Structural-break models under mis-specification: implications for forecasting

This paper shows that in the presence of model mis-specification, the conventional inference procedures for structural-break models are invalid. In doing so, we establish new distribution theory for structural break models under the relaxed assumption that our structural break model is the best linear approximation of the true but unknown data generating process. Our distribution theory involves cube-root asymptotics and it is used to shed light on forecasting practice. We show that the conventional forecasting methods do not necessarily produce the best forecasts in our setting. We also propose a new forecasting strategy, which incorporates our new distribution theory, and apply our forecasting method to numerous macroeconomic data. The performance of various contemporary forecasting methods is compared to ours.