Post-Processing Improvements to an Ensemble Forecast Using an Archive of Past Forecasts and Verifications

Ensemble forecasts are the primary tool used operationally to assess forecast uncertainty. Studies of ensemble forecasts, however, have shown that forecast verifications too frequently lie outside of the ensemble's range of possibilities, meaning that uncorrected ensemble forecasts suggest more confidence than is justified. To make ensemble forecasts more representative of the actual range of possibilities, we present a technique to post-process ensemble forecasts by replacing member forecasts with verifications of what actually occurred when past forecasts were similar. To maximize the information that can be extracted from an archive of past forecasts and verifications, we allow analogs to come from different locations in space. We evaluated our procedure to post-process NCEP ensemble precipitation forecasts for the United States for 15-day periods in July 2005 and January 2006. Our analog correction technique significantly improved the ensemble's ability to forecast the probability of precipitation, in particular correcting the NCEP Global Ensemble's ``wet' bias at low precipitation amounts. Brier Skill Scores for 6-hour accumulated precipitation during the winter indicated that uncorrected ensemble forecasts were less skillful at predicting the probability of precipitation than forecasting zero precipitation as indicated by negative Brier Skill Scores (roughly -2.5). Post processed forecasts had Brier Skill Scores as high as 0.34. The tendency of the ensemble to underforecast heavy precipitation events, however, was not well corrected by our post-processing technique. Examinations of analog locations during heavy precipitation events indicated that analogs were taken from regions where precipitation patterns differed from those at the forecast point. This indicates that analogs must be chosen using more information than merely the similarity of ensemble precipitation forecasts to past forecasts.