Statistical Post-Processing of GEFS Ensemble Forecasts for Precipitation Accumulations

We present a post-processing method that generates full predictive probability distributions for precipitation accumulations by fitting shifted, left-censored gamma distributions to statistics of the raw ensemble forecasts. This distribution type is shown to be adequate for modeling the distribution of observed precipitation accumulations given the ensemble forecasts both in situations with good predictability (e.g. at short lead times) and decreased predictability (e.g. at longer lead times or during summer months). When the forecasts are only available on a coarser grid than the verifying observations, our approach can perform downscaling in addition to calibration.

The proposed method will be demonstrated with GEFS precipitation reforecasts over the conterminous United States and verified against an 1/8-degree climatology-calibrated precipitation analyses using common metrics (skill, reliability, and so forth). We also discuss the effect of training sample size on the calibration of the post-processed predictions and show how an intelligent pooling of training data across different grid points can partially compensate for a reduction of the length of the reforecast data set used for model fitting.