Robust forecast methods and monitoring during structural change

We examine how to forecast after a recent break. We consider a new approach, monitoring for change and then combining forecasts from two models, one using the full sample and the other solely data from after the identified break point. We compare this to some robust techniques: rolling regressions, forecast averaging over all possible windows and exponentially weighted forecasts. We examine the efficacy of these methods with Monte Carlo experiments where there are single deterministic or multiple stochastic location shifts, and for a large number of UK and US macroeconomic series. No single method is uniformly superior. Monitoring brings only small improvements in forecast performance, so that robust methods are preferred. In some cases, forecast averaging is the best option, with only a small loss of forecast performance in the absence of breaks.