A MIXTURE-MODEL APPROACH TO COMBINING FORECASTS

A multiprocess mixture-model approach to combining forecasts from alternative sources is proposed. This approach extends the Granger-Ramanathan method by allowing the weights used in producing the combination forecast to vary over time. In addition, the procedure discounts outlying data points that arise during time periods when all of the competing forecasts miss the mark. An empirical comparison with traditional and more recently proposed combination methods demonstrates that the proposed methodology outperforms these.