betaFeedback

WorksEconomics, Econometrics and Finance

Portfolio-Size Constraints on Default Correlation: Exact Coherence Tests and Sharp Asymptotic Bounds

Abstract

An absorbing default process imposes joint restrictions on cumulative default distributions across horizons. We characterize these restrictions for finite homogeneous portfolios. For binomial mixtures whose mixing distribution functions cross at most once, stochastic ordering of the entire default count is equivalent to two inequalities: the probability of no defaults must be nonincreasing and the probability of all defaults must be nondecreasing. In the one-factor Gaussian model, these inequalities give an exact interval of admissible later correlations, obtained from two scalar equations. The intervals are nested as portfolio size increases. If the earlier and later default probabilities are p0<p1p_0<p_1 and the earlier latent correlation is ρ0∈(0,1)ρ_0\in(0,1), both interval boundaries approach ρ0ρ_0 at rate (log⁡n)−1/2(\log n)^{-1/2}, with common leading magnitude ρ02(1−ρ0)[Φ−1(p1)−Φ−1(p0)]ρ_0\sqrt{2(1-ρ_0)}[Φ^{-1}(p_1)-Φ^{-1}(p_0)]. We also give a constructive absorbing coupling, an exact sequential repair, differential conditions for smooth horizon curves, and a finite-size certificate of incompatibility. Log-domain quadrature, bracketed root finding, and deterministic numerical experiments evaluate the intervals and verify the restrictions on the full count distribution. The framework provides a feasibility analysis of prescribed horizon laws.

Introduction

An individual default forecast gives the probability that a borrower defaults by a specified horizon. A portfolio forecast describes how many borrowers may default together. Consistency across horizons requires cumulative default counts to be nondecreasing along each realized path. This pathwise restriction implies that every upper-tail probability of the count distribution is nondecreasing in maturity.

This restriction arises naturally when portfolio risk is calibrated across maturities. A borrower model can supply a cumulative default-probability curve, while a dependence model supplies a correlation at each horizon. Together they determine a sequence of portfolio default-count distributions. The feasibility of those distributions as one cumulative default process depends on their joint stochastic ordering. In particular, increasing expected defaults does not guarantee that all portfolio loss probabilities evolve consistently.

A homogeneous one-factor Gaussian model provides a tractable setting in which to study this issue. Its single-horizon count distribution is a binomial mixture, determined by a cumulative default probability and a latent correlation parameter. A finite pool permits changes in that parameter across horizons while preserving the existence of an absorbing default process. The admissible changes depend on the number of borrowers, the movement in cumulative default probability, and the earlier correlation.

The analysis is related to several strands of credit-risk research. The Gaussian default-time construction of Li [7] connects individual default curves through a common dependence structure. Andersen [1] studies term-structure and conditional-survival restrictions in factor models and their implications for intertemporal portfolio losses. Torresetti et al. [13] develops consistency conditions for expected tranched loss surfaces, and Jackson et al. [6] considers temporal dependence in dynamic factor copulas. The large-pool interpretation of the factor model is central to Vasicek [14] and Gordy [5].

Our formulation starts from prescribed exchangeable horizon laws and asks whether a joint absorbing process exists with those laws. It permits arbitrary couplings between horizons. The resulting finite-portfolio characterization combines a two-event criterion for single-crossing binomial mixtures with exact correlation intervals in the Gaussian model. It also yields a constructive absorbing coupling, a sequential projection for incompatible correlation curves, and a differential characterization for smooth curves. The dependence of the feasible interval on portfolio size admits the expansion

ρn±=ρ0±ρ02(1−ρ0)[Φ−1(p1)−Φ−1(p0)]log⁡n+O(log⁡log⁡nlog⁡n).(1)\rho_n^{\pm} = \rho_0 \pm\frac{\rho_0\sqrt{2(1-\rho_0)}\left[\Phi^{-1}(p_1)-\Phi^{-1}(p_0)\right]}{\sqrt{\log n}} + O\left(\frac{\log\log n}{\log n}\right). \tag*{(1)}

The common leading constant identifies the rate at which the finite-portfolio freedom in correlation contracts. It also determines the critical size at which a fixed small correlation change becomes incompatible with absorption.

The mathematical ingredients have established precedents. Strassen [12] provides the classical monotone-coupling framework; Shaked and Shanthikumar [11] develops stochastic-order results; Boland et al. [3] studies crossing properties of mixture distributions; and Yu [15] provides ordering criteria for exponential-family mixtures. Gaussian orthant monotonicity is related to Plackett [9], and high-dimensional orthant asymptotics are studied by Rinott and Rotar [10]. The present results connect these tools to a finite-pool maturity-feasibility problem, with exact scalar correlation bounds and a sharp asymptotic characterization of their endpoints.

Recent credit portfolio comparisons also emphasize dependence uncertainty. Ansari and Lütkebohmert [2] derives convex-order comparisons and risk bounds for Bernoulli mixture models, including Gaussian thresholds. Those results address dispersion and portfolio-risk bounds under model uncertainty. For cumulative default forecasts, first-order stochastic order characterizes maturity feasibility of the count laws. The single-crossing criterion below makes that order test computationally inexpensive for the homogeneous Gaussian family.

Network-based borrower prediction provides an additional motivation. Óskarsdóttir and Bravo [8] and Zandi et al. [16] develop predictive models using multilayer borrower relationships. Their outputs can inform horizon-specific cumulative default probabilities, while a portfolio dependence model determines the associated count distributions. The Gaussian benchmark offers a transparent feasibility diagnostic for this latter step.

We work with exchangeable borrowers and equal loss amounts per default. All horizon laws refer to the same forecasting origin and information set. The correlation parameter is the latent Gaussian correlation. The numerical section combines constructed examples, exact finite-size calculations, and deterministic sensitivity analyses, with scripts provided for replication.

Default forecasting and temporal compatibility

Cumulative default probabilities and portfolio distributions

Let τi\tau_i be the first default time of borrower ii. Its cumulative default indicator is Di(t)=1{τi≤t}D_i(t)=1_{\{\tau_i\le t\}}, so Di(s)≤Di(t)D_i(s)\le D_i(t) whenever s<ts<t. This is the sense in which default is absorbing. The indicator records default history through time tt, including borrowers who subsequently cure or repay. A current delinquency indicator records the borrower’s status at the observation date and can return to zero after a cure.

All horizon distributions in this paper are assessed at one forecasting origin, under a common information set. Thus pi(t)=P(τi≤t)p_i(t) = \mathbb{P}(\tau_i \le t) must be nondecreasing in tt. Forecast updating introduces an evolving information set: a revised estimate of default probability by a fixed terminal date can rise or fall as information arrives. Our compatibility conditions describe the cumulative horizon laws associated with each fixed information set.

For a homogeneous portfolio, write pi(t)=p(t)p_i(t) = p(t) and K(t)=∑i=1nDi(t)K(t) = \sum_{i=1}^{n} D_i(t). Marginal probabilities determine EK(t)=np(t)\mathbb{E}K(t) = np(t). The joint law of the default indicators determines the remaining features of the count distribution. With equal exposures and a common deterministic loss per default, cumulative portfolio loss is proportional to K(t)K(t), so the same ordering restrictions apply to its distribution.

Absorption requires every event {K(s)≥k}\{K(s) \ge k\} to be contained in {K(t)≥k}\{K(t) \ge k\} along a common realization. Hence every count-tail probability must be nondecreasing in the horizon. This is first-order stochastic dominance, expressed through the finite set of thresholds k=1,…,nk = 1,\ldots,n. It orders the expectations of all increasing functions of the count and, in particular, its mean; see Shaked and Shanthikumar [11].

Expected tranche losses provide an economic interpretation. Torresetti et al. [13] use maturity and attachment-point restrictions to construct expected tranched loss surfaces from CDO quotes. For the normalized count loss L(t)=K(t)/nL(t) = K(t)/n, a tranche spanning [(k−1)/n,k/n][(k-1)/n,k/n] has loss

min⁡{(L(t)−k−1n)+,1n}=1n1{K(t)≥k},k=1,…,n.(2)\min\left\{\left(L(t)-\frac{k-1}{n}\right)_{+},\frac{1}{n}\right\}=\frac{1}{n}\mathbf{1}_{\{K(t)\ge k\}},\qquad k=1,\ldots,n. \tag*{(2)}

Thus count-tail ordering is equivalent to nondecreasing expected losses for every single-default-width tranche. The two extreme events in our criterion correspond to the first and last of these tranches. The Gaussian characterization gives conditions under which their monotonicity implies ordering across the entire count grid. A practical measure of an inconsistency between two horizons is

Vn=max⁡1≤k≤n[P(K0≥k)−P(K1≥k)]+.(3)V_n=\max_{1\le k\le n}\left[\mathbb{P}(K_0\ge k)-\mathbb{P}(K_1\ge k)\right]_{+}. \tag*{(3)}

It is the largest decline in a count-tail probability and, for a common deterministic loss amount per default, the corresponding maximum decline in expected loss across single-default-width layers divided by that loss amount. Exact feasibility corresponds to Vn=0V_n=0.

