NWC REU 2004
May 24 - July 31



Ensemble Forecast Bias Correction

Angeline Pendergrass and Kimberly Elmore



This study investigates two bias correction methods, lagged average and lagged linear regression, for individual members of ensemble forecasts. Both methods use the forecast bias from previous forecasts to predict the bias of the current forecast at every station. Also considered is the training period length that results in the smallest forecast error.


Ensemble forecast and verification data span 23 July through 15 September 2003. The data are organized into a mini-ensemble composed of 5 models and 30 days. This mini-ensemble is corrected using each method of correction for training period lengths between 3 and 12 days. The resulting bias, mean absolute error, RMS error, and interquartile range of the corrected forecasts are then compared.


Forecasts corrected with the lagged linear regression method are less biased but have more variance than those corrected with the lagged average method.

Full Paper [PDF]