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…

By the Truvace deskOctober 17, 20262 min read1 sourcePeer-reviewedTRV-2026-1743· cite & verify

In brief

Researchers tested whether trade openness, governance, environmental taxation and macroeconomic indicators could classify countries into low, medium or high per-capita CO2 emission levels from 2000 to 2023. They compared four standard classifiers and two ordinal models across several validation designs, using World Development Indicators and OECD data.

The findings matter because they show how validation design can overstate the reliability of emission-classification tools used for climate monitoring. The authors stress that the predictors are not causal determinants of emissions, so the models support risk monitoring and further policy evaluation but not causal policy assessment.

Main points

Nested country-grouped cross validation gave mean accuracies from 0.637 for ordinal XGBoost to 0.676 for ordered logit, with overlapping confidence intervals.
A random country-year split reached 0.91 accuracy, a 12.5 percentage point gap for XGBoost.
Chronological holdout and 1-year-ahead forecasting accuracy fell in the region of 0.65 to 0.74.
Government effectiveness, urban population growth and environmental tax revenue were the most influential features across SHAP, LIME and permutation importance.

The problem

Random country-year splits inflate accuracy to 0.91 because models recognize seen country profiles, while country-grouped validation yields only about 0.64 to 0.68 accuracy for unseen countries.

The rundown

The study divided CO2 emissions per capita into three ordinal levels using tertile cut points estimated only on training data within each validation split. Models were tuned and compared under country-grouped holdout, grouped 5-fold cross validation, chronological holdout, 1-year-ahead forecasting and a random split retained as a reference.

No single classifier clearly outperformed the others under country-grouped validation. The support vector machine, multinomial logit and ordered logit scored slightly higher than the tree-based models, though confidence intervals overlapped.

Explainability analysis found trade openness contributed moderately and GDP growth least. The feature ranking was stable across importance methods, learning algorithms and cross validation folds. Accuracy fell less under time-based validation, suggesting generalization is better across time within a country than across countries.