An artificial neural network (ANN) is used to calibrate probabilistic quantitative precipitation forecasts (PQPF’s) from the ECMWF Ensemble Prediction System (EPS). Twenty-four hour precipitation accumulations to 10 days are calibrated for four years of EPS output. Rain gauge data for the contiguous U.S. are used for verification.

Post-processing of the EPS PQPF fields yields Brier scores that are significantly more skillful than the uncorrected forecasts, especially for low thresholds. Decomposition of the Brier score indicates that the improvement primarily comes from the mitigation of conditional biases. Calibration of just the EPS precipitation output does not significantly alter the ability to discriminate events after day 1.

The results indicate that ANN calibration can mitigate biases in ensemble PQPF fields after a short training period. However, major improvements in forecast specificity, especially for heavy precipitation amounts, may have to await the screening of dynamic-thermodynamic predictors from a larger training sample and more accurate model fields from improved versions of the EPS.