This paper illustrates the improvement in skill obtained in seasonal rainfall outlooks for the Australian region from the optimal combination of partially independent statistical prediction schemes. The linear combination technique was applied to two Southern Oscillation Index (SOI) based seasonal rainfall forecast techniques currently used in the National Climate Centre; a linear discriminant analysis (LDA) technique and a SOI phase analogue technique. Optimal weighting of the individual probability predictions of each prediction scheme is achieved using a least-squares variational reduction of the forecast errors with respect to the weighting factors. The combination of the two prediction schemes shows a significant improvement in forecast skill over a twenty-year evaluation period from 1972 to 1991. Mean square errors (half-Brier scores) of 0.213 for the SOI phase technique and 0.221 for the LDA model were reduced to a half-Brier score of 0.197 for the combination over the evaluation period.
Trevor M. Casey (Fri,) studied this question.