Gender-Differentiated Credit Algorithms using Machine Learning
Woman selling fruits on the street | Photo Credit Adobe Stock
Context
Low-income women disproportionately lack access to credit, often because they lack credit histories, property rights, and formal earnings. This study seeks to understand 1) whether gender-differentiated credit scoring models using non-traditional data (e.g. mobile phone call detail records) can increase women’s access to formal credit, 2) whether women rejected by standard credit scoring models would benefit from credit access along the dimensions of asset ownership, labor supply, risk coping, intra-household bargaining power, and intra-partner violence, and 3) how these benefits compare to the benefits of credit access for women selected by standard models.
Study Design
Partnering with a large telecommunications company and a large bank in the Dominican Republic, this study combines novel algorithm development with an impact evaluation of credit allocated based on the novel algorithm. Using existing credit clients’ data on bill payment histories, call detail records, and default, the researchers constructed a gender-differentiated credit scoring model using machine learning to jointly optimize objectives on two stratified populations (men and women).
Results and Policy Lessons
The researchers used the same data their partner uses to screen loan applicant ability and loan outcomes to simulate the effect of three retroactive credit scoring models: the traditional model with data for men and women pooled and an indicator variable for gender included in the model, as well as two separate male and female-only models. In this retrospective analysis, 48% of the women who applied for loans would be approved under both credit scoring models, 14% would be approved only under the female-only model, 13% would be approved only under the pooled model, and 24% would be rejected by both models. In other words, more than one-third of women who would be rejected by the traditional pooled credit scoring model would instead be approved for credit using a gender-differentiated model. Based on the promise of these results, the research team is taking a similar approach to a new project in Mexico.
