In preparing medium-range forecasts, forecasters use ensemble mean and variance maps of postprocessd variables and also probabilistic products which are based on the count of ensemble members in which a variable exceeds a threshhold value. These products are affected by model biases, and the variance and probability products are also contaminated by differences between the ensemble variance (spread) and the variance of forecast errors.

We are experimenting with statistical correction of postprocessed fields to reduce both the bias of the ensemble mean and the under- or over-prediction of ensemble spread, using decaying averages of recent forecast error statistics. This provides unbiased estimates of ensemble mean fields, first-order calibration of probabilistic products, and corrections to spread-based estimates of forecast error. The variance correction also provides us with the statistical basis for evaluating and improving the process by which we introduce ensemble perturbations. We present verification graphics and statistics to evaluate the extent to which these corrections achieve their stated purposes.