Sampling Errors in Seasonal Forecasting

Abstract

The limited numbers of start dates and ensemble sizes in seasonal forecasts lead to sampling errors in predictions. Defining the magnitude of these sampling errors would be useful for end users as well as informing decisions on resource allocation to minimize total system error. A numerical experiment has been designed to measure them, and results indicate that sampling errors are substantial in state-of-the-art seasonal forecast systems. The standard solution of increasing sample sizes is of limited benefit in seasonal forecasting because of restrictions imposed by resource costs and nonstationary observations. Alternative options, based on the postprocessing of forecast and hindcast data, are presented in this paper. The spatial and temporal aggregations of data together with the appropriate use of theoretical distributions can reduce the effect of sampling errors on forecast quantities by an amount equivalent to increasing samples sizes by a factor of 4 of more, with insignificant losses of forecast information. These postprocessing techniques can be viewed as cost-effective methods of reducing the effects of sampling errors in seasonal forecast quantities.