Using structural break inference for forecasting time series

Rather than relying on a potentially poor point estimate of a coefficient break date when forecasting, this paper proposes averaging forecasts over sub-samples indicated by a confidence interval or set for the break date. Further, we examine whether explicit consideration of a possible variance break and the use of a two-step methodology improves forecast accuracy compared with using heteroskedasticity robust inference. Our Monte Carlo results and empirical application to US productivity growth show that averaging using the likelihood ratio-based confidence set typically performs well in comparison with other methods, while two-step inference is particularly useful when a variance break occurs concurrently with or after any coefficient break.