The Gaussian benchmark and the coupling question

The Gaussian copula construction links individual default curves through latent normal variables; see Li [7]. In the homogeneous one-factor specification, let V,ϵ1,…,ϵnV,\epsilon_1,\ldots,\epsilon_n be independent standard normals and set

Xi=ρV+1−ρϵi,Ui=Φ(Xi),X_i=\sqrt{\rho}V+\sqrt{1-\rho}\epsilon_i,\qquad U_i=\Phi(X_i),
Di(t)=1{Ui≤p(t)},0≤ρ<1.(4)D_i(t)=\mathbf{1}_{\{U_i\le p(t)\}},\qquad0\le\rho<1. \tag*{(4)}

Each UiU_i is uniform, giving the required marginal probability p(t)p(t). Using the same UiU_i at every horizon makes each default indicator nondecreasing whenever p(t)p(t) is nondecreasing. Conditional on the factor, the indicators at a fixed horizon are independent, and their count is binomial. Integrating over the factor produces the binomial mixture studied below, as in the conditional-independence framework of Vasicek [14].

Allowing a separate correlation at each horizon changes the distribution of the conditional default probability. A fixed-factor construction requires monotonicity at every factor realization, as in the term-structure analysis of Andersen [1]. The general coupling formulation allows a joint law across horizons with the prescribed horizon marginals. It treats the temporal connection between those marginals as a modeling choice subject to absorption. Proposition 3.2 provides an explicit construction for exchangeable portfolios.

Absorbing default laws for an exchangeable portfolio

Fix ordered horizons t0<⋯<tmt_0 < \cdots< t_m and a portfolio size nn. At horizon tjt_j, let νj\nu_j be an exchangeable probability law on {0,1}n\{0,1\}^n. Write DijD_{ij} for the default indicator of borrower ii and Kj=∑i=1nDijK_j = \sum_{i=1}^n D_{ij} for the default count.

Definition 3.1. The laws ν0,…,νm\nu_0,\ldots,\nu_m are absorbing-compatible if there is a joint distribution with these horizon marginals such that Di,j+1≥DijD_{i,j+1} \ge D_{ij} almost surely for every ii and jj.

An exchangeable binary law is determined by the distribution of its count: conditional on Kj=kK_j=k, each subset of kk defaulted names has probability (nk)−1\binom{n}{k}^{-1}. For integer-valued random variables supported on {0,…,n}\{0,\ldots,n\}, we write K≤stK′K \le_{\mathrm{st}} K' if P(K≥k)≤P(K′≥k)\mathbb{P}(K \ge k) \le\mathbb{P}(K' \ge k) for k=1,…,nk=1,\ldots,n.

Proposition 3.2. The laws ν0,…,νm\nu_0,\ldots,\nu_m are absorbing-compatible if and only if K0≤st⋯≤stKmK_0 \le_{\mathrm{st}} \cdots\le_{\mathrm{st}} K_m. If this condition holds, an exchangeable absorbing coupling can be constructed using one uniform random variable and one independent uniform permutation of the borrowers.

Proof. Necessity follows by summing Di,j+1≥DijD_{i,j+1} \ge D_{ij}. For sufficiency, let FjF_j be the count distribution function and define Kj=Fj−1(U)K_j=F_j^{-1}(U) for one UU uniform on (0,1)(0,1), using the left quantile. Stochastic ordering gives K0≤⋯≤KmK_0 \le\cdots\le K_m almost surely. Let R1,…,RnR_1,\ldots,R_n be the ranks of an independent uniform permutation, and put Dij=1{Ri≤Kj}D_{ij}=\mathbf{1}_{\{R_i\le K_j\}}. Conditional on Kj=kK_j=k, the defaulted set is uniform among all subsets of size kk. Thus the horizon law is νj\nu_j, and every indicator is nondecreasing in jj.

The coupling defines a default-time model on the given grid by assigning each borrower the first horizon at which its indicator becomes one. Borrowers already in default at t0t_0 can be assigned time t0t_0, and those surviving tmt_m a time after tmt_m. A given family of horizon distributions generally admits multiple temporal laws.

Example 3.3. Let n=3n=3, let K0=2K_0=2 surely, and let K1K_1 equal 1 or 3 with probabilities 2/52/5 and 3/53/5. Both no-default probabilities are zero, while the all-default probability rises from zero to 3/53/5. The mean count rises from 2 to 11/511/5, and P(K1≥2)=3/5<1=P(K0≥2)\mathbb{P}(K_1\ge2)=3/5<1=\mathbb{P}(K_0\ge2). This interior-tail inequality violates the ordering required for absorption. Uniform defaulted subsets conditional on these counts give exchangeable binary laws with the stated properties.

The following section gives the additional structure under which the two extreme events do suffice.

A two-event criterion for binomial mixtures

For a∈{0,1}a\in\{0,1\}, suppose

Ka∣Qa∼Bin⁡(n,Qa),0<Qa<1,(5)K_a\mid Q_a \sim\operatorname{Bin}(n,Q_a), \qquad0<Q_a<1, \tag*{(5)}

and denote the distribution function of QaQ_a by HaH_a. The variable QaQ_a represents the default probability conditional on the common environment. If pa=EQap_a=\mathbb{E}Q_a, the mixture has

EKa=npa,Var⁡(Ka)=npa(1−pa)+n(n−1)Var⁡(Qa).(6)\mathbb{E}K_a=np_a,\qquad\operatorname{Var}(K_a)=np_a(1-p_a)+n(n-1)\operatorname{Var}(Q_a). \tag*{(6)}

The second term records the contribution of common uncertainty to count dispersion. Different mixing laws can therefore preserve the mean while changing the likelihood of both mild and severe default outcomes. The criterion below compares their entire count distributions. A function has at most one sign change if it has constant sign, or changes once from nonnegative to nonpositive, or once in the reverse direction. Zeros are allowed.

Theorem 4.1. Assume that H0−H1H_0-H_1 has at most one sign change on (0,1)(0,1). Then K0≤stK1K_0 \le_{\mathrm{st}} K_1 if and only if

E(1−Q1)n≤E(1−Q0)n,EQ1n≥EQ0n.(7)\mathbb{E}(1-Q_1)^n \le\mathbb{E}(1-Q_0)^n,\qquad\mathbb{E}Q_1^n \ge\mathbb{E}Q_0^n. \tag*{(7)}

Proof. Necessity follows from the count thresholds 11 and nn. For sufficiency, put g=H0−H1g=H_0-H_1 and wk(q)=qk−1(1−q)n−kw_k(q)=q^{k-1}(1-q)^{n-k}. Integration by parts gives

Tk:=P(K1≥k)−P(K0≥k)=n(n−1k−1)∫01wk(q)g(q) dq.(8)T_k:=\mathbb{P}(K_1\ge k)-\mathbb{P}(K_0\ge k)=n\binom{n-1}{k-1}\int_0^1 w_k(q)g(q)\,\mathrm{d}q. \tag*{(8)}

Indeed, the derivative of the binomial tail with respect to qq is n(n−1k−1)wk(q)n\binom{n-1}{k-1}w_k(q), and the boundary terms vanish. Conditions (7) say T1,Tn≥0T_1,T_n\ge0.

