Model forecasts contain unavoidable biases that result from model deficiencies. Methods exist to explicitly remove biases from short-term forecasts used in data assimilation systems, thereby avoiding biased analyses, but satisfactory bias correction on a grid has not been accomplished for the case of a true forecast. Here we extend some basic analysis concepts to remove biases from free-running forecasts of a model recently described by Lorenz (not yet published).

The approach involves two steps: estimating the bias at observation locations from the running-mean forecast error, and then spreading that correction to a grid via the statistical analysis equation. Covariances for the spreading reflect the uncertainty in the bias computation, and are computed from the same sample of forecasts that produce the bias estimate at the observation locations.

Statistical characteristics of the bias correction algorithm are explored with the new Lorenz model. It contains two scales, and is an improvement over earlier models because the statistics are temporally and spatially correlated. A biased version of the model is created by adding an additional forcing term with spatial variability. It is shown that under certain conditions, which are reasonable to expect in near-surface flows, the forecast bias can be effectively reduced or eliminated in this model. The forecasts thus show improved predictive skill.