Enhancing electricity price forecasting accuracy: A novel filtering strategy for improved out-of-sample predictions

Abstract

Reliable electricity price forecasts are key for energy sector strategy. The presence of market volatility and price spikes may negatively affect the accuracy of predictions if not properly addressed. In this study, we introduced a novel filtering strategy designed to enhance the accuracy of electricity price forecasting by effectively identifying and replacing extreme price spikes. Our approach is grounded in the application of robust statistical techniques within a rolling window framework, allowing for the systematic cleansing of input data used for forecasting models. We validated the efficiency and accuracy of our method using state-of-the-art statistical and deep learning models within an open-access dataset framework encompassing six different energy markets. The comparison of accuracy metrics and the outcome of statistical tests consistently demonstrated improvements in forecast accuracy when using our filtered data, with gains of up to 4% for certain models with respect to the predictions obtained with unfiltered inputs. Finally, the proposed filtering strategy exhibits reasonable and affordable computational requirements, making it suitable for practical applications in a real-world market setting.