f 0 > 0. (37)
Proposition 9 (Onset with a power-law cost tail). Let q 0 = 2 f 0 / η q_0=\sqrt{2f_0/\eta} q 0 = 2 f 0 / η and μ 0 = η q 0 / s \mu_0=\eta q_0/s μ 0 = η q 0 / s . As h ↓ 0 h\downarrow0 h ↓ 0 ,
m ( μ 0 + h ) ∼ { ( s / c ) h , 0 < α < 1 , 2 s f 0 C η + 2 c f 0 C h , α = 1 , M α h α , α > 1 , M α = q 0 C ( s q 0 ) α . (38) m(\mu_0+h)\sim
\begin{cases}
(s/c)h, & 0<\alpha<1,\\
\frac{2sf_0C}{\eta+2cf_0C}h, & \alpha=1,\\
M_\alpha h^\alpha, & \alpha>1,
\end{cases}
\qquad M_\alpha=q_0C(sq_0)^\alpha. \tag*{(38)} m ( μ 0 + h ) ∼ ⎩ ⎨ ⎧ ( s / c ) h , η + 2 c f 0 C 2 s f 0 C h , M α h α , 0 < α < 1 , α = 1 , α > 1 , M α = q 0 C ( s q 0 ) α . (38) For e > 0 e>0 e > 0 , a threshold-centered binary refinement is harmful for all sufficiently small displacements when 0 < α < 2 0<\alpha<2 0 < α < 2 , and beneficial when α > 2 \alpha>2 α > 2 . At α = 2 \alpha=2 α = 2 its second-order coefficient is ( e μ 0 M 2 − d 2 / Λ ) / 2 (e\mu_0M_2-d^2/\Lambda)/2 ( e μ 0 M 2 − d 2 /Λ ) /2 , whose zero case requires higher-order information.
The proof in Appendix § 5.5 .§ 1.1 compares the gain above the cost floor with the inventory feedback in the equilibrium equation. A sufficiently flat density can remove the local adverse effect, which explains why smoothing need not preserve it at every information scale.
A diffuse distribution of coarse forecasts
To study information across a range of market states, take a random variable X X X with continuously differentiable compactly supported density p p p , an independent symmetric sign B B B , and the fixed target D = X + δ B D = X + \delta B D = X + δ B . The coarse signal reveals X X X . A further noisy observation of B B B with accuracy ( 1 + ϵ / δ ) / 2 (1 + \epsilon/\delta)/2 ( 1 + ϵ / δ ) /2 gives refined posterior means X ± ϵ X \pm\epsilon X ± ϵ . This construction keeps the target law unchanged and implements a feasible refinement for every 0 < ϵ < δ 0 < \epsilon< \delta 0 < ϵ < δ .
Theorem 10 (Diffuse-forecast information effect). Suppose ψ \psi ψ is continuous and piecewise C 2 C^{2} C 2 on a neighborhood of the support of p p p , with finitely many derivative jumps J τ J_{\tau} J τ and bounded piecewise second derivatives there. Then
lim ϵ ↓ 0 Δ C ϵ 2 = 1 2 { ∫ p ( x ) ψ r e g ′ ′ ( x ) d x + ∑ τ p ( τ ) J τ } . (39) \lim_{\epsilon\downarrow0}\frac{\Delta C}{\epsilon^{2}}=\frac{1}{2}\left\{\int p(x)\psi_{\mathrm{reg}}^{\prime\prime}(x)\,\mathrm{d}x+\sum_{\tau}p(\tau)J_{\tau}\right\}. \tag*{(39)} ϵ ↓ 0 lim ϵ 2 Δ C = 2 1 { ∫ p ( x ) ψ reg ′′ ( x ) d x + τ ∑ p ( τ ) J τ } . (39) In particular, let p w ( x ) = w − 1 p 0 ( ( x − τ ) / w ) p_{w}(x)=w^{-1}p_{0}((x-\tau)/w) p w ( x ) = w − 1 p 0 (( x − τ ) / w ) , where p 0 p_{0} p 0 is a continuously differentiable density supported on [ − 1 , 1 ] [-1,1] [ − 1 , 1 ] and p 0 ( 0 ) > 0 p_{0}(0)>0 p 0 ( 0 ) > 0 . If J τ > 0 J_{\tau}>0 J τ > 0 , the support includes no other boundary, and ∣ ψ r e g ′ ′ ∣ ≤ M |\psi_{\mathrm{reg}}^{\prime\prime}|\le M ∣ ψ reg ′′ ∣ ≤ M there, then the limit is positive whenever J τ p 0 ( 0 ) / w > M J_{\tau}p_{0}(0)/w>M J τ p 0 ( 0 ) / w > M . If the support lies in a smooth region where ψ ′ ′ ≤ − r < 0 \psi^{\prime\prime}\le-r<0 ψ ′′ ≤ − r < 0 , the limit is at most − r / 2 -r/2 − r /2 .
Proof. The distributional second derivative is the locally finite signed measure ν = ψ r e g ′ ′ ( x ) d x + ∑ τ J τ δ τ \nu=\psi_{\mathrm{reg}}^{\prime\prime}(x)\,\mathrm{d}x+\sum_{\tau}J_{\tau}\delta_{\tau} ν = ψ reg ′′ ( x ) d x + ∑ τ J τ δ τ . Twice integrating the derivative gives
ψ ( x + ϵ ) + ψ ( x − ϵ ) − 2 ψ ( x ) = ∫ ( ϵ − ∣ x − y ∣ ) + ν ( d y ) . \psi(x+\epsilon)+\psi(x-\epsilon)-2\psi(x)=\int(\epsilon-|x-y|)_{+}\nu(\mathrm{d}y). ψ ( x + ϵ ) + ψ ( x − ϵ ) − 2 ψ ( x ) = ∫ ( ϵ − ∣ x − y ∣ ) + ν ( d y ) .
After multiplication by p ( x ) / ( 2 ϵ 2 ) p(x)/(2\epsilon^{2}) p ( x ) / ( 2 ϵ 2 ) and integration, Fubini applies on the fixed compact neighborhood because ν \nu ν has finite total variation there. The triangular kernels ϵ − 2 ( ϵ − ∣ ⋅ ∣ ) + \epsilon^{-2}(\epsilon-|\cdot|)_{+} ϵ − 2 ( ϵ − ∣ ⋅ ∣ ) + integrate to one, and their convolution with the continuous p p p converges uniformly to p p p . This proves (39) . The integral of the regular part lies between − M -M − M and M M M , while the sole atomic contribution is J τ p 0 ( 0 ) / w J_{\tau}p_{0}(0)/w J τ p 0 ( 0 ) / w , giving the sufficient condition. The final assertion follows by integrating ψ ′ ′ ≤ − r \psi^{\prime\prime}\le-r ψ ′′ ≤ − r against a probability density.
A diffuse coarse forecast places only order-ϵ \epsilon ϵ probability within order-ϵ \epsilon ϵ of a boundary, so its aggregate boundary effect is second order in displacement. An atom at a harmful boundary instead restores a first-order contribution proportional to its probability mass. The distinction concerns the distribution of coarse beliefs, and prevents interpreting the threshold-centered expansion as a universal elasticity with respect to information precision.
Exact finite populations and terminal alternatives
The symmetric Bayesian equilibrium at finite size
For N N N potential inventory allocations, trader i i i purchases q i / N q_i/N q i / N , pays temporary charge η q i 2 / ( 2 N ) \eta q_i^2/(2N) η q i 2 / ( 2 N ) and activation cost f i / N f_i/N f i / N , and liquidates the same position at the event. Costs are conditionally independent across traders and independent of ( D , Y ) (D,Y) ( D , Y ) . The population scaling preserves the total mass of potential allocations as N N N changes, with normalized profit equal to N N N times an individual’s cash profit. The institution chooses before it observes the realized entries, simultaneously with the traders, so its first-order condition holds their strategies fixed.
Theorem 11 (Exact symmetric finite-population equilibrium). Under Assumption 1, for each N ≥ 2 N \ge2 N ≥ 2 there is a unique symmetric Bayesian equilibrium in entry and normalized quantities, up to zero-measure indifferent types. Put
η N = η + 2 b N , c N = c − b N > 0. (40) \eta_N=\eta+\frac{2b}{N},\qquad c_N=c-\frac{b}{N}>0. \tag*{(40)} η N = η + N 2 b , c N = c − N b > 0. (40) Its entry probability, common entrant quantity, mean aggregate inventory, and institutional pretrade satisfy
a N = F ( η N s 2 μ 2 2 ( η N + c N a N ) 2 ) , q N = s μ η N + c N a N , (41) a_N=F\left(\frac{\eta_Ns^2\mu^2}{2(\eta_N+c_Na_N)^2}\right),\qquad q_N=\frac{s\mu}{\eta_N+c_Na_N}, \tag*{(41)} a N = F ( 2 ( η N + c N a N ) 2 η N s 2 μ 2 ) , q N = η N + c N a N s μ , (41) m ˉ N = a N q N , z N = d μ − b m ˉ N A . \bar{m}_N=a_Nq_N,\qquad z_N=\frac{d\mu-b\bar{m}_N}{A}. m ˉ N = a N q N , z N = A d μ − b m ˉ N . All equilibrium positions have finite second moments. The realized aggregate has conditional variance a N ( 1 − a N ) q N 2 / N a_N(1-a_N)q_N^2/N a N ( 1 − a N ) q N 2 / N , and expected institutional cost is
C N ( H ) = λ E [ D 2 ] + E [ ψ N ( E [ D ∣ H ] ) ] , ψ N ( μ ) = − d 2 μ 2 2 A + e μ m ˉ N − k m ˉ N 2 . (42) C_N(\mathcal{H})=\lambda\mathbb{E}[D^2]+\mathbb{E}[\psi_N(\mathbb{E}[D\mid\mathcal{H}])],\qquad\psi_N(\mu)=-\frac{d^2\mu^2}{2A}+e\mu\bar{m}_N-k\bar{m}_N^2. \tag*{(42)} C N ( H ) = λ E [ D 2 ] + E [ ψ N ( E [ D ∣ H ])] , ψ N ( μ ) = − 2 A d 2 μ 2 + e μ m ˉ N − k m ˉ N 2 . (42) Proof. Against symmetric rival entry and quantity ( a , q ) (a,q) ( a , q ) , an entering trader choosing normalized quantity u u u obtains conditional expected normalized gross profit
u { λ μ − b z − b ( 1 − 1 / N ) a q } − η N 2 u 2 . u\{\lambda\mu-bz-b(1-1/N)aq\}-\frac{\eta_N}{2}u^2. u { λ μ − b z − b ( 1 − 1/ N ) a q } − 2 η N u 2 . Strict concavity makes the entered quantity independent of private activation cost, and the optimal gross profit is η N q 2 / 2 \eta_Nq^2/2 η N q 2 /2 . The institution’s objective is affine in random aggregate inventory, which is independent of the target innovation conditional on Y Y Y , so its expected objective for every possible deviation is (7) evaluated at m = a q m=aq m = a q . Its unique best response is therefore z = ( d μ − b a q ) / A z=(d\mu-baq)/A z = ( d μ − ba q ) / A . Substituting this into the minor’s first-order condition yields η N q = s μ − c N a q \eta_Nq=s\mu-c_Naq η N q = s μ − c N a q . Since c − b / 2 = b η 0 / ( 2 A ) > 0 c-b/2=b\eta_0/(2A)>0 c − b /2 = b η 0 / ( 2 A ) > 0 , c N > 0 c_N>0 c N > 0 for all N ≥ 2 N\ge2 N ≥ 2 , and the monotone-root proof of Theorem 1 applies to (41). This proves existence and uniqueness within the symmetric Bayesian class and verifies global deviations in both quantity and entry. The bounds ∣ q N ∣ ≤ s ∣ μ ∣ / η N ≤ s ∣ μ ∣ / η |q_N|\le s|\mu|/\eta_N\le s|\mu|/\eta ∣ q N ∣ ≤ s ∣ μ ∣/ η N ≤ s ∣ μ ∣/ η give integrability. Independent Bernoulli entry gives the variance formula, and completing the institution’s square after taking the conditional expectation gives (42).
The square in (42) is the square of expected inventory. The institution chooses before learning which traders enter. Using the expected square of realized inventory would give it additional information and describe a different game. The theorem establishes uniqueness within the symmetric Bayesian class, without claiming uniqueness of asymmetric equilibria. Its monotonicity argument uses N ≥ 2 N\ge2 N ≥ 2 .
Corollary 12 (A nonuniform information limit). Assume a positive cost floor with F ( f 0 + x ) = ℓ 0 x + O ( x 2 ) F(f_0+x)=\ell_0x+O(x^2) F ( f 0 + x ) = ℓ 0 x + O ( x 2 ) and e > 0 e>0 e > 0 . The finite entry boundary and right inventory slope are
μ 0 , N = 2 η N f 0 s > μ 0 , L N = 2 s f 0 ℓ 0 η N + 2 c N f 0 ℓ 0 . (43) \mu_{0,N}=\frac{\sqrt{2\eta_Nf_0}}{s}>\mu_0,\qquad L_N=\frac{2sf_0\ell_0}{\eta_N+2c_Nf_0\ell_0}. \tag*{(43)} μ 0 , N = s 2 η N f 0 > μ 0 , L N = η N + 2 c N f 0 ℓ 0 2 s f 0 ℓ 0 . (43)
A binary refinement centered at μ 0 , N \mu_{0,N} μ 0 , N changes cost by e μ 0 , N L N ϵ / 2 + O ( ϵ 2 ) > 0 e\mu_{0,N}L_N\epsilon/2 + O(\epsilon^2) > 0 e μ 0 , N L N ϵ /2 + O ( ϵ 2 ) > 0 locally. At the unchanged continuum center μ 0 \mu_0 μ 0 , however, every fixed N N N has
Δ N ( ϵ ) = − d 2 2 A ϵ 2 < 0 if 0 < ϵ < min { μ 0 , μ 0 , N − μ 0 } . (44) \Delta_N(\epsilon) = -\frac{d^2}{2A}\epsilon^2 < 0 \quad\text{if } 0 < \epsilon< \min\{\mu_0,\mu_{0,N}-\mu_0\}. \tag*{(44)} Δ N ( ϵ ) = − 2 A d 2 ϵ 2 < 0 if 0 < ϵ < min { μ 0 , μ 0 , N − μ 0 } . (44)
For each fixed finite displacement the cost comparison converges to the continuum comparison as N → ∞ N \to\infty N → ∞ , but
lim N → ∞ lim ϵ ↓ 0 Δ N ( ϵ ) ϵ = 0 , lim ϵ ↓ 0 lim N → ∞ Δ N ( ϵ ) ϵ = e μ 0 L 0 2 > 0. (45) \lim_{N\to\infty}\lim_{\epsilon\downarrow0}\frac{\Delta_N(\epsilon)}{\epsilon}=0,\qquad\lim_{\epsilon\downarrow0}\lim_{N\to\infty}\frac{\Delta_N(\epsilon)}{\epsilon}=\frac{e\mu_0L_0}{2}>0. \tag*{(45)} N → ∞ lim ϵ ↓ 0 lim ϵ Δ N ( ϵ ) = 0 , ϵ ↓ 0 lim N → ∞ lim ϵ Δ N ( ϵ ) = 2 e μ 0 L 0 > 0. (45)
Proof. Apply the endpoint expansion with ( η N , c N ) (\eta_N,c_N) ( η N , c N ) in place of ( η , c ) (\eta,c) ( η , c ) . Below μ 0 , N \mu_{0,N} μ 0 , N neither finite branch enters, and (42) reduces to the no-entry quadratic, proving (44) . For fixed μ \mu μ , any subsequential limit of the roots in the compact interval [ 0 , 1 ] [0,1] [ 0 , 1 ] solves the continuum equation by continuity. Uniqueness gives convergence of the whole sequence, and the quantities and costs then converge. The two iterated limits follow from (44) and Theorem 4 . □
Individual price impact raises the forecast threshold at which entry first becomes profitable. Near that threshold, a small shift can reverse the information comparison. An approximate equilibrium payoff bound cannot settle the sign of an information effect smaller than the bound. The exact equilibrium gives that comparison directly, without imposing a uniform position bound.
Corollary 13 (Joint information and population scale). Under Corollary 12, define
ξ 0 = b μ 0 η > 0 , γ 0 = e μ 0 L 0 2 > 0. \xi_0=\frac{b\mu_0}{\eta}>0,\qquad\gamma_0=\frac{e\mu_0L_0}{2}>0. ξ 0 = η b μ 0 > 0 , γ 0 = 2 e μ 0 L 0 > 0. Then μ 0 , N = μ 0 + ξ 0 / N + O ( N − 2 ) \mu_{0,N}=\mu_0+\xi_0/N+O(N^{-2}) μ 0 , N = μ 0 + ξ 0 / N + O ( N − 2 ) . For the unchanged binary prior centered at μ 0 \mu_0 μ 0 and information displacement ϵ N = x / N \epsilon_N=x/N ϵ N = x / N , with fixed x > 0 x>0 x > 0 ,
lim N → ∞ N Δ N ( x / N ) = γ 0 ( x − ξ 0 ) + . (46) \lim_{N\to\infty}N\Delta_N(x/N)=\gamma_0(x-\xi_0)+. \tag*{(46)} N → ∞ lim N Δ N ( x / N ) = γ 0 ( x − ξ 0 ) + . (46) If x < ξ 0 x<\xi_0 x < ξ 0 , the change is exactly − d 2 x 2 / ( 2 A N 2 ) -d^2x^2/(2AN^2) − d 2 x 2 / ( 2 A N 2 ) for all sufficiently large N N N . If x > ξ 0 x>\xi_0 x > ξ 0 , it is positive of order N − 1 N^{-1} N − 1 . At equality the first-order limit is zero and does not determine the sign.
Proof. Expand μ 0 , N = μ 0 1 + 2 b / ( η N ) \mu_{0,N}=\mu_0\sqrt{1+2b/(\eta N)} μ 0 , N = μ 0 1 + 2 b / ( η N ) . The fixed endpoint expansion of F F F and coefficients converging to strictly positive limits give, uniformly for small h ≥ 0 h\geq0 h ≥ 0 , m ˉ N ( μ 0 , N + h ) = L N h + O ( h 2 ) \bar m_N(\mu_{0,N}+h)=L_Nh+O(h^2) m ˉ N ( μ 0 , N + h ) = L N h + O ( h 2 ) , with L N → L 0 L_N\to L_0 L N → L 0 . For large N N N , the center and lower branch are inactive. The upper branch consequently has inventory L N ( x − ξ 0 ) + / N + O ( N − 2 ) L_N(x-\xi_0)+/N+O(N^{-2}) L N ( x − ξ 0 ) + / N + O ( N − 2 ) , including an O ( N − 2 ) O(N^{-2}) O ( N − 2 ) bound at equality. Inserting this into (42) proves (46) . When x < ξ 0 x<\xi_0 x < ξ 0 , both branches are below the finite threshold eventually, giving the exact quadratic expression. □
The relevant resolution near onset is N ϵ N\epsilon N ϵ . This limit explains the interaction of information and strategic size, but it is not an exact moderate-N N N sign rule. Under the principal numerical parameters, ξ 0 ≈ 0.953145 \xi_0\approx0.953145 ξ 0 ≈ 0.953145 and γ 0 ≈ 0.0333783 \gamma_0\approx0.0333783 γ 0 ≈ 0.0333783 . At x = 1 x=1 x = 1 , N Δ N ( 1 / N ) N\Delta_N(1/N) N Δ N ( 1/ N ) is negative at N = 100 N=100 N = 100 and positive at N = 1,000 N=1{,}000 N = 1 , 000 , approaching 0.00156393 0.00156393 0.00156393 .
A signed outside completion facility
Allow the institution, after observing D D D , to submit any signed auction order j ∈ R j \in\mathbb{R} j ∈ R and complete D − z − j D-z-j D − z − j through an outside provider charging χ ( D − z − j ) 2 / 2 \chi(D-z-j)^2/2 χ ( D − z − j ) 2 /2 , where χ > 0 \chi>0 χ > 0 . The provider’s price anchor is the unaffected benchmark, independent of inherited auction displacement. Anticipators still liquidate in the auction at H ( z + m ) + λ ( j − m ) H(z+m)+\lambda(j-m) H ( z + m ) + λ ( j − m ) . Orders may offset across the two venues, even when the total residual is zero. The admissible set thus permits unrestricted signed routing, rather than requiring each venue to receive the same-direction share of a residual order. The unique terminal choice is
j ∗ ( D , z , m ) = χ ( D − z ) − H ( z + m ) + λ m χ + 2 λ . (47) j^*(D,z,m)=\frac{\chi(D-z)-H(z+m)+\lambda m}{\chi+2\lambda}. \tag*{(47)} j ∗ ( D , z , m ) = χ + 2 λ χ ( D − z ) − H ( z + m ) + λm . (47)
Appendix E derives coefficients ( A χ , d χ , b χ , s χ , c χ , l χ ) (A_\chi,d_\chi,b_\chi,s_\chi,c_\chi,l_\chi) ( A χ , d χ , b χ , s χ , c χ , l χ ) for the resulting continuum game, whose equilibrium retains the scalar form whenever A χ > 0 A_\chi>0 A χ > 0 , s χ > 0 s_\chi>0 s χ > 0 , and c χ ≥ 0 c_\chi\geq0 c χ ≥ 0 .