If g≥0g\ge0, every integral is nonnegative. If g≤0g\le0, T1≥0T_1\ge0 and positivity of w1w_1 on (0,1)(0,1) force g=0g=0 almost everywhere, so every integral vanishes. Suppose next that gg is nonnegative below a crossing point q∗q_* and nonpositive above it. The ratio wn/wk=[q/(1−q)]n−kw_n/w_k=[q/(1-q)]^{n-k} is nondecreasing. Consequently

wn(q)g(q)≤wn(q∗)wk(q∗)wk(q)g(q),0<q<1.(9)w_n(q)g(q)\le\frac{w_n(q_*)}{w_k(q_*)}w_k(q)g(q),\qquad0<q<1. \tag*{(9)}

Since the integral on the left is nonnegative, the integral of wkgw_kg is also nonnegative. If gg is nonpositive below q∗q_* and nonnegative above it, use instead the nondecreasing ratio wk/w1=[q/(1−q)]k−1w_k/w_1=[q/(1-q)]^{k-1} to obtain

wk(q)g(q)≥wk(q∗)w1(q∗)w1(q)g(q).(10)w_k(q)g(q)\ge\frac{w_k(q_*)}{w_1(q_*)}w_1(q)g(q). \tag*{(10)}

Now T1≥0T_1\ge0 implies the required inequality. The cases in which the crossing occurs only at a boundary reduce to constant sign. Hence Tk≥0T_k\ge0 for every kk.

The proof uses a variation-diminishing argument for the binomial kernel, in the tradition of Boland et al. [3]. The condition is stated directly for the mixing distribution functions, which are available in closed form for location-scale probit mixtures. The conclusion replaces the nn separate count-tail comparisons with two extreme-event probabilities.

Corollary 4.2. Let ZZ have a continuous strictly increasing distribution function FF on R\mathbb{R}, let G:R→(0,1)G:\mathbb{R}\to(0,1) be a continuous strictly increasing bijection, and let Qa=G(μa+σaZ)Q_a=G(\mu_a+\sigma_a Z) with σa>0\sigma_a>0. The equivalence in Theorem 4.1 holds.

Proof. The distribution function is Ha(q)=F((G−1(q)−μa)/σa)H_a(q)=F((G^{-1}(q)-\mu_a)/\sigma_a). Because FF is strictly increasing, the sign of H0(q)−H1(q)H_0(q)-H_1(q) is the sign of

(1σ0−1σ1)G−1(q)−μ0σ0+μ1σ1.(11)\left(\frac{1}{\sigma_0}-\frac{1}{\sigma_1}\right)G^{-1}(q)-\frac{\mu_0}{\sigma_0}+\frac{\mu_1}{\sigma_1}. \tag*{(11)}

This affine function of G−1(q)G^{-1}(q) has at most one sign change.

The result applies to both probit-normal and logit-normal mixing. It also applies with a nonnormal full-support factor, provided the same factor distribution and link are used at the two horizons. This common location-scale representation supplies the single-crossing property used in the theorem. More general factor families and heterogeneous borrower probabilities require corresponding conditions on the mixing distributions and count laws.

Exact Gaussian correlation intervals

Let ZZ be standard normal and define

Q(p,ρ)=Φ(c(p)+ρZ1−ρ),c(p)=Φ−1(p),0<p<1,0<ρ<1.(12)Q(p,\rho)=\Phi\left(\frac{c(p)+\sqrt{\rho}Z}{\sqrt{1-\rho}}\right), \qquad c(p)=\Phi^{-1}(p), \qquad0<p<1,\quad0<\rho<1. \tag*{(12)}

Then EQ(p,ρ)=p\mathbb{E}Q(p,\rho)=p. For instance, if ϵ\epsilon is an independent standard normal, P(1−ρϵ−ρZ≤c(p))=p\mathbb{P}(\sqrt{1-\rho}\epsilon-\sqrt{\rho}Z\le c(p))=p. The resulting binary vector is the homogeneous one-factor Gaussian default model introduced in Section 2. The sign of ZZ is immaterial.

The parameter ρ\rho is the correlation of the latent normals. For distinct borrowers, the resulting correlation of their default indicators is

Corr⁡(Di,Dj)=Φ2(c(p),c(p);ρ)−p2p(1−p),i≠j,(13)\operatorname{Corr}(D_i,D_j)=\frac{\Phi_2(c(p),c(p);\rho)-p^2}{p(1-p)}, \qquad i\ne j, \tag*{(13)}

where Φ2(⋅,⋅;ρ)\Phi_2(\cdot,\cdot;\rho) denotes the bivariate standard normal distribution function with correlation ρ\rho. Thus a constant latent correlation can coexist with changing default-indicator correlation as the cumulative probability pp increases. The interval restrictions below concern the latent parameter. The conditional default probability in (12) has probit location and scale

μ(p,ρ)=c(p)1−ρ,σ(ρ)=ρ1−ρ.(14)\mu(p,\rho)=\frac{c(p)}{\sqrt{1-\rho}}, \qquad\sigma(\rho)=\sqrt{\frac{\rho}{1-\rho}}. \tag*{(14)}

Define the extreme-event probabilities

An(p,ρ)=EQ(p,ρ)n,Sn(p,ρ)=E[1−Q(p,ρ)]n=An(1−p,ρ).(15)A_n(p,\rho)=\mathbb{E}Q(p,\rho)^n,\qquad S_n(p,\rho)=\mathbb{E}[1-Q(p,\rho)]^n=A_n(1-p,\rho). \tag*{(15)}

At ρ=0\rho=0 these extend continuously to An(p,0)=pnA_n(p,0)=p^n and Sn(p,0)=(1−p)nS_n(p,0)=(1-p)^n.

Lemma 5.1. For n≥2n\ge2, An(p,ρ)A_n(p,\rho) and Sn(p,ρ)S_n(p,\rho) are continuous and strictly increasing in ρ∈[0,1)\rho\in[0,1). At fixed ρ<1\rho<1, AnA_n is strictly increasing in pp, while SnS_n is strictly decreasing in pp.

Proof. The assertions in pp follow directly from (12). For correlation, AnA_n is the probability that all coordinates of an equicorrelated standard normal vector lie below c(p)c(p). The identity of Plackett [9] expresses the derivative with respect to any off-diagonal correlation as the bivariate normal density at the two corresponding thresholds, multiplied by the conditional probability that the remaining coordinates lie below their thresholds. Both factors are strictly positive for a nonsingular correlation matrix and finite thresholds. Along the equicorrelation path, the derivative is the sum of these positive quantities over pairs. Continuity at zero follows by dominated convergence in (12). Apply the same argument at 1−p1-p for SnS_n.

Fix a correlation cap ρˉ<1\bar{\rho}<1, an earlier pair (p0,ρ0)(p_0,\rho_0) with 0<ρ0<ρˉ0<\rho_0<\bar{\rho}, and a later cumulative probability p1≥p0p_1\ge p_0. Let In(p0,ρ0;p1)I_n(p_0,\rho_0;p_1) be the set of later correlations in [0,ρˉ][0,\bar{\rho}] for which the two Gaussian horizon laws are absorbing-compatible.

Theorem 5.2. For n≥2n\ge2, In(p0,ρ0;p1)=[ℓn,un]I_n(p_0,\rho_0;p_1)=[\ell_n,u_n], where

ℓn={0,p1n≥An(p0,ρ0),the unique r∈(0,ρ0] with An(p1,r)=An(p0,ρ0),otherwise,(16)\ell_n= \begin{cases} 0, & p_1^n\ge A_n(p_0,\rho_0),\\ \text{the unique }r\in(0,\rho_0]\text{ with }A_n(p_1,r)=A_n(p_0,\rho_0), & \text{otherwise}, \end{cases} \tag*{(16)}

and

un={ρˉ,Sn(p1,ρˉ)≤Sn(p0,ρ0),the unique r∈[ρ0,ρˉ] with Sn(p1,r)=Sn(p0,ρ0),otherwise.(17)u_n= \begin{cases} \bar{\rho}, & S_n(p_1,\bar{\rho})\le S_n(p_0,\rho_0),\\ \text{the unique }r\in[\rho_0,\bar{\rho}]\text{ with }S_n(p_1,r)=S_n(p_0,\rho_0), & \text{otherwise}. \end{cases} \tag*{(17)}

The interval contains ρ0\rho_0. If p1=p0p_1=p_0, it is the singleton {ρ0}\{\rho_0\}. If p1>p0p_1>p_0, it contains an open neighborhood of ρ0\rho_0. For any finite sequence of such Gaussian horizon laws, absorbing compatibility is equivalent to satisfaction of these intervals at every adjacent pair of horizons.

Proof. For positive correlations, Corollary 4.2 gives equivalence to An(p1,r)≥An(p0,ρ0)A_n(p_1,r) \ge A_n(p_0,\rho_0) and Sn(p1,r)≤Sn(p0,ρ0)S_n(p_1,r) \le S_n(p_0,\rho_0). Lemma 5.1 yields the stated roots and the interval between them. The fixed-correlation choice r=ρ0r=\rho_0 satisfies both inequalities because p1≥p0p_1 \ge p_0. At equality of the probabilities, strict monotonicity in correlation forces r=ρ0r=\rho_0. For p1>p0p_1>p_0, both inequalities are strict at ρ0\rho_0, so continuity gives a neighborhood.

