Detecting and Predicting Forecast Breakdowns

We propose a theoretical framework for detecting and predicting forecast breakdowns, defined as situations where the out-of-sample performance of a forecast method, judged by some loss function, is systematically worse than its in-sample performance. Our framework can be used to establish the robustness of a forecast method in the past, and to predict whether the forecast method will break down at a future date. We show that main causes of forecast breakdowns are overfitting and structural breaks in the data generating process that have some probability of occurring again in the future (”recurring breaks”). Finally, we find evidence of a forecast breakdown in the Phillips’ curve forecasts of U.S. inflation over the past three decades, and link the breakdown to changes in the monetary policy reaction function of the Fed and in inflation variability. J.E.L. Codes: C22, C52, C53 Acknowledgements: We would like to thank F. Diebold, G. Elliott, M. McCracken, N. Meddahi, and seminar participants to the 2005 Missouri Economics Conference, the Duke Financial Econometrics Group, the 2005 CIRANO-CIREQ Conference on Forecasting in Macroeconomics and Finance.