Proposition 14 (Information incidence with outside completion). Let B χ = χ + 2 λ B_\chi=\chi+2\lambda B χ = χ + 2 λ and suppose the stated facility coefficient restrictions hold. Optimized institutional cost relative to no anticipators is e χ μ m − k χ m 2 e_\chi\mu m-k_\chi m^2 e χ μ m − k χ m 2 , where
e χ = d χ b χ A χ − χ ( λ − H ) B χ = χ A e B χ A χ , k χ = b χ 2 2 A χ + ( λ − H ) 2 2 B χ > 0. (48) e_\chi=\frac{d_\chi b_\chi}{A_\chi}-\frac{\chi(\lambda-H)}{B_\chi}=\frac{\chi A e}{B_\chi A_\chi},\qquad
k_\chi=\frac{b_\chi^2}{2A_\chi}+\frac{(\lambda-H)^2}{2B_\chi}>0. \tag*{(48)} e χ = A χ d χ b χ − B χ χ ( λ − H ) = B χ A χ χ A e , k χ = 2 A χ b χ 2 + 2 B χ ( λ − H ) 2 > 0. (48) Thus every admissible facility preserves the sign of initial inventory incidence. Fixed-roster information remains beneficial, and when e > 0 e>0 e > 0 the harmful-onset theorem holds at the facility’s own entry threshold.
Proof. After optimizing j j j , the constant target term is l χ E [ D 2 ] l_\chi\mathbb{E}[D^2] l χ E [ D 2 ] . Completing the remaining square in z z z gives (48) , with the algebra detailed in Appendix E. At fixed mass, write m = t μ m=t\mu m = t μ with t = a s χ / ( η + c χ a ) ≥ 0 t=a s_\chi/(\eta+c_\chi a)\geq0 t = a s χ / ( η + c χ a ) ≥ 0 . The coefficient of E [ μ 2 ] \mathbb{E}[\mu^2] E [ μ 2 ] after extracting the constant target term is
− χ ( λ − H ) B χ t − ( λ − H ) 2 2 B χ t 2 − ( d χ − b χ t ) 2 2 A χ < 0. -\frac{\chi(\lambda-H)}{B_\chi}t-\frac{(\lambda-H)^2}{2B_\chi}t^2-\frac{(d_\chi-b_\chi t)^2}{2A_\chi}<0. − B χ χ ( λ − H ) t − 2 B χ ( λ − H ) 2 t 2 − 2 A χ ( d χ − b χ t ) 2 < 0.
The inequality is strict for t > 0 t>0 t > 0 from its first term and for t = 0 t=0 t = 0 from d χ > 0 d_\chi>0 d χ > 0 . The positive multiplier relating e χ e_\chi e χ to e e e then extends the onset proof.
Numerical implications of the characterization
Normalization and the two mechanisms
Choose a reference target quantity D ∗ D_* D ∗ and normalize cash by λ D ∗ 2 \lambda D_*^2 λ D ∗ 2 , inventory by D ∗ D_* D ∗ , and price-impact and temporary-cost coefficients by λ \lambda λ . The decay parameter is the physical decay rate multiplied by the pre-event horizon. All results below are constructed scenarios in these dimensionless units. The onset scenario uses
λ = κ = 1 , ρ = 0.1 , η 0 = η = 0.05 , F = Uniform [ 0.02 , 0.04 ] , ϕ t = 1. \lambda=\kappa=1,\qquad\rho=0.1,\qquad\eta_0=\eta=0.05,\qquad F=\operatorname{Uniform}[0.02,0.04],\qquad\phi_t=1. λ = κ = 1 , ρ = 0.1 , η 0 = η = 0.05 , F = Uniform [ 0.02 , 0.04 ] , ϕ t = 1.
It has H = 0.951626 H=0.951626 H = 0.951626 , K = 0.483742 K=0.483742 K = 0.483742 , e = 0.452291 e=0.452291 e = 0.452291 , and μ 0 = 0.089562 \mu_0=0.089562 μ 0 = 0.089562 . Its impact half-life is log ( 2 ) / 0.1 = 6.93 \log(2)/0.1=6.93 log ( 2 ) /0.1 = 6.93 pre-event horizons, making it a deliberately persistent environment. The saturation scenario instead uses the zero-pre-event-impact parameters in (30) , with threshold 2 \sqrt{2} 2 .
At ϵ = 0.01 \epsilon= 0.01 ϵ = 0.01 , the onset cost increase is 0.00030485 and the saturation increase is 0.00044338. At saturation, inventory is beneficial for every target mean. Cost nevertheless rises because the refinement changes participation asymmetrically. Both comparisons reoptimize institutional pretrade and retain the posterior-variance term through the constant-second-moment representation. The upper panels show an affine-detrended conditional-mean cost curve. The lower panels compare public experiments, which is a different calculation.
Table 2 distinguishes raw cost changes from their economic scale. The onset effect is about 5.38% of the coarse institutional bill, compared with 0.084% at full participation, despite the latter’s larger raw gap. These are percentages of constructed execution cost, not basis points of assets or estimates of investor losses.
Table 2. Information magnitudes at ϵ = 0.01 \epsilon= 0.01 ϵ = 0.01 . Coarse cost includes target variance under D = μ 0 ± 0.04 D = \mu_0 \pm0.04 D = μ 0 ± 0.04 and D = 2 ± 0.1 D = \sqrt{2} \pm0.1 D = 2 ± 0.1 , respectively. The percentage denominator is that scenario’s own coarse cost.
One prior, increasing precision
Figure 2 follows a Gaussian target D ∼ N ( μ ˉ , σ D 2 ) D \sim\mathcal{N}(\bar{\mu},\sigma_D^2) D ∼ N ( μ ˉ , σ D 2 ) as a public Gaussian signal explains an increasing fraction r r r of its variance. Each curve holds both prior parameters fixed. Formula (A.1) includes the same target second moment at every precision, and (A.2) relates the path’s slope to regular curvature and boundary contributions.
Figure 2. Gaussian precision under unchanged priors. Each curve retains its indicated mean and the panel’s target standard deviation. The boundary τ \tau τ is μ 0 \mu_0 μ 0 on the left and μ 1 \mu_1 μ 1 on the right. Costs are relative to each curve’s own no-information cost. Full revelation is r = 1 r = 1 r = 1 , and the cost scales differ across panels.
At the onset boundary, full revelation raises cost by 0.00060202. At the full-participation boundary, small precision gains initially raise cost, but revelation lowers it by 0.00370761. On the reported grid the latter path peaks at r = 0.0694 r = 0.0694 r = 0.0694 , with increase 0.00055083, or 0.10393% of its coarse bill. This is a grid maximum, not a proved optimizer. With mean zero instead, full-revelation changes are −0.00076215 and −0.00666667, respectively, and Proposition 6 proves that both centered paths decrease throughout. Local harm consequently does not by itself determine the value of complete information.
Strategic size and smooth heterogeneity
Figure 3 solves the exact symmetric finite game with the onset technology. With the prior held at the continuum center and ϵ = 0.01 \epsilon= 0.01 ϵ = 0.01 , the cost change is −0.00001503 at N = 100 N = 100 N = 100 , +0.00014455 at N = 200 N = 200 N = 200 , and +0.00030485 in the continuum. At N = 100 N = 100 N = 100 , the threshold is 0.098634, exceeding the continuum threshold by about 0.009072. Recentering the prior at that finite threshold gives +0.00033884 instead, illustrating why a threshold-recentered experiment must be distinguished from the unchanged-prior comparison. The numerical N N N counts potential inventory allocations and is not an estimate of the number of trading firms.
Figure 3. Finite strategic size changes the information comparison. The left panel gives the exact entry boundary. The right fixes ϵ = 0.01 \epsilon= 0.01 ϵ = 0.01 and binary target half-spread 0.04, comparing a prior centered at the continuum threshold with priors centered at each finite threshold. Each individual coarse/refined pair has the same target law, while the recentered family changes that law across N N N .
The full-support lognormal construction in Proposition 7 gives cost increases 0.0000022010 and 0.0002278461 at displacements 0.001 and 0.01 around its quarter-participation state. The target remains D = μ ∗ ± 0.02 D = \mu_* \pm0.02 D = μ ∗ ± 0.02 , and its coarse institutional cost is 0.35090830. The two increases are therefore 0.000627% and 0.064930% of that same execution bill. Existence of an adverse region need not imply a large finite effect. Appendix § 5 .§ 1.2 reports the denominators, execution-parameter sensitivity, and one equal-charge example with positive decay. Table IA.Table 1 reports both positive and negative curvature states. At the principal technology’s unchanged onset prior, replacing uniform costs by moment-matched gamma or lognormal laws leaves positive gaps of 0.00007863 and 0.00009884, while the analogous full-participation gaps become negative. Thus the two boundary examples have different distributional sensitivity. Appendix § 6 supplies these comparisons, multiplicative smoothing, and the diffuse-forecast display on an unclipped symmetric-log scale.
Table 1. Position relative to the closest literature. Reference names link to the cited articles or working-paper versions. The last column identifies the question or result added by the present paper.
Terminal technology and a consistent schedule comparison
The outside-facility scenarios in Table 3 retain the onset parameters and use the unrestricted signed routing of § 2.5.2 . The incidence coefficient stays positive throughout, as Proposition 14 requires. Nevertheless, a fixed finite refinement centered at the original continuum threshold becomes beneficial for a sufficiently inexpensive facility because that technology moves the entry boundary. The recentered comparisons remain positive in the reported cases, and the no-entrant column records the institution’s direct forecasting benefit within each facility environment.
Table 3. Outside completion and information, with ϵ = 0.01 \epsilon= 0.01 ϵ = 0.01 . “Fixed prior” keeps the original onset center, “Recentered” uses μ 0 , χ \mu_{0,\chi} μ 0 , χ , and “No entrants” gives the matched zero-participation information change. All binary priors have half-spread 0.04.
Schedule sensitivity uses early, uniform, and late profiles 2 ( 1 − t ) 2(1-t) 2 ( 1 − t ) , 1 1 1 , and 2 t 2t 2 t . We keep the underlying rate-cost coefficients fixed and recompute both K K K , H H H and the temporary charges, so the early and late schedules multiply each charge by ∫ ϕ 2 = 4 / 3 \int\phi^2 = 4/3 ∫ ϕ 2 = 4/3 . Appendix § 6.2 reports the full comparison at decay rates 0.1, 2, and 8. At slow decay, early and late schedules preserve positive initial incidence but make the unchanged-prior information comparison beneficial by moving the threshold. Faster decay reduces carryover, and the early schedule at ρ = 8 \rho= 8 ρ = 8 has e < 0 e < 0 e < 0 . Some finite recentered displacements are beneficial even with e > 0 e > 0 e > 0 , consistently with the onset theorem’s sufficiently-small-displacement qualification.
The direction of portfolio information
A supplier pays one activation cost for a basket and then chooses quantities across p ≥ 2 p \ge2 p ≥ 2 assets. The vector game in Appendix § 7 has an isotropic specialization with
Λ = λ I p , K = K s I p , H = H s I p , E 0 = η 0 I p , E = η I p , \Lambda= \lambda I_p,\quad K = K_s I_p,\quad H = H_s I_p,\quad E_0 = \eta_0 I_p,\quad E = \eta I_p, Λ = λ I p , K = K s I p , H = H s I p , E 0 = η 0 I p , E = η I p ,
where ( λ , K s , H s , η 0 , η , F ) (\lambda, K_s, H_s, \eta_0, \eta, F) ( λ , K s , H s , η 0 , η , F ) satisfy Assumption 1. Each entrant pays one private fixed cost for the basket. Equilibrium quantities are parallel to the posterior mean μ \mu μ , and the optimized mean-dependent cost is Ψ ( μ ) = ψ ( ∥ μ ∥ ) \Psi(\mu) = \psi(\lVert\mu\rVert) Ψ ( μ ) = ψ (∥ μ ∥) , using the scalar function in (16) . At a nonzero smooth state, put r = ∥ μ ∥ r = \lVert\mu\rVert r = ∥ μ ∥ and n = μ / r n = \mu/r n = μ / r . Then
∇ 2 Ψ ( μ ) = ψ ′ ′ ( r ) n n T + ψ ′ ( r ) r ( I p − n n T ) . (49) \nabla^2\Psi(\mu) = \psi^{\prime\prime}(r)nn^{\mathsf T} + \frac{\psi^\prime(r)}{r}(I_p - nn^{\mathsf T}). \tag*{(49)} ∇ 2 Ψ ( μ ) = ψ ′′ ( r ) n n T + r ψ ′ ( r ) ( I p − n n T ) . (49)
The radial direction changes basket magnitude. Orthogonal directions change composition to first order. The following feasible experiments turn that local distinction into a cash comparison under one prior.
Corollary 15 (Equal forecast accuracy, opposite execution values). Let n , v ∈ R p n, v \in\mathbb{R}^p n , v ∈ R p be orthonormal, and choose a smooth state r > 0 r > 0 r > 0 with ψ ′ ′ ( r ) > 0 \psi^{\prime\prime}(r) > 0 ψ ′′ ( r ) > 0 . Fix δ > 0 \delta> 0 δ > 0 and the target law
D = r n + δ ( B 1 n + B 2 v ) , (50) D = rn + \delta(B_1n + B_2v), \tag*{(50)} D = r n + δ ( B 1 n + B 2 v ) , (50) where B 1 , B 2 B_1, B_2 B 1 , B 2 are independent symmetric signs. For 0 < ϵ < min { δ , r } 0 < \epsilon< \min\{\delta,r\} 0 < ϵ < min { δ , r } , let U 1 , U 2 U_1, U_2 U 1 , U 2 be independent signs, independent of the target signs and activation costs, with Pr ( U j = 1 ) = ( 1 + ϵ / δ ) / 2 \Pr(U_j = 1) = (1 + \epsilon/\delta)/2 Pr ( U j = 1 ) = ( 1 + ϵ / δ ) /2 . Observing T 1 = B 1 U 1 T_1 = B_1U_1 T 1 = B 1 U 1 alone or T 2 = B 2 U 2 T_2 = B_2U_2 T 2 = B 2 U 2 alone gives posterior means r n ± ϵ n rn \pm\epsilon n r n ± ϵ n or r n ± ϵ v rn \pm\epsilon v r n ± ϵ v , respectively, each pair equiprobable. Both refine no information and reduce ordinary squared forecast error by ϵ 2 \epsilon^2 ϵ 2 and Λ \Lambda Λ -weighted squared error by λ ϵ 2 \lambda\epsilon^2 λ ϵ 2 . Their cost changes are
Δ C m a g n i t u d e = 1 2 { ψ ( r + ϵ ) + ψ ( r − ϵ ) } − ψ ( r ) = 1 2 ψ ′ ′ ( r ) ϵ 2 + o ( ϵ 2 ) , (51) \Delta C_{\mathrm{magnitude}} = \frac{1}{2}\{\psi(r+\epsilon) + \psi(r-\epsilon)\} - \psi(r) = \frac{1}{2}\psi^{\prime\prime}(r)\epsilon^2 + o(\epsilon^2), \tag*{(51)} Δ C magnitude = 2 1 { ψ ( r + ϵ ) + ψ ( r − ϵ )} − ψ ( r ) = 2 1 ψ ′′ ( r ) ϵ 2 + o ( ϵ 2 ) , (51) Δ C c o m p o s i t i o n = ψ ( r 2 + ϵ 2 ) − ψ ( r ) < 0. \Delta C_{\mathrm{composition}} = \psi\left(\sqrt{r^2+\epsilon^2}\right) - \psi(r) < 0. Δ C composition = ψ ( r 2 + ϵ 2 ) − ψ ( r ) < 0. The magnitude signal raises cost for sufficiently small ϵ > 0 \epsilon> 0 ϵ > 0 . The composition signal lowers cost for every 0 < ϵ < δ 0 < \epsilon< \delta 0 < ϵ < δ .
The proof is in Appendix § 7 .§ 2.3 . Both signals refine the same coarse experiment, although they need not be Blackwell ordered relative to each other. Table 4 evaluates the comparison at the certified smooth state of Proposition 7. The signals remove the same amount of statistical and execution-weighted forecast error. Their different cost effects arise from the direction in which they change the forecast.
Table 4. Two signals under one portfolio target law, with δ = .02 \delta= .02 δ = .02 and ϵ = .01 \epsilon= .01 ϵ = .01 . The isotropic economy has λ = 1 \lambda= 1 λ = 1 . The common coarse bill is .351308300279, including both target shocks. Costs are normalized model quantities.
For general cross-impact, every fixed-roster refinement is favorable across square-integrable target laws exactly when the symmetric cost matrix P a P_a P a in Proposition 1A.3 is negative semidefinite.
Simultaneous orthogonal diagonalization is sufficient, not necessary. Appendix § 7 .§ 2.2 verifies failure of the ranking outside that class from original cash payments.
Information in leveraged-fund target proxies
Sample and measurement
The five-minute study uses a common observation period from November 9, 2021 through November 13, 2026. SPY, QQQ, IWM, and DIA represent four broad-index families. The start is the latest first observation among the included benchmark and geared products, and applies to feature construction and training as well as evaluation. The common forecasting sample contains 717 normal sessions in 2023–2025 and 219 in 2026. SOXX enters the joint-market extension in Appendix § 10 .
Prices and activity use completed bars in New York time. The primary convention treats labels as interval starts. An end-stamped sensitivity shifts every relevant reference consistently. Price windows require valid prices and positive transaction counts. Volume-weighted windows also require positive volume. The five-minute volume field enters only relative measures and within-window weights. A separately sourced daily panel provides share volume, dollar turnover, and split- and distribution-adjusted returns for the appendix analyses. Appendix § 8.2 defines the sample and timing conventions.
Mean prediction and uncertainty are separate questions
At a fixed noon decision time, the response is the simple return from the last completed bar to the final regular-session reference. The primary convention uses the 11:55 and 15:55 labels. The return predictors include the already observed reset-period return, measured from the immediately preceding session’s actual final regular bar, and the overnight return. Range features describe morning variation and close location. Activity features use past-only same-clock normalization, with volume and count-plus-size as alternatives.
The feature ablation uses a fixed 252-observation rolling ridge regression, with penalty 0.1 after training-only standardization. The benchmarks are zero remaining return and the rolling historical mean. All models share 717 evaluation dates in 2023–2025 and 219 in 2026, with four families per date. The common reference sample starts on February 10, 2023 because activity normalization and regression each require a complete prior window. We fixed the specification before evaluating the longer sample. Earlier exploratory analyses informed the design, and some 2026 prices had already been examined. The 2026 comparison is therefore a retrospective temporal evaluation, not an untouched holdout.
Every fitted mean forecast in Table 5 has higher point RMSE than zero remaining return. The baseline RMSEs are 66.94 and 49.47 basis points in the two periods. Adding count and relative size to the range model changes MSE by 8.22 squared basis points in 2023–2025 and −0.37 in 2026. The corresponding date-block intervals are [ − 39.15 , 43.39 ] [-39.15, 43.39] [ − 39.15 , 43.39 ] and [ − 49.12 , 44.68 ] [-49.12, 44.68] [ − 49.12 , 44.68 ] . The alternative volume increments also have intervals containing zero. The end-label convention gives the same qualitative result: no fitted model improves point RMSE over zero. This finding applies to the declared models and does not establish that public information is uninformative. Appendix § 8.2 .§ 2 reports annual results and all paired incremental losses.
Table 5. Noon information: mean forecasts and uncertainty bands. Panel A reports raw pooled RMSE relative to zero remaining return. Panel B reports zero-centered interval coverage, full width, and the negatively oriented 95 percent interval score. The four-family samples are identical within each period.
Range information is more useful for uncertainty in this exercise. We compare symmetric zero-centered bands using raw, range-scaled, or close-return-RV-scaled absolute historical returns. Each band uses the 241st order statistic of 252 prior scores, whose iid continuous-score coverage benchmark is 241 / 253 241/253 241/253 , rather than a conditional guarantee for financial data. Range scaling reduces the average interval score by 42.59 basis points in 2023–2025 and 26.70 in 2026. The respective individual 95 percent five-session date-block intervals are [ − 101.07 , − 3.21 ] [-101.07, -3.21] [ − 101.07 , − 3.21 ] and [ − 52.61 , − 3.28 ] [-52.61, -3.28] [ − 52.61 , − 3.28 ] . The reductions remain negative with block lengths 1, 10, and 20. The direct comparison with RV scaling is smaller: range minus RV is −10.13 basis points, with interval [ − 15.22 , − 5.55 ] [-15.22, -5.55] [ − 15.22 , − 5.55 ] , in 2023–2025, and −0.66 with interval [ − 7.08 , 6.71 ] [-7.08, 6.71] [ − 7.08 , 6.71 ] in 2026. Thus current intraday scale helps characterize uncertainty, while the incremental advantage of the high-low range is supported only in the reference-period comparison.
The score gains are concentrated. The year 2025 accounts for 96.51% of the reference period’s net range-versus-raw improvement. The five largest absolute date contributions account for 65.80% of that net gain and 61.51% in 2026. All valid observations remain in the primary estimates. Appendix § 9.1 .§ 1.2 reports the annual contributions, selected-date sensitivities, and the distinction between narrower bands and smaller exceedance penalties.
