Can Machine Learning Improve Consumer Lending? We Think So.

Bill Byrnes, Managing Director Risk and Compliance
Suresh Baral, Managing Director Advanced Analytics

The ready availability of large volumes of internal, external and social media data, along with advances in analytics and the advent of machine learning (ML), appear to have created the perfect opportunity for improving consumer lending decisions. But how big of a role will – or should – machine learning play in lending decisions? Can it make the processes that financial institutions use to ensure they are meeting both the stringent regulatory requirements and the ever-changing consumer demands more efficient, while reducing credit risk? These questions were among the topics discussed during a recent webinar conducted by Protiviti’s advanced analytics practice leaders and attended by more than 200 people.

During the webinar session, we asked the attendees to weigh in on whether machine learning is currently used in their consumer lending business. Just 10.9 percent said yes, while half indicated they did not know whether their organization is using machine learning. This, along with our conversations with industry participants, suggests to us that many lenders may not be fully aware of the potential of machine learning as a viable tool for optimizing lending decisions – but that they should be.

Where Can Machine Learning Add Value?

Generally, machine learning refers to the ability of computers  to “learn” (i.e., progressively improve performance on a specific task) with data, without being explicitly programmed. Machine learning can detect patterns or apply known rules to predict outcomes, detect anomalies and yield insights.

In lending decisions, machine learning can produce tremendous value when applied to complex nonlinear problems where there is a large amount of data, particularly unstructured data. As an organization’s lending processes grow to incorporate new ways of capturing data (from online applications and social media vs. paper application forms, for example), the use cases for incorporating machine learning continue to grow in number. Those use cases include:

  • Credit risk: Clustering algorithms can help an organization understand how a consumer’s credit history and transaction volume/profile has changed over time and can help the lending institution make better credit decisions for both new and existing customers. Machine learning can be used to enhance feature selection to improve credit scoring.
  • Fraud detection: Pattern recognition algorithms like Long Short-Term Memory (LSTM) can identify a specific sequence of transactions within highly unorganized data that could signal fraud or money laundering.
  • Churn protection: Deep learning algorithms like Recurrent Neural Networks (RNN) and LSTM can help businesses understand how transaction volumes and profiles have changed over time and also identify which customers may be about to leave, allowing the business to take steps to retain customers.
  • Marketing: Association-based algorithms can be used to identify consumer spending patterns to target consumers for advertising campaigns. These targeted marketing and sales efforts help financial institutions offer customers the services they need, rather than a “one size fits all” approach that barrages customers with every possible loan service.

Transparency of Credit Decisions Still a Requirement

Banks normally use transaction, payment history and credit bureau data to build credit scores to make lending decisions. Machine learning, which can handle a wider variety of data types in large volumes and analyze it quickly, makes it possible to add other types of data to the mix for decision making – data from social media activity, mobile phone usage and other unstructured data types. This can produce faster and better-informed credit decisions. Machine learning models have the potential both to lower the cost of assessing credit risks for some prospects and increase the number of prospects for whom banks can measure credit risk.

However, one limitation of machine learning is the lack of transparency and explainability of the credit decision. Most credit decisions require compliance with the Fair Credit Reporting Act and Equal Credit Opportunity Act, which require an explanation of negative actions (adverse action codes). This requirement for transparency and explainability has been an impediment to the wider use of machine learning in consumer lending.

In addition to complying with technical disclosure requirements, lenders must review AI/ML models to ensure they do not inadvertently increase exposure to fair lending claims by providing less favorable treatment to protected classes.

Many organizations are working on techniques to improve the explainability of decisions reached through machine learning models. In the meantime, it is possible to combine machine learning and traditional scorecard approaches to achieve better performance and explainability. Machine learning can be used to identify newer attributes (not picked up by traditional scoring models), which, when combined with a scorecard model, can translate into increased transparency while improving prediction.


Machine learning can improve the accuracy of all predictive models used in consumer lending, such as a customer’s likelihood of responding, approval, delinquency, default and other behaviors. Better models are a critical factor in improving decision-making, enabling lenders to maximize revenue while holding credit losses below a specified threshold and thereby optimize the net income line, which is what credit management is all about. As the role of consumer lending in attracting new customers and building long-term consumer loyalty continues to grow, machine learning stands to significantly improve the meeting of increasingly aggressive revenue targets while reducing lenders’ credit risk.

To learn more, listen to our recorded webinar.

















1 comment