Machine learning and phone data can improve targeting of humanitarian aid

Summary

Peer-reviewed study evaluating machine-learning algorithms trained on mobile phone CDR data to predict individual poverty and target COVID-19 emergency cash transfers in Togo's Novissi programme. Reports that the ML approach reduces exclusion errors by 4–21% relative to geographic targeting and reached beneficiaries substantially poorer than the average population. Covers methodology (gradient boosting on ~150 CDR-derived features, trained against phone survey ground truth for 5.7 million subscribers), fairness analysis, and comparison with alternative targeting regimes.