Overall coverage conceals differences across market states. In 2026, the raw band covers 98.76% of low-range observations but 93.02% of high-range observations. Range scaling gives 95.29% and 96.62%, while reducing average full width from 222.79 to 214.97 basis points. The reference period shows a similar state imbalance. Table IA.15 reports the annual comparisons. Appendix § 9.2 adds HAC inference, explicit multiple-testing families, conditional calibration moments, and alternative training windows. The reference-period range-versus-RV score advantage survives the familywise adjustment. The range-versus-raw comparison is more sensitive to the inferential method. These tests do not establish coverage in every information state.
The connection to committed demand
Under the idealized no-flow balance sheet, the forecasted target per initial dollar is
d ^ t = β ( β − 1 ) { R 0 , t + ( 1 + R 0 , t ) r ^ t , T } . \hat{d}_t = \beta(\beta- 1)\{R_{0,t} + (1 + R_{0,t})\hat{r}_{t,T}\}. d ^ t = β ( β − 1 ) { R 0 , t + ( 1 + R 0 , t ) r ^ t , T } .
Here time zero is the previous reset reference. Zero remaining return retains the already observed component of the target, just as the directional construction in § 2.3.5 conditions on information already known.
The same accounting translates uncertainty into target units. Put a β = β ( β − 1 ) a_\beta= \beta(\beta- 1) a β = β ( β − 1 ) , c t = 10 4 a β R 0 , t c_t = 10^4a_\beta R_{0,t} c t = 1 0 4 a β R 0 , t , and k t = a β ( 1 + R 0 , t ) > 0 k_t = a_\beta(1 + R_{0,t}) > 0 k t = a β ( 1 + R 0 , t ) > 0 for the geared objectives considered here. If Y Y Y is the remaining return in basis points, then
D t b p = c t + k t Y , [ l t , u t ] ⟼ [ c t + k t l t , c t + k t u t ] . (52) D_t^{\mathrm{bp}} = c_t + k_tY,\qquad[l_t,u_t] \longmapsto[c_t + k_tl_t,c_t + k_tu_t]. \tag*{(52)} D t bp = c t + k t Y , [ l t , u t ] ⟼ [ c t + k t l t , c t + k t u t ] . (52)
Coverage is unchanged, while widths and interval scores multiply by k t k_t k t . We transform each observation before averaging. Table 6 reports mean half-widths as percentages of hypothetical initial assets. The larger inverse-fund coefficient explains the factor of two between −3 and +3. These labels use the same four benchmark paths and add no independent observations. Historical dollar orders would additionally require assets, flows, routing, and prior executions.
Table 6. Uncertainty in the normalized reset target. Entries are average interval half-widths in percent of initial assets, after the observation-level transformation in (52). The no-flow, no-fee balance sheet defines the target proxy.
The distinction between mean and scale is also central to interpreting the model. Holding the environment and posterior mean fixed, posterior variance changes conditional expected expenditure but leaves risk-neutral equilibrium positions unchanged. Better range-based uncertainty forecasts therefore do not establish additional activation. Range and activity may also reveal the environment Z Z Z in § 2.3.2 . Adding these variables to a forecast does not by itself implement a target-only Blackwell refinement with execution conditions held fixed.
We also examine close-related price movements with direction frozen at 15:30, fresh window prices, a buffered post-close interval, and a paired 14:00 placebo. Across common SPY/QQQ/IWM dates, the signed post move is −1.23 basis points in 2022–2025 but −0.12 in 2026. Both periods lack a clearly positive preceding close leg, and the pooled paired clock-difference intervals include zero. Appendix § 9 therefore presents these paths, alternative anchors, and predetermined activity interactions as descriptive evidence. They neither identify auction-price reversal nor test a post-event price equation of the game.
The appendix examines related-market information and daily factor exposures. Accounting for dependence improves joint uncertainty scores for QQQ/SPY/IWM and QQQ/SOXX relative to independent marginals. In both systems, however, a correlation estimate frozen before evaluation has better point scores than rolling correlation. The comparison supports modeling dependence, without showing a benefit from updating correlations. Joint mean forecasts also fail to improve on their own-market counterparts. The volume analysis separates daily turnover into constant-denomination share activity and a price-related component. Factor attribution separates broad return exposure from deviations against a financed benchmark position. Together, these results describe target uncertainty and the benchmark approximation. They do not identify structural cross-impact or historical reset orders.
Conclusion
A public target forecast changes both the institution’s preparation and the inventory that other traders choose to supply. The shape of that supply response determines whether the institution benefits. Information can attract inventory whose impact raises cost, withdraw helpful participation near full entry, or change smooth supply where its curvature and marginal cost effect reinforce each other. These mechanisms coexist with precise favorable comparisons for fixed rosters, centered targets, and exponential activation costs. The exact finite game adds a quantitative qualification: near entry, information resolution must be assessed relative to individual strategic size.
For portfolios, the direction of information matters as well as its amount. A matrix criterion governs the fixed-roster ranking. Commuting execution matrices preserve that ranking and the centered-target comparison, whereas general cross-impact can reverse both. With shared basket activation, signals that reduce ordinary and execution-weighted forecast error equally can have opposite effects on institutional cost, even under isotropic execution. Changes in participation and the direction of the inventory response thus provide distinct sources of adverse information value.
The conclusions apply to a fixed-schedule, risk-neutral quantity game with a linear terminal supply rule and a target external to affected prices. They concern institutional cash expenditure, not social welfare. The price application separates mean prediction from uncertainty calibration. Its small and variable close-related movements do not identify activation, causal impact, or institutional execution costs. A natural extension would allow traders to wait for another signal before activating inventory. This would link information value to the execution time remaining and require an additional strategic stage.
Internet Appendix
Trading Before a Committed Rebalance
Public Information and Optional Liquidity Supply
This appendix develops the supporting theory and empirical analyses. The first part contains supporting scalar mathematics and cash accounts. The portfolio extension follows, including information rankings and feasible directional comparisons. Empirical methods, joint forecasts, econometric robustness, turnover, and factor attribution form the final part. Returns and price changes are in basis points unless stated otherwise.
Smooth-interior certification and favorable information
A reproducible curvature certificate
For the lognormal construction in Proposition 7 , the reduced coefficients are
b = 101 200 , A = 51 50 , d = 101 100 , s = 10199 20400 , c = 10403 40800 , e = 9997 20400 , k = 10201 81600 . b = \frac{101}{200},\quad A = \frac{51}{50},\quad d = \frac{101}{100},\quad s = \frac{10199}{20400},\quad c = \frac{10403}{40800},\quad e = \frac{9997}{20400},\quad k = \frac{10201}{81600}. b = 200 101 , A = 50 51 , d = 100 101 , s = 20400 10199 , c = 40800 10403 , e = 20400 9997 , k = 81600 10201 .
With z = ( log g − ℓ f ) / σ f z = (\log g - \ell_f)/\sigma_f z = ( log g − ℓ f ) / σ f , the density and its derivative are
f F ( g ) = φ ( z ) σ f g , f F ′ ( g ) = − f F ( g ) g ( 1 + z σ f ) . f_F(g) = \frac{\varphi(z)}{\sigma_f g},\qquad f_F'(g) = -\frac{f_F(g)}{g}\left(1+\frac{z}{\sigma_f}\right). f F ( g ) = σ f g φ ( z ) , f F ′ ( g ) = − g f F ( g ) ( 1 + σ f z ) .
The quantile z ∗ = Φ − 1 ( 1 / 4 ) z_*= \Phi^{-1}(1/4) z ∗ = Φ − 1 ( 1/4 ) lies in [ − 0.674490 , − 0.674489 ] [-0.674490,-0.674489] [ − 0.674490 , − 0.674489 ] . To certify the bracket, for x = − z > 0 x=-z>0 x = − z > 0 evaluate
Φ ( − x ) = 1 2 − 1 2 π ∑ j = 0 12 ( − 1 ) j x 2 j + 1 2 j j ! ( 2 j + 1 ) + R ( x ) , ∣ R ( x ) ∣ ≤ x 27 2 π 2 13 13 ! 27 . \Phi(-x) = \frac{1}{2} - \frac{1}{\sqrt{2\pi}}\sum_{j=0}^{12}\frac{(-1)^j x^{2j+1}}{2^j j!(2j+1)} + R(x),\qquad|R(x)| \le\frac{x^{27}}{\sqrt{2\pi}2^{13}13!\,27}. Φ ( − x ) = 2 1 − 2 π 1 j = 0 ∑ 12 2 j j ! ( 2 j + 1 ) ( − 1 ) j x 2 j + 1 + R ( x ) , ∣ R ( x ) ∣ ≤ 2 π 2 13 13 ! 27 x 27 .
The terms decrease at both bracket endpoints, so the alternating-series remainder applies. Outward interval evaluation at 50 decimal digits gives Φ ( − 0.674490 ) < 0.249999921 \Phi(-0.674490)<0.249999921 Φ ( − 0.674490 ) < 0.249999921 and Φ ( − 0.674489 ) > 0.250000238 \Phi(-0.674489)>0.250000238 Φ ( − 0.674489 ) > 0.250000238 . Evaluating the density formulas and (18) over that bracket, with F ( g ∗ ) = 1 / 4 F(g_*)=1/4 F ( g ∗ ) = 1/4 exactly, encloses the curvature in
[ 4.40037496282478 , 4.40039330270280 ] . [4.40037496282478,\ 4.40039330270280]. [ 4.40037496282478 , 4.40039330270280 ] .
where the displayed endpoints have been rounded outward. This proves the deliberately looser bound in the proposition. The replication code records the full intervals, precision, series remainder, and coefficient inputs. It uses no estimated parameters.
Table IA.1. Smooth full-support costs under the interior technology. The gamma and lognormal laws both have mean 0.021 and variance 0.002 2 / 12 0.002^2/12 0.00 2 2 /12 . Positive curvature occurs at one-quarter participation but is absent at the displayed median states.
Finite gaps and execution-parameter sensitivity
Table IA.Table 2 holds each binary target law D = μ ∗ ± 0.02 D=\mu_* \pm0.02 D = μ ∗ ± 0.02 fixed within its coarse/refined comparisons. The first two rows use the certified lognormal economy. The third uses the same cost law and a uniform schedule with ( λ , κ , ρ , η 0 , η ) = ( 1 , 1.5 , 1 , 4 , 4 ) (\lambda,\kappa,\rho,\eta_0,\eta)=(1,1.5,1,4,4) ( λ , κ , ρ , η 0 , η ) = ( 1 , 1.5 , 1 , 4 , 4 ) . Its carryover is H = 0.948181 H=0.948181 H = 0.948181 , its integrated impact is K = 0.551819 K=0.551819 K = 0.551819 , and its impact half-life is 0.693147 pre-event horizons. At quarter participation, μ ∗ = 0.477822 \mu_*=0.477822 μ ∗ = 0.477822 and ψ ′ ′ ( μ ∗ ) ≈ 0.103310 \psi''(\mu_*)\approx0.103310 ψ ′′ ( μ ∗ ) ≈ 0.103310 . Equal temporary charges and positive decay therefore admit a positive numerical gap as well. Concentrated activation costs and high carryover remain features of this illustration. The additional example is verified from the original cash objective at 60-digit precision, but has no directed-rounding curvature certificate.
Table IA.2. Economic scale of smooth information effects. Coarse cost includes the fixed prior variance. The final column divides the gap by that row’s coarse institutional bill. Costs are normalized cash units, not basis points of assets.
Figure IA.Figure 1 varies one primitive at a time with ϵ = 0.001 \epsilon= 0.001 ϵ = 0.001 and δ = 0.02 \delta= 0.02 δ = 0.02 . Solid curves retain the original target center. Dashed curves recenter at quarter participation in each economy, changing the prior across economies while preserving it within every information comparison. Along the unchanged-prior paths, the numerical zero crossings are H ≈ 0.93753 H \approx0.93753 H ≈ 0.93753 , cost coefficient of variation 0.13847 0.13847 0.13847 , and institutional temporary charge η 0 ≈ 0.03650 \eta_0 \approx0.03650 η 0 ≈ 0.03650 . The respective root brackets are [0.90, 0.95], [0.10, 0.20], and [0.02, 0.05]. These are one-dimensional calculations under the stated restrictions, not global parameter boundaries or calibration estimates.
Figure 1. Two information mechanisms. The upper panels subtract the left tangent from the posterior-mean cost function at each boundary, preserving every information comparison while exposing the slope change, and the lower panels compare optional with fixed participation. Each column holds its own binary target law fixed: D = μ 0 ± 0.04 D = \mu_0 \pm0.04 D = μ 0 ± 0.04 at onset and D = 2 ± 0.1 D = \sqrt{2} \pm0.1 D = 2 ± 0.1 at saturation. The common horizontal scale is the percentage of target variance explained, 100 ( ϵ / δ ) 2 100(\epsilon/\delta)^2 100 ( ϵ / δ ) 2 , where δ \delta δ is each fixed prior’s half-spread. Cost units are normalized and differ across the two scenarios.
Figure IA.1. Smooth-interior sensitivity. Each panel changes only its indicated primitive from the certified economy. Left: zero decay with κ = H \kappa= H κ = H and K = H / 2 K = H/2 K = H /2 . Middle: lognormal dispersion with mean cost fixed at 0.021. Right: institutional temporary charge with supplier charge fixed at 4. The vertical scale is percent of each comparison’s own coarse cost. The curves are numerical illustrations.
Components of the finite information cost
Table IA.Table 3 evaluates (23) for the same lognormal economy and unchanged scalar prior D = μ ∗ ± .02 D = \mu_* \pm.02 D = μ ∗ ± .02 . The inventory response outweighs the direct forecasting and institutional readjustment savings. Its coarse bill is .350908300279. This differs from the portfolio bill in Table 4 by the additional target-shock variance, although the radial cost gaps coincide. Values are high-precision numerical evaluations of the original cash objective.
Table IA.3. Exact accounting components of a finite refinement. All columns use normalized scalar cash units. The first and third terms are savings, and their sum with inventory incidence equals the total change. Intermediate accounting steps are not separate equilibria.
A globally beneficial economy
Proposition IA.1 (Exponential costs without pre-event impact). Let λ = η 0 = η = 1 \lambda= \eta_0 = \eta= 1 λ = η 0 = η = 1 , K = H = 0 K = H = 0 K = H = 0 , and F ( g ) = 1 − exp ( − g / θ ) F(g) = 1 - \exp(-g/\theta) F ( g ) = 1 − exp ( − g / θ ) on g ≥ 0 g \ge0 g ≥ 0 , where θ > 0 \theta> 0 θ > 0 . Then ψ \psi ψ is globally strictly concave. Every feasible public refinement weakly lowers institutional cost, strictly whenever the posterior means change with positive probability.
Proof. For μ > 0 \mu> 0 μ > 0 , set x = g / θ x = g/\theta x = g / θ , F = 1 − e − x F = 1 - e^{-x} F = 1 − e − x , B = 1 + ( 2 x − 1 ) e − x B = 1 + (2x - 1)e^{-x} B = 1 + ( 2 x − 1 ) e − x , and u = B / ( 3 + 2 B ) u = B/(3 + 2B) u = B / ( 3 + 2 B ) . Formula (20) gives
ψ ′ ′ = − 4 3 − 2 u 3 − u 2 3 − 6 x ( 1 + F ) ( 3 − 2 x ) e − x ( 3 + 2 B ) 3 . \psi^{\prime\prime} = -\frac{4}{3} - \frac{2u}{3} - \frac{u^2}{3} - \frac{6x(1+F)(3-2x)e^{-x}}{(3+2B)^3}. ψ ′′ = − 3 4 − 3 2 u − 3 u 2 − ( 3 + 2 B ) 3 6 x ( 1 + F ) ( 3 − 2 x ) e − x .
For 0 < x ≤ 3 / 2 0 < x \le3/2 0 < x ≤ 3/2 , all terms are nonpositive and the first is strictly negative. For x > 3 / 2 x > 3/2 x > 3/2 , B ≥ 1 B \ge1 B ≥ 1 and 1 + F ≤ 2 1 + F \le2 1 + F ≤ 2 . The only positive term is at most
12 125 x ( 2 x − 3 ) e − x ≤ 108 e − 3 125 < 0.044 , \frac{12}{125}x(2x-3)e^{-x} \le\frac{108e^{-3}}{125} < 0.044, 125 12 x ( 2 x − 3 ) e − x ≤ 125 108 e − 3 < 0.044 ,
because the function x ( 2 x − 3 ) e − x x(2x - 3)e^{-x} x ( 2 x − 3 ) e − x attains its maximum on [ 3 / 2 , ∞ ) [3/2,\infty) [ 3/2 , ∞ ) at x = 3 x = 3 x = 3 . Thus ψ ′ ′ < 0 \psi^{\prime\prime} < 0 ψ ′′ < 0 there as well. At zero, m = O ( μ 3 ) m = O(\mu^3) m = O ( μ 3 ) , so ψ \psi ψ has matching first derivatives and ψ ′ ′ ( 0 ) = − 4 / 3 \psi^{\prime\prime}(0) = -4/3 ψ ′′ ( 0 ) = − 4/3 . Reflection proves global strict concavity. Conditional strict Jensen then gives the information ranking.
Gaussian precision under an unchanged prior
For D ∼ N ( μ ˉ , σ D 2 ) D \sim N(\bar{\mu},\sigma_D^2) D ∼ N ( μ ˉ , σ D 2 ) , σ D > 0 \sigma_D > 0 σ D > 0 , observing D D D plus independent Gaussian noise of variance σ D 2 ( 1 − r ) / r \sigma_D^2(1-r)/r σ D 2 ( 1 − r ) / r gives posterior mean distribution N ( μ ˉ , r σ D 2 ) N(\bar{\mu},r\sigma_D^2) N ( μ ˉ , r σ D 2 ) , 0 < r < 1 0 < r < 1 0 < r < 1 . Lower precision can be implemented by adding independent noise to a higher-precision signal. The endpoints r = 0 , 1 r = 0,1 r = 0 , 1 are no information and revelation, respectively. For Z ∼ N ( 0 , 1 ) Z \sim N(0,1) Z ∼ N ( 0 , 1 ) ,
C ( r ) = λ ( μ ˉ 2 + σ D 2 ) + E [ ψ ( μ ˉ + σ D r Z ) ] . (53) C(r) = \lambda(\bar{\mu}^2 + \sigma_D^2) + \mathbb{E}[\psi(\bar{\mu} + \sigma_D\sqrt{r}Z)]. \tag*{(53)} C ( r ) = λ ( μ ˉ 2 + σ D 2 ) + E [ ψ ( μ ˉ + σ D r Z )] . (53)
Under the uniform-cost specifications, ψ \psi ψ has quadratic growth, bounded piecewise curvature, and finitely many slope jumps. The Gaussian density p r p_r p r obeys ∂ r p r = ( σ D 2 / 2 ) ∂ x x p r \partial_r p_r = (\sigma_D^2/2)\partial_{xx}p_r ∂ r p r = ( σ D 2 /2 ) ∂ xx p r . Differentiation and distributional integration by parts, with vanishing Gaussian boundary terms, give for $r > 0
C ′ ( r ) = σ D 2 2 [ ∫ R p r ( x ) ψ r e g ′ ′ ( x ) d x + ∑ τ p r ( τ ) J τ ] . (54) C'(r) = \frac{\sigma_D^2}{2}\left[\int_{\mathbb{R}} p_r(x)\psi^{\prime\prime}_{\mathrm{reg}}(x)\,\mathrm{d}x + \sum_{\tau}p_r(\tau)J_\tau\right]. \tag*{(54)} C ′ ( r ) = 2 σ D 2 [ ∫ R p r ( x ) ψ reg ′′ ( x ) d x + τ ∑ p r ( τ ) J τ ] . (54)
The result is the usual Gaussian smoothing identity applied to the cost function. Numerical integration splits at all four participation boundaries, with independent checks in both posterior-mean and standard-normal coordinates.
Sensitivity to estimated posterior means
For fixed primitives write L m = s / c L_m = s/c L m = s / c and t 1 = s / ( η + c ) t_1 = s/(\eta+ c) t 1 = s / ( η + c ) . The monotone equilibrium equation implies ∣ m ( u ) − m ( v ) ∣ ≤ L m ∣ u − v ∣ |m(u) - m(v)| \le L_m|u-v| ∣ m ( u ) − m ( v ) ∣ ≤ L m ∣ u − v ∣ , including nondifferentiable cost laws, and ∣ m ( u ) ∣ ≤ t 1 ∣ u ∣ |m(u)| \le t_1|u| ∣ m ( u ) ∣ ≤ t 1 ∣ u ∣ . To see the first assertion, write μ = ( η q + c m ) / s \mu= (\eta q + cm)/s μ = ( η q + c m ) / s and note that q q q and m m m are increasing together, also across zero. The second follows from m / μ = a s / ( η + c a ) m/\mu= as/(\eta+ ca) m / μ = a s / ( η + c a ) and a ≤ 1 a \le1 a ≤ 1 . Consequently,
∣ ψ ( u ) − ψ ( v ) ∣ ≤ C ψ ( ∣ u ∣ + ∣ v ∣ ) ∣ u − v ∣ , (55) |\psi(u) - \psi(v)| \le C_{\psi}(|u| + |v|)|u - v|, \tag*{(55)} ∣ ψ ( u ) − ψ ( v ) ∣ ≤ C ψ ( ∣ u ∣ + ∣ v ∣ ) ∣ u − v ∣ , (55)
C ψ = d 2 2 A + ∣ e ∣ 2 ( L m + t 1 ) + k t 1 L m . C_{\psi} = \frac{d^2}{2A} + \frac{|e|}{2}(L_m + t_1) + kt_1L_m. C ψ = 2 A d 2 + 2 ∣ e ∣ ( L m + t 1 ) + k t 1 L m .
