November 11, 2024
In banking, lenders must make decisions on whether to accept or reject loan applications based on credit policies and credit risk models. However, for future modeling and strategy development, rejected applicants often remain an untapped source of valuable data, as their repayment outcomes are unknown. This creates a challenge: how do we account for these rejected applicants when building better credit risk models for approval decisions?
Stratyfy’s approach, called probabilistic modeling with proportional sampling, offers a way to improve credit decision-making by using this hidden data to refine predictions, allowing lenders to assess credit risk more accurately.
What is Probabilistic Modeling with Proportional Sampling?
At its core, this technique helps lenders make better lending decisions by considering rejected application’s assumed performance in the credit model. Historically, the use of credit models built with data from only accepted applications leads to bias (and an underestimation of future risk) because the outcomes for rejected applicants (whether they would repay or default) remain unknown. From a business perspective, it is not possible to improve on previously built credit models and strategies using only previously accepted applicants.
Probabilistic modeling helps fill this gap by estimating the likelihood of repayment or default for rejected applicants. These estimates are then used to build a more complete dataset, which combines both accepted and rejected applications for training a stronger, more inclusive credit model.
How Does Probabilistic Modeling Work?
The process begins by using a model that is trained on a combination of accepted and rejected applicants. This model is used to predict the probability of whether rejected applicants would have been “Good” (repaying) or “Bad” (defaulting) and is called semi-supervised learning. The joined dataset is then used for rule mining, and the labeled dataset for model calibration.
For models that can’t use probabilities directly, a second step called proportional sampling is applied. In this step, the rejected applicants are assigned either a “Good” or “Bad” label based on their probabilities. These completed observations are then added to the training dataset, allowing for a fuller picture of potential applicants and improved decision-making.
Refining the Model: Iterative Approach
One of the unique aspects of this technique is its ability to improve its predictions by breaking the dataset into parts and estimating one at a time. In each round, the model focuses on rejected applicants who closely resemble those who have already had their outcomes predicted. This helps to refine the estimates, making the model more accurate and capable of handling different types of applicants.
By repeating this process, lenders can improve their understanding of how likely rejected applicants are to repay their loans. This iterative refinement helps reduce the risk of inaccurate predictions while improving the overall accuracy of the credit risk model.
Case Study: Unsecured Product Underwriting
In the context of unsecured underwriting (such as personal loans or credit cards), this approach has shown significant benefits. By using probabilistic modeling, a lender was able to improve their credit risk model without sacrificing accuracy for accepted applicants. The model also offered more realistic projections for rejected applicants, helping the lender manage portfolio risk better when expanding their approval criteria.
For instance, the enhanced model offered a more accurate prediction of delinquency rates when the approval rate increased by 1%. This led to better risk control, ensuring that the lender could grow their mortgage portfolio without taking on excessive risk.
This methodology has also been successfully replicated in secured products such as mortgages.
Why Should Lenders Care About Reject Inference?
- Improved Decision-Making: By incorporating data from rejected applicants, lenders can make more informed lending decisions for expansion. The model becomes more robust and generalizes better to future applicants, including those who might have been rejected under older models.
- Enhanced Risk Management: Reject inference through this method results in better-calibrated predictions for default risk. This allows lenders to set more accurate approval thresholds while maintaining control over default rates.
- Preserves Data Balance: The technique ensures that the proportion of “Good” and “Bad” loans in the dataset remains balanced, reducing bias in the final model.
- Consistency in Accepted Applicants’ Performance: Lenders can improve their model’s performance on rejected applicants while maintaining the accuracy of predictions for accepted ones. This offers the best of both worlds—expanded loan approvals with controlled risk.
Probabilistic modeling with proportional sampling is a powerful tool for lenders looking to enhance their credit models. By making use of reject inference, this technique offers a more complete view of potential future borrowers, allowing for better predictions and lending decisions. Whether you’re looking to expand your portfolio or simply tighten your risk management, this method provides a flexible and effective approach for modern lending.
Interested in more original insights from our team? Subscribe to our quarterly newsletter.