[Paper Review] Solving Stochastic Differential Equations with Jump-Diffusion Efficiently: Applications to FPT Problems in Credit Risk

Di Zhang, Roderick Melnik|ArXiv.org|Feb 28, 2007

Credit Risk and Financial Regulations14 references3 citations

TL;DR

This paper proposes an efficient Monte Carlo method for solving first passage time (FPT) problems in multivariate jump-diffusion processes, crucial for credit risk modeling. By combining fast one-dimensional FPT simulation with correlated multivariate sampling via the multivariate uniform (UNIF) method, the approach achieves significant computational speedups—up to 190x faster than conventional Monte Carlo—while maintaining accuracy in default probability and correlation estimates.

ABSTRACT

The first passage time (FPT) problem is ubiquitous in many applications. In finance, we often have to deal with stochastic processes with jump-diffusion, so that the FTP problem is reducible to a stochastic differential equation with jump-diffusion. While the application of the conventional Monte-Carlo procedure is possible for the solution of the resulting model, it becomes computationally inefficient which severely restricts its applicability in many practically interesting cases. In this contribution, we focus on the development of efficient Monte-Carlo-based computational procedures for solving the FPT problem under the multivariate (and correlated) jump-diffusion processes. We also discuss the implementation of the developed Monte-Carlo-based technique for multivariate jump-diffusion processes driving by several compound Poisson shocks. Finally, we demonstrate the application of the developed methodologies for analyzing the default rates and default correlations of differently rated firms via historical data.

Motivation & Objective

• Address the computational inefficiency of conventional Monte Carlo methods when simulating first passage times (FPT) in multivariate jump-diffusion processes.
• Develop a fast, scalable Monte Carlo technique for FPT problems under correlated, multivariate jump-diffusion processes driven by multiple compound Poisson shocks.
• Enable accurate and efficient estimation of default probabilities and default correlations for credit risk analysis using historical data.
• Validate the proposed method against conventional Monte Carlo in terms of accuracy and computational performance.
• Provide a practical computational framework applicable to option pricing and credit risk modeling under realistic market dynamics with jumps.

Proposed method

• Adapt a fast one-dimensional FPT Monte Carlo method for jump-diffusion processes to the multivariate case using correlated Brownian motion and Lévy processes.
• Implement the multivariate uniform (UNIF) sampling method to generate correlated first passage times across multiple assets, ensuring joint distribution consistency.
• Model asset values using a log-price process with drift, correlated Brownian motion, and compound Poisson jumps with normally distributed jump sizes.
• Use kernel density estimation to non-parametrically approximate the first passage time density function from simulated paths.
• Apply the method to simulate default events and correlations by generating inter-jump times from an exponential distribution and jump sizes from a normal distribution.
• Validate results by comparing default correlation estimates and CPU times against conventional Monte Carlo simulations with identical parameters.

Experimental results

Research questions

• RQ1 How can first passage time problems in multivariate jump-diffusion processes be solved more efficiently than with conventional Monte Carlo methods?
• RQ2 What is the impact of correlated jumps and diffusion components on default correlation estimates in credit risk models?
• RQ3 Can the proposed method achieve significant computational speedups while preserving accuracy in default probability and correlation estimation?
• RQ4 How do different credit ratings (e.g., A-rated vs. Ba-rated firms) affect default rate trends and volatility under jump-diffusion dynamics?
• RQ5 To what extent does the multivariate UNIF method reproduce the same default correlation results as conventional Monte Carlo, given the same input parameters?

Key findings

• The multivariate UNIF method reduced CPU time from 0.119668 seconds (conventional Monte Carlo) to 0.000621 seconds for A-rated firms, achieving a 193x speedup.
• For Ba-rated firms, the UNIF method reduced CPU time from 0.119675 to 0.000622 seconds, a speedup of approximately 192x.
• The A-rated firm’s default density function showed an increasing trend, suggesting a rising default rate over time, while the Ba-rated firm’s density decreased, indicating a slowing or stable default rate.
• The default correlation between two firms converged to a stable value over long horizons, consistent with theoretical expectations.
• The proposed method produced default correlation estimates nearly identical to those from conventional Monte Carlo, confirming its validity and accuracy.
• The model revealed that Ba-rated firms exhibit higher jump volatility (σ_Z) and larger jump size means (μ_Z), indicating greater sensitivity to sudden economic shocks compared to A-rated firms.