Bridging accuracy and interpretability: Comparative insights from interpretable and black-box models for CO₂ emission forecasting

The accurate and understandable carbon dioxide (CO2) emission prediction is necessary in developing effective climate policies especially in fast developing nations such as India. Although some highly developed machine learning (ML) models (e.g., XGBoost and LSTM) have a high predictive accuracy, they are black-box models and do not permit application directly in policy making. To fill this gap, this paper explores the possibility of interpretable ML models to predict CO2 emission with a small yet critical set of predictors: total energy production (TEP) and total energy consumption (TEC). Decision Trees, Explainable Boosting Machines (EBMs), and Generalized Additive Models (GAMs) were constructed to compare annual 1990–2023 data and compare them against traditional black-box solutions. These findings indicate that, in terms of accuracy and interpretability, EBMs and GAMs outperform traditional models, and their error measurements prove their high level of performance. SHAP (SHapley Additive Explanations) analysis also presented the fact that the increasing TEP and TEC have a great impact that contributes to the emissions, so it is necessary to consider renewable energy and energy-efficient solutions on a large scale. This paper, which combines strong forecasting with clear understanding, can assist in replicable analysis of applying interpretable models to climate policy, to achieve more specific interventions and effective monitoring of the reduction of emissions.