Financial Stress Testing Powered by Machine Learning

Bud Mishra, Professor of Computer Science and Mathematics, devises novel method for stress testing and causality analysis

In the wake of the 2008 financial crisis, stress testing became a popular way to measure the strength of financial institutions and instruments. Stress testing analyzes or simulates rare, catastrophic scenarios and determines the response of specific portfolios or institutions. While many approaches to stress testing exist, they usually follow the same two fundamental steps—the generation of the stressed scenario, and then the portfolio/balance sheet projection after the scenario.

CDS’s Bud Mishra, with fellow researchers, has developed a novel stress testing method and algorithm using Suppes Bayes Causal Networks (SBCNs). Mishra’s method differentiates itself from existing methods by discovering the causality structure of financial events. This enables analysts to glean direct and ancillary causes of financial losses from the results of stress tests using SBCNs. Mishra’s SCBNs establish a causality structure by incorporating causality theory which stipulates that the cause of an event must occur before the event (also known as temporal priority).

This means that the network is arranged according to an intuitive time flow of events. Component branches of the network are also classified as “risky” or “profitable,” which produces decision trees to reduce computational complexity. Machine learning tools allow the system to identify which scenarios would be most detrimental to specific portfolios and further mitigate the computational complexity of these types of problems.

While Mishra had intended to apply this method to the study of cancer progression, he realized during development that, combined with machine learning, the method was an ideal approach to financial risk analysis.

Mishra’s SBCNs are trained on simulated historical data to combine three widely-used stress testing scenario generation methods (historical, hypothetical, and portfolio-specific). Mishra evaluated the performance of his method by embedding causal relationships in the simulated data and determining the capabilities of an SBCN to identify those causes. The model initially returned high numbers of false positives and negatives, but after “bootstrapping” (repeating many times) datasets the SBCNs far outperformed traditional Bayesian Networks.

Despite the advanced machine learning capabilities embedded in their method, however, the role of experts is still important. The researchers note that SBCNs can identify stressed scenarios that would cause adverse financial results, but experts are needed to determine which of those scenarios are actually plausible. Experts can also use SBCNs to test a particular stressed scenario.

Mishra and collaborators plan to test their algorithms on real data and compare the results with the research of human experts. Their new approach to stress testing offers financial institutions an efficient approach to fulfill federal and international regulations as well as their own internal evaluations.

By Paul Oliver