An interpretable Comprehensive Capital Analysis and Review (CCAR) neural network model for portfolio loss forecasting and stress testing

Abstract

This paper proposes an interpretable nonlinear neural network (NN) model that translates business regulatory requirements into model constraints. The model is then compared with linear and nonlinear NN models without the constraint for Comprehensive Capital Analysis and Review (CCAR) loss forecasting and scenario stress testing. Based on a monthly time series data set of credit card portfolio chargeoffs, the model outperforms the benchmark linear model in mean squared errors, and the improvement increases with network architecture complexity. However, the NN models could be vulnerable to overfitting, which could make the model uninterpretable. The constrained NN model ensures model interpretability at a small cost to model performance. Thus, it is insufficient to measure the model’s statistical performance without ensuring model interpretability and clear CCAR scenario narratives.