Indeed, use u m ( u ) − v m ( v ) = { ( u + v ) [ m ( u ) − m ( v ) ] + [ m ( u ) + m ( v ) ] ( u − v ) } / 2 um(u) - vm(v) = \{(u + v)[m(u) - m(v)] + [m(u) + m(v)](u - v)\}/2 u m ( u ) − v m ( v ) = {( u + v ) [ m ( u ) − m ( v )] + [ m ( u ) + m ( v )] ( u − v )} /2 for the middle term of ψ \psi ψ and factor both squares. Cauchy–Schwarz gives, for square-integrable U , V U,V U , V ,
∣ E ψ ( U ) − E ψ ( V ) ∣ ≤ C ψ ( ∥ U ∥ 2 + ∥ V ∥ 2 ) ∥ U − V ∥ 2 . |\mathbb{E}\psi(U) - \mathbb{E}\psi(V)| \le C_{\psi}(\lVert U\rVert_2 + \lVert V\rVert_2)\lVert U - V\rVert_2. ∣ E ψ ( U ) − E ψ ( V ) ∣ ≤ C ψ (∥ U ∥ 2 + ∥ V ∥ 2 ) ∥ U − V ∥ 2 .
Apply this inequality to both the coarse and refined estimates and add the bounds. The target second moment cancels in a same-prior cost difference. This controls sensitivity to a given mean-error bound, without estimating that error or the structural coefficients from ETF bars.
B Notation
Table 7.
Cash accounts, units, and round trips
Derivation of the two objectives
Let Φ t = ∫ 0 t ϕ s d s \Phi_t = \int_0^t \phi_s\,\mathrm{d}s Φ t = ∫ 0 t ϕ s d s . The institution has inventory z Φ t z\Phi_t z Φ t before the event, an entrant has q Φ t q\Phi_t q Φ t , and impact is ( z + m ) κ ∫ 0 t e − ρ ( t − s ) ϕ s d s (z+m)\kappa\int_0^t e^{-\rho(t-s)}\phi_s\,\mathrm{d}s ( z + m ) κ ∫ 0 t e − ρ ( t − s ) ϕ s d s . Integration of the pre-event cash purchases gives K z ( z + m ) Kz(z+m) K z ( z + m ) for the institution and K q ( z + m ) Kq(z+m) K q ( z + m ) for an entrant. Adding each temporary charge and the terminal purchase or sale gives
C 0 r e a l i z e d = K z ( z + m ) + η 0 z 2 / 2 + ( D − z ) p A , C_0^{\mathrm{realized}} = Kz(z+m) + \eta_0 z^2/2 + (D-z)p_A, C 0 realized = K z ( z + m ) + η 0 z 2 /2 + ( D − z ) p A ,
π r e a l i z e d ( q ) = q p A − K q ( z + m ) − η q 2 / 2. \pi^{\mathrm{realized}}(q) = qp_A - Kq(z+m) - \eta q^2/2. π realized ( q ) = q p A − K q ( z + m ) − η q 2 /2.
Taking expectations and using E [ ( D − z ) 2 ∣ Y ] = ( μ − z ) 2 + v \mathbb{E}[(D-z)^2 \mid Y] = (\mu-z)^2+v E [( D − z ) 2 ∣ Y ] = ( μ − z ) 2 + v produces (7) and (9) . In particular, replacing the institution’s completion payment by a diagonal residual penalty would remove the terms involving m m m and would change the game.
The unscaled institution and mass-one anticipators use compatible aggregate units. The expected gross anticipator profit is a η q 2 / 2 a\eta q^2/2 a η q 2 /2 , and total fixed entry expenditure is ∫ 0 a q 2 / 2 f d F ( f ) \int_0^{a q^2/2} f\,\mathrm{d}F(f) ∫ 0 a q 2 /2 f d F ( f ) . If these expenditures are interpreted as resource costs, net trading profit subtracts that integral. If they are fees paid to another service provider, the recipient must also be included before making a welfare statement.
Counterparty accounting
Define the outside supplier’s marked revenue as its pre-event receipts, temporary charges, and auction receipts, less inventory valued at the unaffected benchmark. Conditional expected revenue is
Π D = K ( z + m ) 2 + η 0 z 2 / 2 + a η q 2 / 2 + H ( z + m ) ( μ − z − m ) + λ { ( μ − z − m ) 2 + v } . (56) \begin{aligned}
\Pi_D &= K(z+m)^2+\eta_0z^2/2+a\eta q^2/2 \\
&\quad+H(z+m)(\mu-z-m)+\lambda\{(\mu-z-m)^2+v\}. \tag*{(56)}
\end{aligned} Π D = K ( z + m ) 2 + η 0 z 2 /2 + a η q 2 /2 + H ( z + m ) ( μ − z − m ) + λ {( μ − z − m ) 2 + v } . (56)
Direct subtraction of the cash accounts gives
C 0 − a π ( q ) = Π D . (57) C_0-a\pi(q)=\Pi_D. \tag*{(57)} C 0 − aπ ( q ) = Π D . (57)
This identity is checked numerically for every baseline equilibrium in the reported grid. It is a transfer identity, and it does not establish a social objective because dealer costs, capital requirements, and risk preferences are not modeled. In particular, adding an independently chosen temporary-displacement penalty to the identity would define a new welfare criterion rather than complete the accounting.
Admissibility and zero-target deviations
When D = 0 D=0 D = 0 , m = 0 m=0 m = 0 , and the institution trades z z z before undoing that trade at the event, its cash cost is A z 2 / 2 ≥ 0 Az^2/2 \ge0 A z 2 /2 ≥ 0 . A sole finite minor with actual quantity u / N u/N u / N has normalized profit − ( η / 2 + b / N ) u 2 ≤ 0 -(\eta/2+b/N)u^2 \le0 − ( η /2 + b / N ) u 2 ≤ 0 before its nonnegative fixed cost. Thus the retained technology does not reward these mechanical round trips. The statement is limited to the prescribed schedule and terminal convention, and does not claim absence of manipulation for arbitrary multivenue or multi-asset trading paths.
For each public state, square completion gives the institution a finite lower bound because A > 0 A > 0 A > 0 , and the minor a finite upper bound because η > 0 \eta> 0 η > 0 . Theorem 1 gives ∣ q ∣ ≤ s ∣ μ ∣ / η |q| \le s|\mu|/\eta ∣ q ∣ ≤ s ∣ μ ∣/ η , ∣ m ∣ ≤ s ∣ μ ∣ / ( η + c ) |m| \le s|\mu|/(\eta+ c) ∣ m ∣ ≤ s ∣ μ ∣/ ( η + c ) , and
∣ z ∣ ≤ d + b s / η A ∣ μ ∣ . |z| \le\frac{d + bs/\eta}{A}|\mu|. ∣ z ∣ ≤ A d + b s / η ∣ μ ∣.
Since conditional expectation contracts L 2 L^2 L 2 , all quadratic cash terms are integrable. The expected entry-cost integral is bounded by a η q 2 / 2 a\eta q^2/2 a η q 2 /2 , so it is integrable even when the unconditional entry-cost distribution has an infinite first moment. These bounds establish the admissibility of the proposed strategies without an ellipticity condition or a state-space compactness assumption.
Endpoint expansions and cost-tail regularity
For a positive posterior mean, parameterize equilibrium by the entrant quantity q q q , with g = η q 2 / 2 g = \eta q^2/2 g = η q 2 /2 and
μ = q s { η + c F ( g ) } , m = q F ( g ) . \mu= \frac{q}{s}\{\eta+ cF(g)\}, \qquad m = qF(g). μ = s q { η + c F ( g )} , m = q F ( g ) .
At onset, q 0 = 2 f 0 / η q_0 = \sqrt{2f_0/\eta} q 0 = 2 f 0 / η and μ 0 = η q 0 / s \mu_0 = \eta q_0/s μ 0 = η q 0 / s . Write x = g − f 0 ↓ 0 x = g - f_0 \downarrow0 x = g − f 0 ↓ 0 . Expansion of the square root and the endpoint CDF gives
μ − μ 0 = ( 1 s q 0 + c q 0 ℓ 0 s ) x + O ( x 2 ) , m = q 0 ℓ 0 x + O ( x 2 ) . \mu- \mu_0 = \left(\frac{1}{s q_0} + \frac{c q_0 \ell_0}{s}\right)x + O(x^2), \qquad m = q_0\ell_0x + O(x^2). μ − μ 0 = ( s q 0 1 + s c q 0 ℓ 0 ) x + O ( x 2 ) , m = q 0 ℓ 0 x + O ( x 2 ) .
The coefficient of x x x is positive, so inversion gives m ( μ 0 + h ) = L 0 h + O ( h 2 ) m(\mu_0 + h) = L_0h + O(h^2) m ( μ 0 + h ) = L 0 h + O ( h 2 ) with the constant in Theorem 4 . Below the boundary the unique equilibrium has a = m = 0 a = m = 0 a = m = 0 .
At saturation let q 1 = 2 f 1 / η q_1 = \sqrt{2f_1/\eta} q 1 = 2 f 1 / η and write x = f 1 − g ↓ 0 x = f_1 - g \downarrow0 x = f 1 − g ↓ 0 . Then q = q 1 − x / ( η q 1 ) + O ( x 2 ) q = q_1 - x/(\eta q_1) + O(x^2) q = q 1 − x / ( η q 1 ) + O ( x 2 ) and F ( g ) = 1 − ℓ 1 x + O ( x 2 ) F(g) = 1 - \ell_1x + O(x^2) F ( g ) = 1 − ℓ 1 x + O ( x 2 ) , so
μ 1 − μ = ( η + c s η q 1 + c q 1 ℓ 1 s ) x + O ( x 2 ) , \mu_1 - \mu= \left(\frac{\eta+ c}{s\eta q_1} + \frac{c q_1\ell_1}{s}\right)x + O(x^2), μ 1 − μ = ( sη q 1 η + c + s c q 1 ℓ 1 ) x + O ( x 2 ) ,
m ( μ 1 ) − m ( μ ) = ( 1 η q 1 + q 1 ℓ 1 ) x + O ( x 2 ) . m(\mu_1) - m(\mu) = \left(\frac{1}{\eta q_1} + q_1\ell_1\right)x + O(x^2). m ( μ 1 ) − m ( μ ) = ( η q 1 1 + q 1 ℓ 1 ) x + O ( x 2 ) .
The ratio is s ( 1 + 2 f 1 ℓ 1 ) / [ η + c ( 1 + 2 f 1 ℓ 1 ) ] = L 1 s(1 + 2f_1\ell_1)/[\eta+ c(1 + 2f_1\ell_1)] = L_1 s ( 1 + 2 f 1 ℓ 1 ) / [ η + c ( 1 + 2 f 1 ℓ 1 )] = L 1 , proving the left expansion and its quadratic remainder without assuming global smoothness of F F F . Above the boundary m = t 1 μ m = t_1\mu m = t 1 μ exactly. The symmetric-difference expansion of ψ \psi ψ therefore has the remainder asserted in (29) .
For completeness, in the binary experiment write S ∈ { − 1 , 1 } S \in\{-1,1\} S ∈ { − 1 , 1 } for the reported sign and impose Pr ( S = B ∣ B ) = ( 1 + ϵ / δ ) / 2 \Pr(S = B \mid B) = (1 + \epsilon/\delta)/2 Pr ( S = B ∣ B ) = ( 1 + ϵ / δ ) /2 . Then Pr ( S = 1 ) = 1 / 2 \Pr(S = 1) = 1/2 Pr ( S = 1 ) = 1/2 , E [ B ∣ S ] = S ϵ / δ \mathbb{E}[B \mid S] = S\epsilon/\delta E [ B ∣ S ] = S ϵ / δ , and Var ( D ∣ S ) = δ 2 − ϵ 2 \operatorname{Var}(D \mid S) = \delta^2 - \epsilon^2 Var ( D ∣ S ) = δ 2 − ϵ 2 . Forgetting S S S recovers the coarse experiment. This verifies feasibility under the unchanged target distribution, including the variance component of institutional expenditure.
Proof of the power-law onset classification
Let h = μ − μ 0 > 0 h = \mu- \mu_0 > 0 h = μ − μ 0 > 0 and x = g − f 0 x = g - f_0 x = g − f 0 . Continuity and uniqueness of equilibrium imply x ↓ 0 x \downarrow0 x ↓ 0 with h ↓ 0 h \downarrow0 h ↓ 0 , and the parametrization above gives
h = x s q 0 ( 1 + o ( 1 ) ) + c q 0 C s x α ( 1 + o ( 1 ) ) , m = q 0 C x α ( 1 + o ( 1 ) ) . h = \frac{x}{s q_0}(1 + o(1)) + \frac{c q_0 C}{s}x^\alpha(1 + o(1)), \qquad m = q_0 Cx^\alpha(1 + o(1)). h = s q 0 x ( 1 + o ( 1 )) + s c q 0 C x α ( 1 + o ( 1 )) , m = q 0 C x α ( 1 + o ( 1 )) .
If α > 1 \alpha> 1 α > 1 , the first term dominates, yielding x ∼ s q 0 h x \sim sq_0h x ∼ s q 0 h and the stated M α M_\alpha M α . If α < 1 \alpha< 1 α < 1 , the second term dominates and m ∼ s h / c m \sim sh/c m ∼ s h / c . When α = 1 \alpha= 1 α = 1 , both terms contribute and their sum yields 2 s f 0 C / ( η + 2 c f 0 C ) 2sf_0C/(\eta+ 2cf_0C) 2 s f 0 C / ( η + 2 c f 0 C ) . In the lower posterior branch entry is zero, so the exact cost change is
− d 2 2 A ϵ 2 + 1 2 { e ( μ 0 + ϵ ) m ( μ 0 + ϵ ) − k m ( μ 0 + ϵ ) 2 } . -\frac{d^2}{2A}\epsilon^2+\frac{1}{2}\{e(\mu_0+\epsilon)m(\mu_0+\epsilon)-km(\mu_0+\epsilon)^2\}. − 2 A d 2 ϵ 2 + 2 1 { e ( μ 0 + ϵ ) m ( μ 0 + ϵ ) − k m ( μ 0 + ϵ ) 2 } .
Substituting the three equivalents proves Proposition 9 . In the case 1 < α < 2 1 < \alpha< 2 1 < α < 2 , the leading positive term is of order ϵ α \epsilon^\alpha ϵ α , while for α > 2 \alpha> 2 α > 2 all inventory terms are o ( ϵ 2 ) o(\epsilon^2) o ( ϵ 2 ) . At α = 2 \alpha= 2 α = 2 the displayed coefficient follows, with no sign conclusion at equality because the hypotheses specify only a little-o o o remainder.
An outside completion facility
Conditional on terminal information, the institution chooses j j j to minimize
j { H ( z + m ) + λ ( j − m ) } + χ 2 ( D − z − j ) 2 . j\{H(z+m)+\lambda(j-m)\}+\frac{\chi}{2}(D-z-j)^2. j { H ( z + m ) + λ ( j − m )} + 2 χ ( D − z − j ) 2 .
The curvature is χ + 2 λ > 0 \chi+2\lambda>0 χ + 2 λ > 0 , and its unique minimizer is (47) . Put B χ = χ + 2 λ B_\chi=\chi+2\lambda B χ = χ + 2 λ and
L χ = χ μ − ( χ + H ) z + ( λ − H ) m . L_\chi=\chi\mu-(\chi+H)z+(\lambda-H)m. L χ = χ μ − ( χ + H ) z + ( λ − H ) m .
After substitution, conditional expected total institutional expenditure is
C χ = K z ( z + m ) + η 0 2 z 2 + χ 2 { ( μ − z ) 2 + v } − L χ 2 + χ 2 v 2 B χ . (58) C_\chi=Kz(z+m)+\frac{\eta_0}{2}z^2+\frac{\chi}{2}\{(\mu-z)^2+v\}-\frac{L_\chi^2+\chi^2v}{2B_\chi}. \tag*{(58)} C χ = K z ( z + m ) + 2 η 0 z 2 + 2 χ {( μ − z ) 2 + v } − 2 B χ L χ 2 + χ 2 v . (58)
Define the institutional coefficients
A χ = 2 K + η 0 + χ − ( χ + H ) 2 B χ , d χ = χ ( 2 λ − H ) B χ , A_\chi=2K+\eta_0+\chi-\frac{(\chi+H)^2}{B_\chi}, \qquad d_\chi=\frac{\chi(2\lambda-H)}{B_\chi}, A χ = 2 K + η 0 + χ − B χ ( χ + H ) 2 , d χ = B χ χ ( 2 λ − H ) ,
b χ = K + ( χ + H ) ( λ − H ) B χ , l χ = λ χ B χ , b_\chi=K+\frac{(\chi+H)(\lambda-H)}{B_\chi}, \qquad l_\chi=\frac{\lambda\chi}{B_\chi}, b χ = K + B χ ( χ + H ) ( λ − H ) , l χ = B χ λ χ ,
and the minor’s two aggregate price loadings
b z , χ = K − H + λ ( χ + H ) B χ , b m , χ = K − H + λ ( χ + λ + H ) B χ . b_{z,\chi}=K-H+\frac{\lambda(\chi+H)}{B_\chi}, \qquad b_{m,\chi}=K-H+\frac{\lambda(\chi+\lambda+H)}{B_\chi}. b z , χ = K − H + B χ λ ( χ + H ) , b m , χ = K − H + B χ λ ( χ + λ + H ) .
The institution’s first-order condition is A χ z = d χ μ − b χ m A_\chi z=d_\chi\mu-b_\chi m A χ z = d χ μ − b χ m . An entrant’s expected profit is
q ( l χ μ − b z , χ z − b m , χ m ) − η q 2 / 2. q(l_\chi\mu-b_{z,\chi}z-b_{m,\chi}m)-\eta q^2/2. q ( l χ μ − b z , χ z − b m , χ m ) − η q 2 /2.
Thus, with
s χ = l χ − b z , χ d χ A χ , c χ = b m , χ − b z , χ b χ A χ , s_\chi=l_\chi-\frac{b_{z,\chi}d_\chi}{A_\chi}, \qquad c_\chi=b_{m,\chi}-\frac{b_{z,\chi}b_\chi}{A_\chi}, s χ = l χ − A χ b z , χ d χ , c χ = b m , χ − A χ b z , χ b χ ,
the proof of Theorem 1 applies whenever A χ > 0 A_\chi>0 A χ > 0 , s χ > 0 s_\chi>0 s χ > 0 , and c χ ≥ 0 c_\chi\geq0 c χ ≥ 0 , replacing ( s , c , A , d , b ) (s,c,A,d,b) ( s , c , A , d , b ) by the appropriate displayed coefficients. These inequalities are explicit primitive checks, not estimated stability constants. The numerical routine rejects a parameter set failing them, and all reported facility cases pass.
The optimized expression can be written directly as
C χ ∗ = l χ ( μ 2 + v ) − χ ( λ − H ) B χ μ m − ( λ − H ) 2 2 B χ m 2 − ( d χ μ − b χ m ) 2 2 A χ . C_{\chi}^{*}=l_{\chi}(\mu^{2}+v)-\frac{\chi(\lambda-H)}{B_{\chi}}\mu m-\frac{(\lambda-H)^{2}}{2B_{\chi}}m^{2}-\frac{(d_{\chi}\mu-b_{\chi}m)^{2}}{2A_{\chi}}. C χ ∗ = l χ ( μ 2 + v ) − B χ χ ( λ − H ) μ m − 2 B χ ( λ − H ) 2 m 2 − 2 A χ ( d χ μ − b χ m ) 2 .
Its no-entry value sets m = 0 m=0 m = 0 , giving the decomposition in Proposition 14 . To verify the incidence identity without assuming its sign, write r = λ − H r=\lambda-H r = λ − H and d = 2 λ − H d=2\lambda-H d = 2 λ − H . Since d χ = χ d / B χ d_{\chi}=\chi d/B_{\chi} d χ = χ d / B χ ,
e χ = χ B χ A χ ( d b χ − r A χ ) . e_{\chi}=\frac{\chi}{B_{\chi}A_{\chi}}(db_{\chi}-rA_{\chi}). e χ = B χ A χ χ ( d b χ − r A χ ) .
Substitution of the displayed coefficients cancels the χ \chi χ terms and gives
d b χ − r A χ = H K + r H − r η 0 = H ( K + r ) − r η 0 = H b − ( λ − H ) η 0 = A e . db_{\chi}-rA_{\chi}=HK+rH-r\eta_{0}=H(K+r)-r\eta_{0}=Hb-(\lambda-H)\eta_{0}=Ae. d b χ − r A χ = H K + r H − r η 0 = H ( K + r ) − r η 0 = H b − ( λ − H ) η 0 = A e .
This proves the identity and its sign under χ , A χ > 0 \chi,A_{\chi}>0 χ , A χ > 0 .
For completeness, A χ → A A_{\chi}\to A A χ → A , d χ → d d_{\chi}\to d d χ → d , b χ → b b_{\chi}\to b b χ → b , l χ → λ l_{\chi}\to\lambda l χ → λ , and both minor loadings tend to b b b as χ → ∞ \chi\to\infty χ → ∞ . The unique roots of the continuous scalar entry equations converge by compactness and uniqueness, and the corresponding quantities and cash costs then converge. This establishes the baseline limit without claiming a capacity-constrained auction theorem.
