Inflation forecasting with rolling windows: An appraisal

Abstract

We examine the performance of rolling windows procedures in forecasting inflation. We implement rolling windows augmented Dickey–Fuller (ADF) tests and then conduct a set of Monte Carlo experiments under stylized forms of structural breaks. We find that as long as the nature of inflation is either stationary or non‐stationary, popular varying‐length window techniques provide little advantage in forecasting over a conventional fixed‐length window approach. However, we also find that varying‐length window techniques tend to outperform the fixed‐length window method under conditions involving a change in the inflation process from stationary to non‐stationary, and vice versa. Finally, we investigate methods that can provide early warnings of structural breaks, a situation for which the available rolling windows procedures are not well suited.