betaFeedback

WorksEconomics, Econometrics and Finance

Adapted Wasserstein Distance for Credit Risk: Rating Information and Model-Transfer Hedging

Abstract

Two credit models can quote the same bond prices today and still disagree on how information about the issuer unfolds over time. We quantify this disagreement by means of the adapted Wasserstein distance, an optimal transport distance between laws of processes consistent with the flow of information associated with them. For defaultable bond price processes, the squared distance is the smallest probability by which the two default indicators must disagree. This yields closed-form distances for rating reviews and a dynamic programming principle for multi-period rating migration. For a single review, the distance can be estimated from observed posterior survival probabilities at the parametric rate without losing calibration. For intensity models, we obtain an explicit formula for the adapted distance between bond prices that controls the error of bootstrap hedges along a quoted term structure, determine its sharp range over hazard schedules calibrated to the same prices, and show that it is stable under perturbations of the hazard rates. Finally, we complement the theoretical findings with applications to S&P transition data and Monte Carlo experiments to illustrate the size of these effects.

Introduction

A credit model is calibrated to the prices at which the bonds and credit default swaps of an issuer trade today. With zero recovery and deterministic interest rates, these prices fix the survival probabilities of the issuer at the quoted maturities. However, they determine neither the survival curve between these maturities nor, more importantly, how these probabilities are revised as information about the issuer reaches the market. Two models can therefore agree on every quoted price and still describe very different markets: in one of them the credit quality of the issuer is revealed gradually, in the other it is revealed in a single rating review. Consequently, in one of them a desk that hedges a long bond by rolling short bonds never has to inject cash; in the other it does. In this paper, we quantify this kind of disagreement, computing it explicitly in the models used in credit practice, and we relate it to the hedging errors it produces.

The basic object is the price of a defaultable zero-coupon bond with zero recovery. With zero interest rates, its price at a date tt before maturity NN, on survival of the issuer, is the probability of surviving until NN given the information available at tt. We refer to it as the defaultable bond price process. It is a martingale starting at price pp quoted today, and at maturity it equals 1 if the issuer has survived and 0 otherwise. A pricing model is the law of this process, together with the information it carries. If two models quote the same price pp, their bond price processes have the same law at maturity, namely a Bernoulli law with parameter pp, and they are distinguished solely by the way their prices move in between.

As a guiding example, consider a bond maturing at date 22 and two models with the same price pp. In the coarse model, no information is unveiled before maturity, and the bond price stays equal to pp until it jumps to 00 or 11. In the rating model, a review at date 11 occurs and moves the bond price to a random level QQ, the posterior survival probability given the outcome of the review, with EQ=p\mathbb{E}Q=p so that the price today is still pp. For instance, we may consider Q=p±δQ=p\pm\delta with probability one half each. The two models quote the same price and predict the same probability of default, but a desk that hedges in the rating model must trade at date 11, while in the coarse model it need not.

A natural way of comparing two pricing models is provided by optimal transport. Optimal transport compares two probability laws by constructing both random objects on a common space, a coupling, and measuring how far apart they are on average. For two pricing models, this means generating a price path of each model jointly and comparing the paths. However, the ordinary Wasserstein distance allows every coupling and hence is agnostic of the time information. In the guiding example, it may first draw the whole path of the rating model, outcome of the review included, and then let the coarse model default exactly when the rating model does. The two default indicators then always agree. The only discrepancy is the price gap ∣Q−p∣|Q-p| at date 11, and the distance is of order δ\delta. This coupling uses information the coarse model does not have: at date 11 the coarse model knows nothing, yet its future default has been chosen to match the outcome of the review.

To overcome this flaw, we admit only couplings that respect the flow of information: at every date, the part of the coupling constructed for one model may depend on the past of the other model, but not on its future. Such couplings are called bicausal [9, 53], and the adapted Wasserstein distance is the optimal transport cost computed over them [5, 12, 64]. As cost we take the squared differences of the price increments, since this is the quantity that controls the profit and loss of a common trading strategy.1 In the guiding example, bicausality forces the default of the coarse model to be independent of the outcome of the review. If the review has moved the survival probability to p+δp+\delta, say, the coarse model still survives with probability pp, so the two defaults must disagree with probability at least δ\delta, and each disagreement costs a jump of size one. The squared adapted distance is therefore of order δ\delta, and the distance itself of order δ\sqrt{\delta}, much larger than δ\delta for small revisions. Since the two models have the same initial price and the same terminal survival law, the whole discrepancy is produced by information.

The same mechanism is at work in the intensity models that are standard in credit practice, where default can occur at any time and arrives with a hazard rate, the instantaneous probability of default per unit of time given survival [31, 50]. For two hazard paths, we show that the squared adapted distance between the two bond price processes is again the smallest probability that their survival indicators disagree at maturity, when both are constructed on a common flow of information. We call it the adapted default mismatch, and we show that it has an explicit formula and is attained by a common-shock construction. It also has a direct meaning for hedging, which makes precise the situation described at the beginning of the introduction. A natural strategy is to fit a bootstrap model, that is, a credit curve with piecewise constant hazard rate that reprices the quoted bonds [62], and to hedge a long bond by rolling short ones, sized by the forward survival factors of the fitted curve. We call this the bootstrap hedge. It replicates the long bond exactly under the bootstrap model. Under a market model with rating reviews, however, it requires cash injections, and the adapted default mismatch controls their size.

Main results

Our results on defaultable bond prices can be summarized as follows.

Theorem 1.1 (Information revealed by credit ratings). Let XX and YY be the defaultable bond price processes of two models with the same initial price pp.

