Automated forecasting systems often can create several forecasts of the same weather parameter (e.g. by generating linear regression equations to predict maximum temperature based on the output of several different numerical weather prediction models). The question therefore arises which forecast should be used? Could they be combined to generate a prediction superior to the separate inputs?

Traditional statistical methods, such as multiple linear regression, can be used to "weight" the inputs. Unfortunately, the colinearity inherent in such forecasts often leads to instability and poorly conditioned design matrices. Another alternative is to use a mixture of experts to determine the best combination. This technique divides the input space into subregions, relying more strongly on the input that is best in each subregion. Once a framework has been established for the mixture, the Expectation Maximization (EM) algorithm can be used to estimate the mixing parameters from a training set.

This method is applied and tested for the problem of predicting typical weather parameters (e.g. temperature) and the results are compared to those obtained via multiple regression. The input forecasts are generated using dynamic model output statistics approaches applied to the output of different NWP models as well as climatology and persistence.