Forecasting in the presence of recent and recurring structural change

Structural change is a major source of forecast failure. Immediately after a break, forecasting problems are particularly severe due to a lack of information about the new data generation process. Techniques exist for monitoring for structural change in real time, but the optimal post-break strategy is unexplored. We consider two approaches. First, monitoring for change and then combining forecasts from models that do and do not use data before the change; second, using methods robust to structural change. Robust methods include rolling regressions, forecast averaging over different windows and exponentially weighted moving average (EWMA) forecasting. We derive analytical results for the performance of robust methods relative to a full-sample recursive benchmark. For a model with stochastic breaks there is a ranking where the MSFE of rolling regression < forecast averaging < full sample regression < EWMA. Expressions are also derived for models involving deterministic breaks which may be more appropriate for structural change in small samples. We assess the methods with Monte Carlo experiments. Forecasting based on break monitoring improves performance at low cost; rolling regressions are effective for some cases including large breaks but may reduce performance in other cases; the EWMA is a still riskier strategy; and forecast averaging is a good compromise, improving performance in many cases. We apply the tests to a large number of UK and US macroeconomic series.