A near optimal test for structural breaks when forecasting under square error loss

We propose a near optimal test for structural breaks of unknown timing when the purpose of the analysis is to obtain accurate forecasts under square error loss. A bias-variance trade-off exists under square forecast error loss, which implies that small structural breaks should be ignored. We study critical break sizes, assess the relevance of the break location, and provide a test to determine whether modeling a break will improve forecast accuracy. Asymptotic critical values and near optimality properties are established allowing for a break under the null, where the critical break size varies with the break location. The results are extended to a class of shrinkage forecasts with our test statistic as shrinkage constant. Empirical results on a large number of macroeconomic time series show that structural breaks that are relevant for forecasting occur much less frequently than indicated by existing tests.