Underdetermination and Variability of the Results in Macro-to-Micro Stress Tests – A Machine Learning Approach

We investigate the impact of the uncertainties surrounding the modelling process when conducting a stress test. These uncertainties are due to several choices left to the modeller with regards to, among others, the variables to select, the data samples used for the calibration of the different models and how these models are combined together. We run tests to quantify the impact of these sources of uncertainty by using as an example the FED CCAR 2016 scenario. We conclude that the impact could be non-negligible as it adds substantial variability to the final results. We leverage on Probabilistic Graphical Models - a machine learning technique – to corroborate our findings