Numerical models for weather prediction on a seasonal scale require postprocessing based on previous performance, in order to remove bias and produce realistic uncertainty estimates. For being useful for spatial forecasts, postprocessing techniques ought to account for spatial correlations of the prediction error.
Particularly challenging in the context of seasonal predictions is the shortage of training data, since the number of observed seasons available is often well below 50.
To tackle these issues we propose a multivariate postprocessing approach utilizing covariance tapering, combined with a dimension reduction step based on principal component analysis.
The proposed technique is robust and can be applied on a global scale, even with little training data available. Moreover, it can correctly and efficiently handle non-stationary, non-isotropic and negatively correlated spatial error patterns. Further, a moving average approach to marginal postprocessing is shown to flexibly handle trends in biases caused by global warming.
In an application to global sea surface temperature forecasts issued by the Norwegian Climate Prediction Model (NorCPM), our proposed methodology outperforms known reference methods.