Predicting Carbon Prices with Explainable Time-Varying Feature Selection: Evidence from the EU ETS

He Jiang, Rongyu Cao（通讯作者）

Abstract

While accurate carbon price prediction in the EU Emissions Trading System (EU ETS) is increasingly important, existing studies often rely on static feature sets and thus overlook the time-varying effects of institutional reforms and external shocks. To address this limitation, this paper develops a forecasting approach that incorporates explainable time-varying features into the prediction process. Using daily multidimensional data spanning policy, financial, energy, and commodity variables from 2013 to 2025, a rolling explainable machine learning method is employed to trace shifts in predictor relevance over time. The evidence indicates a gradual transition from a policy-dominant to a more market-driven predictive structure across different phases of the EU ETS. To interpret these shifts more systematically, feature contributions are grouped by economic dimension, and a Systemic Fragility Index is constructed from driver concentration, realized volatility, and tail risk. Building on this evidence, the dynamically screened features are incorporated into carbon price forecasting. Compared with static alternatives, the resulting forecasts achieve higher out-of-sample accuracy and perform well against benchmark models. A supplementary trading exercise further suggests that the forecasts contain economic value under realistic trading frictions. These findings provide empirical evidence on the dynamic evolution of carbon price predictors and offer practical implications for improving carbon market forecasting and regulatory decision-making.

Keywords: carbon price forecasting; feature selection; deep learning; rolling SHAP