At r=0r=0, Q1=p1Q_1=p_1 surely. Its distribution function differs from H0H_0 by a function that is nonnegative below p1p_1 and nonpositive above it. Theorem 4.1 therefore covers this endpoint as well. Finally, pairwise stochastic order is transitive and Proposition 3.2 supplies one joint absorbing coupling on the whole grid.

The root equations admit stable numerical evaluation in logarithmic form. The lower boundary enforces a nondecreasing probability of universal default. The upper boundary enforces a nonincreasing probability of universal survival. At fixed pp, increasing dependence raises both extreme-event probabilities, giving each boundary a distinct role.

Corollary 5.3. If p1≥p0p_1 \ge p_0 and ρ1≥ρ0\rho_1 \ge\rho_0, compatibility is equivalent to Sn(p1,ρ1)≤Sn(p0,ρ0)S_n(p_1,\rho_1) \le S_n(p_0,\rho_0). If p1≥p0p_1 \ge p_0 and ρ1≤ρ0\rho_1 \le\rho_0, it is equivalent to An(p1,ρ1)≥An(p0,ρ0)A_n(p_1,\rho_1) \ge A_n(p_0,\rho_0).

Proof. In the first case the all-default inequality holds automatically by monotonicity in both arguments. In the second case the no-default inequality holds automatically, since increasing pp and decreasing ρ\rho both decrease SnS_n. Apply Theorem 5.2.

Proposition 5.4. The intervals are nested in portfolio size: In+1(p0,ρ0;p1)⊆In(p0,ρ0;p1)I_{n+1}(p_0,\rho_0;p_1) \subseteq I_n(p_0,\rho_0;p_1) for every n≥2n \ge2.

Proof. An absorbing coupling for n+1n+1 borrowers remains absorbing after one borrower is deleted. Each restricted horizon law is still the Gaussian binomial-mixture law with the same (p,ρ)(p,\rho) parameters and nn borrowers. Hence any correlation feasible for n+1n+1 is feasible for nn.

Consequently, for a fixed pair of parameter values, the set of feasible integer portfolio sizes is an initial segment, possibly all positive integers. Portfolio enlargement imposes successively tighter restrictions on the same marginal probabilities and latent correlations.

Sequential repair

Suppose nondecreasing cumulative probabilities p0,…,pmp_0,\ldots,p_m and raw correlations ρ^0,…,ρ^m\hat{\rho}_0,\ldots,\hat{\rho}_m in [0,ρˉ][0,\bar{\rho}] have been supplied. Set ρ0∗=ρ^0\rho_0^*=\hat{\rho}_0. Given ρj−1∗\rho_{j-1}^*, calculate the interval [ℓj,uj][\ell_j,u_j] for the transition from (pj−1,ρj−1∗)(p_{j-1},\rho_{j-1}^*) to pjp_j, and define

ρj∗=min⁡{uj,max⁡{ℓj,ρ^j}},j=1,…,m.(18)\rho_j^*=\min\{u_j,\max\{\ell_j,\hat{\rho}_j\}\}, \qquad j=1,\ldots,m. \tag*{(18)}

The formulas extend to an earlier correlation of zero by the same extreme-event inequalities.

Proposition 5.5. The repaired sequence is absorbing-compatible and preserves every pjp_j. Conditional on the preceding repaired pair, (18) is the unique minimizer of ∣r−ρ^j∣|r-\hat{\rho}_j| over the feasible interval for every ρ>0\rho>0.

Proof. The interval is nonempty because retaining the preceding correlation is feasible. Projection selects a point in it. Theorem 5.2 and Proposition 3.2 give compatibility. The individual probability equals pjp_j for every correlation in (12). The last assertion is the elementary closest-point property of a closed interval.

Each projection minimizes the current adjustment conditional on the preceding repaired pair. Earlier choices determine the feasible intervals at later maturities. A global calibration would optimize those choices jointly across the full horizon grid, and selection of a temporal law would introduce further modeling criteria.

Smooth horizon curves

The finite-grid characterization has a differential counterpart. Suppose μ(t)\mu(t) and σ(t)>0\sigma(t)>0 are continuously differentiable on a compact interval, with σ\sigma bounded away from zero. Let ZZ be standard normal, Qt=Φ(μ(t)+σ(t)Z)Q_t=\Phi(\mu(t)+\sigma(t)Z), and Kt∣Qt∼Bin⁡(n,Qt)K_t\mid Q_t\sim\operatorname{Bin}(n,Q_t). Define

Jk(t)=E[ϕ(μ+σZ)Qtk−1(1−Qt)n−k],mk(t)=E[Zϕ(μ+σZ)Qtk−1(1−Qt)n−k]Jk(t).(19)J_k(t)=\mathbb{E}\left[\phi(\mu+\sigma Z)Q_t^{k-1}(1-Q_t)^{n-k}\right],\qquad m_k(t)=\frac{\mathbb{E}\left[Z\phi(\mu+\sigma Z)Q_t^{k-1}(1-Q_t)^{n-k}\right]}{J_k(t)}. \tag*{(19)}

Here and below μ\mu and σ\sigma inside expectations are evaluated at tt. These quantities are finite and Jk(t)>0J_k(t)>0.

Proposition 6.1. All count tails P(Kt≥k)\mathbb{P}(K_t\ge k) are nondecreasing in tt if and only if, at every tt,

μ˙(t)+σ˙(t)m1(t)≥0,μ˙(t)+σ˙(t)mn(t)≥0.(20)\dot{\mu}(t)+\dot{\sigma}(t)m_1(t)\ge0,\qquad\dot{\mu}(t)+\dot{\sigma}(t)m_n(t)\ge0. \tag*{(20)}

Moreover, m1(t)<⋯<mn(t)m_1(t)<\cdots<m_n(t) for n≥2n\ge2.

Proof. Differentiation under the integral is justified by Gaussian integrability and boundedness of the parameters and their derivatives on the compact interval. The binomial-tail derivative gives

ddtP(Kt≥k)=n(n−1k−1)Jk(t)[μ˙(t)+σ˙(t)mk(t)].(21)\frac{d}{dt}\mathbb{P}(K_t\ge k)=n\binom{n-1}{k-1}J_k(t)\left[\dot{\mu}(t)+\dot{\sigma}(t)m_k(t)\right]. \tag*{(21)}

Consider the probability measure with density proportional to the integrand defining JkJ_k. Passing from kk to k+1k+1 tilts it by the strictly increasing function Qt/(1−Qt)Q_t/(1-Q_t) of ZZ. Hence the mean of ZZ strictly increases. More explicitly, for an integrable increasing tilt hh, the change in the mean is Cov⁡(Z,h(Z))/Eh(Z)>0\operatorname{Cov}(Z,h(Z))/\mathbb{E}h(Z)>0; the covariance is half the expectation of (Z−Z′)[h(Z)−h(Z′)](Z-Z')[h(Z)-h(Z')] for independent copies under the tilted base law. All required moments are finite here. Thus mkm_k increases strictly. An affine function of mkm_k is nonnegative at every kk if and only if it is nonnegative at the two endpoints. Equation (21) proves the equivalence. □\square

For the Gaussian credit parametrization, c(t)=Φ−1(p(t))c(t)=\Phi^{-1}(p(t)) and (14) give an equivalent pair of inequalities:

c˙(t)+ρ˙(t)2(1−ρ(t))[c(t)+mk(t)ρ(t)]≥0,k∈{1,n}.(22)\dot{c}(t)+\frac{\dot{\rho}(t)}{2(1-\rho(t))}\left[c(t)+\frac{m_k(t)}{\sqrt{\rho(t)}}\right]\ge0,\qquad k\in\{1,n\}. \tag*{(22)}

Thus smooth changes in a marginal default curve place exact restrictions on the speed of changes in dependence. The inequalities characterize horizon curves under the fixed initial information used throughout the paper.

Large portfolios and sharp boundary rates

Compatibility at every portfolio size

The familiar large-portfolio interpretation of the Gaussian model is that conditional idiosyncratic uncertainty diversifies away; see (2002). This gives a general compatibility criterion before any Gaussian calculations are used.

