Neural network (NN) models represent an opportunity to improve the credit loss forecasting and stress testing in Comprehensive Capital Analysis and Review (CCAR), which can be estimated as a function of macroeconomic variables using ARIMA-type models. For reference, please refer to: Capital Planning at Large Bank Holding Companies: Supervisory Expectations and Ranges of Current Best Practices, 2013. However, there remain challenges for the application of NN models such as model interpretability per regulatory requirements and a tendency to overfitting. This session will discuss this research and highlight the following:
- By leveraging a credit card firm’s monthly write-off data for over 15 years, a parsimonious NN model can be developed, which outperforms the traditional regression model with ARIMA errors in Mean Squares Error (MSE).
- Two macroeconomic variables with lags are selected from a pool of 500 by the combination of LASSO and stepwise regression algorithms, which enables the NN model to be interpretable for the CCAR scenario narratives. The sign or direction of estimated input weights should be consistent with or constrained by business intuition.
- This study also found that the NN model could be vulnerable to overfitting. The stress testing is sensitive to the design of network architect.
