Loss-Given-Default Modeling by Post-Last Passage Time Process
Listen
Radio episode about this paper
Transcript
Introduction to the show: ident: Quantitative Finance. The day in quantitative finance: models of markets, risk and prices.
Ruth: We're looking at the paper: "Loss-Given-Default Modeling by Post-Last Passage Time Process".
Title and authors: Ruth: Moving into the title and authors for "Loss-Given-Default Modeling by Post-Last Passage Time Process." These are Masahiko Egami and Rusudan Kevkhishvili. It sounds like a deep dive into using advanced stochastic calculus to tackle a real problem in credit risk modeling.
Warren: It certainly does. When you hear diffusion processes and last passage time, you know we're looking at some heavy mathematics that’s trying to describe the firm value dynamics in a very specific way.
Iris: The authors are experts in this area, which is why they can delve into the details of transient linear diffusions and occupation times. They are taking a known mathematical concept and applying it to model a more complex financial phenomenon like LGD distribution.
Ruth: The implication here is that they’re suggesting that the way we model default isn't just a simple trigger, but involves how long the asset-to-debt ratio process spends in different states before it hits that point.
Warren: So it moves beyond just looking at the default probability and starts looking at the entire path taken by the firm value process leading up to that event.
Iris: Exactly. It’s about capturing what happens before default, using concepts like Doob’s h-transform and Martin boundary theory, even though last passage time isn't technically a stopping time.
Ruth: That's a subtle distinction they make, but it shows they are using these advanced probability literature tools to give their financial model more depth than just a basic Merton framework.
Warren: It’s about adding the path information to the risk assessment. I get that; knowing where the process was before default helps you understand the risk exposure better.
The paper's summary: Ruth: So, summarizing what they are actually doing in this paper, they propose a stochastic model for LGD distribution that uses last passage time of a linear diffusion to model the firm value. This is based on credit market and company-specific financial conditions.
Warren: It’s basically saying we can get the LGD distribution based on those two inputs, which is usually hard because you only have the market data which assumes a fixed sixty percent loss rate for all firms.
Iris: Right, and they are addressing that by building this model that uses the last passage time to bridge the gap between what we observe and what we want to predict about LGD distribution.
Ruth: They use this last passage time mechanism to create a hybrid model where you can use the firm-value approach before and then switch over to an intensity-based approach after default occurs.
Warren: So that way, they can use the information from the firm value process up to that point, and then switch gears based on what happens in that post default period.
Iris: And under minimal assumptions, they explicitly obtain the distributions of default time and LGD, which is a major theoretical win for this approach.
Ruth: That means they're not just giving us a general idea; they are giving us the actual math for both things directly from their framework.
Warren: So it’s providing a concrete tool rather than just another high-level concept. How does that translate to something useful in the real world?
Iris: It translates into being able to compute the loss given default distribution density, which is what risk managers really want, not just an expected value.
Ruth: Because they derive the density function of K sub D of xi using corollary four point two, which gives them a probability density function for the loss rate on debt.
The paper's improvements: Warren: So what are the specific improvements they suggest to this modeling approach that make it better than just sticking to an old model? Where do they actually gain something here?
Iris: One improvement is using last passage time and occupation time analysis gives insights into how much time the asset-to-debt ratio process spends above level alpha before it hits a killing boundary.
Ruth: That means you get this detailed view of the process's behavior before default, which is useful for understanding risk management. It shows how much time is spent in those high states.
Warren: So, that’s more than just the probability; it’s about the persistence in high-risk states.
Iris: And they also show that when you compare the last passage time and occupation time densities for exceptional companies like DAL and KSS highlights different aspects of credit risk management.
Ruth: That means they can see how volatility plays a role in distinguishing between these two types of risk management even when the underlying asset dynamics look similar.
Warren: So, that’s about finding a way to separate the idiosyncratic factors that matter for specific firms.
Iris: And they also show that the density function of K sub D of xi is given by four point thirteen based on estimated parameters, which is a key result because it's explicit and computationally simple.
Ruth: That simplicity makes the model practical for risk management applications. It’s not just some abstract theory; it’s something you can actually run and get results from.
Conclusion: Ruth: So, wrapping up this paper "Loss-Given-Default Modeling by Post-Last Passage Time Process." The main implication is that they provide a model that estimates the LGD distribution based on both market data and company specifics.
Warren: They manage to do it with a hybrid approach that combines firm value before the last passage time and an intensity-based approach after.