(i) The squared adapted Wasserstein distance between the two models is the smallest probability, over bicausal couplings, that their default indicators disagree at maturity, see Proposition 2.9 for a precise statement.

(ii) If one model reveals nothing before maturity and the other reveals the posterior QQ at an intermediate date, then AW22=E∣Q−p∣AW_2^2 = \mathbb{E}|Q-p|, while the ordinary Wasserstein distance satisfies E∣Q−p∣2≤W22≤2E∣Q−p∣2\mathbb{E}|Q-p|^2 \le W_2^2 \le2\mathbb{E}|Q-p|^2, see Theorems 3.6 and 3.8 for a precise statement.

(iii) If both models reveal the same posterior, but at different dates, then AW22=E∣Q−Q~∣AW_2^2 = \mathbb{E}|Q-\widetilde{Q}|, where Q~\widetilde{Q} is an independent copy of QQ (see Theorem 3.13 for a precise statement).

(iv) If both defaultable bond prices are driven by finite-state rating chains, the distance is the value of a backward recursion over pairs of current ratings. If migration is monotone in credit quality, the recursion is solved by driving both chains with the same uniform random variables, see Theorems 3.16 and 3.18 for a precise statement.

In particular, close to the coarse model, that is, when the review revises the survival probability only slightly, the two distances behave very differently. For a symmetric revision Q=p±δQ = p \pm\delta, part (ii) gives W2W_2 of order δ\delta but AW2AW_2 of order δ\sqrt{\delta}, so that the ratio AW2/W2AW_2/W_2 is of order δ−1/2\delta^{-1/2} and blows up as δ→0\delta\to0. A small rating revision is therefore almost invisible to the ordinary Wasserstein distance, while the adapted distance registers it at a much larger scale.

Parts (i) and (iv) build on known results. That, for martingale laws, the adapted cost reduces under bicausal couplings to the squared difference of the terminal values is used by Backhoff-Veraguas, Bartl, Beiglböck and Eder in their computation for geometric Brownian motions [5 Proposition 3.6]. What part (i) adds is that for defaultable bond prices the terminal values are default indicators, so that the terminal cost is a probability of disagreement, which can be reduced once more to the prices at the last date before maturity. Part (iv) is a consequence of part (i), of the dynamic programming principle of [9], and of the optimality of the synchronous, or Knotthe-Rosenblatt, coupling between stochastically monotone Markov processes, which goes back to Rüschendorf [67] and is proved in the bicausal setting in [9 Proposition 5.3]. Parts (ii) and (iii), and more generally the explicit formulas for the information structures of rating-based models, follow from the second reduction in part (i). Together with part (iv), they turn the adapted distance into an exactly computable measure of how differently two calibrated rating systems release information, for instance, a point-in-time and a through-the-cycle rating philosophy.

In the model with one review, the distance can also be estimated from data. If the law of the posterior is replaced by the empirical law of observed posteriors, rescaled so that it still reproduces the price pp, the squared adapted distance is estimated with expected error at most n−1/2n^{-1/2}, see Theorem 3.10.

For intensity models, our results read as follows.

Theorem 1.2 (Default mismatch and model-transfer hedging). Let aa and bb be hazard paths with cumulative hazards At=∫0tas dsA_t = \int_0^t a_s\,\mathrm{d}s, Bt=∫0tbs dsB_t = \int_0^t b_s\,\mathrm{d}s and Mt=∫0tmax⁡{as,bs} dsM_t = \int_0^t \max\{a_s,b_s\}\,\mathrm{d}s.

(i) If aa and bb are deterministic, the squared adapted Wasserstein distance between the price processes of the zero-recovery bonds maturing at tt in the two Cox models is the smallest probability, over bicausal couplings, that their survival indicators disagree at tt, and it equals

Ψt(a,b)=e−At+e−Bt−2e−Mt.\Psi_t(a,b) = e^{-A_t} + e^{-B_t} - 2e^{-M_t}.

It is attained for all maturities simultaneously by a common shock with intensity min⁡{a,b}\min\{a,b\}. For random intensities, among couplings that preserve the doubly stochastic structure of both models, the problem reduces to an adapted transport problem between the factor laws with cost Ψ\Psi, see Theorem 4.3 and Theorem 4.11 for a precise statement.

(ii) The expected absolute cash residual of the bootstrap roll, implemented under a rating-driven market model, equals the expected absolute difference of the two survival factors at maturity, and is bounded by EΨ\mathbb{E}\Psi. The gap is produced by crossings of the two hazard paths inside the bucket and is given explicitly, see Theorem 4.16 for a precise statement.

(iii) Along a quoted term structure, the expected absolute cash residuals of the rolling bootstrap are bounded bucket by bucket by the adapted default mismatch between the market model and a model that switches to the bootstrap one bucket ahead, with equality when ratings are reviewed at the calibration dates and select flat hazards for the following buckets, see Theorem 4.26 for a precise statement.

Part (i) is the continuous-time counterpart of Theorem 1.1(i). In both settings, the squared adapted distance between two defaultable bond prices is the minimal bicausal probability that the two defaults disagree, and for deterministic hazards this probability is available in closed form. The bounds in parts (ii) and (iii) are therefore bounds by adapted Wasserstein distances between bond prices.

Calibration to the quoted prices leaves the adapted default mismatch largely undetermined. If the hazard rates are bounded, we compute its exact range over all hazard schedules with prescribed survival probabilities at the ends of a bucket, see Theorem 4.7. In particular, two schedules with the same survival probability can have a mismatch as large as the intensity bound allows, see Corollary 4.8. The mismatch is moreover Lipschitz in the hazard rates with respect to the L1L^1 norm, with sharp constant one, which controls the error made when hazard paths are approximated, see Theorem 4.9.

