J-PAL Agarwal et al. 2019 India — mobile / social-footprint ML predicts loan defaults more effectively than credit-only models

Claim. Agarwal et al. (2019) India RCT-evidence study found that mobile-phone and social-graph footprint signals fed into ML models predict loan defaults more effectively than models restricted to traditional credit scores.

Quote.

"Predict loan defaults more effectively than models that only use credit scores."

Source. Abdul Latif Jameel Poverty Action Lab (J-PAL), summary of Agarwal et al. (2019) India fintech-lending study, povertyactionlab.org (accessed 2026-06-21).

Confidence. Verified. J-PAL is a methodologically conservative research organisation; default-rate prediction is a clean outcome measure.

Caveats. India context, with mobile / social-graph data availability that may differ from Canadian privacy regime ([[reference-candid-kb-content-pattern]]); replicate the mechanism claim, not the data inputs, for Canadian application.

Implication / use. Third leg of the alt-data credit evidence stack (with NBER and IFC). The convergence across three independent literatures — primary academic, multilateral, RCT-tradition — is what makes the risk-domain claim materially harder to wave away than the retention magnitudes.