Iris: They show consistency by comparing spreads per one percent loss given default against market CDS spreads, which is a huge piece of evidence for their methodology.
Ruth: The paper shows how the model captures company-specific financial conditions on LGD, which is the main contribution here.
Warren: It’s a method to estimate whole distribution density of LGD, not merely an expected value. That capability is appreciated from banks’ and investors’ credit risk management point of view.
Iris: The explicit formula for the distribution of default time in proposition four point one is computationally simple, which makes the model practical for risk management applications.
Ruth: This paper gives us a method to estimate whole distribution density of LGD, not merely an expected value, and it’s based on last passage time modeling.
Warren: We're getting ready for the next paper on some different topic that might be interesting next.
Graduate School of Economics, Kyoto University · Faculty of Economics and Business, Hokkaido University
q-fin.RM
Submitted: 2020-09-02
Updated: 2026-11-17
Project page: https://www.nasdaq.com/market-activity/stocks/screener
License: http://arxiv.org/licenses/nonexclusive-distrib/1.0/
Importance score: 85/100
The gist: This study proposes a stochastic model for loss-given-default (LGD) which provides the LGD distribution based on credit market and company-specific financial conditions The model utilizes last…
Key concepts
- Linear Diffusion (Firm Value)
- This models the firm's value as a process that moves randomly over time, similar to how stock prices fluctuate. It is used to represent the asset side of the credit risk calculation, showing how firm value changes continuously.
- Last Passage Time
- This concept identifies the specific point in time when a financial process (like firm value) first reaches a certain threshold level. In this model, reaching this threshold signals that default is about to occur as a sudden event.
- LGD Distribution Density
- Instead of just predicting an average loss rate, the model calculates the full probability distribution density of loss given default (K_D(xi)). This allows risk managers to understand the range and likelihood of different loss outcomes, offering a more nuanced view than fixed recovery assumptions.
Terminology
Summary
This study proposes a stochastic model for loss-given-default (LGD) which provides the LGD distribution based on credit market and company-specific financial conditions The model utilizes last passage time of a linear diffusion (representing firm value) to a certain threshold point, after which default occurs as a surprising event By treating the post-last passage time process in a continuum of the original process, we are able to use firm-value approach before and intensity-based approach after the last passage time, leading to a hybrid model Under minimal and standard assumptions, we obtain the distributions of default time and LGD explicitly We provide a computationally simple estimation procedure and real-world examples of estimated LGD distribution implied in CDS market. p. 0
Model Formulation and Methodology
The core idea is to derive the distribution of the ratio of the firm’s total assets over its total debt upon default, denoted by Y, which is modeled as a linear diffusion The process starts with a geometric Brownian motion for firm value V and debt B, leading to the asset-to-debt ratio process Yt = V/B The model focuses on the corporate-level loss distribution, allowing credit managers to calculate loss rates for individual debt obligations The mathematical setting involves a linear regular canonical diffusion X with state space I = (l,r) and an assumed transience condition where s(l) > −∞ and s(r) = +∞.
Key Mathematical Results
The paper presents two main mathematical results concerning last passage time for general transient linear diffusions Proposition 3. p. 21 provides the Laplace transform of the occupation time above a certain level, which is related to the distribution of default time. p. 7 Proposition 3.2 gives the transition density for the post-last passage time process, which is given by formula (3. p. 84). These results allow for a consistent analysis of pre- and post-last passage time states within the framework. p. 6
LGD Distribution Derivation
The LGD distribution is obtained by focusing on the post-last passage time process Z, which follows dynamics (4.1) The default time ξ is defined using a random time τ modeled via an exponential random variable J with rate 1. p. 7 The LGD on debt B is defined as K B(ξ) = 1−Yξ = 1−Vξ/Bξ. p. 2 The final distribution of the LGD is derived in Corollary 4. p. 32, which depends on the distribution of Zτ. p. 0
Empirical Validation and Consistency Check
The empirical study applies the model to eight NYSE companies to estimate parameters such as µ, σ, M, and implied current Y0. The calibration of the threshold level α is performed so that the model-implied default probability matches the market counterpart A crucial check involves comparing the model-implied CDS spread with the market quotes by examining spreads per 1% loss given default
ρ. The analysis shows consistency, with a percentage difference of less than 7. p. 142% for six out of eight companies
Conclusion and Implications