Additional information comparisons
Cost distributions and diffuse forecasts
Moment matching and small CDF perturbations answer different robustness questions. Table IA.Table 4 fixes mean cost at 0.03 and variance at 0.02 2 / 12 0.02^{2}/12 0.0 2 2 /12 , replacing the original uniform law by gamma and lognormal alternatives. Table IA.Table 5 instead uses U exp ( σ Z ) U\exp(\sigma Z) U exp ( σ Z ) , with U ∼ Uniform [ 0.02 , 0.04 ] U\sim\operatorname{Uniform}[0.02,0.04] U ∼ Uniform [ 0.02 , 0.04 ] and independent standard normal Z Z Z . This smooths both endpoints, but its moments change with σ \sigma σ . Each comparison retains the original target law and ϵ = 0.01 \epsilon=0.01 ϵ = 0.01 . Neither exercise implies a distribution-independent sign.
Table IA.4. Moment-matched full-support costs under unchanged priors. Positive entries indicate harmful information.
Figure IA.Figure 2 uses p w ( x ) = 15 [ 1 − ( ( x − ξ ) / w ) 2 ] 2 / ( 16 w ) p_w(x)=15[1-((x-\xi)/w)^2]^2/(16w) p w ( x ) = 15 [ 1 − (( x − ξ ) / w ) 2 ] 2 / ( 16 w ) on ∣ x − ξ ∣ ≤ w |x-\xi|\le w ∣ x − ξ ∣ ≤ w , with coarse information revealing X X X and D = X + δ B D=X+\delta B D = X + δ B as in Theorem 10. Each grid point changes the coarse-forecast distribution and defines its own target law, held fixed within its refinement. The displayed coefficient is evaluated by quadrature, splitting at all participation boundaries. Table IA.Table 6 gives untransformed values at the boundary center, where the inverse-width jump contribution is most visible.
Table IA.5. Multiplicative smoothing under unchanged priors. The cost law is U exp ( σ Z ) U\exp(\sigma Z) U exp ( σ Z ) , and entries are refined-minus-coarse cost.
Figure IA.2. Diffuse coarse forecasts. Red denotes a positive local coefficient and blue a negative one, with a black zero contour. Each colorbar uses symmetric-log normalization, linear on [ − 0.05 , 0.05 ] [-0.05, 0.05] [ − 0.05 , 0.05 ] and logarithmic outside, spanning the full data range without clipping. The panels have separate ranges. The vertical axis is the support half-width.
Table IA.6. The local coefficient lim ϵ ↓ 0 Δ C / ϵ 2 \lim_{\epsilon\downarrow0}\Delta C/\epsilon^2 lim ϵ ↓ 0 Δ C / ϵ 2 at a coarse-forecast center equal to the participation boundary.
Smooth interior with finite strategic size
The strict smooth-interior comparison also survives finite populations eventually under an unchanged prior. Theorem 11 gives convergence at each of the three posterior means used by a fixed binary refinement, so the finite cost gap converges to its positive continuum value. Table IA.7 uses the lognormal construction of Proposition 7 , center μ ∗ = 0.8250600786 \mu^* = 0.8250600786 μ ∗ = 0.8250600786 , target half-spread 0.02, and displacement 0.001. Selected positive finite values do not establish positivity at every larger integer, and N N N counts normalized potential allocations rather than observed trading firms.
Table IA.7. Smooth-interior information under one unchanged prior.
Joint finite-size scale
Table IA.8 holds the prior at the continuum onset boundary and sets ϵ N = 1 / N \epsilon_N = 1/N ϵ N = 1/ N . The coefficient γ 0 ( 1 − ξ 0 ) \gamma_0(1-\xi_0) γ 0 ( 1 − ξ 0 ) is positive, but the exact cost sign at N = 100 N = 100 N = 100 is negative. This illustrates the difference between the joint asymptotic scale and a finite-N N N decision rule.
Table IA.8. Convergence on the joint scale, under the principal onset technology.
Numerical implementation and sensitivity
The uniform schedule has K = κ ( ρ − 1 + e − ρ ) / ρ 2 K = \kappa(\rho- 1 + e^{-\rho})/\rho^2 K = κ ( ρ − 1 + e − ρ ) / ρ 2 and H = κ ( 1 − e − ρ ) / ρ H = \kappa(1-e^{-\rho})/\rho H = κ ( 1 − e − ρ ) / ρ , with continuous limits at zero. The scalar participation equation is solved by a bracketed root method. Independent verification reconstructs the cash accounts and checks institutional and minor deviations, rather than relying only on the scalar residual. The smoothing CDF is
F σ ( f ) = 1 f 1 − f 0 ∫ f 0 f 1 Φ ( log ( f / u ) σ ) d u , f > 0 , F_\sigma(f) = \frac{1}{f_1-f_0}\int_{f_0}^{f_1}\Phi\left(\frac{\log(f/u)}{\sigma}\right)\,\mathrm{d}u,\qquad f>0, F σ ( f ) = f 1 − f 0 1 ∫ f 0 f 1 Φ ( σ log ( f / u ) ) d u , f > 0 ,
evaluated with 256-node Gauss–Legendre quadrature and checked at 512 nodes. Diffuse-forecast integration splits at every participation boundary before adding the derivative-jump terms, and is independently checked against finite symmetric differences.
Table IA.9. Schedule comparisons with rate-cost coefficients 0.05 multiplied by ∫ ϕ 2 \int\phi^2 ∫ ϕ 2 and displacement 0.01. Priors are fixed at the original onset or recentered as labeled. Positive e e e alone does not fix the sign at a finite displacement.
Joint markets and basket activation
Related benchmarks can share inventory suppliers. A trader who prepares Nasdaq-100 inventory may also trade the broad market or semiconductors, while a common activation expense makes the participation decision depend on the whole basket. Cross-impact gives a second source of interaction. Portfolio execution with cross-impact has been studied by Min et al. [27 ] . Here the additional question is whether a shared entry decision remains tractable when each entrant chooses several quantities.
Under the idealized no-flow balance sheet, a benchmark family’s aggregate exposure adjustment is W b R b W_bR_b W b R b , where W b = ∑ i : b ( i ) = b A i β i ( β i − 1 ) W_b = \sum_{i:b(i)=b} A_i\beta_i(\beta_i - 1) W b = ∑ i : b ( i ) = b A i β i ( β i − 1 ) . Dividing by a fixed reference price expresses this quantity in the benchmark units used below. Long and inverse geared funds both contribute positive coefficients to W b W_b W b . Cross-family return dependence therefore gives correlated target proxies, while shared inventory suppliers can link their implementation. The game below describes one institution with a vector target. It does not assume that independent fund managers coordinate their execution, and it does not infer the unobserved historical A i A_i A i from secondary-market volume.
A vector quantity game
Let D ∈ R p D \in\mathbb{R}^p D ∈ R p be the institution’s target in fixed benchmark units, with posterior mean μ \mu μ and covariance V V V . Quantities z z z , q q q , m m m are vectors. Entrants pay one private fixed cost to activate the basket, after which every coordinate can be traded. Let Λ \Lambda Λ denote terminal inverse depth and let K K K , H H H denote schedule-integrated impact and terminal carryover. Assume these matrices are symmetric, K , H ⪰ 0 K,H \succeq0 K , H ⪰ 0 , Λ − H ≻ 0 \Lambda-H \succ0 Λ − H ≻ 0 , and the temporary-charge matrices E 0 , E ≻ 0 E_0,E \succ0 E 0 , E ≻ 0 . The common scalar decay specification gives K = k Γ K=k\Gamma K = k Γ and H = h Γ H=h\Gamma H = h Γ for a symmetric positive-semidefinite impact matrix Γ \Gamma Γ . No simultaneous diagonalization with Λ , E 0 , E \Lambda,E_0,E Λ , E 0 , E is assumed.
The terminal price displacement is
p A = H ( z + m ) + Λ ( D − z − m ) . p_A = H(z+m) + \Lambda(D-z-m). p A = H ( z + m ) + Λ ( D − z − m ) .
Pre-event payments are z T K ( z + m ) + z T E 0 z / 2 z^{\mathsf{T}}K(z+m) + z^{\mathsf{T}}E_0z/2 z T K ( z + m ) + z T E 0 z /2 for the institution and q T K ( z + m ) + q T E q / 2 q^{\mathsf{T}}K(z+m) + q^{\mathsf{T}}Eq/2 q T K ( z + m ) + q T E q /2 for an entrant. Define
B = Λ + K − H , A = E 0 + 2 B , L = 2 Λ − H , C = B − B A − 1 B , S = Λ − B A − 1 L . (59) \begin{aligned}
B &= \Lambda+ K - H, & A &= E_0 + 2B, & L &= 2\Lambda- H, \\
C &= B - BA^{-1}B, & S &= \Lambda- BA^{-1}L. \tag*{(59)}
\end{aligned} B C = Λ + K − H , = B − B A − 1 B , A S = E 0 + 2 B , = Λ − B A − 1 L . L = 2Λ − H , (59)
The matrix S S S need not be symmetric. Direct expansion of the institution’s cash account gives
C 0 ( z , m ∣ Y ) = tr ( Λ V ) + μ ⊤ Λ μ − μ ⊤ ( Λ − H ) m + 1 2 z ⊤ A z − z ⊤ ( L μ − B m ) . (60) C_0(z,m\mid Y)=\operatorname{tr}(\Lambda V)+\mu^{\top}\Lambda\mu-\mu^{\top}(\Lambda-H)m+\frac{1}{2}z^{\top}Az-z^{\top}(L\mu-Bm). \tag*{(60)} C 0 ( z , m ∣ Y ) = tr ( Λ V ) + μ ⊤ Λ μ − μ ⊤ ( Λ − H ) m + 2 1 z ⊤ A z − z ⊤ ( Lμ − B m ) . (60)
An entrant’s gross expected profit is
π ( q ) = q ⊤ { Λ μ − B ( z + m ) } − 1 2 q ⊤ E q . \pi(q)=q^{\top}\{\Lambda\mu-B(z+m)\}-\frac{1}{2}q^{\top}Eq. π ( q ) = q ⊤ { Λ μ − B ( z + m )} − 2 1 q ⊤ E q .
As in the scalar game, the institution and entrants choose simultaneously after observing the common signal. The institution does not first commit to a reaction function. Its first-order condition is substituted only when solving the simultaneous conditions.
Proposition IA.2 (Unique basket-activation equilibrium). Suppose E ∥ D ∥ 2 < ∞ \mathbb{E}\lVert D\rVert^2<\infty E ∥ D ∥ 2 < ∞ and the nonnegative activation cost has a continuous distribution function F F F . There is a unique common entrant quantity and participation mass, with
q ( a ) = ( E + a C ) − 1 S μ , g ( a ) = 1 2 q ( a ) ⊤ E q ( a ) , q(a)=(E+aC)^{-1}S\mu,\qquad g(a)=\frac{1}{2}q(a)^{\top}Eq(a), q ( a ) = ( E + a C ) − 1 S μ , g ( a ) = 2 1 q ( a ) ⊤ E q ( a ) , a = F ( g ( a ) ) , m = a q ( a ) , z = A − 1 ( L μ − B m ) . (61) a=F(g(a)),\qquad m=aq(a),\qquad z=A^{-1}(L\mu-Bm). \tag*{(61)} a = F ( g ( a )) , m = a q ( a ) , z = A − 1 ( Lμ − B m ) . (61) The scalar equation for a a a has exactly one root in [ 0 , 1 ] [0,1] [ 0 , 1 ] . Quantities have at most linear growth in ∥ μ ∥ \lVert\mu\rVert ∥ μ ∥ , and the quadratic cash accounts are integrable.
Proof. The assumptions imply B ≻ 0 B\succ0 B ≻ 0 and A ≻ 2 B A\succ2B A ≻ 2 B . Inversion reverses the positive-definite order, giving B A − 1 B ≺ B / 2 BA^{-1}B\prec B/2 B A − 1 B ≺ B /2 and therefore C ≻ B / 2 ≻ 0 C\succ B/2\succ0 C ≻ B /2 ≻ 0 . Strict convexity of (H.2) gives the displayed institutional quantity. Strict concavity of entrant profit gives
E q = Λ μ − B ( z + m ) = S μ − C m . Eq=\Lambda\mu-B(z+m)=S\mu-Cm. E q = Λ μ − B ( z + m ) = S μ − C m . All entering traders consequently choose the same quantity, and m = a q m=aq m = a q gives (H.3). At the optimum, gross profit is q ⊤ E q / 2 q^{\top}Eq/2 q ⊤ E q /2 , so entry occurs exactly when the fixed cost does not exceed g ( a ) g(a) g ( a ) , up to a null set.
Put W = E − 1 / 2 C E − 1 / 2 ≻ 0 W=E^{-1/2}CE^{-1/2}\succ0 W = E − 1/2 C E − 1/2 ≻ 0 and x = E − 1 / 2 S μ x=E^{-1/2}S\mu x = E − 1/2 S μ . Then
g ( a ) = 1 2 x ⊤ ( I + a W ) − 2 x , g ′ ( a ) = − x ⊤ W ( I + a W ) − 3 x ≤ 0. g(a)=\frac{1}{2}x^{\top}(I+aW)^{-2}x,\qquad
g'(a)=-x^{\top}W(I+aW)^{-3}x\le0. g ( a ) = 2 1 x ⊤ ( I + aW ) − 2 x , g ′ ( a ) = − x ⊤ W ( I + aW ) − 3 x ≤ 0. Thus a − F ( g ( a ) ) a-F(g(a)) a − F ( g ( a )) is continuous and strictly increasing, with a nonpositive value at zero and a nonnegative value at one. Existence and uniqueness follow. Moreover,
∥ q ( a ) ∥ ≤ ∥ S ∥ λ min ( E ) ∥ μ ∥ , ∥ m ∥ ≤ ∥ q ( a ) ∥ . \lVert q(a)\rVert\le\frac{\lVert S\rVert}{\lambda_{\min}(E)}\lVert\mu\rVert,\qquad
\lVert m\rVert\le\lVert q(a)\rVert. ∥ q ( a )∥ ≤ λ m i n ( E ) ∥ S ∥ ∥ μ ∥ , ∥ m ∥ ≤ ∥ q ( a )∥ . The formula for z z z gives the remaining linear bound. Conditional expectation contracts L 2 L^2 L 2 , establishing integrability. Expected fixed expenditure is bounded by a g ( a ) ag(a) a g ( a ) and is therefore integrable as well.
Information and the source of jointness
Substituting the institutional optimum yields
E [ C 0 ∗ ] = E [ D ⊤ Λ D ] + E [ Ψ ( M ) ] , (62) \mathbb{E}[C_0^*] = \mathbb{E}[D^{\top}\Lambda D] + \mathbb{E}[\Psi(M)], \tag*{(62)} E [ C 0 ∗ ] = E [ D ⊤ Λ D ] + E [ Ψ ( M )] , (62)
Ψ ( μ ) = − μ ⊤ ( Λ − H ) m ( μ ) − 1 2 { L μ − B m ( μ ) } ⊤ A − 1 { L μ − B m ( μ ) } . \Psi(\mu) = -\mu^{\top}(\Lambda-H)m(\mu) - \frac{1}{2}\{L\mu-Bm(\mu)\}^{\top}A^{-1}\{L\mu-Bm(\mu)\}. Ψ ( μ ) = − μ ⊤ ( Λ − H ) m ( μ ) − 2 1 { Lμ − B m ( μ ) } ⊤ A − 1 { Lμ − B m ( μ )} .
For nested public experiments, M 0 = E [ M 1 ∣ H 0 ] M_0=\mathbb{E}[M_1\mid\mathcal{H}_0] M 0 = E [ M 1 ∣ H 0 ] . Concavity of Ψ \Psi Ψ on a convex region containing the posterior means is sufficient for every refinement within that region to benefit the institution. At a smooth interior state, a symmetric displacement ± ϵ u \pm\epsilon u ± ϵ u changes cost by
1 2 ϵ 2 u ⊤ ∇ 2 Ψ ( μ ) u + o ( ϵ 2 ) . \frac{1}{2}\epsilon^2u^{\top}\nabla^2\Psi(\mu)u+o(\epsilon^2). 2 1 ϵ 2 u ⊤ ∇ 2 Ψ ( μ ) u + o ( ϵ 2 ) .
Information can therefore have different local values in different portfolio directions. The scalar fixed-roster and centered-target rankings are examined below. They do not hold automatically for general execution matrices.
Even diagonal impact and charge matrices leave the assets linked through the single equation for a a a , because gross entry profit is the sum of the coordinate profits. Independent scalar games instead require separately paid, separately chosen activation decisions. Conversely, correlated target innovations alone do not create cross-impact. Holding μ \mu μ and the execution matrices fixed, V V V enters the risk-neutral institutional bill through tr ( Λ V ) \operatorname{tr}(\Lambda V) tr ( Λ V ) and does not change equilibrium quantities. This distinction matters when interpreting joint return forecasts.
Favorable portfolio information rankings
The scalar fixed-roster and centered-target rankings require additional restrictions in the portfolio model. The distinction can be stated exactly. For a fixed participation mass a ∈ [ 0 , 1 ] a\in[0,1] a ∈ [ 0 , 1 ] , define
T a = a ( E + a C ) − 1 S , T_a=a(E+aC)^{-1}S, T a = a ( E + a C ) − 1 S ,
P a = − Sym { ( Λ − H ) T a } − 1 2 ( L − B T a ) ⊤ A − 1 ( L − B T a ) , (63) P_a=-\operatorname{Sym}\{(\Lambda-H)T_a\}-\frac{1}{2}(L-BT_a)^{\top}A^{-1}(L-BT_a), \tag*{(63)} P a = − Sym {( Λ − H ) T a } − 2 1 ( L − B T a ) ⊤ A − 1 ( L − B T a ) , (63)
where Sym ( X ) = ( X + X ⊤ ) / 2 \operatorname{Sym}(X)=(X+X^{\top})/2 Sym ( X ) = ( X + X ⊤ ) /2 . Then m = T a μ m=T_a\mu m = T a μ and Ψ a ( μ ) = μ ⊤ P a μ \Psi_a(\mu)=\mu^{\top}P_a\mu Ψ a ( μ ) = μ ⊤ P a μ .
Proposition IA.3 (Information with a fixed portfolio roster). Fix a a a and the execution matrices. For nested public experiments with posterior means M 0 , M 1 M_0,M_1 M 0 , M 1 , the change in expected institutional cost is
Δ C = E [ ( M 1 − M 0 ) ⊤ P a ( M 1 − M 0 ) ] . (64) \Delta C=\mathbb{E}[(M_1-M_0)^{\top}P_a(M_1-M_0)]. \tag*{(64)} Δ C = E [( M 1 − M 0 ) ⊤ P a ( M 1 − M 0 )] . (64) Every refinement weakly lowers cost for every square-integrable target law if and only if P a ⪯ 0 P_a\preceq0 P a ⪯ 0 . If Λ , K , H , E 0 , E \Lambda,K,H,E_0,E Λ , K , H , E 0 , E are simultaneously orthogonally diagonalizable, then P a ≺ 0 P_a\prec0 P a ≺ 0 for every a ∈ [ 0 , 1 ] a\in[0,1] a ∈ [ 0 , 1 ] . In this class, every refinement that changes the posterior mean with positive probability strictly lowers cost at a fixed roster. Under optional basket activation, every signal with a nonzero posterior mean with positive probability strictly improves on no information for a centered vector target.
Proof. Subtract the two instances of (H.4). The target-second-moment term cancels. Since E [ M 1 − M 0 ∣ H 0 ] = 0 \mathbb{E}[M_1-M_0\mid\mathcal{H}_0]=0 E [ M 1 − M 0 ∣ H 0 ] = 0 , the cross term in the quadratic expansion has zero expectation, proving (H.6). Negative semidefiniteness is sufficient. If u ⊤ P a u > 0 u^{\top}P_au>0 u ⊤ P a u > 0 , a target equal to u u u and − u -u − u with equal probabilities and its complete revelation give a strict increase, proving necessity.
For the sufficient class, work in the common orthonormal basis and use lower-case letters for the diagonal entries. Write
b j = λ j + k j − h j , A j = η 0 j + 2 b j , l j = 2 λ j − h j , c j = b j − b j 2 A j , s j = λ j η 0 j + b j h j A j > 0. b_j=\lambda_j+k_j-h_j,\quad A_j=\eta_{0j}+2b_j,\quad l_j=2\lambda_j-h_j,\quad c_j=b_j-\frac{b_j^2}{A_j},\quad s_j=\frac{\lambda_j\eta_{0j}+b_jh_j}{A_j}>0. b j = λ j + k j − h j , A j = η 0 j + 2 b j , l j = 2 λ j − h j , c j = b j − A j b j 2 , s j = A j λ j η 0 j + b j h j > 0.
All b j , c j , l j b_j,c_j,l_j b j , c j , l j are positive, and t j ( a ) = a s j η j + a c j ≥ 0 t_j(a)=\frac{as_j}{\eta_j+ac_j}\geq0 t j ( a ) = η j + a c j a s j ≥ 0 . The j j j th diagonal entry of P a P_a P a is
− ( λ j − h j ) t j ( a ) − { l j − b j t j ( a ) } 2 2 A j < 0. (65) -(\lambda_j-h_j)t_j(a)-\frac{\{l_j-b_jt_j(a)\}^2}{2A_j}<0. \tag*{(65)} − ( λ j − h j ) t j ( a ) − 2 A j { l j − b j t j ( a ) } 2 < 0. (65)
At a = 0 a=0 a = 0 , the squared term is strictly negative. At a > 0 a>0 a > 0 , the first term is strictly negative. This proves the fixed-roster assertion. With optional activation, the same expression holds at the common endogenous mass a ( μ ) a(\mu) a ( μ ) , so Ψ ( μ ) < 0 \Psi(\mu)<0 Ψ ( μ ) < 0 for μ ≠ 0 \mu\ne0 μ = 0 , whereas Ψ ( 0 ) = 0 \Psi(0)=0 Ψ ( 0 ) = 0 . Comparing with the zero posterior mean proves the centered-target assertion.