Finally, a desk that knows a rating-driven market model and can trade bonds of several maturities can remove its hedging error altogether. Indeed, in the time-homogeneous rating-driven market considered below [60], finitely many consecutive maturities span the price of the long bond, see Proposition 4.28. We use this in Section 5.3 to separate two sources of hedging error that are usually confounded: missing instruments and a misspecified model. We also show that, in the model with one review, an adapted Wasserstein ball around the coarse model is exactly a bound on the mean absolute deviation of the posterior, which yields explicit robust prices for options on the bond price at the review date. Moreover, under calibrated perturbations of rating-dependent hazard rates, the ordinary distance grows linearly in the size of the perturbation, while the adapted one grows like its square root, see Theorem 4.35 and Corollary 4.36.

In Section 5 we apply these results to the rating transition data published by S&P Global Ratings [69]. We compare rating systems that agree on today’s bond price but release information differently: U.S. against European rating histories, the long-run average against a single recent year, and a point-in-time system that reacts to the business cycle against a through-the-cycle one. The adapted distance between such systems turns out to be a sizeable fraction of their distance from a model that reveals no information at all. The way in which information is released therefore matters even when the initial price of the bond being compared agrees, while static measures of rating mobility can hardly tell some of these systems apart. Agency matrices are not exactly monotone, but the synchronous coupling of Theorem 1.1(iv) remains optimal up to a negligible error so that the distances can be computed in closed form. We then simulate the bootstrap hedge in a market calibrated to the same data. The bound of Theorem 1.2(iii) is exact when ratings change only at the calibration dates and still captures most of the realised error when they change in between. Finally, when the desk can trade bonds of several maturities, a few bonds suffice to remove the hedging error if the market model is known, whereas the bootstrap hedge improves only slowly. With a rich hedging universe, most of the remaining error therefore comes from the model rather than from missing instruments.

Comments on the proofs. All explicit results on defaultable bond prices rest on a single observation. Under a bicausal coupling, neither model learns anything about the other’s future, so both price processes remain martingales with respect to the joint flow of information, and so does their difference. The martingale isometry then turns the sum of squared increment differences into the squared difference of the terminal values, which for defaultable bond prices is the indicator that the two defaults disagree. Conditioning on the information at the last date before maturity, the two defaults are Bernoulli variables with parameters equal to the current prices, and the elementary fact that two Bernoulli variables with parameters aa and bb disagree with probability at least ∣a−b∣\lvert a-b\rvert reduces the problem further to the distance between the prices at that date (Proposition 2.9). The explicit formulas of Section 3 follow by identifying which variables bicausality forces to be independent, typically because the history of one model is deterministic up to the relevant date.

For rating chains, causality and the Markov property imply that, given the current pair of ratings, the next pair is distributed according to a coupling of the two one-step transition laws, which yields the dynamic programming principle of [9] as a backward recursion over pairs of current states. When migration is monotone, every value function of the recursion has the Monge property, under which the comonotone coupling is optimal at each step, and composing the comonotone couplings gives the synchronous coupling. We include this short argument to keep the paper self-contained.

For intensity models with deterministic hazards, the bond price is a martingale of finite variation that jumps only at default, and its filtration is that of the default indicator. The quadratic variation of the difference of two such prices is the sum of its squared jumps, and the martingale isometry again turns the cost into the probability that the two defaults disagree. Bicausality forces each default indicator to keep its hazard rate in the joint flow of information. The lower bound is then obtained by following the probability that both names survive. Integration by parts expresses its evolution through the two hazard rates and the rate of simultaneous default, which cannot exceed the smaller hazard. A differential inequality and Grönwall’s lemma then bound the joint survival probability by e−Mte^{-M_t}, and three independent Poisson processes, one common and two idiosyncratic, attain the bound. That this construction is bicausal follows from the martingale representation of single-jump processes. For random intensities, conditioning on the factor paths reduces the problem to deterministic hazards. The hedging results all follow the same pattern: by the Cox pricing formula, each cash residual is the conditional expectation, given the current information, of a difference of survival factors under the market model and under a bootstrap or hybrid model. Conditional Jensen’s inequality and the explicit form of Ψ\Psi then separate the realised error, the information arriving inside a bucket, and the adaptation slack.

Related literature and what is new

In this section we place our results in the context of the literature on adapted transport, stability in mathematical finance, explicit solutions of transport problems and credit risk, and we highlight where our main contributions lie.

Adapted transport. The idea of comparing stochastic processes through couplings that respect the flow of information appears independently in several communities. A first construction in this direction is Vershik’s iterated Kantorovich distance, developed for the classification of filtrations [71], and the bicausality condition already appears in Rüschendorf’s work on couplings of Markov processes [67]. Independently, Pflug and Pichler introduced nested distributions and the nested distance, a metric on scenario trees under which the values of multistage programs are continuous, and were the first to exploit it systematically in multistage stochastic optimization [63, 64, 65]. In optimal transport, causal and bicausal couplings were formulated by Lassalle and by Backhoff-Veraguas, Beiglböck, Lin and Zalashko [9, 53]. The latter also derived a dynamic programming principle for bicausal transport in discrete time and showed that the Knothe-Rosenblatt rearrangement is optimal for co-monotone processes. The relation between causal couplings and enlargement of filtrations was clarified in [2]. Backhoff-Veraguas, Bartl, Beiglböck and Eder showed that the resulting topologies all coincide, and that they agree with older notions such as Aldous’ extended weak topology and Hellwig’s information topology [6], while fundamental metric properties of these distances were established in [8]. Bartl, Beiglböck and Pammer identified the completion of the space of laws under the adapted distance with the space of filtered processes, and showed that it is a geodesic space in which martingales form a closed convex subset [12]. Statistical estimation, which requires modified empirical measures because the ordinary empirical measure does not converge in adapted topologies, and numerical methods are treated in [7, 32]. Continuous-time versions of the theory are developed in [10, 13, 16, 20].

