A global machine learning framework for classifying national CO2 emission levels: responsible and explainable artificial intelligence for sustainable climate governance

We investigate whether trade openness, governance, environmental taxation and selected macroeconomic indicators can classify the contemporaneous CO2 emission level of countries observed over 2000 to 2023, using data from the World Development Indicators and the OECD. CO2 emissions per capita are divided into three ordinal levels, low, medium and high, using tertile cut points estimated on the training data only within every validation split. Four classifiers, Random Forest, XGBoost, a support vector machine and a multinomial logistic regression, are tuned and compared alongside two models designed specifically for ordinal outcomes, an ordered logistic regression and an ordinal decomposition of XGBoost, under four validation designs: a country grouped holdout and country grouped 5 fold cross validation, in which no country appears in both the training and test sets; a chronological holdout; a 1 year ahead forecasting task; and a random stratified country year split, retained only as a reference point. Under country grouped cross validation, no classifier achieves a decisively higher accuracy than the others: nested cross validation gives a mean accuracy of 0.644 for XGBoost, 0.648 for Random Forest, 0.671 for the support vector machine, 0.674 for the multinomial logit, 0.676 for the ordered logit and 0.637 for the ordinal XGBoost model, with overlapping confidence intervals throughout. This is markedly lower than the 0.91 accuracy obtained under a random country year split with the same tuned pipeline, a gap of 12.5 percentage points for XGBoost, confirming that a substantial part of the high accuracy obtainable under random splitting reflects the model recognizing previously seen country profiles rather than generalizing to unseen countries. Accuracy falls less sharply under the chronological holdout and the 1 year ahead forecasting task, in the region of 0.65 to 0.74, indicating that the classifiers generalize better across time within a country than across countries. Permutation importance, SHAP and LIME consistently identify government effectiveness, urban population growth and environmental tax revenue as the features most strongly associated with model predictions, with a moderate contribution from trade openness and the smallest from GDP growth, a ranking that is stable across importance methods, learning algorithms and cross validation folds. These variables are predictors of the model's output, not estimated causal determinants of emissions, and the results are best read as evidence for climate risk monitoring and further policy evaluation rather than as a causal policy assessment.