Forecasting Non-Stationary Economic Time Series

In their second book on economic forecasting, Michael Clements and David Hendry askwhy some practices seem to work empirically despite a lack of formal support from theory. Afterreviewing the conventional approach to economic forecasting, they look at the implications forcausal modeling, present a taxonomy of forecast errors, and delineate the sources of forecastfailure. They show that forecast-period shifts in deterministic factors--interacting with modelmisspecification, collinearity, and inconsistent estimation--are the dominant source of systematicfailure. They then consider various approaches for avoiding systematic forecasting errors, includingintercept corrections, differencing, co-breaking, and modeling regime shifts; they emphasize thedistinction between equilibrium correction (based on cointegration) and error correction(automatically offsetting past errors). Their results on forecasting have wider implications for theconduct of empirical econometric research, model formulation, the testing of economic hypotheses,and model-based policy analyses.

Michael P. Clements is Research Fellow in Economics at the Universityof Warwick, UK.