Stability in mathematical finance. The paper closest to ours in spirit is [5]. There, Backhoff-Veraguas, Bartl, Beiglböck and Eder define the adapted distance between semimartingale laws through the quadratic variation of the difference of the martingale parts and the total variation of the difference of the drifts, and explain why this cost, rather than the nested distance, which uses the cost on the path values, has the right scaling for a continuous-time limit [5 Section 3.3]. They prove that under bicausal couplings the coordinate processes keep their semimartingale decompositions [5 Lemma 3.1], which is the fact behind our terminal representation, and they bound the difference of the stochastic integrals of a common bounded strategy [5 Lemma 3.9]. From these estimates, they derive Lipschitz stability of superhedging errors, both when the hedge is adapted to the new model and when the same Lipschitz strategy is used in both [5 Theorem 1.5], of average value at risk, of utility maximization and of utility indifference prices, with bounds that are sharp for Brownian models and call options. They also compute the distance explicitly for diffusions with suitably simple coefficients, where the synchronous coupling is optimal, and for geometric Brownian motions, and show by example that the synchronous coupling is not optimal in general [5 Section 3.2]. Earlier, Glanzer, Pflug and Pichler bounded acceptability prices of contingent claims through the nested distance [38]. Bartl and Wiesel computed the first-order sensitivity of multiperiod optimization problems with respect to adapted perturbations [15], and Acciaio, Backhoff-Veraguas and Pammer quantified the fundamental theorem of asset pricing under adapted perturbations [1]. These results are bounds of the form ∣V(P)−V(Q)∣≤C AW(P,Q)|V(P)-V(Q)| \le C\,AW(P,Q) for broad classes of value functionals. We use them rather than reprove them: the stability of prices and of a common hedge for defaultable bond prices, Remark 3.21 and Proposition 3.22, is a direct instance of [5], recorded with the constants we need. Robust pricing over ordinary Wasserstein balls is studied in [14, 21, 59], model uncertainty more broadly in [28, 39], and the performance of hedges under a misspecified model in [33, 35].

Explicit solutions. Explicit solutions of adapted transport problems are available only in a few cases. Besides the diffusion examples of [5] and their extension to stochastic differential equations in [10], where the synchronous coupling is optimal, the most developed case is the Gaussian one. Gunasingam and Wong computed the adapted Wasserstein distance between Gaussian processes in discrete time in closed form [41] and extended the analysis to filtered Gaussian processes [42]. Acciaio, Hou and Pammer obtained closed forms for multidimensional Gaussian processes and for the entropic regularization of the problem [4], and the geometry of the resulting adapted Bures-Wasserstein space is studied in [3]. Barycenters of Gaussian processes were characterized in [40], using a decomposition of the squared adapted distance into classical squared Bures-Wasserstein distances between covariance contributions of successive innovations. For Markov processes, Rüschendorf identified the optimal coupling between stochastically monotone chains [67], and for stationary finite-state Markov chains O’Connor, McGoff and Nobel compute optimal transition couplings by policy iteration [61]. The present paper adds explicit solutions of a different kind. The laws involved are martingales with binary terminal values, which are very different from Gaussian laws, and the closed forms come from the terminal representation rather than from linear algebra. In continuous time, Theorem 4.3 gives the adapted distance between the price processes of two Cox bonds with deterministic hazards in closed form, with a common-shock optimal coupling. We present this as an explicit solution for a particular class of single-jump credit martingales, without claiming priority over the broader coupling literature on jump processes.

Credit risk. Intensity-based models represent default as the first jump of a process with a given hazard rate, possibly driven by economic factors, and are the standard framework for pricing and hedging defaultable claims [18, 19, 31, 34, 47, 50, 51, 68]. Rating-migration models describe credit quality by a Markov chain on rating classes [45, 48], and the role of the information available to investors, which is the object of this paper, has been emphasized in models with incomplete or delayed information [30, 36, 46]. Wasserstein distances have been used to proxy credit curves from static default distributions [58], a setting in which information plays no role. Relative entropy, the usual alternative for comparing or pooling models [57], is infinite in the direction from a nondegenerate one-review model to the coarse benchmark, because the former is not absolutely continuous with respect to the latter (Remark 3.5). The same obstruction motivates the transport criteria of [70] for diffusion models. The optimal coupling in Theorem 4.3 is a Marshall-Olkin common shock [56], and couplings of killing times of this type are classical in probability, where they appear as optimal co-adapted couplings [24, 27, 54]. In credit risk, however, adapted transport does not seem to have been used so far. to compare models or to quantify the hedging errors of credit strategies, and this is the gap that the present paper addresses.

What is new. In view of the above, the contributions of this paper are the following. First, the identification of the squared adapted distance between defaultable bond prices with a minimal probability of default disagreement, and its reduction to the last pre-maturity prices, which yield closed forms for one review, for a change of its date and for the no-information benchmark. Second, the combination of this representation with the classical synchronous coupling, which turns the comparison of calibrated rating systems into an exact computation, carried out on agency data. Third, in continuous time, the closed-form adapted Wasserstein distance between the price processes of Cox bonds with deterministic hazards, attained by a common shock, and the reduction of factor-driven intensities to an adapted transport problem between factor laws. Fourth, exact identities and bounds for the hedging errors of bootstrap strategies, in a single bucket, along a term structure and with several instruments, together with their empirical evaluation. The spanning of rating-driven markets by finitely many bonds is a discrete-time instance of known completeness results [19, 60], and what is new there is the quantitative comparison of bootstrap and market-based strategies on the same hedging universe. The stability bounds of [5] are used, not reproved, and serve as a benchmark against which the exact identities can be compared.