Simultaneous diagonalization is sufficient, not necessary. The criterion is P a ⪯ 0 P_a\preceq0 P a ⪯ 0 . The proposition does not assert that every optional-entry refinement helps, even in the diagonal class, and it does not establish a vector Gaussian-precision ranking. Proposition 6’s Gaussian statement concerns the scalar model.
Execution-weighted forecast error. For a fixed roster, put Q = E [ ( M 1 − M 0 ) ( M 1 − M 0 ) T ] Q=\mathbb{E}[(M_1-M_0)(M_1-M_0)^{\mathsf T}] Q = E [( M 1 − M 0 ) ( M 1 − M 0 ) T ] and
G Λ = E [ ( D − M 0 ) T Λ ( D − M 0 ) ] − E [ ( D − M 1 ) T Λ ( D − M 1 ) ] = tr ( Λ Q ) ≥ 0. G_\Lambda=\mathbb{E}[(D-M_0)^{\mathsf T}\Lambda(D-M_0)]-\mathbb{E}[(D-M_1)^{\mathsf T}\Lambda(D-M_1)]=\operatorname{tr}(\Lambda Q)\geq0. G Λ = E [( D − M 0 ) T Λ ( D − M 0 )] − E [( D − M 1 ) T Λ ( D − M 1 )] = tr ( Λ Q ) ≥ 0.
If θ − \theta_- θ − and θ + \theta_+ θ + are the extreme eigenvalues of Λ − 1 / 2 P a Λ − 1 / 2 \Lambda^{-1/2}P_a\Lambda^{-1/2} Λ − 1/2 P a Λ − 1/2 , then
θ − G Λ ≤ Δ C = tr ( P a Q ) ≤ θ + G Λ . (66) \theta_-G_\Lambda\leq\Delta C=\operatorname{tr}(P_aQ)\leq\theta_+G_\Lambda. \tag*{(66)} θ − G Λ ≤ Δ C = tr ( P a Q ) ≤ θ + G Λ . (66)
The two Loewner inequalities θ − Λ ⪯ P a ⪯ θ + Λ \theta_-\Lambda\preceq P_a\preceq\theta_+\Lambda θ − Λ ⪯ P a ⪯ θ + Λ , multiplied by Q ⪰ 0 Q\succeq0 Q ⪰ 0 and traced, prove the bound. Under an invertible change of quantity units D ~ = R D \widetilde{D}=RD D = R D , the matrices become Λ ~ = R − T Λ R − 1 \widetilde{\Lambda}=R^{-\mathsf T}\Lambda R^{-1} Λ = R − T Λ R − 1 , P ~ a = R − T P a R − 1 \widetilde{P}_a=R^{-\mathsf T}P_aR^{-1} P a = R − T P a R − 1 , and Q ~ = R Q R T \widetilde{Q}=RQR^{\mathsf T} Q = R Q R T . Both traces and the generalized eigenvalues are unchanged. The comparison therefore measures execution incidence per unit of forecast improvement without depending on asset denomination. Feasibility remains determined by the prior. The bound does not assert that every Q ⪰ 0 Q\succeq0 Q ⪰ 0 is attainable, and it concerns a fixed roster rather than a state-dependent activation mass.
A portfolio counterexample with original cash payments
The unrestricted matrix class admits a different mechanism. Consider
Λ = ( 26.1 15 15 9.1 ) , H = ( 25 15 15 9 ) , K = H / 2 , \Lambda=\begin{pmatrix}26.1&15\\15&9.1\end{pmatrix},\qquad H=\begin{pmatrix}25&15\\15&9\end{pmatrix},\qquad K=H/2, Λ = ( 26.1 15 15 9.1 ) , H = ( 25 15 15 9 ) , K = H /2 ,
E 0 = diag ( .1 , 10 ) , E = diag ( 2 , 10 ) , a = 1 2 , u = ( .4 , − 1 ) T . (67) E_0=\operatorname{diag}(.1,10),\qquad E=\operatorname{diag}(2,10),\qquad a=\frac{1}{2},\qquad u=(.4,-1)^{\mathsf T}. \tag*{(67)} E 0 = diag ( .1 , 10 ) , E = diag ( 2 , 10 ) , a = 2 1 , u = ( .4 , − 1 ) T . (67)
Here H = ( 5 , 3 ) T ( 5 , 3 ) ⪰ 0 H=(5,3)^{\mathsf T}(5,3)\succeq0 H = ( 5 , 3 ) T ( 5 , 3 ) ⪰ 0 and Λ − H = diag ( 1.1 , .1 ) ≻ 0 \Lambda-H=\operatorname{diag}(1.1,.1)\succ0 Λ − H = diag ( 1.1 , .1 ) ≻ 0 . A uniform schedule with zero decay and instantaneous impact matrix H H H gives K = H / 2 K=H/2 K = H /2 . Both charge matrices are positive definite. Thus the example satisfies the common-schedule interpretation as well as the displayed matrix assumptions.
Let D D D equal u u u or − u -u − u with probability one-half. No information gives zero quantities and expected expenditure u ⊤ Λ u = 1.276 u^{\top}\Lambda u = 1.276 u ⊤ Λ u = 1.276 . Under revelation at D = u D = u D = u , the equilibrium quantities are
q ≈ ( − .428237463 , − .099612436 ) ⊤ , m = q / 2 , z ≈ ( .014887660 , − .082724781 ) ⊤ . q \approx(-.428237463,-.099612436)^{\top}, \qquad m=q/2, \qquad z \approx(.014887660,-.082724781)^{\top}. q ≈ ( − .428237463 , − .099612436 ) ⊤ , m = q /2 , z ≈ ( .014887660 , − .082724781 ) ⊤ .
All quantities change sign at D = − u D=-u D = − u , leaving the cash payments unchanged. Table IA.10 evaluates pre-event impact, temporary charges, and terminal completion directly. With the rational primitives in (H.9), the exact cost increase is
u ⊤ P 1 / 2 u = 9196314467527330257 235903720523056824500 > 0. (68) u^{\top}P_{1/2}u=\frac{9196314467527330257}{235903720523056824500}>0. \tag*{(68)} u ⊤ P 1/2 u = 235903720523056824500 9196314467527330257 > 0. (68)
This is a 3.0551% increase relative to the no-information bill. The generalized eigenvalues in (H.8) are approximately − .27450906 -.27450906 − .27450906 and .03974858 .03974858 .03974858 . Here G Λ = 1.276 G_{\Lambda}=1.276 G Λ = 1.276 and Δ C / G Λ = .03055121 \Delta C/G_{\Lambda}=.03055121 Δ C / G Λ = .03055121 . The positive gap persists under sufficiently small admissible perturbations. These are normalized model costs, not estimated ETF execution losses.
Table IA.10. Original institutional cash payments in the portfolio counterexample. The centered target law and the participation mass a = 1 / 2 a=1/2 a = 1/2 are unchanged between experiments. The exact identity (H.10) verifies the positive total difference.
The same target law yields a harmful comparison with optional entry. Set g ∗ = q ⊤ E q / 2 ≈ .233000511835 g_* = q^{\top}Eq/2 \approx.233000511835 g ∗ = q ⊤ E q /2 ≈ .233000511835 at the displayed half-roster equilibrium and choose the fixed cost law F F F to be uniform on [ 0 , 2 g ∗ ] [0,2g_*] [ 0 , 2 g ∗ ] . This one law is used in both experiments. Under revelation, F ( g ∗ ) = 1 / 2 F(g_*)=1/2 F ( g ∗ ) = 1/2 at either state. Under no information, gross profit and participation are zero. Proposition IA.2 gives uniqueness, and the original cash difference remains (H.10).
The adverse fixed-roster comparison comes from inventory direction. The term − μ ⊤ ( Λ − H ) m -\mu^{\top}(\Lambda-H)m − μ ⊤ ( Λ − H ) m in (H.4) need not be nonpositive. Cross-impact and unequal trading charges can rotate supply. Here entrants sell the first asset despite its positive target component. This mechanism operates at fixed participation. Noncommutativity permits it without making every economy adverse.
Proof and magnitudes for directional information
Section 9 states the isotropic model and Corollary 15. Basket profit depends only on the posterior-mean radius. The scalar equilibrium equation therefore applies at r = ∥ μ ∥ r=\lVert\mu\rVert r = ∥ μ ∥ , and (H.4) reduces to Ψ ( μ ) = ψ ( r ) \Psi(\mu)=\psi(r) Ψ ( μ ) = ψ ( r ) , giving the Hessian in (49) .
Proof of Corollary 15. Bayes’ rule gives E [ B j ∣ T j ] = ( ϵ / δ ) T j \mathbb{E}[B_j\mid T_j]=(\epsilon/\delta)T_j E [ B j ∣ T j ] = ( ϵ / δ ) T j and E [ B 3 − j ∣ T j ] = 0 \mathbb{E}[B_{3-j}\mid T_j]=0 E [ B 3 − j ∣ T j ] = 0 , hence the stated posterior means, each with probability one-half. Conditional on either signal, the posterior variance trace falls from 2 δ 2 2\delta^2 2 δ 2 to 2 δ 2 − ϵ 2 2\delta^2-\epsilon^2 2 δ 2 − ϵ 2 . The Λ \Lambda Λ -weighted reduction is also equal, at λ ϵ 2 \lambda\epsilon^2 λ ϵ 2 . Equation (H.4) cancels the unchanged target-second-moment term. Isotropy gives the two exact cost differences, and Taylor expansion yields the radial expression. Strict decrease of ψ \psi ψ on the positive half-line, from Proposition 6 , proves the finite tangential inequality. At smooth positive states the scalar proof also gives ψ ′ ( r ) < 0 \psi'(r) < 0 ψ ′ ( r ) < 0 , so the tangential eigenvalue in (49) is negative.
The signals need not be Blackwell comparable to each other. Each refines the same coarse information. The result shows why the amount of forecast-error reduction alone does not determine execution value across portfolio directions. A rotation to n = ( 1 , 1 ) T / 2 n = (1,1)^{\mathsf{T}}/\sqrt{2} n = ( 1 , 1 ) T / 2 and v = ( 1 , − 1 ) T / 2 v = (1,-1)^{\mathsf{T}}/\sqrt{2} v = ( 1 , − 1 ) T / 2 gives a common-basket direction and a spread direction without changing the isotropic economy.
Use the lognormal cost law and execution coefficients of Proposition 7 , with r ≈ .825060078573 r \approx.825060078573 r ≈ .825060078573 and δ = .02 \delta= .02 δ = .02 . The radial and tangential Hessian eigenvalues are approximately 4.40038637 and − .41909807 -.41909807 − .41909807 . The common coarse bill is .351308300279, including both target shocks. Table IA.11 reports direct high-precision evaluations of the original cash objective. The existing interval certificate proves the local radial sign, while these finite magnitudes are numerical evaluations.
Table IA.11. Two signals under one unchanged target law. Coarse MSE is 0.0008. Both signals attain the reported fine MSE. Cost changes use normalized model units. Percent changes divide by the same coarse institutional bill, including both shocks.
Figure IA.3. Direction of information and execution value, with ϵ = .01 \epsilon= .01 ϵ = .01 . Left: the four target states and the two alternative pairs of posterior means, shown relative to the coarse mean. Right: radial and tangential changes in the original institutional bill. Both signals reduce MSE by 0.0001. The economy is constructed from the smooth-interior example, without fitting market data.
Accounting and identification checks
The error in substituting intraday variation for a daily return square
Taylor expansion of log ( 1 + β R ) − β log ( 1 + R ) \log(1+\beta R)-\beta\log(1+R) log ( 1 + β R ) − β log ( 1 + R ) cancels the linear term and gives (2) . For fixed β \beta β and ∣ R ∣ ≤ r |R| \le r ∣ R ∣ ≤ r with max ( 1 , ∣ β ∣ ) r < 1 \max(1,|\beta|)r < 1 max ( 1 , ∣ β ∣ ) r < 1 , the third derivative is bounded, so the remainder is uniformly O ( ∣ R ∣ 3 ) O(|R|^3) O ( ∣ R ∣ 3 ) . Applying the same calculation at each five-minute reset instead gives − β ( β − 1 ) ∑ k r k 2 / 2 -\beta(\beta-1)\sum_k r_k^2/2 − β ( β − 1 ) ∑ k r k 2 /2 plus the sum of the third-order remainders. The two approximations target different portfolios, even if their first-order benchmark exposures agree at the start.
In the numerical round trip, the benchmark rises eight percent in log-price units and returns to its initial value. The endpoint daily-reset multiplier is exactly one, while the repeated-reset multiplier is ∏ k ( 1 + β r k ) \prod_k(1+\beta r_k) ∏ k ( 1 + β r k ) . This is an accounting example selected independently of the empirical outcomes, with no claim about measured aggregate investor losses. The same-path calculation isolates reset accounting. The empirical target calculation uses the actual preceding-session reference.
What a price-driven target forecast measures
Let R 0 , t R_{0,t} R 0 , t be the return from the previous reset reference to time t t t and let r t , T r_{t,T} r t , T be the remaining simple return. The exact composition is
R 0 , T = R 0 , t + ( 1 + R 0 , t ) r t , T . R_{0,T}=R_{0,t}+(1+R_{0,t})r_{t,T}. R 0 , T = R 0 , t + ( 1 + R 0 , t ) r t , T .
Under the idealized mechanical rule, a forecast r ^ t , T \hat{r}_{t,T} r ^ t , T implies
d ^ t = β ( β − 1 ) { R 0 , t + ( 1 + R 0 , t ) r ^ t , T } , \hat{d}_t=\beta(\beta-1)\{R_{0,t}+(1+R_{0,t})\hat{r}_{t,T}\}, d ^ t = β ( β − 1 ) { R 0 , t + ( 1 + R 0 , t ) r ^ t , T } ,
per assumed initial dollar, with forecast error β ( β − 1 ) ( 1 + R 0 , t ) ( r t , T − r ^ t , T ) \beta(\beta-1)(1+R_{0,t})(r_{t,T}-\hat{r}_{t,T}) β ( β − 1 ) ( 1 + R 0 , t ) ( r t , T − r ^ t , T ) . This transforms remaining-return uncertainty into mechanical target uncertainty under stated normalization. It does not include uncertainty about fund assets, flows, derivatives routing, or prior executions, and multiple leverage labels applied to the same benchmark return do not create independent observations.
Data and empirical methods
Common sample and product scope
All five-minute histories begin on November 9, 2021 and end on November 13, 2026 for this study. The start is the latest first observation among the 26 geared ETFs and five benchmark proxies in the sample. Earlier observations are excluded from normalization and estimation. Table IA.12 records this common sample boundary, rather than each product’s inception or first raw observation. The main forecast exercise uses SPY, QQQ, IWM, and DIA. The joint-market extension adds SOXX, with SOXL and SOXS included in the daily volume and factor analysis.
The family assignments use issuer benchmark descriptions [13 ] (ProShares, Direxion, Invesco, Black-Rock, and State Street 2026). A proxy ETF’s market return need not equal its benchmark index return. The semiconductor extension uses SOXX as an observable sector proxy and does not treat USD/SSG as exact SOXX exposures. Products in commodity, rates, international, and unrelated sector families do not enter these comparisons. Geared products sharing a benchmark are not additional independent realizations of that benchmark’s return.
Time, reference prices, and data provenance
All UTC timestamps are converted to America/New_York. The exchange calendar supplies session dates and actual closing times, including ten early closes in each core series. The primary convention treats labels as interval starts, making the last available bar at noon the 11:55 bar and at 15:30 the 15:25 bar. The final regular reference is the 15:55 bar on a normal session and the 12:55 bar on an early close. The preceding reset reference always comes from the immediately preceding exchange session, including a half-day, rather than from the previous eligible full day. Normal-close days alone enter the main outcome samples.
Table IA.12. Table IA.12. Geared products and their common five-minute sample. Every row starts on November 9, 2021 and ends on November 13, 2026. The same start applies to the five benchmark proxies. The semiconductor family is used in the extensions.
The five-minute records contain prices, volume, weighted average price, and transaction count. The later historical record takes precedence at common symbol–timestamp keys. The original extraction settings and absolute volume units are undocumented, so this panel is used for price features, relative activity, and within-window weights. A valid bar has positive, internally ordered OHLC prices and positive count. Weighted-price windows additionally require positive volume and average price within the OHLC range, allowing 0.02 basis points for rounding. These rules are applied uniformly. The separately sourced daily data in Appendix § 11 identify share and dollar units directly.
The end-stamped sensitivity shifts every relevant label forward by five minutes, including prior-session references, decision marks, and outcome windows. It tests an alternative provisional alignment. It does not imply that both alignments are correct. Neither convention identifies an official auction print, NAV, or institutional fill. Prices may also reflect distributions, splits, and historical filtering that are undocumented in the records.
Forecast definitions and chronological evaluation
Let P − P_{-} P − be the previous session’s final regular reference, O O O today’s first regular open, P t P_t P t the completed noon mark, and P T P_T P T the final regular reference. The response is Y = 10 4 ( P T / P t − 1 ) Y = 10^{4}(P_T/P_t - 1) Y = 1 0 4 ( P T / P t − 1 ) . The return block contains 10 4 ( P t / P − − 1 ) 10^{4}(P_t/P_{-} - 1) 1 0 4 ( P t / P − − 1 ) and 10 4 ( O / P − − 1 ) 10^{4}(O/P_{-} - 1) 1 0 4 ( O / P − − 1 ) . The range block adds
v t = 10 4 ∑ b ≤ t ( log C b − log C b − 1 ) 2 , h t = 10 4 ∑ b ≤ t ( log H b − log L b ) 2 4 log 2 , v_t = 10^{4}\sqrt{\sum_{b \le t}(\log C_b-\log C_{b-1})^2}, \qquad h_t = 10^{4}\sqrt{\frac{\sum_{b \le t}(\log H_b-\log L_b)^2}{4\log2}}, v t = 1 0 4 b ≤ t ∑ ( log C b − log C b − 1 ) 2 , h t = 1 0 4 4 log 2 ∑ b ≤ t ( log H b − log L b ) 2 ,
where C 0 = O C_0 = O C 0 = O , and the close’s location in the morning high-low span, centered at one half. Both scales have a one-basis-point floor. The high-low statistic is a feature, not an automatically unbiased variance estimator under microstructure noise [25 ] . Activity is the log ratio of cumulative count, volume, or volume/count to its previous 60-observation same-clock median. All 60 observations must exist. Volume and count-plus-size are alternative specifications because their logged levels have an exact accounting dependence.
For each family, the model uses the preceding 252 eligible observations. Training-only centering and scaling precede ridge regression, with an unpenalized intercept and penalty 0.1 ∥ w ∥ 2 0.1\lVert w\rVert^2 0.1 ∥ w ∥ 2 added to mean squared error. We retain a historical-mean benchmark and zero remaining return. The training window, feature order, and penalty are fixed rather than chosen after examining 2026 outcomes. The reference evaluation is 2023–2025 and the temporal check is January 2–November 13, 2026. Exploratory 2022–2025 analyses informed the design, and some 2026 prices were examined before the specification was fixed. This chronology supports a retrospective temporal comparison, not a wholly untouched holdout or preregistration.
All forecasts use the same family-dates. Raw pooled RMSE ratios divide by the zero forecast’s RMSE, while paired differences use mean squared loss in squared basis points.
Table IA.13. Annual raw RMSE ratios relative to zero remaining return, on common four-family dates.
Table IA.14. Paired incremental mean-forecast loss. MSE differences and intervals have units of squared basis points. Each row compares the indicated forecast with its baseline on identical dates.
For scale evaluation, symmetric zero-centered bands use the 241st order statistic of 252 prior absolute scores. Raw scores are ∣ Y ∣ |Y| ∣ Y ∣ , range scores are ∣ Y ∣ / h t |Y|/h_t ∣ Y ∣/ h t , and RV scores are ∣ Y ∣ / v t |Y|/v_t ∣ Y ∣/ v t . Multiplication by the current scale gives the half-width. If these scores were iid and continuous, their rank benchmark would be 241/253 = 95.26%. The financial scores need not satisfy that assumption, and we claim no distribution-free conditional guarantee. With α = .05 \alpha= .05 α = .05 , the interval score is
IS α ( l , u ; y ) = ( u − l ) + 2 α ( l − y ) 1 { y < l } + 2 α ( y − u ) 1 { y > u } , \operatorname{IS}_{\alpha}(l,u;y) = (u-l) + \frac{2}{\alpha}(l-y)1_{\{y<l\}} + \frac{2}{\alpha}(y-u)1_{\{y>u\}}, IS α ( l , u ; y ) = ( u − l ) + α 2 ( l − y ) 1 { y < l } + α 2 ( y − u ) 1 { y > u } ,
following [22 ] . Lower values favor narrow intervals while penalizing misses. Low and high range states use each family’s preceding 252-observation median, without a future threshold. State coverage and width weight family-date observations equally. The interval-score difference used for inference averages the available state observations within each date first, so a state-specific difference need not equal the difference of pooled state scores.
