University of Washington - Department of Statistics
Linear pooling is the most popular method for forecast combination. However, if calibrated density forecasts are linearly aggregated, the combined density forecast is necessarily uncalibrated and overdispersed. The overdispersion of the linear pool can be addressed by spread adjustments to the density components, as implemented in the deflated linear pool (DLP), or via nonlinear recalibration transforms, such as the beta transformed linear pool (BLP). Both methods can be used effectively to combine calibrated as well as uncalibrated sources. The effects and techniques are demonstrated theoretically, in simulation examples and a case study on density forecasts for daily maximum temperature at Seattle-Tacoma Airport.