Open problems

We conclude this section with some questions that arise in view of our results.

  1. Nonlinear credit claims. Our exact identities exploit the fact that the value of a zero-coupon bond, and more generally of a claim with deterministic cash flows such as a coupon bond or a leg of a credit default swap, is a linear function of the survival curve. The cash residual of a hedge is then a difference of survival factors, and its size is controlled by the adapted default mismatch. Many traded credit claims are not of this form. Callable and putable bonds contain exercise decisions. Step-up coupons and rating triggers make payments depend on future ratings, and options on bonds or on credit spreads depend nonlinearly on the conditional survival curve. For such claims, the hedging error is no longer determined by the default mismatch alone, since it also depends on how the two models revise the value of the claim over time. A natural goal is to bound, or ideally to compute, the largest model-transfer error over a class of such claims by a transport problem. Because the value of a nonlinear claim depends on conditional laws of the future, the relevant cost should be of weak type, that is, a function of conditional distributions rather than of paths. The open question is whether this problem admits a dynamic programming principle that makes it as tractable as the one used here for linear claims.

  1. Continuous information. In our rating models information arrives at finitely many dates, typically rating reviews. In practice, credit information is also revealed continuously, through the observation of spreads, accounting data or the issuer’s asset value, and in filtering models the price of a defaultable bond evolves in continuous time [30, 36]. The terminal representation only uses the martingale property, and Theorem 4.3 extends it to Cox bond prices with deterministic hazards, but computing the distance is harder when prices move continuously before default. A natural route is to discretize the observation dates and pass the multi-date formulas to the limit. This requires tightness of almost optimal bicausal couplings, a proof that the bicausality constraint survives in the limit, and control of the coupling of the two default times at maturity. In filtering models,

the natural state variable is the conditional law of the unobserved credit factor, so that the problem becomes a transport problem between laws of measure-valued processes, in the continuous-time framework of [10, 13]. Explicit formulas in simple filtering models, comparable to those obtained here for rating reviews, would already be of interest.

(3) Estimation. The rating systems compared in § 5 are estimated from finite samples of rating histories. Transitions out of investment-grade classes, and defaults in particular, are rare, histories are censored when ratings are withdrawn, and the estimated matrices are historical rather than risk neutral [37, 44, 52]. Part of the adapted distance between two estimated systems may therefore reflect sampling noise rather than a genuine difference in the way information is released. Separating the two requires nonasymptotic bounds, or confidence intervals, for the adapted distance between a Markov chain and its estimate. In the finite-state case, such bounds should follow from the concentration of transition counts, in the spirit of the estimation results of [7]. A related question is how sensitive the distances are to the choice of the adjustment from historical to risk-neutral migration.

(4) Portfolios and aggregation. Our results compare a single name under two models, whereas credit indices, tranches and counterparty valuation adjustments depend on dependent defaults across many names. Comparing two portfolio models requires coupling vectors of default indicators, and it is not clear whether the multivariate Marshall-Olkin construction [56] remains optimal when defaults are dependent within each model. A related question is the aggregation of several credit models, such as different rating systems, into a consensus model. Candidates are the arithmetic pooling of experts in the spirit of [70] and adapted barycenters [12, 40]. Since the adapted distance controls hedging errors, the distance of the consensus to each expert would bound the error of hedging with it when one of the experts is correct.

(5) Couplings beyond the doubly stochastic class. For factor-driven intensities, Theorem 4.11 minimizes the default mismatch over couplings that preserve the Cox structure of both models. Bicausal couplings of the full processes may also let the joint coupling of future factor innovations depend on the observed default history, while preserving each marginal conditional law, as explained in Remark 4.12. Whether such feedback can strictly reduce the mismatch is open.

Outline of the paper

Section 2 develops the filtered price-law setting and the terminal transport reduction. Section 3 treats one-review, timing and migration models, including calibration-preserving statistical estimation and dynamic hedging. Section 4 establishes the Cox mismatch formulas, sharp hazard envelopes and perturbation bounds, and analyzes model-transfer hedging, robust pricing and term-structure effects. Section 5 contains the empirical illustrations.

Defaultable bond prices and adapted distances

Standing assumptions

Throughout the paper, unless specified differently, we assume that the time is discrete, the interest rates are zero, and the recovery rate is zero. Moreover, we assume that all probability spaces are standard Borel spaces, so that regular conditional distributions exist.

Defaultable bond prices

First, we make precise the objects described in the introduction. Fix an integer N≥2N \ge2 and index time by {0,1,…,N}\{0,1,\ldots,N\}. Let (Ω,F,F,Q)(\Omega,\mathcal{F},\mathbb{F},\mathbb{Q}) be a filtered probability space with filtration F=(Ft)t=0N\mathbb{F} = (\mathcal{F}_t)_{t=0}^{N} with F0={∅,Ω}\mathcal{F}_0 = \{\varnothing, \Omega\}, where Q\mathbb{Q} is a pricing measure and Ft\mathcal{F}_t represents the information available to the market at date tt. We take the money market account as numeraire and normalize it to one, so that interest rates are zero and all prices are discounted. The default time of the issuer is an F\mathbb{F}-stopping time, with τ=+∞\tau= +\infty if the issuer never defaults.

