Electricity price forecasting models are typically estimated via rolling windows, i.e. by using only the most recent observations. Nonetheless, the current literature does not provide much guidance on how to select the size of such windows. This paper shows that determining the appropriate window prior to estimation dramatically improves forecasting performances. In addition, it proposes a simple two-step approach to choose the best performing models and window sizes. The value of this methodology is illustrated by analyzing hourly datasets from two large power markets with a selection of ten different forecasting models. Incidentally, our empirical application reveals that simple models, such as the linear regression, can perform surprisingly well if estimated on extremely short samples.

Keywords: electricity price forecasting, day-ahead market, parameter instability, bandwidth selection, artificial neural networks.

JEL classification: C22, C45, C51, C53, Q47.

(Published version on Energy  Journal: https://doi.org/10.5547/01956574.41.4.cfez)