Forecasting carbon dioxide emissions using macroeconomic indicators: a machine learning approach

초록

This study employs 56 macroeconomic indicators to forecast the annual carbon dioxide (CO2) emissions in Korea. The 56 macroeconomic indicators are analogous to those utilized in FRED-MD and represent Korean data from 1985 to 2022. This study uses various machine learning techniques, including XGBoost, complete subset regression, and random forest models. The Boruta algorithm, a variable selection technique based on the random forest model, is applied to identify the variables that exert the greatest influence on annual CO2 emissions forecasting. Furthermore, out-of-sample multi-period forecasts are evaluated for each forecasting model to select that which best predicts annual CO2 emissions. The results demonstrate that the random forest model and the random forest model with the Boruta algorithm exhibit the most favourable out-of-sample performance in short-term forecasting. In selecting variables to predict CO2 emissions using the Boruta algorithm, the production index and shipment index of producers in the electricity, gas, steam, and air conditioning supply industry are identified as significant for enhancing the predictive capacity of the model. The findings provide key policy insights for carbon emissions management. Leveraging real-time macroeconomic indicators can help policymakers address delays in CO ${}_2$ 2 emissions data, enabling adaptive regulations that balance economic growth and environmental sustainability.

키워드