Consider a zero-recovery defaultable zero-coupon bond maturing at NN, with terminal payoff D=1{τ>N}D = \mathbf{1}_{\{\tau>N\}}. Its price process is

St=EQ[D∣Ft],0≤t≤N.(1)S_t = \mathbb{E}_{\mathbb{Q}}[D \mid\mathcal{F}_t], \qquad0 \leq t \leq N. \tag*{(1)}

Then SS is a bounded Q\mathbb{Q}-martingale with values in [0,1][0,1], and SN=D∈{0,1}S_N = D \in\{0,1\}. The initial price is p=S0=Q[τ>N]p = S_0 = \mathbb{Q}[\tau> N]. Throughout we assume p∈(0,1)p \in(0,1), since the degenerate values p∈{0,1}p \in\{0,1\} carry no default uncertainty.

Let ΩN=[0,1]N+1\Omega_N = [0,1]^{N+1} with canonical process Xt(x)=xtX_t(x) = x_t and canonical filtration FtX=σ(X0,…,Xt)\mathcal{F}_t^X = \sigma(X_0,\ldots,X_t). For p∈(0,1)p \in(0,1) define

MN(p)={P∈P(ΩN):X0=p P-a.s., X is a P-martingale, XN∈{0,1} P-a.s.}.(2)\mathcal{M}_N(p) = \left\{\mathbb{P} \in\mathcal{P}(\Omega_N) : X_0 = p\ \mathbb{P}\text{-a.s.},\ X\text{ is a }\mathbb{P}\text{-martingale},\ X_N \in\{0,1\}\ \mathbb{P}\text{-a.s.}\right\}. \tag*{(2)}

Every price process of a zero-recovery defaultable zero-coupon bond induces a law in MN(p)\mathcal{M}_N(p). The following lemma shows that all these laws coincide at maturity, so that the initial price alone determines the law of the survival indicator at maturity.

Lemma 2.1 (Terminal law). Every P∈MN(p)\mathbb{P} \in\mathcal{M}_N(p) satisfies P[XN=1]=p\mathbb{P}[X_N = 1] = p and P[XN=0]=1−p\mathbb{P}[X_N = 0] = 1-p.

Proof. By the martingale property, EP[XN]=EP[X0]=p\mathbb{E}_{\mathbb{P}}[X_N] = \mathbb{E}_{\mathbb{P}}[X_0] = p. Since XN∈{0,1}X_N \in\{0,1\}, EP[XN]=P[XN=1]\mathbb{E}_{\mathbb{P}}[X_N] = \mathbb{P}[X_N = 1]. □\square

Lemma 2.1 formalizes the calibration issue: once the time-zero bond price pp is fixed, all pricing laws in MN(p)\mathcal{M}_N(p) share the same terminal Bernoulli law. At the same time, their intermediate conditional survival probabilities and filtrations may differ.

