The Stability of Empirical Long-Range Forecast Techniques: A Case Study

Abstract

The stability of simple linear regression equations for the long-range prediction of Australian spring rainfall was studied. Specifically, the way in which the accuracy of the forecasts depends on the number of years of data used to derive the equations and the length of the period between the end of the data used in forecast equation derivation and the application of the equations in prediction were examined. An optimum period of data of about 15 years was found; the use of more or less data in deriving the forecast equations led to deterioration in the forecasts. The forecasts also deteriorated if the equations were used for more than a few years after the end of the period of data from which they were derived, suggesting a need for routine updating of the forecast equations. The lack of stability in the forecast equations presumably reflects nonstationarity in the data series possibly resulting from changes in the general circulation patterns. If this is so, the results might be applicable to similar statistical long-range forecast methods in other areas.