Forecasting model selection

At the heart of any forecasting system is usually a statistical model, often a very simple one. But the model should be as accurate as possible, capturing the key features of data such as seasonality and trend. The model’s parameters need to be estimated efficiently. The Centre’s recent research has examined problems of model and parameter selection from simple exponential smoothing models to complex neural networks.

The selected model should provide accurate, reliable and robust forecasts. Focusing solely on accuracy may result in very volatile forecasts that are not practical for organisations; therefore model selection has to balance all these targets. Naturally, model selection is closely related to parameter selection for the different candidate models.

Current research is investigating i) robust model parameter optimisation (Kourentzes and Kaparis, 2013), ii) model selection rules and the effect of data properties on their performance (Fildes and Kourentzes, 2013) and iii) robust model selection for long term forecasting using temporal aggregation (Kourentzes, Petropoulos and Trapero, 2013).

A full list of publications can be found here.