A salient benefit of an ensemble prediction system (EPS) is its ability to provide a means of estimating forecast error. This study tests the error prediction skill of three EPS features: spread (standard deviation) among ensemble members, consistency between MEX/MOS output and the ensemble mean, and consistency between consecutive 24-hour runs of an EPS. For 27 stations throughout the northeastern United States, 15-day high and low temperature forecasts from calibrated ECMWF ensemble output issued Feb 1 through May 31 2007 are examined. For each of the 15 forecast days, the significance (from r2 statistic) and slope of the error predictor-forecast error relationship is used to determine how valuable each feature is in estimating forecast error.
For high temperature forecasts in the northeast, ensemble spread and run-to-run consistency are most effective at predicting forecast error through 9-day lead times. Although error prediction skill in both drops off in the longer ranges, ensemble spread is more useful beyond 9-day lead times. Spread and run-to-run consistency are more effective for low temperature forecast error prediction, with ensemble spread still performing best beyond 9-day lead times. Model-to-model consistency is only a moderate to weak error predictor for high temperature forecasts in the short range. For the sake of comparison of error predictability between two regions with disparate climates, data from 21 stations in the southwestern United States are also examined. In this region, all three error predictors are consistently effective in anticipating forecast uncertainty for both high and low temperatures.