Enhancing Interpretability in Credit Risk Modeling: A Synergistic Approach Integrating SHAP and Counterfactual Explanations

Every time a bank decides whether to lend money, it relies on a model to weigh a person's financial history against the risk of a default. For decades, these models were simple enough that a human could trace the logic: if you missed payments, your score went down. But as banks began using more powerful computer systems to find subtle patterns in vast amounts of data, the logic became hidden inside a "black box." These advanced systems could predict risk with great accuracy, but they could not explain why they made a specific decision. This created a serious problem. Laws in many places now require lenders to explain their decisions to applicants, and without a clear reason, a denial of credit feels arbitrary and unfair. The challenge for scientists has been to find a way to keep the high accuracy of these complex systems while opening the black box to reveal the reasoning inside.

Two new tools have emerged to help solve this puzzle. The first is a method that acts like a magnifying glass, showing exactly which pieces of information—such as the balance in a checking account or the length of a loan—pushed the model toward a specific conclusion. The second tool is more like a "what-if" simulator. It asks the computer to imagine the smallest possible changes a person could make to their financial profile to flip a rejection into an approval. While each tool is useful on its own, researchers at Arizona State University wondered if combining them would create a clearer picture. They set out to test whether using these two methods together could not only explain why a decision was made but also provide a practical, fair path forward for the person who was turned down.

To test this idea, the researchers used a well-known collection of data containing one thousand real credit applications from Germany. They built two different types of prediction models to see how they behaved. One model used a straightforward, linear approach, while the other used a more complex system that builds many decision trees to find patterns. They then applied their combined explanation method to both models. The results showed that the complex model did not actually perform much better than the simple one in terms of overall accuracy, but the two models made different kinds of mistakes. The simple model was better at catching people who were likely to default, while the complex model was better at avoiding false alarms. More importantly, the combined explanation method worked remarkably well. By using the "magnifying glass" tool to identify the most important factors first, the researchers were able to guide the "what-if" simulator to find the best solutions much faster.

The study found that this combined approach made the explanations significantly more efficient. When the researchers asked the computer to find the minimal changes needed to change a rejection into an approval, the old method required an applicant to imagine changing about three different things at once. With the new, guided method, the applicant only needed to imagine changing about two things. This reduction in complexity makes the advice much easier for a human to understand and act upon. The system successfully generated valid suggestions one hundred percent of the time, proving that it could reliably find a path to a different outcome without breaking the rules of the model.

Beyond just making the process clearer, the researchers used this framework to look for unfairness. They discovered that the models treated different groups of people differently, even when the applicants had similar financial profiles. For instance, single men were flagged as high-risk more often than married or widowed men, even when their financial situations were comparable. The analysis showed that the model was stricter with the single men, requiring them to make larger financial improvements to get approved. By using the combined tools, the researchers could pinpoint exactly where this bias happened. They saw that the lack of a checking account hurt the single men's scores more than it hurt the married men's scores, suggesting the model had learned a hidden bias from the historical data it was trained on.

The researchers also found that the most powerful factors in these decisions were not demographic details like age or gender, but concrete financial realities. The status of a checking account and the history of credit repayment were the dominant forces driving the decisions. This is a reassuring finding, as it suggests the models are focusing on the actual ability to repay rather than irrelevant personal characteristics. However, the study also highlighted that even when the main factors are fair, the way the model weighs them can still create unequal outcomes for different groups. The combined explanation method was able to expose these subtle inequalities, showing exactly how much extra effort one group had to exert compared to another to achieve the same result.

Ultimately, this work suggests that we do not have to choose between a smart model and a fair, understandable one. By weaving together the ability to explain the past with the ability to suggest a better future, financial institutions can build systems that are both accurate and transparent. The study demonstrates that when we guide the search for solutions using the most important factors, we can give applicants clear, actionable advice that respects their specific circumstances. This approach moves beyond simply telling a person they were denied; it offers a concrete, fair roadmap for how they might improve their standing, ensuring that the technology serving our financial lives remains accountable to the people it affects.