09-03-2026
Machine learning (ML) has transformed how financial institutions evaluate loan applicants. By analyzing hundreds of variables simultaneously, AI-driven credit models can identify patterns that traditional credit scoring methods often miss, helping lenders make faster and more accurate lending decisions.
But these models have a blind spot: they often struggle to evaluate applicants with limited financial histories, including college students, recent immigrants, refugees and young professionals, who can be denied credit not because they’re riskier, but because there isn’t enough data to judge them accurately.
New research by the Daniels School's Cecilia Ying , who worked as a director of Credit Risk Analytics in a leading Canadian financial institution before joining academia, and Haokun Du from the University of Auckland proposes a solution. Their study, “ Policies for Mitigating Subgroup Differences in Machine Learning for Credit Assessment ,” introduces a post-processing framework that helps lenders reduce these disparities without redesigning or retraining their existing machine learning models.
The effectiveness of ML models depends heavily on the quality and completeness of the underlying data.
Using a dataset of more than 1.53 million loan applications containing 465 applicant features, the researchers evaluated four widely used machine learning algorithms: Decision Trees, Random Forest, LightGBM and XGBoost. Applicants were grouped according to the completeness of their financial histories.
The results revealed a consistent pattern: applicants with sparse financial histories were evaluated less accurately across every major machine learning model tested. Performance gaps reached as high as 20%, depending on the model and evaluation metric. These differences were statistically significant across nearly every comparison, suggesting the issue is systemic rather than incidental.
Importantly, the researchers emphasize that these applicants are not necessarily riskier borrowers. Instead, incomplete financial histories reduce the models' ability to distinguish between qualified and unqualified applicants.
Rather than requiring lenders to collect additional customer information or replace their machine learning systems, the researchers developed Subgroup Threshold Optimization (STO ), a model-agnostic post-processing technique.
Traditional credit models typically apply a single approval threshold. STO instead assigns different approval thresholds to borrower groups based on the completeness of their financial histories. This post-hoc adjustment approach can be implemented with any model without modifying or retraining.
The framework allows lenders to optimize lending policies according to different strategic priorities, balancing fairness and predictive performance. The researchers evaluated three approaches:
One of the study's most surprising findings challenges a common assumption in AI governance that improving fairness inevitably reduces financial performance.
The fairness-focused policy reduced subgroup performance gaps by nearly 100%. Although this approach lowered predictive scores for some groups, financial analysis showed it generated the highest overall lender profits among the approaches evaluated.
The explanation lies in identifying borrowers who had previously been rejected because of incomplete financial histories rather than genuine default risk. By more accurately approving creditworthy applicants from underrepresented groups, lenders expanded lending opportunities while increasing profitable loans.
The balanced policy also produced meaningful improvements, reducing performance disparities by roughly 49% while maintaining or improving baseline predictive performance across borrower groups. Meanwhile, the performance-focused policy still reduced disparities by approximately 31%, demonstrating that subgroup-specific thresholds can improve both fairness and operational effectiveness.
Financial institutions face growing pressure from regulators, customers and investors to ensure AI systems produce equitable outcomes.
Because STO operates after predictions are generated, it offers organizations a comparatively simple way to improve fairness while preserving existing machine learning infrastructure. The framework can also be customized to match an institution's risk tolerance, profitability goals and regulatory requirements.
Beyond lending, the researchers suggest the framework could be adapted to other machine learning classification problems where a single decision threshold may unintentionally disadvantage certain populations. To that end, business leaders evaluating AI-powered decision systems should consider several lessons from the research:
As AI becomes increasingly central to lending decisions, the research suggests organizations need not choose between fairness and performance. With thoughtful policy design and targeted optimization, financial institutions may be able to improve both.
