Model-driven multimodal LSTM-CNN for unbiased structural forecasting of European Union allowances open-high-low-close price

European Union allowances (EUAs), the “currency in circulation” of the EU Emissions Trading Scheme (ETS), have spawned a great deal of speculative trading. This study proposes a model-driven long short-term memory network (LSTM)-convolutional neural network (CNN) hybrid model that integrates numerical data features and candlestick features to achieve accurate and unbiased structural prediction of EUA futures open-high-low-close (OHLC) prices during the four phases of the EU ETS. During EU ETS Phase IV, the out-of-sample prediction outcomes of the LSTM-CNN model exhibited a mean absolute percentage error (MAPE) of 0.942%, a mean absolute error (MAE) of 0.877, a root mean squared error (RMSE) of 1.157, a goodness-of-fit (R2) of 0.953, an accuracy ratio (AR) of 0.544, and a forecast correct rate of ups and downs (UP) of 0.579. In comparison to Naive methods, vector autoregression (VAR) combined with vector error correction model (VECM), multiple linear regression (MLR), partial least squares (PLS), support vector regression (SVR), and standalone LSTM, the LSTM-CNN approach demonstrated a notable enhancement in the average MAPE across the four stages of the EU ETS—specifically, an improvement of 21.66%, 43.15%, 15.73%, 15.72%, 10.45%, and 5.91%, respectively. Drawing from the unbiased structural forecasts of OHLC data, this study proposes fruitful intraday trading strategies that attain substantive investment returns in the realm of EUA futures trading. The multimodal forecasting methodology and intraday trading strategies advanced in this study hold considerable promise within the domain of energy finance.