LAUSR

A recent electricity price forecasting study has shown that the Seasonal Component AutoRegressive (SCAR) modeling framework, which consists of decomposing a series of spot prices into a trend-seasonal and a stochastic component, modeling them independently and then combining their forecasts, can yield more accurate point predictions than an approach in which the same autoregressive model is calibrated to the prices themselves. Here, we show that further accuracy gains can be achieved when the explanatory variables (load forecasts) are deseasonalized as well. More importantly, considering a novel extension of the SCAR concept to probabilistic forecasting and applying two methods of combining predictive distributions we find that (i) SCAR-type models nearly always significantly outperform the autoregressive benchmark but are in turn outperformed by combined SCAR forecasts, (ii) predictive distributions computed using Quantile Regression Averaging (QRA) outperform those obtained from historical simulation and bootstrap methods, and (iii) averaging over predictive distributions generally yields better probabilistic forecasts of electricity spot prices than averaging over quantiles.