Combining Predictive Schemes in Short-Term Forecasting

Abstract

In this article, the theory is presented for a linear combination of two independent predictive techniques (either probabilistic or binary). It is shown that substantial gains might be expected for optimal weighting of the combination. The theory is general but also is applied to several special cases which may be useful for both short-term weather prediction and long-range forecasting. Using data from a recent operational evaluation of techniques for the short-term predicting of rainfall, a linear combination of two independent predictive techniques gives, in practice, improvement in skill compared with the techniques used individually. In the present case, a Markov chain and a numerical weather prediction (NWP) model were combined. The half-Brier wore of the linear combination was 0.142 compared with individual scores of 0.164 for the Markov chain model and 0.258 for the NWP model. The combined Markov-NWP scheme may provide a possible simple alternative to the MOS approach for predictions up to 12 hours ahead.