كنفرانس بين المللی مهندسی صنايع و سيستم ها

صفحه اصلی / دوازدهمین كنفرانس بين المللی مهندسی صنايع و سيستم ها

Short Term Forecasting of European Union Allowance Prices Using XGBoost and Permutation Based Feature Selection

نویسندگان :

Yasin Heidari Soochelmaei¹ Nasser Safaie² Majid Mirzaee Ghazani³

1- دانشگاه صنعتی خواجه نصیرالدین طوسی 2- دانشگاه صنعتی خواجه نصیرالدین طوسی 3- دانشگاه صنعتی خواجه نصیرالدین طوسی

کلمات کلیدی :

European Union Allowance،permutation feature importance،XGBoost،feature selection،EU ETS،Carbon Price Prediction

چکیده :

Accurate short-term forecasting of European Union Allowance (EUA) prices is essential for supporting carbon-market participants in compliance planning, investment decisions, and risk management. However, forecasting performance may deteriorate when redundant or weakly informative explanatory variables are included. This study proposes a concise and interpretable forecasting framework that integrates permutation feature importance with the Extreme Gradient Boosting (XGBoost) algorithm to identify the most informative predictors of EUA prices. Daily observations of EUA prices together with five exogenous variables, including Brent crude oil, Henry Hub natural gas, gold, the S&P 500 index, and the Geopolitical Risk (GPR) index, were analyzed over the period from January 2021 to June 2025. A strictly chronological data-partitioning strategy was adopted to prevent temporal information leakage during feature selection and model evaluation. The results identified geopolitical risk, the S&P 500 index, and gold as the three exogenous variables with the highest permutation importance scores. Using these variables together with the one-day lagged EUA price, the proposed XGBoost model achieved an MAE of 1.2057, an RMSE of 1.4584, a MAPE of 1.5595%, and an R² of 0.8985 on an independent out-of-sample test set. Furthermore, The selected-feature model outperformed the model using all available exogenous variables, the naive benchmark, and the linear regression benchmark, indicating that appropriate feature selection can improve predictive accuracy while maintaining model simplicity and interpretability. These findings demonstrate that a compact feature set can provide reliable short-term carbon price forecasts without relying on unnecessarily complex modeling frameworks. Furthermore, the Diebold-Mariano test yielded a test statistic of 3.8531 (p-value = 0.0001), confirming that the predictive superiority of the proposed framework over the Naive benchmark is statistically significant at the 1% level.