The study concludes that the model is consistent with the market, noting that for companies whose loss rate distribution is concentrated in a region far from 60%, pricing based on the distribution of K D(ξ) will provide significantly different results The analysis of last passage time and occupancy time provides insights into how much time the asset-to-debt ratio process spends above level α before it hits a killing boundary. The comparison between the last passage time and occupation time densities for exceptional companies like 'DAL' and 'KSS' highlights different aspects of credit risk management. The model successfully estimates the LGD distribution implied in the current credit market. p. 2
Pruning redundant reasoning tokens
The model is discussed in Section 2 and an empirical study is performed in Section 5 The essential rationale of using last passage time is as follows While the firm-value approach for credit risk analysis proposed in Merton (1974) is a fundamental method (as is the intensitybased approach) in estimating the default probability, one has some difficulties in calculating the LGD under this approach The post-last passage time process well captures this reality and is suitable for estimating LGD distribution
How the training signal is built
The model check proposed in Section 4.3 inspects whether the model-implied default time and LGD distribution are consistent with the credit market and company-specific financial conditions Specifically, we check that the CDS spread implied from the estimated LGD distribution is consistent with the quoted spread in the market The procedure for estimating parameters involves using maximum likelihood estimation of Duan (1994), Duan (2000), and Lehar (2005)
Where the gains come from
The explicit formula for the distribution of default time in Proposition 4.1 is computationally simple The resulting density function of K D(ξ) is given by Corollary 4. p. 42 This distribution allows for a more nuanced understanding of loss rates compared to the standard constant recovery rate assumption The comparison between Model ρ and Market ρ demonstrates that the model captures the impact of company-specific financial conditions on LGD
The last passage time is an important subject in the probability literature due to being closely related to the concepts of transience/recurrence, Doob’s h-transform, time-reversed processes, and the Martin boundary theory While the last passage time is not a stopping time, it has been utilized widely in financial applications as discussed in Nikeghbali and Platen (2013) The transition density (with respect to the speed measure m(dy)) of the post-last passage time process is given by formula (3.4)
The analysis of Figure 2 and Figure 3 accurately capture high credit risk The density of both λα∗ and λ A,c α∗ in the case of ‘DAL’ is concentrated around 6 years, around the same levels of the other (six) companies while ‘KSS’ has concentration at much lower level This suggests that K D(ξ) and λα∗ (or λ A,c α∗) provide different aspects of credit risk management
The final observation can be confirmed by the densities with c = 0 (in blue) and c = 2/3 α (in orange): they are almost identical This means that the process 1σln(Y) spends most of the time above level α∗ up until the last passage time λα∗ The higher initial values Y0 as compared to α, especially for the companies with low CDS spreads, contribute to this observation
The model check is performed by comparing “spreads per 1% loss given default” because the market CDS spreads are quoted under the assumption that the LGD is fixed at 60% for all firms The two CDS spreads cannot be compared directly The deviation of ‘DAL’ and ‘KSS’ is understandable due to high volatility σ = 0.4433
The estimated CDS spread (5Y) based on the model-implied LGD distribution is compared to the market spread The final procedure involves matching the premium leg to the default leg to obtain the model-implied CDS spread
The estimated CDS spread (5Y) based on our model is close enough to the market counterpart, with a percentage difference of less than 7.42% for six out of eight companies This indicates that the model is consistent with the market The closer the distribution of K D(ξ) is to 60% loss rate, the estimated spread per 1% loss given default (which is the definition of ρ) is sufficiently close to the market counterpart
The density function of Zτ is obtained by differentiating (4.11) with respect to Q <alpha∗ The density of the transform 1−eσZτ is given by Pα∗ (K B(ξ) ∈ dx) = f(ln(1−x)/σ)/(sigma(1-x))dx, 1-eσα∗<x<1 This density function is derived from the properties of the process Zτ
The final result for K B(ξ) is given by (4.5) The expected value of K D(ξ) is calculated as Eα∗ [K D(ξ)] = 1−eσα∗/1+σ 2/2 +b2σM This provides a quantitative measure of the expected loss rate based on the model
The model's ability to estimate LGD distribution based on market information and company-specific financial conditions is its primary contribution The hybrid approach, combining firm-value and intensity-based approaches, provides a robust framework for modeling LGD The explicit formula for the distribution of default time in Proposition 4.1 is computationally simple This simplicity makes the model practical for risk management applications
The process Zt follows the dynamics (4.1) right below; that is, the post-la dynamics of the normalized asset-to-debt ratio process The starting point of Z is α∗:= 1/σln(α), t > 0 This process Zt follows the dynamics (4. p. 2 p. 21) right below; that is, the post-la dynamics of the normalized asset-to-debt ratio process
