Top Down Stress Testing: An Application of Adaptive Lasso to Forecasting Credit Loss Rates

Abstract

The aim of this thesis is to determine the data requirements and feasibility of data-driven top-down stress testing for credit loss rates. To that end, we use the Adaptive Lasso method to simultaneously select and estimate parsimonious linear models from a very large set of potential model specifications. Adaptive Lasso is a penalized regression method which can accurately and uniquely select substantially relevant predictors and has attractive asymptotic properties. The selected models are able to give accurate forecasts in baseline and severely adverse macro-economic scenarios for the United States. We find that the loan data needs to be divided into a minimum of five categories to adequately capture the link between the macro-economy and credit loss rates. For reliable forecasts, roughly 20 years of credit loss data is required, or at least one complete business cycle must be present in the data.