Proposition 7.1. Let Q0,Q1Q_0,Q_1 have fixed laws on [0,1][0,1], and let Ka,n∣Qa∼Bin⁡(n,Qa)K_{a,n}\mid Q_a\sim\operatorname{Bin}(n,Q_a). Then K0,n≤stK1,nK_{0,n}\le_{\mathrm{st}}K_{1,n} for every positive integer nn if and only if Q0≤stQ1Q_0\le_{\mathrm{st}}Q_1.

Proof. If Q0≤stQ1Q_0\le_{\mathrm{st}}Q_1, couple them by common quantiles so Q0≤Q1Q_0\le Q_1. Conditional on this pair, use the same independent uniforms to generate the Bernoulli indicators in both counts. This orders the counts for every nn. Conversely, Ka,n/nK_{a,n}/n converges in distribution to QaQ_a, for example by conditional variance and E[(Ka,n/n−Qa)2]≤1/(4n)\mathbb{E}[(K_{a,n}/n-Q_a)^2]\le1/(4n). Stochastic order is preserved under weak limits, as is seen by taking expectations of bounded continuous increasing functions. Hence ordering for every nn implies Q0≤stQ1Q_0\le_{\mathrm{st}}Q_1.

Corollary 7.2. Let 0<p0,p1<10<p_0,p_1<1 and 0<ρ0,ρ1<10<\rho_0,\rho_1<1. The Gaussian horizon laws are absorbing-compatible at every portfolio size if and only if ρ1=ρ0\rho_1=\rho_0 and p1≥p0p_1\ge p_0.

Proof. By Proposition 7.1, the mixing variables must be stochastically ordered. Their distribution functions have the form in Corollary 4.2, with F=G=ΦF=G=\Phi. If the scales differ, the affine difference in (11) takes both signs because Φ−1(q)\Phi^{-1}(q) ranges over R\mathbb{R}. Stochastic ordering is therefore impossible. Equality of scales is equivalent to equality of ρ\rho, and with this scale fixed the order is equivalent to c(p1)≥c(p0)c(p_1)\ge c(p_0). Sufficiency also follows directly from the common-factor construction.

The corollary gives an all-size formulation of the Gaussian rigidity discussed by Andersen [1]. It applies to the prescribed homogeneous Gaussian binomial-mixture horizon laws, with arbitrary temporal couplings allowed for each finite nn. Within this family, compatibility for all portfolio sizes requires a common latent correlation across horizons.

A uniform orthant estimate

We record the logarithmic estimate needed for the interval boundaries. It is a direct extreme-value argument for an equicorrelated normal probability, a problem treated in greater generality by Rinott and Rotar [10].

Lemma 7.3. Uniformly for pp and ρ\rho in compact subsets of (0,1)(0,1), as n→∞n\to\infty,

log⁡An(p,ρ)=−1−ρρlog⁡n+c(p)2(1−ρ)ρlog⁡n+O(log⁡log⁡n).(23)\log A_n(p,\rho)=-\frac{1-\rho}{\rho}\log n+\frac{c(p)\sqrt{2(1-\rho)}}{\rho}\sqrt{\log n}+O(\log\log n). \tag*{(23)}

Proof. Let ϵ1,…,ϵn,Z\epsilon_1,\ldots,\epsilon_n,Z be independent standard normals, Mn=max⁡iϵiM_n=\max_i\epsilon_i, and use μ,σ\mu,\sigma from (14). Conditional independence gives

An(p,ρ)=P(Mn≤μ+σZ)=EΦ‾(Mn−μσ).(24)A_n(p,\rho)=\mathbb{P}(M_n\le\mu+\sigma Z)=\mathbb{E}\overline{\Phi}\left(\frac{M_n-\mu}{\sigma}\right). \tag*{(24)}

Put L=log⁡nL=\log n, b=2Lb=\sqrt{2L}, and d=3log⁡L/bd=3\log L/b. Standard normal tail bounds give

P(Mn>b+d)≤nΦ‾(b+d)=O(L−7/2),P(Mn<b−d)≤exp⁡{−aL5/2}(25)\mathbb{P}(M_n>b+d)\le n\overline{\Phi}(b+d)=O(L^{-7/2}),\qquad\mathbb{P}(M_n<b-d)\le\exp\{-aL^{5/2}\} \tag*{(25)}

for some constant a>0a>0 and sufficiently large nn. For the second inequality use Φ(b−d)n≤exp⁡[−nΦ‾(b−d)]\Phi(b-d)^n\le\exp[-n\overline{\Phi}(b-d)] and the lower normal tail bound. These statements do not involve pp or ρ\rho. Since Φ‾\overline{\Phi} is decreasing, (24) is bounded below by P(Mn<b+d)Φ‾((b+d−μ)/σ)\mathbb{P}(M_n<b+d)\overline{\Phi}((b+d-\mu)/\sigma) and above by Φ‾((b−d−μ)/σ)+exp⁡{−aL5/2}\overline{\Phi}((b-d-\mu)/\sigma)+\exp\{-aL^{5/2}\}. On the stated compact sets, μ\mu is bounded and σ\sigma is bounded above and away from zero. The additive exponential error is uniformly negligible relative to the normal tail in the upper bound. The normal tail expansion log⁡Φ‾(x)=−x2/2−log⁡x−12log⁡(2π)+O(x−2)\log\overline{\Phi}(x)=-x^2/2-\log x-\frac{1}{2}\log(2\pi)+O(x^{-2}) therefore yields

log⁡An=−Lσ2+μ2Lσ2+O(log⁡L).(26)\log A_n=-\frac{L}{\sigma^2}+\frac{\mu\sqrt{2L}}{\sigma^2}+O(\log L). \tag*{(26)}

Substitution of (14) proves (23), with a uniform remainder.

The correlation boundary expansion

Theorem 7.4. Fix 0<p0<p1<10 < p_0 < p_1 < 1 and 0<ρ0<ρˉ<10 < \rho_0 < \bar{\rho} < 1. Let [ℓn,un][\ell_n,u_n] be the admissible interval of Theorem 5.2, and put

C=ρ02(1−ρ0)[c(p1)−c(p0)]>0.(27)C = \rho_0\sqrt{2(1-\rho_0)}[c(p_1)-c(p_0)] > 0. \tag*{(27)}

For all sufficiently large nn, both boundaries are interior to [0,ρˉ][0,\bar{\rho}], and

ℓn=ρ0−Clog⁡n+O(log⁡log⁡nlog⁡n),un=ρ0+Clog⁡n+O(log⁡log⁡nlog⁡n).(28)\ell_n = \rho_0-\frac{C}{\sqrt{\log n}}+O\left(\frac{\log\log n}{\log n}\right), \qquad u_n = \rho_0+\frac{C}{\sqrt{\log n}}+O\left(\frac{\log\log n}{\log n}\right). \tag*{(28)}

In particular, the interval width is 2C/log⁡n+O(log⁡log⁡n/log⁡n)2C/\sqrt{\log n}+O(\log\log n/\log n).

Proof. Write a(r)=(1−r)/ra(r)=(1-r)/r and b(p,r)=c(p)2(1−r)/rb(p,r)=c(p)\sqrt{2(1-r)}/r. For any fixed sufficiently small ε>0\varepsilon>0, Lemma 7.3 shows that the sign of

log⁡An(p1,r)−log⁡An(p0,ρ0)=[a(ρ0)−a(r)]L+[b(p1,r)−b(p0,ρ0)]L+O(log⁡L)(29)\log A_n(p_1,r)-\log A_n(p_0,\rho_0)=[a(\rho_0)-a(r)]L+[b(p_1,r)-b(p_0,\rho_0)]\sqrt{L}+O(\log L) \tag*{(29)}

at r=ρ0±εr=\rho_0\pm\varepsilon is respectively positive and negative. The same conclusion holds with (p0,p1)(p_0,p_1) replaced by (1−p0,1−p1)(1-p_0,1-p_1). Thus the roots defining both boundaries are interior for large nn and converge to ρ0\rho_0.

