Principles of Forecasting

Volume 30 of the series International Series in Operations Research & Management Science pp 303-362

Econometric Forecasting

  • P. Geoffrey AllenAffiliated withDepartment of Resource Economics, University of Massachusetts
  • , Robert FildesAffiliated withDepartment of Management Science, University of Lancaster

* Final gross prices may vary according to local VAT.

Get Access


Several principles are useful for econometric forecasters: keep the model simple, use all the data you can get, and use theory (not the data) as a guide to selecting causal variables. Theory, however, gives little guidance on dynamics, that is, on which lagged values of the selected variables to use. Early econometric models failed in comparison with extrapolative methods because they paid too little attention to dynamic structure. In a fairly simple way, the vector autoregression (VAR) approach that first appeared in the 1980s resolved the problem by shifting emphasis towards dynamics and away from collecting many causal variables. The VAR approach also resolves the question of how to make long-term forecasts where the causal variables themselves must be forecast. When the analyst does not need to forecast causal variables or can use other sources, he or she can use a single equation with the same dynamic structure. Ordinary least squares is a perfectly adequate estimation method. Evidence supports estimating the initial equation in levels, whether the variables are stationary or not. We recommend a general-to-specific model-building strategy: start with a large number of lags in the initial estimation, although simplifying by reducing the number of lags pays off. Evidence on the value of further simplification is mixed. If there is no cointegration among variables, then error-correction models (ECMs) will do worse than equations in levels. But ECMs are only sometimes an improvement even when variables are cointegrated. Evidence is even less clear on whether or not to difference variables that are nonstationary on the basis of unit root tests. While some authors recommend applying a battery of misspecification tests, few econometricians use (or at least report using) more than the familiar Durbin-Watson test. Consequently, there is practically no evidence on whether model selection based on these tests will improve forecast performance. Limited evidence on the superiority of varying parameter models hints that tests for parameter constancy are likely to be the most important. Finally, econometric models do appear to be gaining over extrapolative or judgmental methods, even for short-term forecasts, though much more slowly than their proponents had hoped.


Econometric forecasting error correction model forecast comparisons specification testing vector autoregression