Bicausal couplings

References

  1. [1]B. Acciaio, J. Backhoff-Veraguas, and G. Pammer. Quantitative fundamental theorem of asset pricing. Mathematical Finance, 35:636–660, 2025.
  2. [2]B. Acciaio, J. Backhoff-Veraguas, and A. Zalashko. Causal optimal transport and its links to enlargement of filtrations and continuous-time stochastic optimization. Stochastic Processes and their Applications, 130(5):2918–2953, 2020.arxiv.org/abs/1611.02610
  3. [3]B. Acciaio, D. Bartl, A. Grass, S. Hou, and G. Pammer. The geometry of the adapted Bures–Wasserstein space. arXiv:2602.00623, 2026.arxiv.org/abs/2602.00623
  4. [4]B. Acciaio, S. Hou, and G. Pammer. Entropic adapted Wasserstein distance on Gaussians. Electronic Communications in Probability, 30:1–14, 2025.
  5. [5]J. Backhoff-Veraguas, D. Bartl, M. Beiglböck, and M. Eder. Adapted Wasserstein distances and stability in mathematical finance. Finance and Stochastics, 24(3):601–632, 2020.
  6. [6]J. Backhoff-Veraguas, D. Bartl, M. Beiglböck, and M. Eder. All adapted topologies are equal. Probability Theory and Related Fields, 178:1125–1172, 2020.
  7. [7]J. Backhoff-Veraguas, D. Bartl, M. Beiglböck, and J. Wiesel. Estimating processes in adapted Wasserstein distance. The Annals of Applied Probability, 32(1):529–550, 2022.
  8. [8]J. Backhoff-Veraguas, M. Beiglböck, M. Eder, and A. Pichler. Fundamental properties of process distances. Stochastic Processes and their Applications, 130(9):5575–5591, 2020.arxiv.org/abs/1701.03955
  9. [9]J. Backhoff-Veraguas, M. Beiglböck, Y. Lin, and A. Zalashko. Causal transport in discrete time and applications. SIAM Journal on Optimization, 27(4):2528–2562, 2017.arxiv.org/abs/1606.04062
  10. [10]J. Backhoff-Veraguas, S. Källblad, and B. A. Robinson. Adapted Wasserstein distance between the laws of stochastic differential equations. Stochastic Processes and their Applications, 189:104689, 2025.
  11. [11]A. Bangia, F. X. Diebold, A. Kronimus, C. Schagen, and T. Schuermann. Ratings migration and the business cycle, with application to credit portfolio stress testing. Journal of Banking & Finance, 26(2–3):445–474, 2002.DOI
  12. [12]D. Bartl, M. Beiglböck, and G. Pammer. The Wasserstein space of stochastic processes. Journal of the European Mathematical Society, 28(1):393–454, 2026.
  13. [13]D. Bartl, M. Beiglböck, G. Pammer, S. Schrott, and X. Zhang. The Wasserstein space of stochastic processes in continuous time. arXiv:2501.14135, 2025.arxiv.org/abs/2501.14135
  14. [14]D. Bartl, S. Drapeau, J. Oblój, and J. Wiesel. Sensitivity analysis of Wasserstein distributionally robust optimization problems. Proceedings of the Royal Society A, 477(2256):20210176, 2021.DOI
  15. [15]D. Bartl and J. Wiesel. Sensitivity of multiperiod optimization problems with respect to the adapted Wasserstein distance. SIAM Journal on Financial Mathematics, 14(2):704–720, 2023.DOI
  16. [16]M. Beiglböck, S. Pflügl, and S. Schrott. A probabilistic view on the adapted Wasserstein distance. Stochastic Processes and their Applications, 201:105032, 2026.
  17. [17]A. Ben-Tal and E. Hochman. More bounds on the expectation of a convex function of a random variable. Journal of Applied Probability, 9(4):803–812, 1972.
  18. [18]T. R. Bielecki, M. Jeanblanc, and M. Rutkowski. Hedging of defaultable claims. In Paris–Princeton Lectures on Mathematical Finance 2003, volume 1847 of Lecture Notes in Mathematics, pages 1–132. Springer, Berlin, 2004.DOI
  19. [19]T. R. Bielecki and M. Rutkowski. Credit Risk: Modeling, Valuation and Hedging. Springer, Berlin, 2002.
  20. [20]J. Bion-Nadal and D. Talay. On a Wasserstein-type distance between solutions to stochastic differential equations. The Annals of Applied Probability, 29(3):1609–1639, 2019.
  21. [21]J. Blanchet and K. Murthy. Quantifying distributional model risk via optimal transport. Mathematics of Operations Research, 44(2):565–600, 2019.arxiv.org/abs/1604.01446
  22. [22]P. Brémaud. Point Processes and Queues: Martingale Dynamics. Springer, New York, 1981.
  23. [23]P. Brémaud and M. Yor. Changes of filtrations and of probability measures. Zeitschrift für Wahrscheinlichkeitstheorie und verwandte Gebiete, 45:269–295, 1978.
  24. [24]K. Burdzy and W. S. Kendall. Efficient Markovian couplings: examples and counterexamples. The Annals of Applied Probability, 10(2):362–409, 2000.
  25. [25]R. E. Burkard, B. Klinz, and R. Rudolf. Perspectives of Monge properties in optimization. Discrete Applied Mathematics, 70(2):95–161, 1996.DOI
  26. [26]S. Cambanis, G. Simons, and W. Stout. Inequalities for Ek(X,Y) when the marginals are fixed. Zeitschrift für Wahrscheinlichkeitstheorie und verwandte Gebiete, 36:285–294, 1976.DOI
  27. [27]S. Connor and S. Jacka. Optimal co-adapted coupling for the symmetric random walk on the hypercube. Journal of Applied Probability, 45(3):703–713, 2008.arxiv.org/abs/0801.1220
  28. [28]R. Cont. Model uncertainty and its impact on the pricing of derivative instruments. Mathematical Finance, 16(3):519–547, 2006.
  29. [29]G. Dall’Aglio. Sugli estremi dei momenti delle funzioni di ripartizione doppia. Annali della Scuola Normale Superiore di Pisa, Classe di Scienze, Serie 3, 10(1–2):35–74, 1956.
  30. [30]D. Duffie and D. Lando. Term structures of credit spreads with incomplete accounting information. Econometrica, 69(3):633–664, 2001.
  31. [31]D. Duffie and K. J. Singleton. Modeling term structures of defaultable bonds. The Review of Financial Studies, 12(4):687–720, 1999.DOI
  32. [32]S. Eckstein and G. Pammer. Computational methods for adapted optimal transport. The Annals of Applied Probability, 34(1A):675–713, 2024.arxiv.org/abs/2203.05005
  33. [33]N. El Karoui, M. Jeanblanc-Picqué, and S. E. Shreve. Robustness of the Black and Scholes formula. Mathematical Finance, 8(2):93–126, 1998.DOI
  34. [34]R. J. Elliott, M. Jeanblanc, and M. Yor. On models of default risk. Mathematical Finance, 10(2):179–195, 2000.DOI