Table IA.15. Annual interval calibration. Low and high refer to the past-only range state. Scores are basis points.
Inference samples 10,000 circular blocks of five normal exchange sessions within each reported period, preserving the original calendar grid and missingness masks. The same date draws carry every family. Primary forecasts use complete four-family dates, and pooled close diagnostics use complete SPY/QQQ/IWM dates. Individual proxy and subgroup statistics retain their own eligible masks. Intervals are percentile intervals for the declared date-block approximation, not finite-sample coverage guarantees or a familywise testing procedure. The replication outputs give block lengths 1, 10, and 20 for the principal paired comparisons, all predictions and training endpoints, and the state and date counts. Table IA.16 adds the direct paired range–RV comparison using the same draws. The reference calendar contains 744 normal sessions, with 27 warm-up dates retained as missing, and the 2026 calendar contains 219. The seed is 20261115 + n n n + 100L L L , where n n n is calendar length and L = 5 L = 5 L = 5 . § 9.1 .§ 1.2 examines concentration without changing the primary sample.
Table IA.16. Paired interval-score differences on common four-family dates. Negative values favor the first method. Units are basis points of prediction loss. Intervals are individual retrospective 95 percent date-block intervals.
Close-related price movements
Direction, windows, and paired clock comparison
The event direction is the sign of P 15 : 30 / P − − 1 P_{15:30}/P_{-}-1 P 15 : 30 / P − − 1 , using the completed 15:30 mark under the primary convention. It is fixed before either price-response window. For both positive and negative leverage objectives, β ( β − 1 ) > 0 \beta(\beta-1)>0 β ( β − 1 ) > 0 , so inverse products do not reverse the inferred direction of idealized underlying reset demand. Define A A A as the volume-weighted average field over 15:50–16:00 and P P P as its counterpart over 16:10–16:30. Every bar must have positive count, positive volume, valid prices, and an in-range average. These window calculations assume that average is a WAP and relative volume weights are meaningful.
The pre, post, and cumulative responses are respectively 10 4 s log ( A / P 15 : 30 ) 10^{4}s\log(A/P_{15:30}) 1 0 4 s log ( A / P 15 : 30 ) , 10 4 s log ( P / A ) 10^{4}s\log(P/A) 1 0 4 s log ( P / A ) , and their sum. The earlier-clock placebo fixes its own direction at 13:30, uses a 13:50–14:00 anchor, and a 14:10–14:30 post window. The clock difference pairs eligible dates and preserves both directions in the bootstrap. It is an observational clock contrast with different information sets, not an exogenous treatment effect.
Nasdaq announced rollout of its Extended Trading Close on March 7, 2022, allowing eligible Nasdaq-listed interest to match at the closing-cross price through 16:05 [30 ] . We separate the immediate neighborhood from the main buffered window rather than interpreting a 16:00 bar as an auction. The qualification is venue- and date-specific. The game itself ends at terminal completion and does not predict a post-16:00 decay path.
Table IA.17. Predetermined-direction close diagnostics. Pre means use the same dates as the corresponding post mean. n p n_p n p counts dates eligible for both clocks. Intervals use date blocks.
Freshness, anchors, and robustness
A shared noisy closing endpoint enters the pre return positively and the post return negatively. Their covariance can therefore be negative even without economic reversal. Predetermined direction, distinct window averages, buffered post windows, and cumulative responses reduce the most direct endpoint-selection problem without eliminating all microstructure noise.
For the high-low sensitivity, take the envelope [ L A , H A ] [L_A,H_A] [ L A , H A ] of the 15:55 and 16:00 labeled bars. Given a fresh post reference P P P and predetermined sign s s s , every anchor in the envelope implies a signed post move between
min a ∈ { L A , H A } 10 4 s log ( P / a ) and max a ∈ { L A , H A } 10 4 s log ( P / a ) . \min_{a\in\{L_A,H_A\}} 10^{4}s\log(P/a) \quad\text{and} \quad\max_{a\in\{L_A,H_A\}} 10^{4}s\log(P/a). a ∈ { L A , H A } min 1 0 4 s log ( P / a ) and a ∈ { L A , H A } max 1 0 4 s log ( P / a ) .
A negative upper bound establishes a move opposite the earlier direction for every price in this recorded envelope. It is an auction-price bound only if the official auction price is known to be contained, a premise the recorded bars do not establish. It also does not require a positive preceding close move.
Figure IA.4. Equal-family SPY/QQQ/IWM responses and individual 95 percent date-block intervals. Pre, post, and cumulative use common eligible dates. The clock contrast additionally requires the placebo window. No realized closing return is used to select the direction.
Table IA.18. Additional references and anchor sensitivities on common SPY/QQQ/IWM dates. Each horizon additionally requires the main post window, so counts can differ across rows. End-label and last-price sensitivities are paired with the primary eligibility rule.
Additional diagnostics retain year and sign splits, the large-move group defined by the prior 252-observation 80th percentile with at least 126 observations, leave-one-family-out and leave-one-year-out comparisons, and missingness by earlier direction, range, and activity. Half-days are excluded from the 16:00 exercise. Their small and holiday-selected sample does not provide an exogenous closing-time experiment. The next-session 10:00 reference additionally spans overnight news and potentially distributions, and remains secondary.
Predetermined activity and path shape
A descriptive regression relates the unsigned post-window log return in basis points to the earlier reset-period return x x x in percentage points, log relative count z z z , x z xz x z , range scale, overnight return, recent 30-minute return, and family indicators. All regressors are available by 15:30. The last three return/scale controls are in percentage points. Three thousand five-session date-block draws give the interaction interval in Table IA.19. The specification keeps activity from the outcome window out of its own explanation.
Table IA.19. Predetermined return–activity interaction. Coefficients are basis points of post-window return per percentage point of earlier return times one log unit of relative count.
Table IA.19.
A further descriptive breakdown compares path efficiency ∣ log ( P t / O ) ∣ / ∑ b ∣ Δ log C b ∣ \lvert\log(P_t/O)\rvert/\sum_b\lvert\Delta\log C_b\rvert ∣ log ( P t / O )∣ / ∑ b ∣ Δ log C b ∣ , using denominator floor 10 − 8 10^{-8} 1 0 − 8 , above and below 0.25, within fixed absolute-return bins of 0–50, 50–100, and more than 100 basis points and past-only range states. These small cells are reported in the replication outputs, not selected to support a mechanism. Transaction count is not the number of strategic entrants, high-low range is not a bid-ask spread, and gross activity does not identify F F F , N N N , or causal inverse depth. Historical window prices likewise do not identify executable profits or the counterfactual impact of a changed institutional schedule.
Additional empirical evidence
Forecast stability across families and time
Pooled forecast accuracy can obscure both family differences and brief periods of apparent predictability. Table IA.20 gives each family’s RMSE relative to the zero remaining-return forecast on the common sample. Figure IA.5 repeats the comparison in trailing 126-session windows. Every point uses forecasts already formed from earlier training observations. The moving evaluation window is descriptive and does not select a model for subsequent trading.
Table IA.20. Family-level RMSE ratios on the common evaluation dates. A ratio below one favors the fitted forecast. All predictors and training windows are defined in Section J.3.
Table IA.20.
Uncertainty calibration beyond a pooled score
Methods and calendar masks are defined in Section J.3. The range-minus-raw mean score difference is −7.90 basis points in 2023, +2.72 in 2024, and −119.30 in 2025. Removing 2025 leaves a mean of −2.27 basis points. Thus the reference-period average largely reflects 2025 rather than a uniform annual advantage.
The interval score is full width plus forty times the exceedance beyond the nearest endpoint. In 2023–2025, the 42.59-basis-point improvement divides into 18.90 from narrower bands and 23.68 from reduced exceedance penalties. The corresponding components in 2026 are 7.82 and 18.89.
Figure IA.5. Trailing 126-session RMSE ratios. The dotted line is the zero-forecast benchmark. These overlapping evaluation windows are highly dependent. Short periods below one do not establish a stable forecasting advantage.
Figure IA.6 displays this decomposition and the largest absolute-contribution date in the reference period. On April 9, 2025, the four remaining returns range from 803 to 1,123 basis points. The raw bands have mean full width 254.61 and the range bands 849.60 basis points. All four range bands still miss. Their smaller losses reflect reduced exceedance, not containment of the stress move.
Figure IA.6. Sources of interval-score improvement. Left: mean raw-minus-range width and exceedance-penalty differences, in basis points of prediction loss. Right: realized noon-to-close returns and symmetric half-widths on April 9, 2025. Every return exceeds the range band’s upper endpoint. The date is selected retrospectively by its absolute contribution.
Table IA.21 ranks dates by the absolute four-family mean score difference, allowing both favorable and adverse dates to enter. April 9, 2025 contributes 37.02% of the reference period’s net gain. June 11, 2026 contributes 30.38% of the temporal period’s gain. Omitting the five largest absolute contributions leaves negative point differences in both periods, but both intervals include zero. These omissions are conditional sensitivity summaries following outcome-based selection. They are not selection-adjusted tests or preferred trimmed estimates. Every valid date remains in the primary comparison.
Table IA.21. Table IA.21. Concentration by date. Share is the selected dates’ contribution divided by the full period’s net gain. The last columns omit those dates while retaining their calendar positions as missing. The 95 percent intervals condition on the retrospectively selected omission set.
Table IA.22 applies (52) to all five geared objectives. Positive affine transformation preserves coverage observation by observation and scales squared prediction error by k t 2 k_t^2 k t 2 . Averaging k t k_t k t first would generally give a different result. The observed target component is included in the center, even though the remaining-return forecast is zero.
Table IA.22. Table IA.22. Normalized target units across leverage objectives. Known component is the mean absolute already-observed target. Other entries are mean interval half-widths. All entries are percent of hypothetical initial assets.
The interval comparison keeps the center fixed and asks whether morning variation improves the scale estimate. Figure IA.7 reports the date-level range-minus-raw score difference through time. The upper panel uses a trailing quarter, while the lower panel averages all evaluation dates available up to each point. A negative value favors range scaling. The local series can change sign even when the overall mean is negative. This distinguishes an average improvement from uniform performance across market conditions.
A finer state classification ranks today’s morning range against the preceding 252 observations for the same family. Fixed rank intervals of width 0.2 define five states without using future outcomes. The replication results retain state-level coverage, counts, and widths. These are descriptive summaries and do not assume independent observations across states.
Figure IA.7. Temporal stability of the interval-score difference. Families receive equal weight each day. Both panels report prediction loss, not trading returns. The expanding calculation starts at the first common evaluation date.
Range-scaled intervals widen on mornings with large price movements, but large remaining returns can still exceed them. An overall coverage rate also leaves the dependence between misses unspecified. Figure IA.8 examines the concentration of score improvement by omitting each family and, in the reference sample, each year in turn. These comparisons leave the original forecast and primary sample definition unchanged.
Intraday activity
Figure IA.9 reports median intraday shares of regular-session activity. Each day’s volume and count are divided by that day’s corresponding total. The full-day denominator is used only for this descriptive clock profile, never for a trading decision or noon predictor. Volume and count can have different shapes because average transaction size changes through the day. Neither observable estimates the model’s participation mass.
Close-related paths, annual variation, and anchor uncertainty
The main close comparison uses two prespecified windows. Figure IA.10 adds the intervening path while keeping direction fixed at 15:30. Each family uses dates with valid, positive-count bars throughout the plotted interval. This balanced path sample is selected for descriptive comparability and is not an executable strategy sample. Its last-price observations also differ from the window-weighted references used in the principal tables.
Annual estimates in Figure IA.11 expose variation concealed by a long-period mean. The intervals preserve within-family serial dependence through date blocks. They do not identify a causal effect. of leveraged-fund demand, and their number makes isolated exclusions of zero weak evidence without an explicit multiple-testing design. DIA’s estimates use their own eligible event sample.
Figure IA.8. Influence of individual families and years on the interval-score comparison. Whiskers are individual 95 percent circular five-session block intervals. The omitted family or year is removed before averaging the remaining families by date. These overlapping sensitivity comparisons are not separate confirmatory tests.
Figure IA.9. Median within-day activity shares on normal sessions, 2022–2026. Each five-minute bar’s activity is divided by its regular-session total before taking the median across dates. A median profile need not sum to one. Units cancel within each day.
The joint distribution of the pre and post legs also deserves care. They share the close reference with opposite signs. Noise in this reference can generate negative covariance even without an economic reversal. Figure IA.12 reports the observed distribution but assigns no structural interpretation to its slope. Separating decision, reference, and post windows limits overlapping trades, while neither that separation nor the scatterplot identifies the efficient close price.
Finally, Figure IA.13 asks whether the sign survives every reference in the high-low envelope of the two bars around 16:00. The middle category contains observations for which the inferred sign depends on the chosen reference. The envelope is a sensitivity set built from recorded bars. It is not a confidence interval for the official auction price and need not contain that price.
Figure IA.10. Signed last-price paths from the completed 15:30 mark through 17:00. The sign is the earlier reset-period return. Shading gives individual pointwise 95 percent date-block intervals. Sample counts are fixed along each family path. The vertical line marks 16:00 and does not identify an official auction observation.
Figure IA.11. Annual signed moves from the 15:50–16:00 reference to the 16:10–16:30 reference. Whiskers are individual 95 percent five-session block intervals. Horizontal offsets distinguish families within each year.
Figure IA.12. Joint distribution of signed pre and post legs for SPY, QQQ, and IWM. Hexagon intensity is logarithmic in observation count. Both legs use the same 15:30 direction and share the close reference, so negative association alone cannot establish reversion.
Figure IA.13. Sensitivity of the signed post move to the close reference. Blue and red indicate a negative or positive move for every reference in the high-low envelope. Gray indicates that the sign changes within the envelope. Eligibility requires fresh envelope bars and the primary post window.
Econometric robustness
Serial dependence and multiple comparisons
The loss comparisons retain the same dates and forecasts as Table 5 . For each method pair, let d i t d_{it} d i t be its interval-score difference for family i i i and let d t = 1 4 ∑ i d i t d_t = \frac{1}{4}\sum_i d_{it} d t = 4 1 ∑ i d i t . Equal weighting within a date prevents four correlated market observations from being treated as four independent experiments. The sign convention is always first method minus second method, so negative values favor the first.
We supplement the circular-block intervals with a Bartlett-kernel HAC covariance estimate [33 ] . For a vector of date-level moments g t g_t g t , define u t = g t − g ˉ u_t = g_t - \bar{g} u t = g t − g ˉ and
Ω ^ L = 1 n ∑ t u t u t T + 1 n ∑ k = 1 L ( 1 − k L + 1 ) ∑ t = k + 1 n ( u t u t − k T + u t − k u t T ) . (69) \widehat{\Omega}_L = \frac{1}{n}\sum_t u_tu_t^{\mathsf{T}} + \frac{1}{n}\sum_{k=1}^{L}\left(1-\frac{k}{L+1}\right)\sum_{t=k+1}^{n}\left(u_tu_{t-k}^{\mathsf{T}} + u_{t-k}u_t^{\mathsf{T}}\right). \tag*{(69)} Ω L = n 1 t ∑ u t u t T + n 1 k = 1 ∑ L ( 1 − L + 1 k ) t = k + 1 ∑ n ( u t u t − k T + u t − k u t T ) . (69)
The estimated covariance of g ˉ \bar{g} g ˉ is Ω ^ L / n \widehat{\Omega}_L/n Ω L / n . The main lag is ten evaluation sessions, with five and twenty as sensitivities. Dates are ordered within each reported evaluation period. These calculations allow contemporaneous dependence across families and short-range dependence across dates. Their normal and chi-square references are asymptotic approximations, requiring suitable dependence and moment conditions. Heavy-tailed losses and a short temporal sample remain relevant to their accuracy.
Table IA.23 reports two-sided tests with Holm adjustments over six declared hypotheses: the three interval-method contrasts in each of the two periods. We apply the adjustment separately at each lag. The lag choices are sensitivity checks, not independent confirmations. The reference-period range-versus-RV difference remains significant after adjustment. The range-versus-raw improvements are less decisive under HAC inference than under the percentile bootstrap. The reference-period HAC interval includes zero. This disagreement is consistent with the concentration of score gains in Appendix § 9.1 .§ 1.2 and limits the strength of that comparison.
Table IA.23. Interval-score comparisons with ten-session HAC inference. Means and intervals are in basis points of prediction loss. Holm adjustment covers all six rows.
Conditional loss and calibration moments
A useful forecast comparison asks whether observable conditions predict the difference in losses, not merely whether its unconditional mean is zero [21 ] . Let H i t H_{it} H i t indicate a morning range above its own preceding 252-observation median. For each method pair we test
E [ 1 4 ∑ i d i t ( H i t − 1 2 ) ] = 0. \mathbb{E}\left[\frac{1}{4}\sum_i d_{it}\left(H_{it}-\frac{1}{2}\right)\right] = 0. E [ 4 1 i ∑ d i t ( H i t − 2 1 ) ] = 0.
Figure IA.14. Sensitivity to the HAC lag. Whiskers are individual asymptotic 95 percent intervals, before multiple-testing adjustment. The same forecasts and dates enter every lag choice.
The two-moment Wald statistic uses (M.1) with L = 10 L = 10 L = 10 . A rejection means that these moments distinguish the forecast procedures. It does not imply that one method dominates at every state. The reference-period range-versus-RV contrast and temporal range-versus-raw contrast reject after the six-comparison adjustment in Table IA.24.
Table IA.24. Conditional predictive-loss moments. The instruments are a constant and the observable morning range state. The reference distribution is χ 2 2 \chi^2_2 χ 2 2 , and Holm adjustment covers six tests.
For calibration, write I i t = 1 { ∣ Y i t ∣ > h i t } I_{it} = 1\{|Y_{it}| > h_{it}\} I i t = 1 { ∣ Y i t ∣ > h i t } for a miss by the relevant symmetric band. We test moments of I i t − .05 I_{it} - .05 I i t − .05 against three instruments: a constant, H i t − 1 / 2 H_{it} - 1/2 H i t − 1/2 , and the preceding evaluation date’s average absolute noon-to-close return, expressed in percentage points and centered within the evaluation period. The last instrument is already known at today’s noon decision. The first date of each period is excluded from this particular test, leaving 716 and 218 dates. All four family contributions are averaged before forming the HAC covariance. The joint null is assessed with three degrees of freedom, and Holm adjustment covers the three bands in each of the two periods.
The raw band rejects these calibration moments in both periods. Range scaling does not reject after the six-test adjustment, with adjusted p p p -values 0.083 and 0.814. Thus the tests find departures from the specified raw-band restrictions but insufficient evidence to reject the range-band restrictions. Failure to reject does not establish conditional validity, independent errors, or an exact 95 percent guarantee.
Training-window sensitivity
The scale comparison is repeated with 126, 252, and 504 preceding eligible observations. For window length n n n , the empirical quantile is order ⌈ .95 ( n + 1 ) ⌉ \lceil.95(n + 1)\rceil ⌈ .95 ( n + 1 )⌉ . Every method and window is evaluated on identical four-family dates. The longest window determines the first comparison date, February 15, 2024, leaving 465 reference dates and 219 temporal dates. Thus the differences in Table IA.26 do not reflect changing evaluation composition.
Table IA.25. Conditional coverage moments. Coverage is calculated on the moment-test sample, which excludes the first date of each period. W 3 W_3 W 3 tests a constant, the morning range state, and lagged market movement.
Range scaling has a lower point score than the raw band for every window. The reference-period inference remains sensitive to the window and tail observations. These are fixed design sensitivities, not a selection rule for choosing the most favorable window. The 252-observation specification remains the primary one.
Table IA.26. Alternative training windows on common evaluation dates. Coverage is in percent and score differences are range minus raw, in basis points. The HAC intervals are individual ten-session intervals.
Joint benchmark forecasts
Own-market and joint conditional means
We study two systems: QQQ/SPY/IWM and QQQ/SOXX. The first captures common broad-equity variation with a small-cap component. The second contrasts Nasdaq-100 and semiconductor exposure. Each equation’s own-market predictor vector contains elapsed and overnight return, close-return RV, high-low range, close location, relative count, and relative transaction size. The joint specification adds the same completed-noon features from every other market in its system. Both use the preceding 252 common eligible dates, training-only standardization, an unpenalized intercept, and the same ridge penalty 0.1. The joint specification has more predictors, but no contemporaneous afternoon outcome or future normalization enters training.
The two systems are reported on common evaluation dates beginning February 13, 2023. Each own-versus-joint comparison uses identical target observations. The own and joint columns in Table IA.27 are pooled raw RMSE divided by the zero-forecast RMSE. For inference, squared-error differences are divided by each equation’s training-window target variance, then averaged across markets within the date. The resulting four period-by-system tests use ten-session HAC inference and a four-test Holm adjustment.
Joint mean forecasts do not improve point RMSE in either system or period. In the three-market reference comparison, the standardized loss increase remains significant after adjustment. These results give no empirical basis for using the larger predictor set as a return-trading rule. A different forecast procedure would require its own independent validation.
