<![CDATA[Purpose – This study aims to develop and test a comparative machine learning evaluation framework for predicting poverty status and educational risk as a methodological basis for evidence-based social policy. Design/methods/approach – A comparative simulation study was conducted using a controlled simulated dataset of 10,000 observations, sixteen input features, and two binary targets: poverty status and educational risk. Five supervised classification models were evaluated: Logistic Regression, Decision Tree, Random Forest, XGBoost, and LightGBM. The models were assessed using accuracy, F1-score, AUC, Brier Score, per-class performance, cross-validation stability, explainability, and a proposed Policy Readiness Index. The dataset included predefined prevalence assumptions, missing values, outliers, and simulated nonlinear and interaction effects. Findings - Within the controlled simulation setting, XGBoost achieved the strongest technical performance across both prediction tasks, with the highest accuracy, F1-score, AUC, and calibration quality. However, Random Forest obtained the highest Policy Readiness Index because it provided the best balance between predictive performance, cross-validation stability, and interpretable feature attribution. The findings show that the technically best model is not automatically the most policy-ready model. Research implications/limitations – The study offers a structured decision-support approach for comparing machine learning models in poverty and education policy contexts. However, all results are derived from simulated data and should be interpreted as a methodological proof of concept rather than empirical evidence for a specific real-world population. Originality/value – This study contributes a policy-oriented machine learning evaluation framework that integrates predictive quality, calibration, stability, explainability, and policy usability into a transparent Policy Readiness Index.]]>
