In the last three decades the vast majority of electricity price forecasting (EPF) research has concerned day-ahead markets. However, the rapid expansion of renewable generation—mostly wind and solar—have shifted the focus to intraday markets, which can be used to balance the deviations between positions taken in the day-ahead market and the actual demand and renewable generation. A recent EPF study claims that the German intraday, continuous-time market for hourly products is weak-form efficient, that is, that the best predictor for the so-called ID3-Price index is the most recent transaction price. Here, we undermine this claim and show that we can beat the naïve forecast by combining it with a prediction of a parameter-rich model estimated using the least absolute shrinkage and selection operator (LASSO). We further argue, that that if augmented with timely predictions of fundamental variables for the coming hours, the LASSO-estimated model itself can significantly outperform the naïve forecast.

Deep neural networks are rapidly gaining popularity. However, their application requires setting multiple hyper-parameters, and the performance relies strongly on this choice. We address this issue and propose a robust ex-ante hyper-parameter selection procedure for the day-ahead electricity price forecasting that, when used jointly with a tested forecast averaging scheme, yields high performance throughout three-year long out-of-sample test periods in two distinct markets. Being based on a grid search with models evaluated on long samples, the methodology mitigates the noise induced by local optimization. Forecast averaging across calibration window lengths and hyper-parameter sets allows the proposed methodology to outperform a parameter-rich least absolute shrinkage and selection operator (LASSO)-estimated model and a deep neural network (DNN) with non-optimized hyper-parameters in terms of the mean absolute forecast error.

Recent studies suggest that decomposing a series of electricity spot prices into a trend-seasonal and a stochastic component, modeling them independently, and then combining their forecasts can yield more accurate predictions than an approach in which the same parsimonious regression or neural network-based model is calibrated to the prices themselves. Here, we show that significant accuracy gains can also be achieved in the case of parameter-rich models estimated via the least absolute shrinkage and selection operator (LASSO). Moreover, we provide insights as to the order of applying seasonal decomposition and variance stabilizing transformations before model calibration, and propose two well-performing forecast averaging schemes that are based on different approaches for modeling the long-term seasonal component.