The estimation of parameters involves using data up to a current time point and calibrating α so that default time ξ in (4.4) matches the market-quoted 5-year default probability The estimated LGD distribution is then obtained from Corollary 4. p. 42 This procedure ensures that the model-implied default probability is consistent with the default probability provided by LSEG Workspace
The model provides a method to estimate whole distribution density of LGD, not merely an expected value This capability is appreciated from banks’ and investors’ credit risk management point of view The ability to compute the distribution of default time ξ explicitly is a significant advantage over conventional methods
The model check inspects whether the model-implied default time and LGD distribution are consistent with the credit market and company-specific financial conditions The final procedure involves matching the premium leg to the default leg to obtain the model-implied CDS spread
The parameter estimation uses data of firm’s equity E and debt B available up to the current time, utilizing an option-theoretic approach where equity represents a European call option written on assets with strike price equal to future debt value The log-likelihood function is maximized by differentiating with respect to µ and setting the partial derivative to zero
The model allows for time-varying realized loss rates, which is documented in the empirical literature This necessitates a stochastic model for LGD The hybrid approach, using firm-value before last passage time and intensity-based after, leads to a computationally simple estimation procedure The ability to compute the distribution of LGD distribution implied in the CDS market is its main contribution
The density function of K D(ξ) is given by (4.13) based on estimated parameters The final procedure involves calculating the present value of recovery and spread payments to obtain the model-implied CDS spread
The model provides a method to estimate whole distribution density of LGD, not merely an expected value This capability is appreciated from banks’ and investors’ credit risk management point of view The explicit formula for the distribution of default time in Proposition 4.1 is computationally simple This simplicity makes the model practical for risk management applications
The model allows for time-varying realized loss rates, which is documented in the empirical literature This necessitates a stochastic model for LGD The hybrid approach, using firm-value before last passage time and intensity-based after, leads to a computationally simple estimation procedure
The process Zt follows the dynamics (4.1) right below; that is, the post-la dynamics of the normalized asset-to-debt ratio process The starting point of Z is α∗:= 1/σln(α), t > 0 This process Zt follows the dynamics (4.
Green — the paper says it, and the page is printed. Red — no passage carries it, so treat that sentence as unverified. The same check runs on your own text: check a summary against its source, $5.
Improvements for AI systems
-
Improved LGD distribution modeling: The model uses
last passage time of a linear diffusion (representing firm value) to a certain threshold point, after which default occurs as a surprising event,
leading to a hybrid model that usesfirm-value approach before and intensity-based approach after the last passage time.
-
Enhanced credit risk assessment: The system can estimate the
LGD distribution density
rather than just an expected value, reflectingtime-varying realized loss rates
by incorporating factors like economic cycles and industry distress into recovery rate modeling. -
Market consistency checking: The model includes a procedure to check consistency by comparing
spreads per 1% loss given default
against market CDS spreads quoted under the standard assumption of a constant 60% loss rate, allowing for comparison between model-implied and market risk metrics. -
Empirical parameter estimation: The system can estimate parameters for the asset process using a
maximum likelihood method of Duan (1994), Duan (2000), and Lehar (2005)
to obtain estimates of drift M and diffusion sigma, ensuring model-implied default probability is consistent with market data. -
Distribution prediction: The system can explicitly derive the distribution of default time xi using the formula in Proposition 4.1, which involves a complex expression incorporating functions like Erf and Erfc, allowing for explicit calculation of default time probabilities.
-
Loss rate sensitivity analysis: The system can calculate the
distribution of K D(xi)
by Corollary 4.2, providing a probability density function for the loss rate on debt, enabling risk managers to see how much the estimated spread per 1% loss given default (rho) deviates from market counterparts.
Related papers
- E-backtesting
- Diversified-Risk-Perception Risk Measures and Portfolio Selection
- Robust distortion riskmetrics over Bregman--Wasserstein balls and their intersections
- Distributions or Quantiles? A Convex Variational Theory of Worst-Case Risk
- Learned Monotone Recurrent Features in Governed Credit Scoring: The Price of the Frame and the Necessity of Macro Conditioning
- Correlation Ambiguity, Diversification Fragility, and Insurance Pricing