Univariate and multivariate forecasting of the electricity futures curve using Dynamic Recurrent Neural Networks

In recent years international power markets have witnessed high uncertainty and extraordinary volatility which, given the inherent complexity of the market, has made the Electricity Price Forecasting (EPF) process increasingly difficult. Therefore the development of a proper forecasting framework suitable for both stable and volatile periods has assumed an increasing importance for market players and policymakers in both strategic planning and risk management. At present, the majority of the studies on electricity price forecasting focused on the analysis of spot markets, neglecting the importance of derivative price modeling to mitigate the risks induced by market downturns and turmoil. Our study nests within this research stream and analyzes the potential of a set of state-of-the-art Machine Learning (ML) models for the prediction of the term structure of electricity futures prices. The objective is to define an ML-based framework capable of ensuring high predictive performance of the term structure during both stable and extremely turbulent conditions. In this regard we examined the predictive capabilities of a variety of Dynamic Recurrent Neural Networks (DRNNs) including: Nonlinear Autoregressive Neural Networks (NAR-NNs), NAR with Exogenous Inputs (NARX-NNs), Long Short-Term Memory (LSTM-NNs), Stacked Long Short-Term Memory (ST-LSTM-NNs), Bidirectional Long Short-Term Memory (BI-LSTM-NNs) and Encoder–Decoder Long Short-Term Memory Neural Networks (ED-LSTM-NNs). The models were applied to both low fluctuating and volatile sets of daily futures prices of the European Energy Exchange (EEX) for univariate as well as multivariate forecasting. Additionally, we compared this set of networks to baseline models commonly used in the EPF literature, including classical statistical and ML methods. Empirical results highlighted that DRNN models predictions are consistent with futures prices trends observed under different market regimes and outperform the competitors’ performance. Overall, main outcomes of the study may be summarized as follows: LSTM-based models seem to have the highest predictive power, with robust performance under various conditions. In detail the Multivariate BI-LSTM-NN performs better under quiet market conditions ensuring an accuracy level of 98.11 %, while the Univariate ED-LSTM-NN ensures superior predictive performance in presence of turmoil, achieving a 95.33 % accuracy.