Machine learning and phone data can improve targeting of humanitarian aid

Summary

Peer-reviewed Nature article reporting on the use of machine learning algorithms trained on mobile phone metadata (call detail records) and satellite imagery to target emergency cash transfers in Togo. Documents two-step targeting: (1) geospatial poverty mapping of 397 cantons using satellite and survey data, and (2) individual-level consumption prediction for 5.83 million mobile subscribers using CDR features. Reports that phone-based targeting reduced exclusion errors by 4-21 percent versus feasible geographic alternatives but increased errors by 9-35 percent versus hypothetical PMT. Code publicly available at GitHub.