Joint uncertainty
The covariance comparison holds the mean at zero and scales each outcome by its own observed morning range. Let v t v_t v t be this vector and let
M t = 1 252 ∑ s = t − 252 t − 1 v s v s T , D t = diag ( M t ) , Σ t = .9 M t + .1 D t . M_t = \frac{1}{252}\sum_{s=t-252}^{t-1} v_s v_s^{\mathsf{T}}, \qquad D_t = \operatorname{diag}(M_t), \qquad\Sigma_t = .9M_t + .1D_t. M t = 252 1 s = t − 252 ∑ t − 1 v s v s T , D t = diag ( M t ) , Σ t = .9 M t + .1 D t .
The diagonal model uses D t D_t D t and the joint model uses Σ t \Sigma_t Σ t . Shrinkage is fixed at ten percent. The zero-mean Gaussian negative log score is
ℓ t ( Σ ) = 1 2 { p log ( 2 π ) + log det Σ + v t T Σ − 1 v t } . \ell_t(\Sigma) = \frac{1}{2}\{p\log(2\pi) + \log\det\Sigma+ v_t^{\mathsf{T}}\Sigma^{-1}v_t\}. ℓ t ( Σ ) = 2 1 { p log ( 2 π ) + log det Σ + v t T Σ − 1 v t } .
Restoring the original return units adds the same sum of log scales to both models, so the score difference is unchanged. These are working Gaussian scores for the joint distribution, not an assumption that observed tails are Gaussian. Their evaluation uses the realized out-of-sample return vectors.
Both rolling joint forecasts improve on diagonal covariance, but this comparison does not establish a benefit from updating correlations. We therefore add a matched frozen-correlation forecast. Let Q 0 Q_0 Q 0 be the uncentered second moment of the standardized outcomes on the initial 252 common eligible dates, February 9, 2022 through February 10, 2023, and let D 0 = diag ( Q 0 ) D_0 = \operatorname{diag}(Q_0) D 0 = diag ( Q 0 ) . Define
R 0 = D 0 − 1 / 2 Q 0 D 0 − 1 / 2 , Σ t f r o z e n = D t 1 / 2 ( .9 R 0 + .1 I ) D t 1 / 2 . (70) R_0 = D_0^{-1/2}Q_0D_0^{-1/2}, \qquad\Sigma_t^{\mathrm{frozen}} = D_t^{1/2}(.9R_0 + .1I)D_t^{1/2}. \tag*{(70)} R 0 = D 0 − 1/2 Q 0 D 0 − 1/2 , Σ t frozen = D t 1/2 ( .9 R 0 + .1 I ) D t 1/2 . (70)
The initial correlation estimate stays fixed throughout evaluation, including 2026. Both correlated models retain the same rolling marginal second moments, morning-range scaling, zero mean, ten-percent shrinkage, and evaluation dates. Only the correlation update differs. No evaluation outcome enters R 0 R_0 R 0 . We designed this additional benchmark retrospectively, without selecting an alternative correlation window or shrinkage grid.
Table IA.27 retains the conditional-mean comparison. Table IA.28 separates rolling and frozen dependence from independent marginals. Frozen correlation has the lower point score in both systems and periods. The rolling-minus-frozen differences for QQQ/SPY/IWM are inconclusive. For QQQ/SOXX, they are positive in both periods and remain significant after Holm adjustment over the four system-period comparisons. Each contrast has its own four-test family, with individual ten-session HAC intervals. Table IA.29 translates these covariance forecasts into two fixed portfolio directions.
Persistent correlation alone can produce a substantial improvement over independent marginals. If a zero-mean vector has covariance D 1 / 2 R D 1 / 2 D^{1/2}RD^{1/2} D 1/2 R D 1/2 , with R R R a positive-definite correlation matrix, then
E { ℓ ( D 1 / 2 R D 1 / 2 , v ) − ℓ ( D , v ) } = 1 2 log det R . (71) \mathbb{E}\{\ell(D^{1/2}RD^{1/2},v)-\ell(D,v)\} = \frac{1}{2}\log\det R. \tag*{(71)} E { ℓ ( D 1/2 R D 1/2 , v ) − ℓ ( D , v )} = 2 1 log det R . (71)
The two expected quadratic terms both equal the vector dimension, leaving the log determinant. This expectation identity uses the stated second moments and does not require Gaussian returns. The evidence supports accounting for dependence in the joint uncertainty description. It does not support additional predictive value from updating correlation in these specifications.
Return covariance and structural cross-impact describe different objects. The execution matrices in Appendix H determine how orders affect prices. Holding the posterior mean and execution environment fixed, covariance changes the conditional expected institutional bill through tr ( Λ V ) \operatorname{tr}(\Lambda V) tr ( Λ V ) while leaving equilibrium quantities unchanged. The joint forecast exercise describes target uncertainty and does not estimate cross-impact or activation.
Table IA.27. Own-market and joint conditional means, with the original covariance comparison. RMSE is relative to the zero-forecast RMSE. The final columns report rolling joint minus diagonal negative log score, in natural-log units per return vector. Negative scores favor joint covariance.
Table IA.28. Frozen and rolling correlation on matched dates. Both retain rolling marginal variances. Positive rolling-minus-frozen values favor the frozen estimate. Intervals and Holm values concern that final contrast. The intervals are individual, and the Holm family contains its four system-period tests. Scores are natural-log units per return vector.
Risks of a fixed sum and spread. For QQQ and SOXX, fix w + = ( .5 , .5 ) T w_{+} = (.5, .5)^{\mathsf{T}} w + = ( .5 , .5 ) T and w − = ( .5 , − .5 ) T w_{-} = (.5, -.5)^{\mathsf{T}} w − = ( .5 , − .5 ) T . Let y t y_t y t contain unscaled remaining returns in basis points and S t S_t S t the diagonal of current completed-morning ranges. Return covariance is V ^ t = S t Σ ^ t S t \widehat{V}_t = S_t\widehat{\Sigma}_tS_t V t = S t Σ t S t . Table IA.29 compares observed RMS { ( w T y t ) 2 } 1 / 2 \{(w^{\mathsf{T}}y_t)^2\}^{1/2} {( w T y t ) 2 } 1/2 with forecast RMS { w T V ^ t w } 1 / 2 \{w^{\mathsf{T}}\widehat{V}_tw\}^{1/2} { w T V t w } 1/2 . The latter is the square root of mean variance, not mean daily standard deviation. A diagonal forecast assigns equal variance to sum and spread. Dependence distinguishes the larger common-direction risk from the smaller spread risk.
Table IA.29. Fixed QQQ/SOXX directions on matched evaluation dates. Observed and forecast RMS values are basis points. The final columns are mean projected Gaussian quasi-score differences, with lower scores preferred. Weights are (.5, .5) for the sum and (.5, −.5) for the spread.
The projected score is { log ( 2 π s w , t 2 ) + ( w ⊤ y t ) 2 / s w , t 2 } / 2 \{\log(2\pi s_{w,t}^{2}) + (w^{\top}y_t)^2/s_{w,t}^{2}\}/2 { log ( 2 π s w , t 2 ) + ( w ⊤ y t ) 2 / s w , t 2 } /2 , where s w , t 2 = w ⊤ V ^ t w s_{w,t}^{2}=w^{\top}\widehat{V}_t w s w , t 2 = w ⊤ V t w . Frozen correlation improves on diagonal forecasts in both directions and periods. Rolling correlation scores slightly better than frozen correlation for the sum and worse for the spread, so rankings can differ by projection. These retrospective fixed-weight diagnostics introduce no additional significance tests, conditional-calibration claims, or trading rule. The return directions do not identify the model’s harmful or beneficial signals.
Share volume and dollar turnover
Measurement on a common daily panel
Daily unadjusted OHLC, share volume, VWAP, cash distributions, and splits are obtained from Massive [26 ] . The balanced panel contains all 31 ETFs on 499 dates from November 18, 2024 through November 13, 2026. This is the common available window for the external daily series. It is shorter than the five-minute study, and is used uniformly across every product in the daily turnover analysis.
Let V i t V_{it} V i t be reported shares traded and P i t V W A P P_{it}^{\mathrm{VWAP}} P i t VWAP the daily volume-weighted price. Dollar turnover is
T i t = V i t P i t V W A P . (72) T_{it}=V_{it}P_{it}^{\mathrm{VWAP}}. \tag*{(72)} T i t = V i t P i t VWAP . (72)
Share volumes retain each date’s actual share basis. They therefore change mechanically at a split or reverse split and should not be added across unrelated ETFs as a measure of economic exposure. Dollar turnover avoids that mechanical denomination effect. The data describe trades eligible for the provider’s daily aggregates. They are secondary-market trading measures, not fund creations, redemptions, derivative notionals, or underlying-basket executions.
Table IA.30 reports mean shares, mean dollars, and the upper tail of daily dollar activity for every ETF. Table IA.31 sums geared-product dollar volume within each family, compares it with benchmark-ETF dollar volume, and reports the inverse products’ share. Each family denominator is its proxy ETF’s turnover, rather than total trading in all underlying constituents. The dollar sums measure recorded trading across distinct securities and do not identify the number of distinct investors or independent risk transfers.
The Nasdaq family has mean geared turnover of 8.68 billion dollars per day, compared with 7.10 billion for SOXL and SOXS together. Relative to their benchmark ETF, the semiconductor products have the largest footprint, with a median ratio of 2.92 and a 95th percentile of 5.48. The Nasdaq median ratio is 0.30. The Russell family has the largest average inverse share, about 59.3%. These comparisons establish differences in secondary-market scale. They do not convert turnover into the committed demand D D D .
Activity, market moves, and common shocks
Define relative dollar activity using a strictly lagged denominator,
Z i t = log ( T i t 20 − 1 ∑ s = t − 20 t − 1 T i s ) . Z_{it}=\log\left(\frac{T_{it}}{20^{-1}\sum_{s=t-20}^{t-1}T_{is}}\right). Z i t = log ( 2 0 − 1 ∑ s = t − 20 t − 1 T i s T i t ) .
We relate Z i t Z_{it} Z i t to its benchmark’s absolute daily total return, the negative part of that return, benchmark relative dollar activity, and the interaction of the negative return with an inverse-product indicator. Returns enter in percentage points. ETF fixed effects account for persistent product differences. The first twenty daily observations supply the denominator, leaving a balanced 479-date regression panel. Full-sample, broad-family, and semiconductor regressions use the same dates. A fourth specification removes both ETF and date means from the dependent variable and all regressors, thereby absorbing arbitrary common daily shocks.
Table IA.30. Daily share and dollar volume for all 31 ETFs. Shares and dollars are in millions. P95 is the empirical 95th percentile across the same 499 dates. Raw share quantities retain the contemporaneous share denomination.
Table IA.31. Geared trading by benchmark family. The geared/benchmark ratios are calculated date by date before taking their median or P95. Inverse share is the average date-level fraction of geared dollars.
The covariance calculation first sums each date’s regression scores across products, then applies the ten-session Bartlett weights. This allows arbitrary within-date dependence across ETFs and serial dependence in the aggregated scores. Table IA.32 reports Holm adjustments over the sixteen displayed slope tests. The regressions describe contemporaneous associations. They do not treat today’s return and activity as information available before today’s trading.
Figure IA.15. Twenty-session averages of geared/benchmark dollar-turnover ratios and inverse shares. The left axis is logarithmic. All families use the same daily panel.
Benchmark activity is strongly associated with geared-product activity. After controlling for it, the associations with the magnitude and sign of benchmark returns are smaller and more sensitive to specification. These regressions do not identify how institutional reset orders respond to a benchmark move. The next decomposition separates changes in share activity from changes in price denomination.
Table IA.32. Daily geared-product activity regressions. The outcome is log dollar volume relative to the preceding twenty-day mean. “ETF and date effects” is the within-ETF, within-date specification. Standard errors use date-aggregated HAC scores.
Separating share activity from price denomination
Dollar turnover combines shares and prices. Let c i t c_{it} c i t be the cumulative split ratio in new shares per initial share. Express daily volume and VWAP in a constant denomination as
V i t ∗ = V i t / c i t , P i t ∗ = c i t P i t V W A P , T i t = V i t ∗ P i t ∗ . V_{it}^{*}=V_{it}/c_{it},\qquad P_{it}^{*}=c_{it}P_{it}^{\mathrm{VWAP}},\qquad T_{it}=V_{it}^{*}P_{it}^{*}. V i t ∗ = V i t / c i t , P i t ∗ = c i t P i t VWAP , T i t = V i t ∗ P i t ∗ .
Using the same strictly lagged twenty-session denominator, define
Z i t V = log ( V i t ∗ 20 − 1 ∑ j = 1 20 V i , t − j ∗ ) , Z_{it}^{V}=\log\left(\frac{V_{it}^{*}}{20^{-1}\sum_{j=1}^{20}V_{i,t-j}^{*}}\right), Z i t V = log ( 2 0 − 1 ∑ j = 1 20 V i , t − j ∗ V i t ∗ ) ,
Z i t P = Z i t − Z i t V = log ( P i t ∗ ∑ j = 1 20 V i , t − j ∗ P i , t − j ∗ / ∑ j = 1 20 V i , t − j ∗ ) . (73) Z_{it}^{P}=Z_{it}-Z_{it}^{V}=\log\left(\frac{P_{it}^{*}}{\sum_{j=1}^{20}V_{i,t-j}^{*}P_{i,t-j}^{*}/\sum_{j=1}^{20}V_{i,t-j}^{*}}\right). \tag*{(73)} Z i t P = Z i t − Z i t V = log ( ∑ j = 1 20 V i , t − j ∗ P i , t − j ∗ / ∑ j = 1 20 V i , t − j ∗ P i t ∗ ) . (73)
The second component compares current split-adjusted VWAP with a lagged share-weighted price reference. It is neither simply the current return nor an identified causal price effect.
Table IA.33 uses exactly the same regressors, dates, and fixed-effect transformations for dollar activity, share activity, and the price component. OLS linearity makes every dollar coefficient the sum of the other two. There are 479 dates and 26 geared products in each specification. With ETF and date effects, the negative-return interaction for inverse products is 0.007181 = −0.058678 + 0.065859, up to rounding. Its small positive dollar-activity coefficient therefore combines a negative share-activity component with a positive price-related component. The decomposition is descriptive. It introduces no additional significance claims or causal interpretation.
Table IA.33. Table IA.33. Exact coefficient decomposition of relative dollar activity. Every row uses the same design for all three outcomes. Dollar activity equals share activity plus the price-related component. Return regressors remain in percentage points. No new coefficient-testing family is invoked.
Daily factor exposures and tracking
Factors, returns, and matching
We use the daily US Fama–French market, size, value, profitability, and investment factors, together with momentum and the value-weighted electronic-equipment industry portfolio from the Kenneth French Data Library [16 ] . These factors concern US listed equities and are appropriate broad risk descriptors for the equity benchmark families studied here. The industry series has code Chips, but its SIC definition includes electronic equipment beyond semiconductors. It is an additional exposure control, not a replication of the SOXX benchmark.
The factor files use the September 2026 CRSP database and end on September 30, 2026. Downloaded percentage returns are converted to decimal returns. For ETF i i i , let P i t P_{it} P i t be the unadjusted close, d i t d_{it} d i t the ex-date cash distribution per contemporaneous share, and s i t s_{it} s i t the new-shares-per-old-share split ratio, equal to one on ordinary dates. The total return is
r i t = s i t ( P i t + d i t ) P i , t − 1 − 1. (74) r_{it} = \frac{s_{it}(P_{it}+d_{it})}{P_{i,t-1}} - 1. \tag*{(74)} r i t = P i , t − 1 s i t ( P i t + d i t ) − 1. (74)
Cash amounts retain their ex-date basis. There are no simultaneous split and distribution events in this matched panel. The first available daily close supplies the denominator, so the common regression window is November 19, 2024 through September 30, 2026, with 466 return dates for every ETF. Each factor and ETF return is matched by trading date, without forward filling. Market-price total returns are used throughout, rather than substituting NAV returns or assuming a leveraged ETF equals a fixed multiple of its proxy.
The daily regressions measure ex-post exposure. Their same-day factor returns are unavailable at noon and never enter the intraday forecast experiment. The factor vintage incorporates the French library’s current historical construction, so it is not a point-in-time predictor dataset. All products use the same factor vintage and daily observation window.
Exposure regressions
For each ETF, estimate
r i t − r f t = α i + b i ′ f t + ε i t , (75) r_{it} - r_{ft} = \alpha_i + b_i^{\prime} f_t + \varepsilon_{it}, \tag*{(75)} r i t − r f t = α i + b i ′ f t + ε i t , (75)
where f t f_t f t contains the five Fama–French factors and momentum. The augmented model adds the electronic-equipment return minus the same risk-free return. The intercept and loadings are estimated by OLS, with ten-session HAC standard errors. Table IA.34 reports the six-factor intercept in basis points per day, its HAC statistic, market loading, and R 2 R^2 R 2 before and after the industry control. The industry loading comes from the augmented regression. All 31 products enter the same estimation dates.
For the benchmark ETFs, the six-factor R 2 R^2 R 2 is 99.2% for SPY, 97.6% for IWM, 94.7% for QQQ, 86.4% for DIA, and 72.9% for SOXX. Adding electronic equipment raises SOXX’s R 2 R^2 R 2 to 83.4% and QQQ’s to 96.3%. Broad-market factors therefore cover the general equity exposures well, while sector composition matters materially for the semiconductor family. The increase in in-sample R 2 R^2 R 2 is an attribution result, not an out-of-sample forecast improvement.
No six-factor intercept rejects zero after Holm adjustment across the 31 intercept tests. These intercepts describe factor-model residual returns, not an implementable strategy’s net alpha. We next examine departures from a financed benchmark position, which is a different outcome from the ETF’s total return.
Tracking relative to a financed proxy position
To distinguish systematic exposure from daily implementation differences, define
e i t = r i t − β i r b ( i ) , t − ( 1 − β i ) r f t , (76) e_{it} = r_{it} - \beta_i r_{b(i),t} - (1-\beta_i)r_{ft}, \tag*{(76)} e i t = r i t − β i r b ( i ) , t − ( 1 − β i ) r f t , (76)
where b ( i ) b(i) b ( i ) is the associated benchmark ETF. This is the return difference from a hypothetical daily position with β i \beta_i β i dollars in the proxy and 1 − β i 1-\beta_i 1 − β i dollars in the risk-free asset per initial dollar. It is an accounting benchmark, not an assertion that the fund can finance at the risk-free rate or holds
Table IA.34. Daily factor regressions on 466 common dates. FF6 denotes the five Fama–French factors plus momentum. The R 2 R^2 R 2 columns are percentages. Intercepts and market loadings come from FF6, and the final loading comes from FF6 plus electronic equipment.
Figure IA.16. Factor fit and time variation. Left: full-window R 2 R^2 R 2 with and without electronic equipment. Right: trailing 126-session six-factor market loadings for selected benchmark ETFs. The rolling estimates are descriptive exposures, not selected trading signals.
that cash-ETF portfolio. Proxy fees, index-basis differences, market-price deviations, derivatives, and fund expenses can all enter the residual.
The same-day excess-return regression makes the benchmark comparison exact. Write θ ^ i \hat{\theta}_i θ ^ i for the vector of intercept and loadings from the fund regression, θ ^ b \hat{\theta}_b θ ^ b for its proxy, and θ ^ e \hat{\theta}_e θ ^ e for the tracking difference. With the same design and dates,
θ ^ e = θ ^ i − β i θ ^ b , b ^ e β i = b ^ i β i − b ^ b . (77) \hat{\theta}_e = \hat{\theta}_i - \beta_i\hat{\theta}_b,\qquad\frac{\hat{b}_e}{\beta_i} = \frac{\hat{b}_i}{\beta_i} - \hat{b}_b. \tag*{(77)} θ ^ e = θ ^ i − β i θ ^ b , β i b ^ e = β i b ^ i − b ^ b . (77)
This follows directly from linearity of OLS. Figure IA.17 displays the normalized loading differences for the FF6-plus-electronics specification. It isolates exposures beyond the financed proxy rather than displaying the exposures that the product and proxy share. All products retain the same 466 dates.
Table IA.35 reports the mean and RMS tracking difference and the variation explained by the two factor specifications. The unexplained RMS removes both fitted factors and the OLS intercept. It is therefore not equal to the original uncentered tracking RMS times 1 − R 2 \sqrt{1-R^2} 1 − R 2 unless the tracking difference has zero mean. For TQQQ, FF6 plus electronic equipment explains 22.50% of tracking variation. For SOXL and SOXS, it explains only 9.63% and 2.60%, although the same industry factor materially improves the fit of SOXX’s total return. The two statements concern different dependent variables.
These in-sample attributions distinguish broad exposure from daily implementation and benchmark differences. They do not separate fees, derivatives, financing, index mismatch, or rebalance activity as causes of the remaining variation. Daily tracking in (P.3) also differs from the multi-day compounding identity in (2) .
Table IA.35. Tracking differences from financed benchmark positions and their factor attribution. Mean, RMS, and unexplained RMS are daily basis points. R 2 R^2 R 2 is in percent and concerns tracking variation, not the product’s total return. The unexplained RMS uses FF6 plus electronic equipment and removes the fitted intercept. All 26 products use the same 466 dates.
Table IA.35.
Figure IA.17. Exposures beyond a financed proxy. Each cell is the fund loading divided by signed leverage minus the proxy loading, equivalently the tracking-difference loading divided by signed leverage. The model is FF6 plus electronic equipment. These are in-sample exposure differences, not selected investment signals.
Declaration. This working paper was submitted and accepted to the 2026 INFORMS Conference on Financial Engineering on July 10th, 2026. The latest version reflects updated ETF price-volume data.
References