For bounded real hh, substitute r=ρ0+h/Lr=\rho_0+h/\sqrt{L} in (29). Since a′(ρ0)=−1/ρ02a'(\rho_0)=-1/\rho_0^2, Taylor expansion, uniform for bounded hh, gives

log⁡An(p1,ρ0+h/L)−log⁡An(p0,ρ0)=L[hρ02+[c(p1)−c(p0)]2(1−ρ0)ρ0]+O(log⁡L).(30)\log A_n(p_1,\rho_0+h/\sqrt{L})-\log A_n(p_0,\rho_0)=\sqrt{L}\left[\frac{h}{\rho_0^2}+\frac{[c(p_1)-c(p_0)]\sqrt{2(1-\rho_0)}}{\rho_0}\right]+O(\log L). \tag*{(30)}

Taking h=−C−ηh=-C-\eta and h=−C+ηh=-C+\eta for any fixed η>0\eta>0 brackets the lower root for all sufficiently large nn. Hence L(ℓn−ρ0)\sqrt{L}(\ell_n-\rho_0) is bounded and converges to −C-C. Substituting this bounded value of hh into (30) and using equality at the root gives h=−C+O(log⁡L/L)h=-C+O(\log L/\sqrt{L}), which is the lower expansion.

For the upper boundary use Sn(p,r)=An(1−p,r)S_n(p,r)=A_n(1-p,r) and c(1−p)=−c(p)c(1-p)=-c(p). The difference of the two bb terms changes sign, so the corresponding equation has root h=C+O(log⁡L/L)h=C+O(\log L/\sqrt{L}). This proves the upper expansion and the width formula.

The leading symmetry in (28) emerges as nn grows. At moderate nn, the upper restriction can be substantially weaker than the lower restriction. The lower boundary is governed by an extreme event whose probability may be very small. Assessing economic materiality therefore also requires the magnitude of the resulting tail-probability changes.

Corollary 7.5. Under the assumptions of Theorem 7.4, for nonzero δ\delta sufficiently close to zero let N∗(δ)N_*(\delta) be the largest integer portfolio size for which the horizon pairs (p0,ρ0)(p_0,\rho_0) and (p1,ρ0+δ)(p_1,\rho_0+\delta) are absorbing-compatible. Then N∗(δ)N_*(\delta) is finite and

lim⁡δ→0, δ≠0δ2log⁡N∗(δ)=C2.(31)\lim_{\delta\to0,\ \delta\ne0}\delta^2\log N_*(\delta)=C^2. \tag*{(31)}

Proof. Size one is compatible, and Proposition 5.4 makes the compatible sizes an initial segment. Corollary 7.2 shows that the segment is finite for δ≠0\delta\ne0. Every fixed size is compatible for all sufficiently small ∣δ∣|\delta| by Theorem 5.2; hence N∗(δ)→∞N_*(\delta)\to\infty.

Fix $0<\eta<C$. For δ>0\delta>0, the relevant radius is un−ρ0u_n-\rho_0, which is asymptotic to C/log⁡nC/\sqrt{\log n}. For δ<0\delta<0, use ρ0−ℓn\rho_0-\ell_n, with the same asymptotic equivalent. At the integer sizes nearest exp⁡[(C−η)2/δ2]\exp[(C-\eta)^2/\delta^2] and exp⁡[(C+η)2/δ2]\exp[(C+\eta)^2/\delta^2], these radii are respectively greater and smaller than ∣δ∣|\delta| for sufficiently small ∣δ∣|\delta|. Nesting places N∗(δ)N_*(\delta) between these sizes. Multiply their logarithms by δ2\delta^2 and let first δ→0\delta\to0 and then $\eta\downarrow 0.

The critical-size law in (31) may also be written as log⁡N∗(δ)∼C2/δ2\log N_*(\delta) \sim C^2/\delta^2. The constant CC depends on the earlier latent correlation and on the increase in the normal-transformed cumulative default probability. Consequently, the first portfolio size at which a correlation discrepancy becomes infeasible can change substantially across default-probability curves. The exact interval equations remain relevant for realistic finite pools, where the limiting approximation can be asymmetric.

A finite-size certificate

The all-size result can be made quantitative without a Gaussian assumption. The following bound uses the concentration inequality of Hoeffding [4].

Proposition 7.6. Let Ka,n∣Qa∼Bin⁡(n,Qa)K_{a,n} \mid Q_a \sim\operatorname{Bin}(n,Q_a). Suppose z∈(0,1)z \in(0,1) and 0<ε<min⁡(z,1−z)0 < \varepsilon< \min(z,1-z) satisfy δ=P(Q0≥z+ε)−P(Q1≥z−ε)>0\delta= \mathbb{P}(Q_0 \ge z+\varepsilon) - \mathbb{P}(Q_1 \ge z-\varepsilon) > 0. Then

P(K0,n/n≥z)−P(K1,n/n≥z)≥δ−2e−2nε2.(32)\mathbb{P}(K_{0,n}/n \ge z) - \mathbb{P}(K_{1,n}/n \ge z) \ge\delta- 2e^{-2n\varepsilon^2}. \tag*{(32)}

In particular, compatibility fails whenever n>log⁡(2/δ)/(2ε2)n > \log(2/\delta)/(2\varepsilon^2).

Proof. Conditionally on Q0≥z+εQ_0 \ge z+\varepsilon, the probability that K0,n/n<zK_{0,n}/n < z is at most e−2nε2e^{-2n\varepsilon^2}. Thus P(K0,n/n≥z)≥P(Q0≥z+ε)−e−2nε2\mathbb{P}(K_{0,n}/n \ge z) \ge\mathbb{P}(Q_0 \ge z+\varepsilon) - e^{-2n\varepsilon^2}. Splitting on Q1≥z−εQ_1 \ge z-\varepsilon and applying the opposite conditional tail bound gives P(K1,n/n≥z)≤P(Q1≥z−ε)+e−2nε2\mathbb{P}(K_{1,n}/n \ge z) \le\mathbb{P}(Q_1 \ge z-\varepsilon) + e^{-2n\varepsilon^2}. Subtraction proves the claim. The count threshold is equivalently ⌈nz⌉\lceil nz\rceil.

For continuous mixing distributions that fail stochastic order, such a positive δ\delta exists after choosing a sufficiently small ε\varepsilon around a strict crossing violation. The bound supplies an explicit sufficient size for incompatibility. In the Gaussian model, the two-event test determines the first failing integer size.

Computational methodology and numerical results

Evaluation of the admissible correlation interval

The numerical calculations use one-dimensional adaptive quadrature and bracketed scalar root finding. For n≥2n \ge2, 0<p0≤p1<10 < p_0 \le p_1 < 1, and 0≤ρ0<ρˉ<10 \le\rho_0 < \bar{\rho} < 1, define the log probabilities an(p,r)=log⁡An(p,r)a_n(p,r) = \log A_n(p,r) and sn(p,r)=an(1−p,r)s_n(p,r) = a_n(1-p,r). With μ=μ(p,r)\mu= \mu(p,r) and σ=σ(r)\sigma= \sigma(r) from (14), the orthant integral takes the form

An(p,r)=1σ2π∫−∞∞exp⁡{nlog⁡Φ(x)−(x−μ)22σ2} dx,r>0.(33)A_n(p,r) = \frac{1}{\sigma\sqrt{2\pi}}\int_{-\infty}^{\infty}\exp\left\{n\log\Phi(x)-\frac{(x-\mu)^2}{2\sigma^2}\right\}\,dx,\qquad r>0. \tag*{(33)}

Writing h(x)=nlog⁡Φ(x)−(x−μ)2/(2σ2)h(x)=n\log\Phi(x)-(x-\mu)^2/(2\sigma^2) and x∗=arg⁡max⁡x∈Rh(x)x_*=\arg\max_{x\in\mathbb{R}}h(x), we evaluate

an(p,r)=h(x∗)−log⁡(σ2π)+log⁡∫−∞∞exp⁡{h(x)−h(x∗)} dx.(34)a_n(p,r)=h(x_*)-\log(\sigma\sqrt{2\pi})+\log\int_{-\infty}^{\infty}\exp\{h(x)-h(x_*)\}\,dx. \tag*{(34)}

The function hh is strictly concave and its mode is the unique root of

nϕ(x∗)Φ(x∗)=x∗−μσ2.(35)n\frac{\phi(x_*)}{\Phi(x_*)}=\frac{x_*-\mu}{\sigma^2}. \tag*{(35)}

The scaled integrand in (34) is bounded by one. The implementation evaluates the normal log distribution function directly, brackets the mode, and integrates over [x∗−13σ,x∗+13σ][x_*-13\sigma,x_*+13\sigma]. Strong concavity yields h(x)−h(x∗)≤−(x−x∗)2/(2σ2)h(x)-h(x_*) \le-(x-x_*)^2/(2\sigma^2), which bounds the omitted unnormalized integral by 2σ∫3∞exp⁡(−u2/2) du2\sigma\int_{3}^{\infty}\exp(-u^2/2)\,du. At r=0r=0, the analytic value an(p,0)=nlog⁡pa_n(p,0)=n\log p is used.

