Abstract Poverty remains a persistent challenge that hinders sustainable development, particularly in emerging economies where traditional economic indicators often fail to capture its multidimensional nature. This study aims to advance poverty alleviation strategies by applying machine learning techniques within the framework of development economics. A comprehensive approach was adopted, combining supervised models (Logistic Regression, Random Forest, Gradient Boosting, Neural Networks), unsupervised clustering for poverty typologies, and causal machine learning for intervention impact estimation. Results show that advanced models significantly outperform traditional methods, with Neural Networks achieving the highest accuracy (90%) and robust generalizability across cross-validation folds. Unsupervised clustering revealed hidden dimensions of poverty, while fairness checks confirmed the absence of bias across sensitive groups. The contribution of this study lies in its integrated framework that balances predictive accuracy, fairness, and policy relevance, offering actionable insights for governments and development practitioners to design more effective and equitable poverty reduction programs.
Sanmorino et al. (Thu,) studied this question.