Forecasting European Union Electronic Trading Systems Phase 4 Spot Prices Using Data‐Driven Hybrid Deep Learning Models: Integrating Energy and Market Activity as Controls

Amid the focus on climate change mitigation, this study explores carbon market forecasting. This study uses a hybrid forecasting framework that integrates empirical model decomposition, bidirectional long short‐term memory (BiLSTM) network, and attention mechanism to enhance the predictive performance of carbon spot prices within the European Union (EU) Emissions Trading System (ETS). The model decomposes the nonstationary carbon prices to multiple intrinsic mode functions (IMF) representing each distinct frequency component. The forecasting at IMF level enables learning of temporal dependence and volatility. The final model reconstructs the signals to present overall prediction. The multiple iterations that include a selection of macroeconomic variable led to the final root mean square errors (RMSE) value of 4.59, which shows that the BiLSTM outperforms a conventional long short‐term memory (LSTM) setup. This study also improves the model by including exogenous macroeconomic variables and policy shocks to enhance predictive accuracy. Shapley additive explanations (SHAP) analysis also identified the important features and variables. The visualized confidence interval confirms the reliability of the forecasts. The findings of the study highlight the effectiveness of integrating signal decomposition with deep learning and inclusion of exogenous factors. This study offers practical insights for regulators and researchers who are engaged in the emissions market and climate finance.