Bias adjustment for decadal predictions of precipitation in Europe from CCLM

Abstract

A cross-validated model output statistics (MOS) approach is applied to precipitation data from the high-resolution regional climate model CCLM for Europe. The aim is to remove systematic errors of simulated precipitation in decadal climate predictions. We developed a two-step bias-adjustment approach. In step one, we estimate model errors based on a long-term ‘CCLM assimilation run’ (regionalizing data from a global assimilation run) and observational data. In step two, the resulting transfer function is applied to the complete set of decadal hindcast simulations (285 individual runs). In contrast to lead-time-dependent bias-adjustment approaches, this one is designed for variables with poor decadal prediction skill and without dominant lead-time-dependent bias. In terms of the CCLM assimilation run, MOS is shown to be effective in predictor selection, model skill improvement, and model bias reduction. Yet, the positive effect of MOS correction is accompanied with an underestimation of precipitation variability. After MOS application, an estimated mean square skill score of more than 0.5 is observed regionally. Simulated precipitation in decadal hindcasts is further improved when the MOS is trained on the basis of other decadal hindcasts from the same regional climate model but with a large underestimation in forecast uncertainty. Our results suggest that the MOS system derived from the assimilation run is less effective but allows the potential climate predictability in decadal hindcasts and forecasts to be retained. Using hindcasts itself for training is recommended unless a statistical method is capable of distinguishing biases and predictions within a hindcasts dataset.