Risk Factor Detection with Methods from Explainable ML

The importance of risk management in the financial industry has increased rapidly since the financial crisis, in particular with regard to financial market stability. A particular focus is on stress testing methods, which captures portfolio risk under adverse conditions. Advances in statistical learning and the availability of large, granular data sets offer new methodological possibilities for stress testing. Financial risk management applications such as hedging, scenario analysis and stress testing rely on portfolio models based on risk factors. In addition to observable risk factors, factor models with non-observable, data-based factors offer interesting alternatives. However, the lack of interpretability of the output is limiting. We develop time-dynamic methods for the interpretability of principal components (PCA), which allow to generate aggregated risk factors from existing risk factors. This aggregation makes it possible to plausibly implement less granular and even global stress scenarios.

Download presentation