  35. [35]H. Föllmer and M. Schweizer. Hedging of contingent claims under incomplete information. In M. H. A. Davis and R. J. Elliott, editors, Applied Stochastic Analysis, volume 5 of Stochastics Monographs, pages 389–414. Gordon and Breach, London, 1991.
  36. [36]R. Frey, L. Rösler, and D. Lu. Corporate security prices in structural credit risk models with incomplete information. Mathematical Finance, 29(1):84–116, 2019.
  37. [37]H. Frydman and T. Schuermann. Credit rating dynamics and Markov mixture models. Journal of Banking & Finance, 32(6):1062–1075, 2008.
  38. [38]M. Glanzer, G. C. Pflug, and A. Pichler. Incorporating statistical model error into the calculation of acceptability prices of contingent claims. Mathematical Programming, 174(1–2):499–524, 2019.
  39. [39]P. Glasserman and X. Xu. Robust risk measurement and model risk. Quantitative Finance, 14(1):29–58, 2014.DOI
  40. [40]M. Gunasingam, F. Mattesini, J. Wiesel, and T.-K. L. Wong. Adapted Wasserstein barycenters of Gaussian processes. arXiv:2604.22453, 2026.arxiv.org/abs/2604.22453
  41. [41]M. Gunasingam and T.-K. L. Wong. Adapted optimal transport between Gaussian processes in discrete time. Electronic Communications in Probability, 2025.arxiv.org/abs/2404.06625
  42. [42]M. Gunasingam and T.-K. L. Wong. Adapted optimal transport between filtered Gaussian processes. arXiv:2604.22159, 2026.arxiv.org/abs/2604.22159
  43. [43]J. Jacod and A. N. Shiryaev. Limit Theorems for Stochastic Processes, volume 288 of Grundlehren der mathematischen Wissenschaften. Springer, Berlin, 2nd edition, 2003.
  44. [44]Y. Jafry and T. Schuermann. Measurement, estimation and comparison of credit migration matrices. Journal of Banking & Finance, 28(11):2603–2639, 2004.DOI
  45. [45]R. A. Jarrow, D. Lando, and S. M. Turnbull. A Markov model for the term structure of credit risk spreads. The Review of Financial Studies, 10(2):481–523, 1997.
  46. [46]R. A. Jarrow and P. Protter. Structural versus reduced form models: a new information based perspective. Journal of Investment Management, 2(2):1–10, 2004.
  47. [47]M. Jeanblanc, M. Yor, and M. Chesney. Mathematical Methods for Financial Markets. Springer, London, 2009.
  48. [48]M. Kalkbrenner and N. Packham. A Markov approach to credit rating migration conditional on economic states. Canadian Journal of Statistics, 54:e70039, 2026.arxiv.org/abs/2403.14868
  49. [49]J. Keilson and A. Kester. Monotone matrices and monotone Markov processes. Stochastic Processes and their Applications, 5(3):231–241, 1977.DOI
  50. [50]D. Lando. On Cox processes and credit risky securities. Review of Derivatives Research, 2:99–120, 1998.DOI
  51. [51]D. Lando. Credit Risk Modeling: Theory and Applications. Princeton University Press, Princeton, 2004.
  52. [52]D. Lando and T. M. Skødeberg. Analyzing rating transitions and rating drift with continuous observations. Journal of Banking & Finance, 26(2–3):423–444, 2002.DOI
  53. [53]R. Lassalle. Causal transport plans and their Monge–Kantorovich problems. Stochastic Analysis and Applications, 36(3):452–484, 2018.DOI
  54. [54]T. Lindvall. Lectures on the Coupling Method. Wiley, New York, 1992.
  55. [55]G. G. Lorentz. An inequality for rearrangements. The American Mathematical Monthly, 60(3):176–179, 1953.
  56. [56]A. W. Marshall and I. Olkin. A multivariate exponential distribution. Journal of the American Statistical Association, 62(317):30–44, 1967.
  57. [57]A. Meucci. Fully flexible views: theory and practice. Risk, 21(10):97–102, 2008.
  58. [58]M. Michielon, A. Khedher, and P. Spreij. Proxying credit curves via Wasserstein distances. Annals of Operations Research, 336:1351–1367, 2024.
  59. [59]P. Mohajerin Esfahani and D. Kuhn. Data-driven distributionally robust optimization using the Wasserstein metric: performance guarantees and tractable reformulations. Mathematical Programming, 171:115–166, 2018.arxiv.org/abs/1505.05116
  60. [60]R. Norberg. The Markov chain market. ASTIN Bulletin, 33(2):265–287, 2003.DOI
  61. [61]K. O’Connor, K. McGoff, and A. B. Nobel. Optimal transport for stationary Markov chains via policy iteration. Journal of Machine Learning Research, 23(45):1–52, 2022.
  62. [62]D. O’Kane. Modelling Single-name and Multi-name Credit Derivatives. Wiley, Chichester, 2008.
  63. [63]G. C. Pflug. Version-independence and nested distributions in multistage stochastic optimization. SIAM Journal on Optimization, 20(3):1406–1420, 2009.DOI
  64. [64]G. C. Pflug and A. Pichler. A chance for multistage stochastic optimization models. SIAM Journal on Optimization, 22(1):1–23, 2012.DOI
  65. [65]G. C. Pflug and A. Pichler. Multistage Stochastic Optimization. Springer Series in Operations Research and Financial Engineering. Springer, Cham, 2014.DOI
  66. [66]K. Postek, A. Ben-Tal, D. den Hertog, and B. Melenberg. Robust optimization with ambiguous stochastic constraints under mean and dispersion information. Operations Research, 66(3):814–833, 2018.DOI
  67. [67]L. Rüschendorf. The Wasserstein distance and approximation theorems. Zeitschrift für Wahrscheinlichkeitstheorie und verwandte Gebiete, 70(1):117–129, 1985.DOI
  68. [68]P. J. Schönbucher. Credit Derivatives Pricing Models: Models, Pricing and Implementation. Wiley, Chichester, 2003.
  69. [69]N. Serino and N. W. Kraemer. Default, transition, and recovery: 2024 annual global corporate default and rating transition study. RatingsDirect report, S&P Global Ratings, Mar. 2025. Data from S&P Global Market Intelligence’s CreditPro.
  70. [70]J. Večeř. Geometric and arithmetic likelihood aggregation for diffusions with heterogeneous volatility. arXiv:2609.09470, 2026.arxiv.org/abs/2609.09470
  71. [71]A. M. Vershik. The theory of decreasing sequences of measurable partitions. St. Petersburg Mathematical Journal, 6(4):705–761, 1995. Translated from Algebra i Analiz 6 (1994).

Paper details

vlnc:W_m3qaxnh9
Version
v1 · latest
Published
Fresh — added to the Hub within the last 30 days
Added
License
CC BY 4.0
Usage

views · 5

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