Improving forecasting performance by combining forecasts: The example of road-surface temperature forecasts

A recent paper by J. E. Thornes reported the results of a comparative evaluation of location-specific, road-surface temperature forecasts (RSTFs) formulated by two organisations, the UK Meteorological Office (MO) and Oceanroutes (OR). The present paper broadens the scope of this competition, by considering the possibility of combining the MO and OR forecasts. Combining forecasts can improve forecasting performance when one set of forecasts contains information – in this case, information about observed road-surface temperatures – not contained in the other set of forecasts. Both simple averaging and regression modelling are examined in this study as methods of combining these RSTFs, with the former consisting of averaging the MO and OR forecasts and the latter consisting of regressing these two forecasts on the observed temperatures. The simple average forecasts outperform the MO or OR forecasts, both when these RSTFs are evaluated as forecasts of a continuous variable and when they are evaluated as forecasts of a binary (i.e. frost/no frost) variable. Combining via regression modelling leads to further modest improvements in forecasting performance (based on in-sample results). Overall, simple averaging captures a substantial fraction of the benefits that can be realised by combining the MO and OR forecasts for this small sample of road-surface temperature data. Several issues related to the formulation and evaluation of combined forecasts, as well as some practical implications of the results, are briefly discussed.