Interpretable forecasting of carbon allowance returns using multi-scale trend-aware network with dynamic variable selection and squeeze-excitation

Accurate carbon allowance returns forecasting is crucial for policymakers, investors, and industries seeking to effectively manage carbon emissions and hedge against market uncertainties. However, predicting carbon allowance returns faces the challenges of volatile short-term fluctuations, longer-term trend shifts, and the heterogeneous influence of related market factors. In this paper, we propose a multi-scale trend-aware network (MuSTANet) to capture and match temporal patterns in complex carbon allowance returns. Specifically, we introduce a multi-scale trend-aware self-attention mechanism that leverages parallel convolutional branches with varied receptive fields to discern local slope changes and match temporally relevant points. Furthermore, a dynamic variable selection strategy that adaptively reweights input features at each time step enables the model to focus on the features when their impact on the returns is most pronounced. The squeeze-and-excitation block recalibrates channel-wise responses to amplify important feature channels while suppressing less informative ones. We conduct extensive experiments on 4 carbon trading systems, including the European Union Emissions Trading System (EU ETS) and different Chinese carbon markets. The results demonstrate that MuSTANet reduces RMSE by about 28–35% and increases directional accuracy (DA) by 16–33% in the EU ETS, while achieving about a 15–38% RMSE reduction and 18–70% DA improvement across Chinese carbon markets compared with baselines. It also consistently delivers the best performance in trading simulations against all strategies. Its adaptability and interpretability not only provide robust forecasting tools but also meet policy regulators’ requirements for transparency and provide investors with explainable signals for hedging, portfolio management, and stakeholder reporting.