Improvement of the Multimodel Superensemble Technique for Seasonal Forecasts

Abstract

Abstract The superensemble technique has previously been demonstrated to provide an improved seasonal forecast compared to the bias-removed ensemble of equally weighted models. This paper offers a further improvement to the superensemble method by modifying the regression coefficients used in the weighting of the models for the construction of the superensemble. The improvement is achieved by use of singular value decomposition of the covariance matrix, and selecting only the largest singular value, corresponding to maximal explained variance, for the calculation of the regression coefficients. The results shown here are based on calculations done with 10 yr worth of monthly forecasts from the Atmospheric Model Intercomparison Project (AMIP) dataset, using cross validation.