Data and Analytics

How Neural Network Models Put Financial Services Within Reach

How Neural Network Models Put Financial Services Within Reach

December 16, 2020 | Mike McBurnett

Why Constrained Models are Optimized for Credit Risk

Credit scores are used as a big part of a lender’s credit criteria - higher scores more often mean credit is granted, while lower scores result in declines. In 2015, Equifax pioneered the use of machine learning models in credit scoring after discovering that these models can be better predictors of default risk, while also being able to answer the critical question every consumer has when declined: Why?

Here, we will explore how these smarter models, known as neural network models, are gamechangers for consumers and financial institutions. We will also explain why it is critical to understand how different versions of these models—unconstrained versus constrained—perform in credit risk applications. We will also explore why only one of the versions—the constrained models—best predict default risk and inform the consumer why they have been denied.

Making the Right Decision for Every Consumer

Consumers are all different and so are their credit profiles. For example, two people may both have a 660 credit score, but one may be building new credit, improving and trending up the credit spectrum while another is financially overloaded and trending down the score spectrum. Our scores are based only on the data used to create them.

But new sources of data are becoming available. Equifax is testing a wide variety of new data--alternative data not found in traditional credit reports that allows us to vastly widen the net of consumers who can be appropriately scored with machine learning models. Communications, utility, rental histories, and more non-traditional data is available as inputs in our machine learning models. These data allow our models to capture the same kind of behavior reflected in traditional credit behaviors, but from different sources.

This is important because it opens up a new universe of consumers who were previously overlooked and underserved by financial services providers.

This includes the roughly 66.5 million “thin file” consumers, who have a little traditional credit history but not enough to score up to a prime credit offer. It also includes roughly 25 million “credit invisible” consumers who have no traditional credit, and because of that they are stuck in a “catch-22.” They are denied credit because they have no traditional credit history, and they have no credit history because they are constantly denied.

Our machine learning models are built with financial inclusion in mind. Neural network models can be used to better differentiate and predict credit risk. As a result, lenders can offer higher quality financial services to the right consumers. Just think: better credit terms, higher limits, lower interest rates for a wider audience of consumers, without increasing risk levels. This includes people above and below the line who need credit, loans and mortgages to achieve the dream of home ownership, purchasing a new vehicle, starting a new business or simply buying a new sofa. Put simply, these models can help more consumers live their best financial life.

Using Neural Networks for Credit Risk

Much like consumers, neural network credit risk models are complex and non-linear. However, they can also be more accurate. This is good. It is what enables lenders to identify creditworthy consumers they would not otherwise recognize. However, what we have learned is most critical is how you build the model to allow a logical and actionable explanation to answer the consumer’s first question: Why was I denied?

Since models involving credit activities directly impact consumers, they are governed by regulatory agencies aimed at protecting consumers, such as the Consumer Financial Protection Bureau (CFPB). These agencies enforce laws requiring all credit models—regardless of the technology used to build them—follow a basic set of standard rules. This includes providing consumers with a list of reasons why they were denied credit so they can take steps to remedy the issue(s) and improve their credit.

This was the problem with using neural networks for credit risk, until recently. Due to their intense complexity, people assumed neural networks were impenetrable “black boxes” the inner workings of which were impossible to explain. Thus, these models could not produce reason codes required to support their use. Data scientists in the Equifax Data Science Lab learned how to constrain these models so they provide the required reason codes and can now be used for credit risk decisioning. The reason codes that emerge from these models are always logical and actionable for the consumer. This fundamental achievement is the foundation of what is now known as our patented NeuroDecision® Technology (NDT).

VIDEO: See how NeuroDecision® Technology works: Introduction to NeuroDecision.

To Constrain or Not Constrain Neural Networks

The ability to answer the consumer asking “Why was I denied credit?” implies the ability to evaluate the data in it and tell the person how that data impacts their score. Neural networks are much more complex models than the old industry standard logistic regression, and each datum impacts a score in numerous ways. Our NDT scores can be completely decomposed into the impact of any attribute in the model in terms of its impact on a score.

Given our breakthrough in using neural network models for credit risk, scientists in the Equifax Data Science Lab decided to compare the performance of a constrained neural network model built using NDT versus an unconstrained neural network model built using the same predictive attributes. The two models illustrate differences in their predictions, stability, compliance, and ability to produce logical and actionable reason codes as model explanations.

In the NDT model the impact of each attribute in the final model is constrained to reproduce the directional impact of each attribute considered one-at-a-time. Conversely, in the unconstrained model the requirement is relaxed and it was allowed to fit the development data as closely as possible.

At the end of the study, it was clear that the models—despite both using neural networks and the same predictor attributes—produced very different predictions about specific consumers that did not agree. In fact, the two models generated different reason codes in 99 percent of cases.

Careful examination showed the constrained NDT model consistently produced logical and actionable reason codes that consumers could use over time to improve their credit score. However, the reason codes produced by the unconstrained model were nonsensical in many cases, and we illustrate exactly how these results occur. The worst-case scenario is that a consumer could react logically to reason codes provided by an unconstrained model, and the score may get worse.

Without the constraints imposed by NDT, the unconstrained model overfits the development data. As a result, it produced illogical reason codes that could not—and should not—be acted upon by consumers.

Future Implications for Consumers and Lenders

The findings of this comparison study have important implications for consumers and lenders. Neural network credit risk models are putting financial services within reach for many consumers who have long been overlooked or denied credit in the past. This is clearly good for consumers, but it’s also good for lender portfolios, as it helps grow their customer base without increasing their risk levels for delinquencies and write-offs.

However, it’s important to distinguish between neural network technologies. Given the erroneous, illogical reason codes produced by unconstrained neural network models, these models should never be used to determine credit risk. The primary reason is that reason codes emerging from unconstrained models can mislead consumers. This can negatively impact their credit standing.

On the other hand, constrained neural networks such as NDT are optimized for credit risk. They are uniquely capable of producing highly predictive outcomes and complete model explanations in the form of reason codes that empower consumers. These reasons codes can affect a consumer’s credit behavior to help the consumer achieve the best financial life.

Mike McBurnett

Mike McBurnett

Distinguished Scientist

Michael McBurnett has 30 years of experience building, deploying and monetizing mathematical models of human behavior in the credit risk, counterinsurgency warfare, intelligence, banking, direct marketing, political and academic arenas. He attended Washington University in St. Louis where he earned his M.A., and Ph.D.[...]