We propose two new methods for bias correction in numerical weather prediction, one global and one local. The global method is an elaboration of MOS, combining several modern methods for multiple regression: ACE, CART, and Bayesian model selection. This allows us to represent nonlinear aspects of the bias, and select the important ones in an overall nonlinear but parsimonious model. The local method is the method of neighbors, which estimates the bias as the average bias over the ``neighbors'' of the grid point and time point being forecast, consisting of recent observations at stations that are close geographically and have similar elevation and land use. The methods were applied to 48 hours ahead MM5 forecasts of surface temperature in the US Pacific Northwest. The model parameters were estimated using 2001 data, and the methods were verified using 2002 data. Both methods gave good performance. Combining the two methods using linear regression gave substantially better results than either method on its own, with mean squared forecast error reduced by 44%.