The feasible interval is computed through the monotone equations of Theorem 5.2. The following procedure also gives the one-step projection used in the repair examples.

Algorithm 1 (Feasible correlation band and projection).

  1. Compute a0=an(p0,ρ0)a_0=a_n(p_0,\rho_0) and s0=sn(p0,ρ0)s_0=s_n(p_0,\rho_0) using (34).

  2. Set ℓn=0\ell_n=0 if nlog⁡p1≥a0n\log p_1\ge a_0; otherwise solve an(p1,r)=a0a_n(p_1,r)=a_0 on (0,ρ0](0,\rho_0] by bracketed root finding.

  3. Set un=ρˉu_n=\bar{\rho} if sn(p1,ρˉ)≤s0s_n(p_1,\bar{\rho})\le s_0; otherwise solve sn(p1,r)=s0s_n(p_1,r)=s_0 on [ρ0,ρˉ)[\rho_0,\bar{\rho}).

  4. Given a proposed later correlation ρ^1\hat{\rho}_1, return [ℓn,un][\ell_n,u_n] and ρ1∗=min⁡{un,max⁡{ℓn,ρ^1}}\rho_1^*=\min\{u_n,\max\{\ell_n,\hat{\rho}_1\}\}.

The root brackets follow from Lemma 5.1, and each root is unique when its corresponding endpoint inequality is active. Along a horizon grid, repeating Algorithm 1 with the preceding repaired correlation produces the sequence in (18). The integrals and scalar roots determine the feasible interval without numerical integration of the full count distribution. Full-tail integration is reserved for verification and for the diagnostic VnV_n in (3).

The Python implementation uses absolute and relative quadrature tolerances of 10−1110^{-11} and 3×10−113\times10^{-11} for the mode-centered orthant integral. The scalar roots use a bracketed Brent solver with an absolute tolerance of 5×10−135\times10^{-13}; the mode solver uses 2×10−122\times10^{-12}. The independent full-tail quadrature uses absolute and relative tolerances 10−1210^{-12} and 3×10−123\times10^{-12}. For very large mathematical portfolio sizes the computation retains nn in the log-integrand formulation, avoiding the underflow that would arise from direct evaluation of small orthant probabilities.

Numerical verification

Two closed-form identities serve as analytic benchmarks. At p=ρ=1/2p=\rho=1/2, the mixing variable is uniform on (0,1)(0,1), so An=1/(n+1)A_n=1/(n+1). At n=2n=2 and p=1/2p=1/2, A2=1/4+arcsin⁡(ρ)/(2π)A_2=1/4+\arcsin(\rho)/(2\pi). Nine evaluations of these identities yield a maximum absolute error of 1.8e−151.8e-15 in the probability or log-probability scale, as appropriate. Thirty parameter-pair comparisons at n=8,25,60n=8,25,60 agree between the two-event criterion and an independent integration of all count-tail probabilities. The program also checks the monotonicity of the repaired survival and all-default probabilities across the five-horizon example.

A secondary accuracy run tightens integration tolerances and expands the mode-centered interval. The largest absolute change in the checked log-orthant probabilities is 2.8×10−142.8\times10^{-14}; the largest absolute change in a checked correlation boundary is 1.7×10−161.7\times10^{-16}. These are empirical numerical sensitivity diagnostics. All experiments use constructed parameters, and the computations are deterministic.

Increasing mean defaults with an incompatible distribution

Consider a portfolio of 100 borrowers. The earlier pair is (p0,ρ0)=(0.02,0.10)(p_0,\rho_0)=(0.02,0.10) and the later raw pair is (p1,ρ1)=(0.025,0.40)(p_1,\rho_1)=(0.025,0.40). Table 1 reports the two extreme events. Expected defaults rise from 2 to 2.5, and the probability of no defaults rises from 0.270717 to 0.514092. The probability of at least one default therefore falls by approximately 24.34 percentage points. This decline violates the stochastic ordering required for an absorbing coupling.

Forecastppρ\rhoEKEKP(K=0)\mathbb{P}(K = 0)log⁡10P(K=n)\log_{10}\mathbb{P}(K = n)
Earlier0.0200.1000002.00.270717-27.430
Later raw0.0250.4000002.50.514092-8.146
Later repaired0.0250.1530592.50.270717-18.947

Table 1. A constructed incompatibility and its one-step repair, n=100n = 100. The last column is the base-ten logarithm of the all-default probability.

The closest feasible later correlation is 0.153059. This keeps p1=0.025p_1=0.025 and the expected count 2.5 unchanged, while making the no-default probability equal to its earlier value. Figure 1 shows that every count tail is then nondecreasing. Direct enumeration gives the first incompatibility at n=7n=7 for this fixed parameter pair.

Figure 1. Changes in count-tail probabilities in the constructed example. A negative value contradicts absorption. The repaired curve satisfies all 100 count inequalities; the display focuses on thresholds up to 40.

For scale, consider 100 equal loans with a deterministic loss of $1 million per default. The expected loss in the first $1 million layer equals $1 million times the probability of at least one default. The raw forecast lowers that expected layer loss by approximately $243,375 as maturity increases. This illustrates the economic size of the coherence violation in the constructed scenario. The repaired distribution has V100=0V_{100} = 0, with the earlier marginal probability and the later marginal probability maintained at their stated levels.

Portfolio size and slow convergence

For (p0,p1,ρ0)=(0.02,0.03,0.15)(p_0, p_1, \rho_0) = (0.02, 0.03, 0.15) and cap ρˉ=0.99\bar{\rho} = 0.99, Table 2 reports the exact interval boundaries and the scaled distances predicted to converge to C=0.033826C = 0.033826. At n=100n = 100, the admissible interval is approximately [0.135019,0.253694][0.135019, 0.253694]. At n=106n = 10^6 it has narrowed to [0.142354,0.168643][0.142354, 0.168643].

nnℓn\ell_nunu_nlog⁡n(ρ0−ℓn)\sqrt{\log n}(\rho_0 - \ell_n)log⁡n(un−ρ0)\sqrt{\log n}(u_n - \rho_0)
10110^10.1182620.4956040.0481600.524429
10210^20.1350190.2536940.0321480.222524
10310^30.1388230.1981950.0293770.126669
10610^60.1423540.1686430.0284190.069293
101210^{12}0.1445030.1598040.0288970.051536
104010^{40}0.1468370.1542590.0303540.040875
1016010^{160}0.1483440.1519160.0317950.036777

Table 2. Exact finite-portfolio intervals and the asymptotic scaling. The two scaled columns converge to 0.033826. Very large sizes are mathematical diagnostics.

The upper and lower scaled distances remain visibly different at n=1012n = 10^{12}. Sizes up to 1016010^{160} serve as mathematical diagnostics of the slow asymptotic convergence. The leading-order formula identifies the contraction rate, while the scalar equations provide exact finite-nn boundaries. Figure 2 displays the shrinking interval and the scaled endpoints. All reported very-large-size calculations use the log-integrand representation in (33) to display convergence of the boundary expansion.

Figure 2. Admissible later correlations and their asymptotic scaling for p0=0.02p_0 = 0.02, p1=0.03p_1 = 0.03, ρ0=0.15\rho_0 = 0.15. The left panel displays sizes through 102010^{20}; the right panel includes much larger mathematical sizes to show the logarithmically slow convergence.

Sensitivity to default probabilities and initial correlation

The shape of the admissible interval depends on the default-probability increment, the earlier correlation, and portfolio size. Table 3 reports six parameter configurations, each evaluated at n=50n = 50 and n=500n = 500. All scenarios use the same correlation cap ρˉ=0.99\bar{\rho} = 0.99 and are constructed independently of observed loan data.

p0p_0p1p_1ρ0\rho_0nnℓn\ell_nunu_nWidth
0.0100.0150.10500.08900.27340.1844
0.0100.0150.105000.09270.15660.0638
0.0200.0250.10500.09290.17900.0861
0.0200.0250.105000.09540.12630.0310
0.0200.0300.15500.13270.29340.1608
0.0200.0300.155000.13800.20810.0701
0.0500.0650.15500.13530.22630.0910
0.0500.0650.155000.14010.18350.0434
0.0500.0650.30500.27620.38450.1083
0.0500.0650.305000.28270.34580.0631
0.1000.1250.30500.27380.36940.0956
0.1000.1250.305000.28130.33960.0583

Table 3. Sensitivity of the exact admissible correlation interval to cumulative default probabilities, earlier correlation, and portfolio size. The interval width equals un−ℓnu_n - \ell_n. All entries are deterministic.

