Assessment of model drifts in seasonal forecasting: sensitivity to ensemble size and implications for bias correction

Abstract

Despite its systematic presence in state‐of‐the‐art seasonal forecasts, the model drift (leadtime‐dependent bias) has been seldom studied to date. To fill this gap, this work analyzes its spatiotemporal distribution, and its sensitivity to the ensemble size in temperature and precipitation forecasts. Our results indicate that model continues to drift well beyond the first month after initialization, leading to significant, highly space‐ and time‐varying drifts over vast regions of the world. Nevertheless, small ensembles (less than 10 members) are enough to robustly estimate the mean model drift and its year‐to‐year fluctuations in skillful regions. Differently, in regions of low model skill, larger ensembles are required to appropriately characterize this interannual variability, which is often larger than the drift itself. This points out a necessity to develop new strategies that allow for efficiently dealing with model drift, especially when bias correcting seasonal forecasts—most of the techniques used to this aim rely on the assumption of stationary model errors. We demonstrate here that the use of moving windows can help to remove not only the mean forecast bias but also the unwanted effects coming out from the drift, which can lead to important intraseasonal biases if it is not properly taken into account. The results from this work can help to identify the nature and causes of some of the systematic errors in current coupled models and can have large implications for a wide community of users who need long, continuous unbiased seasonal forecasts to run their impact models.