Integration of AI and machine learning to apply a broader spectrum of appropriate shocks to individual portfolios

By Assad Bouayoun, Senior XVA Quantitative Expert, former HSBC

What, for you, are the benefits of attending a conference like the ‘Stress Testing Europe Summit’ and what can attendees expect to learn from your session?

Stress testing is not only a regulatory requirement but also a different way of observing the risks taken by a financial institution. To benefit fully from the stress testing exercise, you must be able to interpret correctly the results and put in place the right mitigating mechanisms. By coming to this summit, you will be able to learn from industrial experts and share experiences around these topics and ultimately make your institution safer.

How can risk professionals best apply shocks relevant to market and credit risk factors?

We have now a wide range of techniques to generate shocks and to apply them to market and credit risk factors: historical analysis coupled with spectral decomposition, Bayesian nets, copula methods, recurrent neuronal network…

Thanks to the progress of computer science, modelling approaches which seemed unreasonable are becoming possible like agent-based models where we can simulate each transaction of each market. The challenge now is to find a procedure combining these shocks into realistic and adverse scenarios. A combination of big shocks will not necessarily provoke the biggest lose because of the non-linearity and sometimes the non-convexity of some derivative portfolios. Indeed, as the portfolio becomes more complex, embedding products with payoff singularities and partial hedges, the hyper plan representing its aggregated value in function of all risk factors transforms into a rugged landscape. The experience of practitioners remains fundamental, but they need systems that can help discover the landscape and orientate them toward its nadirs, the worst-case scenarios.

Can you provide our readers with an overview on performing analysis on individual portfolios?

When performing a stress on a portfolio we first compute the value and the sensitivities with respect to each risk factor. They can be synthetic (like a short interest rate) or observable (like the price of a swaption). From the sensitivities of first and second orders we can determine if the portfolio is convex. If it is, then the worst-case scenario is easy to find. If it is not the case, then we are dealing with an interesting optimization problem. Non convexity can come from hedged portfolio where the first and second order derivatives are often partially cancelled.  Collateral and initial margin can also accentuate non convexity and also create discontinuities. In any cases we need to make sure that the combination of shocks is producing a realistic scenario. The portfolio specific stress study is necessary but not sufficient. Indeed, results cannot be integrated with analysis done on other portfolios if the risk factors are not modelled consistently across different portfolios. Also, the worst-case scenarios can be different from one portfolio to another.

What are the challenges of offering diversity in scenarios available?

 The main challenge is to assess how realistic the scenario is. The subjectivity of the assessment can be tempered by several heuristics:

• We can measure how often this scenario has happened in the past and assign accordingly a probability.
• We can also check that logical and no arbitrage constraints are not violated.
• If the scenario is breaking some important relation, a qualitative reason must be supplied. For example, shocks in the opposite direction for two pegged currencies can be accepted if the scenario is based on an event provoking a break of the pegging.

What do you predict as the key opportunities and challenges of the next twelve months within AI and machine learning for Stress Testing?

 Machine learning can help at different stage of the stress test process:

• Reducing dimensionality by using grouping algorithm like K-means.
• Generating the combination of shocks: ML algorithms like recurrent neuronal network can help draw predictions of markets which can be reused to specify the amplitude and frequency of the shocks.
• Accelerating valuation and sensitivity computation by learning the valuation function. Finding the worst-case scenario often involves running complex valuation functions millions of times. Learning them and replacing them with the inference algorithm can speed up considerably the search for the worst-case scenario.

You may also be interested in…