Generating forecasts for combination: Applications in inventory and call centre management Event

Event details

There is general consensus among research and practice that combining forecasts is an effective approach for reducing forecast errors and avoiding the model selection challenge. In the forty years since the seminal work of Bates and Granger, two areas of research continue to be of importance in forecast combination, the generation of the individual forecasts to be combined, and the evaluation and application of the combined forecast. The generation of individual forecasts is closely linked to modelling uncertainty, be it about the structure of the model, the estimates of the model parameters, or random unexplained variations in the sample data. This presentation will briefly discuss several methods for generating forecasts based on these different sources of uncertainty. These include the use of different model types, multiple architectures given the same model type, and data driven methods based on bootstrapping, boosting and cross-validation to name a few. The presentation will discuss what has been learnt, and will present results on the forecasting performance of such methods under different properties of the data and combination method.