Deep learning approaches for high-frequency carbon futures price forecasting: A comparative analysis of recurrent, attention-based, convolutional, and hybrid models
TL;DR
This study examines the forecasting performance of standalone and hybrid deep learning models for EU allowance (EUA) futures prices using high-frequency open, high, low and close (OHLC) data and shows that recurrent neural network-based models generally outperform alternative architectures.
Abstract
Carbon markets have expanded rapidly in recent years as part of climate change mitigation efforts. Accurate forecasting of the European Union Emissions Trading System (EU-ETS) futures prices holds significant value for companies, investors, and policymakers, who rely on these forecasts for risk management and strategic planning. This study examines the forecasting performance of standalone and hybrid deep learning models for EU allowance (EUA) futures prices using high-frequency open, high, low and close (OHLC) data. The forecasting models include Transformer, Informer, Temporal Convolutional Network (TCN), Bidirectional Gated Recurrent Unit (BiGRU), TCN-Transformer, TCN-BiGRU, with benchmark models. The empirical results show that recurrent neural network-based models generally outperform alternative architectures. Among the standalone models, BiGRU and Transformer deliver the most robust forecasting performance across different frequencies and price series. Among the hybrid models, TCN-BiGRU consistently outperforms TCN-Transformer and the corresponding standalone models in most cases. The findings highlight the importance of modeling temporal dependencies in high-frequency carbon futures markets.