Although financial agencies and financial instruments vary, they are underpinned by the same risk management methodology: estimate the worst-case hypotheticals to hedge against financial upheavals.
Value at Risk (VaR), one quantitative risk management strategy that emerged as a solid method following the 1987 stock market crash, was heavily trusted prior to the 2008 financial crisis when the holes in it became apparent. With the enactment of the 2010 Dodd-Frank Act, stress testing, used previously in the 90’s as a self-assessment, emerged as a regulatory requirement.
A computer-generated simulation technique that tests firms and portfolios reactions to financial scenarios, stress testing, unlike VaR, explores highly unlikely financial scenarios. By testing financial portfolios in even the most implausible of events, institutions are covering all their bases.
Current stress testing approaches can be based on historical events, hypothetical events that experts deem damaging and plausible, or portfolios–but all of these approaches have weaknesses. While it is necessary to learn from historical events, the past is not always applicable in the present. Likewise, expert judgment of hypothetical events is not enough to determine the relationship between financial risk factors and the portfolios. The portfolio-based method utilizes the most widely used stress test, the Monte Carlo Simulation, but still lacks the capacity to deal with many risk factors.
Bud Mishra, an affiliated professor at CDS, Gelin Gao (NYU Courant), and Daniele Ramazzotti (Stanford) have therefore suggested a new method for stress testing financial portfolios in their newly released paper, Efficient Simulation of Financial Stress Testing Scenarios with Suppes-Bayes Causal Networks (SBCNs).
Their method combines SBCNs probabilistic causation with machine learning classification. SBCNs probabilistic causation confirms causal relationships if the following two conditions are satisfied:
- Temporal priority (any cause occurs before its effect)
- Probability raising (the presence of a cause raises the probability of observing its effect)
Machine learning classification is a type of artificial intelligence categorization of data that learns from the inputs and does not require explicit programming to do so. This is in line with their method as it identifies causal relationships between financial factors and portfolios. In turn, this allows for more accurate simulation of stress testing scenarios.
SBCNs alone, however, would not simulate the ideal stress scenarios relevant to stress testing because of their rarity and scarcity in the data. With machine learning classification, the researchers categorize scenarios as either ‘profitable’ or ‘lossy,’ cuing the program to focus on the lossy, risky, and uncommon scenarios for computation.
Once the framework for stress testing learns SBCNs from the data and classifies entries as either ‘profitable’ or ‘lossy,’ stress tests are simulated. Experts can then rule out implausible or flawed scenarios. They may even have a specific stress scenario in mind, which can also be executed.
Following the release of their paper, the group is now looking forward to presenting their work at the international Conference on Computational Science and being published in Elsevier Procedia Computer Science.
Additionally, with the introduction of their method complete, their focus has shifted to its application through an interface. This is a tool that will be powerful on Wall Street, at FinTech companies, and for artificial intelligence more broadly.
by Nayla Al-Mamlouk