For example, increasing the cumulative probability from 0.020 to 0.025 at ρ0=0.10\rho_0 = 0.10 permits later correlations in approximately [0.0929, 0.1790] for 50 borrowers, and [0.0954, 0.1263] for 500. The upper endpoint moves considerably more than the lower endpoint in this range. The scenarios with higher initial correlations also show material differences in the admissible downward and upward changes. These finite-size asymmetries support direct interval calculation when modelers assess maturity-specific parameter estimates.

Several horizons

Table 4 gives a five-horizon example with increasing individual default probabilities and oscillating raw correlations. The sequential projection keeps each probability unchanged. The repaired no-default probabilities decrease; the all-default probabilities, also checked in the supplied results, increase. Proposition 5.5 therefore gives an absorbing process for the entire repaired grid.

HorizonppRaw ρ\rhoRepaired ρ\rhoRaw survivalRepaired survival
10.0100.1200.1200000.4984120.498412
20.0180.3000.2911980.5062420.498412
30.0260.0800.2692290.1749540.382275
40.0330.2500.2545830.3005310.305172
50.0400.1500.2425140.1524220.244556

Table 4. Sequential repair for a constructed 100-borrower horizon curve. Raw and repaired survival refer to the probability that no borrower has defaulted.

The adjustment at the second horizon affects the feasible intervals at all subsequent horizons. The example shows how successive projections deliver a compatible sequence while carrying forward the dependence on earlier choices. Joint calibration across maturities would optimize those choices over the entire grid.

Modeling implications

The exact interval test can be incorporated into a modular multi-horizon credit-forecasting pipeline. At a common forecast origin, an individual borrower model supplies nondecreasing cumulative default probabilities, and a dependence model supplies an estimated latent correlation at each maturity. The Gaussian count-mixture law follows from these inputs. The two extreme-event probabilities then determine the admissible interval for the next correlation estimate. When a proposed correlation lies outside the interval, a sequential projection produces a coherent sequence while preserving the marginal probabilities.

The test has an interpretation through portfolio loss layers. A count-tail decline at threshold kk implies that the expected loss on the corresponding single-default-width layer decreases with maturity. The maximum decline VnV_n in (3) measures the magnitude of this inconsistency across thresholds. In the example in § 8.3, the most pronounced violation occurs at the first-default threshold and is economically appreciable. Other parameter choices can produce violations concentrated in extremely unlikely events, making the magnitude and location of the breach important alongside the binary feasibility decision.

The construction in Proposition 3.2 supplies an absorbing process on a prescribed horizon grid. The resulting temporal law is determined by its count quantiles and borrower permutation. Additional dynamics may be imposed according to the intended use, including conditional intensities, transition effects, and forward-starting exposures. The feasible correlation intervals provide restrictions on the horizon marginals that such dynamics must preserve.

Exchangeability and equal default losses make count stochastic order sufficient for the present construction. Heterogeneous borrowers can have the same count distribution while differing in their identity-specific default patterns, and unequal exposures alter the portfolio-loss order. Extensions to these settings require the corresponding multivariate and loss-distribution restrictions. The Gaussian specification also imposes a particular dependence tail structure; the single-crossing result offers a route to other mixing families when their distribution functions satisfy its hypothesis.

An empirical application would follow cumulative first-default outcomes for a fixed initial portfolio across several horizons, with explicit treatment of censoring and competing exits. Horizon-specific cumulative default probabilities and dependence parameters could be estimated from those histories, together with their joint uncertainty. The resulting confidence regions could be checked against the feasible correlation intervals. The deterministic calculations here isolate the structural restrictions and provide reproducible benchmarks for such an implementation.

Conclusion

A finite homogeneous portfolio admits maturity-dependent Gaussian correlations within an exact interval determined by the probabilities of no defaults and all defaults. Under a single-crossing condition for the mixing distribution functions, these two events characterize the stochastic order of the entire default-count distribution. The intervals are nested in portfolio size and contract towards the earlier correlation with a common leading boundary displacement of order (log⁡n)−1/2(\log n)^{-1/2}. The corresponding critical portfolio size grows exponentially in the inverse squared correlation discrepancy. The coupling, repair, and smooth-curve results provide complementary constructions for maturity-consistent portfolio forecasts. The numerical sensitivity analysis documents the finite-size asymmetry and its variation across default-probability and correlation settings.

References

  1. [1]Andersen, L. (2006). Portfolio losses in factor models: term structures and intertemporal loss dependence. Journal of Credit Risk, 2(4), 3–31. https://www.risk.net/journal-of-credit-risk/2160604/portfolio-losses-in-factor-models-term-structures-and-intertemporal-loss-dependence.DOI
  2. [2]Ansari, J., and Lütkebohmert, E. (2026). Robust Bernoulli mixture models for credit portfolio risk. Mathematical Finance, 36(3), 528–543. https://doi.org/10.1111/mafi.70020.
  3. [3]Boland, P. J., Proschan, F., and Tong, Y. L. (1989). Crossing properties of mixture distributions. Probability in the Engineering and Informational Sciences, 3(3), 355–366. https://doi.org/10.1017/S0269964800001224.
  4. [4]Hoeffding, W. (1963). Probability inequalities for sums of bounded random variables. Journal of the American Statistical Association, 58(301), 13–30. https://doi.org/10.1080/01621459.1963.10500830.
  5. [5]Gordy, M. B. (2003). A risk-factor model foundation for ratings-based bank capital rules. Journal of Financial Intermediation, 12(3), 199–232. https://doi.org/10.1016/S1042-9573(03)00040-8.
  6. [6]Jackson, K., Kreinin, A., and Zhang, W. (2011). Analytic dynamic factor copula model. Technical report, University of Toronto. https://www.cs.toronto.edu/pub/reports/na/Analytic.Dynamic.Factor.Copula.2011a.Models.pdf.
  7. [7]Li, D. X. (2000). On default correlation: a copula function approach. Journal of Fixed Income, 9(4), 43–54. https://doi.org/10.3905/jfi.2000.319253.
  8. [8]Óskarsdóttir, M., and Bravo, C. (2021). Multilayer network analysis for improved credit risk prediction. Omega, 105, 102520. https://doi.org/10.1016/j.omega.2021.102520.
  9. [9]Plackett, R. L. (1954). A reduction formula for normal multivariate integrals. Biometrika, 41(3–4), 351–360. https://doi.org/10.1093/biomet/41.3-4.351.
  10. [10]Rinott, Y., and Rotar, V. (2001). A remark on quadrant normal probabilities in high dimensions. Statistics & Probability Letters, 51(1), 47–51. https://doi.org/10.1016/S0167-7152(00)00141-3.
  11. [11]Shaked, M., and Shanthikumar, J. G. (2007). Stochastic Orders. Springer Series in Statistics. Springer, New York. https://doi.org/10.1007/978-0-387-34675-5.
  12. [12]Strassen, V. (1965). The existence of probability measures with given marginals. Annals of Mathematical Statistics, 36(2), 423–439. https://doi.org/10.1214/aoms/1177700153.
  13. [13]Torresetti, R., Brigo, D., and Pallavicini, A. (2007). Implied expected tranched loss surface from CDO data. Working paper, revised May 2007. https://doi.org/10.2139/ssrn.933291.
  14. [14]Vasicek, O. A. (2002). Loan portfolio value. Risk, 15(12), 160–162. https://www.risk.net/risk-management/credit-risk/1530287/loan-portfolio-value.
  15. [15]Yu, Y. (2009). Stochastic ordering of exponential family distributions and their mixtures. Journal of Applied Probability, 46(1), 244–254. https://doi.org/10.1239/jap/1238592127.
  16. [16]Zandi, S., Korangi, K., Óskarsdóttir, M., Mues, C., and Bravo, C. (2025). Attention-based dynamic multilayer graph neural networks for loan default prediction. European Journal of Operational Research, 321(2), 586–599. https://doi.org/10.1016/j.ejor.2024.09.025.

Paper details

vlnc:W_s7cyx8zx
Version
v1 · latest
Published
Fresh — added to the Hub within the last 30 days
Added
License
CC BY 4.0
Usage

views · 0

downloads · 0

A view counts once per browser per day, a download each time the PDF is requested.

  • Economics, Econometrics and Finance

Not recorded for this paper

View statement
Looking for related papers…
Contents