Will Machine Learning Models Replace Traditional Credit Risk Models As We Know Them?

Blog Post02/26/2018
Image of Machine Learning Models Replace Traditional Credit Risk Models As We Know Them

Virtual assistants. Self-driving cars. Cancer detection. Machine learning is behind some of the most ground-breaking and potentially life-changing technology around today. For retail credit lenders, machine learning technology presents an opportunity to do more with data than ever before. Yet, so far, there has been little change in the way they build risk assessment models. As fintechs are likely to adopt machine learning models to boost innovation and competitiveness, banks, with the help of regulators, should take a closer look at implementing this technology if they wish to remain relevant. We explore some of the benefits of machine learning technology and the possibilities it enables.

New data needs new analytics

When dealing with high volumes of credit applications, most retail lending institutions rely on logistic regression models to predict the likelihood of a consumer’s delinquency in the next few years.

The output of these models is the credit score that many retail lenders use when making lending decisions. The benefit of these models is that their results—and any decisions based on those results—are easy to explain. That’s because these models clearly show the relationships between consumer behaviours and the resultant credit scores.

But the linear nature of these models is also their downfall. The challenge facing lenders now is being able to analyze a growing volume and variety of data, and traditional analytics methods can’t keep up with this demand.

To uncover new insights, we need new ways of looking at the data. Models that use machine learning can provide the detail and deep understanding needed to improve credit decisions. (Loosely defined, machine learning is a set of computer science and statistical tools and techniques that are used to make predictions.)

More data, more variables, a more robust credit score

Models for assessing risk that are based on machine learning offer several advantages over those that use human judgement or traditional statistical models. One of the main benefits is their ability to run across large volumes of data to predict an outcome. But this is not enough to produce insights. The value of machine learning models lies in their relative lack of limitations.

Most machine learning models are not linearly constrained. This means they are able to look at different combinations of variables to determine the interactions between input variables. They can incorporate large volumes of data to provide a granular view of consumers and reveal previously hidden information about a consumer’s collateral, capacity and character. Some models don’t even need a target variable (called unsupervised learning), so they can discover previously unknown features of certain populations.

The result is a detailed representation of interactions between variables that gives organizations a more discerning view of consumers than credit risk models currently can.

Machine learning models can enhance existing credit scoring tools

Despite these many advantages, machine learning models have not seen wide adoption among retail credit institutions. The most likely reason for this is that it’s difficult to explain how a specific model input results in a particular decision or output.

In many jurisdictions, regulators compel lenders to clearly articulate their decline decisions. Linear models allow lenders and regulators to validate lending decisions with data that clearly maps to the final risk model outcome. With machine learning models, consumers who don’t qualify for credit based on traditional scoring models may prove to be eligible, but the underlying reasoning may not be as apparent. Also, if the models are not derived correctly, they could suffer from overfitting: working so well with the existing data that they don’t perform well with new data.

While these issues may be viewed as something of a drawback, there is still significant merit in using machine learning to enhance existing assessment and decisioning tools.

The benefits are worth the effort

The adoption of any new technology comes with a natural amount of resistance: a few years ago, touch-screen phones seemed to be impractical and difficult to use. Today, just about all mobile technology works off touch-screen devices.

The development of machine learning technology is an ongoing process of improvement, aimed at making life easier for consumers and lenders alike. As machine learning becomes more sophisticated, it will address more of the concerns lenders and regulators may have and enhance the services they are able to offer. Further advances will see us building machine learning models that can test and explain the results of other models, making it easier to substantiate lending decisions.

Machine learning results might take more effort to explain, but the benefits are worth the effort. A machine learning enabled model can help credit providers improve their credit scoring, acquisition, account management and collections strategies.

The sooner retail lenders start applying machine learning technology, the sooner they will reap the benefits of a technology that will inevitably become the accepted norm—and help them stay relevant in an increasingly competitive and automated industry. In short: yes, machine learning models will replace credit risk models, but it will take time for this to happen.

Legal disclaimer

Contact us

* Required field