The Use of Machine Learning for Bank Financial Stress Tests : Prepared for the 3rd Annual Marcus Evans Conference - Best Practices for Stress Testing in Financial Institutions, New York USA

We develop empirical models for the determination of liquidity for U.S. commercial and savings banks (hereafter, banks) with the goal predicting bank liquidity under stressful economic conditions. The data used to develop the liquidity models are taken from income statements and balance sheets banks report to their primary federal regulator. The explanatory variables for our models are measures of banks’ internal drivers of liquidity—capital adequacy, asset quality, management, earnings and funding sources—and external drivers of liquidity—macroeconomic conditions.We model bank liquidity using linear regression and machine learning models for prediction of continuous variables (liquidity measures)—artificial neural networks, gradient boosting, random forest and support vector machines. We use these models to predict bank liquidity by applying estimated models to simulated values of the drivers of liquidity under stressful economic conditions. Following industry practices and regulatory guidance we document each step of model development and provide rational and support for modelling choices. The results of model estimations indicate that machine learning models produced more accurate out-of-sample forecasts of bank liquidity than do linear regression models. Importantly, the performance of machine learning models varies with how one treats correlated variables and data outliers, i.e., data normalization and standardization