Beyond Biases: Five Insights on Enhancing Credit Fairness with Reject Inference Techniques

May 13, 2024

One of the least expensive ways financial institutions can increase their credit portfolio profitability is by improving the accuracy of their approval mechanisms. “Reject inference” techniques help financial institutions to do just that, with implications for offering credit to women who would otherwise be unable to access it.

People seek credit for personal, business, and educational purposes. Financial service providers (FSPs) evaluate these applications through algorithms, loan officers, or a combination of both. However, these evaluation methods can be susceptible to biases and errors, resulting in the unfair rejection of eligible applicants.

Reject inference is a quantitative method that identifies individuals who may be creditworthy but were mistakenly deemed non-creditworthy during credit assessment processes. Women’s World Banking had the opportunity to conduct extensive research on improving reject inference techniques in collaboration with eight financial service providers. This partnership enabled us to produce both a public-facing report and a five-hour course on this topic. This report and course were made possible by PayPal as part of its support of Women’s World Banking’s work using data science to increase financial services for low-income women globally.

“Women’s World Banking is a global force for advancing financial access for women and girls worldwide. We are honored to have contributed to their latest study on how financial service providers around the world can leverage machine learning (ML) and artificial intelligence (AI) to detect reject inference bias in their credit worthiness assessments. This research isn’t just impactful; it has the potential to lead to transformative innovation, especially for low-income women who may not otherwise have access to the critical business funding that financial service providers in their communities can offer. The study offers actionable insights for immediate implementation by those providers, empowering them to be more inclusive and make a lasting difference for their customers.”

Andrea Donkor, SVP, Global Regulatory Relations and Consumer Practices, PayPal

Here in this insight note, we have summarized our main findings and insights:

  1. Reject inference has the potential to mitigate the adverse consequences of the amplified bias effect.
    In credit approvals, understanding feedback loop or amplified bias effect is crucial. This phenomenon occurs when the outcomes of a process are reused as inputs, often reinforcing initial biases or errors. Initial credit rejections, due to biases or mistakes, can adversely affect an individual’s credit history, creating a cycle where these applicants struggle more to obtain future credit because of their now-damaged credit records. Reject inference plays a role in identifying individuals who, despite initial rejections, are likely creditworthy.

  2. Reject inference can enhance the credit assessment processes used by FSPs, without necessitating major alterations to their existing credit evaluation practices.
    FSPs invest considerable financial resources and time in developing their credit assessment methods. When these methods involve developing credit scoring algorithms, the investment becomes even more significant. Major modifications to this model are difficult to undertake. In contrast, reject inference facilitates a smooth integration with current credit assessment methods, maintaining established practices. For FSPs, implementing reject inference techniques is a practical initial step toward enhancing fairness and reducing missed business opportunities.

  3. In saturated markets with numerous FSPs, acquiring new customers is challenging, and mistakenly rejecting potential customers can escalate costs.
    Offering credit in competitive markets comes with unique challenges. The crowded digital credit landscape complicates acquiring and retaining customers. Erroneous rejections, which deny credit to deserving applicants and result in losing potential customers, are therefore particularly expensive to FSPs operating in competitive settings.

  4. Merging matching algorithms and machine learning (ML) techniques can create a powerful and intuitive approach to reject inference.
    Integrating matching algorithms like propensity score matching with ML models presents a robust method for identifying creditworthy applicants mistakenly rejected due to biases or errors. This approach provides a statistically sound and intuitive basis for tackling missed business opportunities using reject inference.

  5. Counterfactual correction opens the door to a new, robust, and explainable class of reject inference techniques.
    Counterfactual correction, a ML technique, can significantly enhance reject inference methods. This method offers clear, human-understandable explanations for automated decisions, especially useful in credit assessments. By identifying the specific attributes that affect credit decisions, it provides actionable feedback to applicants on improving their future creditworthiness. When combined with ML methods designed to detect and correct noisy labels, counterfactual correction introduces a novel and robust approach to reject inference, improving both the fairness and accuracy of credit assessments.

The five insights highlighted point to a clear call to action: If you aim for credit fairness and face a high rejection rate in your portfolio, implementing reject inference techniques and leveraging the power of ML could be the right choice for you.