Time-varying autoregressive distributed lag model with changing volatility for stress test

We present a novel time-varying autoregressive distributed lag (TV-ADL) model that allows for changes in both transmission mechanisms and innovation volatilities. The forecasting performance of the TV-ADL model has been substantially improved by removing the unrealistic traditional assumptions of constant volatility and constant inter-variable relationship. Our model is further adapted to stress tests mandated by the US Federal Reserve to generate conditional forecasts of the pre-provision net revenue of financial holding companies with large assets. The improvement of forecasting performance is demonstrated by the significant reduction of out-of-sample forecast errors at different horizons.