Comparing predictive ability in presence of instability over a very short time

Abstract

We consider forecast comparison in the presence of instability when this affects only a short period of time. We demonstrate that global tests do not perform well in this case, as they were not designed to capture very short-lived instabilities, and their power vanishes altogether when the magnitude of the shock is very large. We then discuss and propose approaches that are more suitable to detect such situations, such as nonparametric methods (S test or MAX procedure). We illustrate these results in different Monte Carlo exercises and in evaluating the nowcast of the quarterly US nominal GDP from the Survey of Professional Forecasters (SPF) against a naive benchmark of no growth, over the period that includes the GDP instability brought by the Covid-19 crisis. We recommend that the forecaster should not pool the sample, but exclude the short periods of high local instability from the evaluation exercise.