The project developed a machine–learning–based debt distress detection tool for microfinance borrowers that makes distress visible and measurable using administrative data already held by lenders. In current practice, lenders observe only repayment outcomes (delinquencies), even though distress often sets in much earlier, with borrowers adopting different repayment strategies and coping mechanisms, such as cutting essential consumption, withdrawing children from school, or foregoing medical care, to continue repaying.
To address this blind spot, we trained and field-tested the model on data from over 250,000 borrowers in partnership with the Robert Bosch Centre for Data Science and AI at IIT Madras and a leading NBFC. We were able to analyse loan characteristics, repayment behaviour, and account-level patterns. The model identifies latent distress and the future probability of delinquency, enabling lenders to detect potential borrower harm early. Field validation showed delinquency prediction accuracy exceeding 95 percent. Distress prediction accuracy of approximately 75 percent also represents a significant advance over prevailing practices where distress remains entirely unmeasured.
As a next step, the detection tool will feed into a broader set of Debt Distress Protocols aimed at testing and institutionalising post-detection interventions, shifting the system from passive observation of harm to structured, evidence-based response strategies.
