While machine learning and artificial intelligence aren’t yet widely used in credit scoring and approvals, they’re gaining a foothold in other areas of the retail credit lifecycle. Tom Galimanas, Senior Consultant at TransUnion, explores how these methods are evolving in the marketplace, helping lenders serve consumers better.
The benefit of machine learning algorithms is they ingest a large amount of data — giving these models an advantage over traditional models and helping lenders:
While algorithms may have clear inputs and outputs, their inner workings aren’t quite as straightforward. This makes it difficult to record the reasons behind scores and adjudication, and comply with regulations for credit scoring and approvals. It also explains why machine learning models don’t feature strongly in these stages of the credit lifecycle. However, other stages of the credit lifecycle — document validation, fraud detection, customer retention and problem resolution — may offer opportunities to benefit from machine learning.
Verify documents in minutes
A time-consuming and often expensive step of the credit-granting process is ensuring documentation is legitimate and belongs to the applicant. This frequently takes place as a manual, back-office function. For example, if an applicant needs to verify their income, it can take up to three days for a lender to confirm documents are genuine and haven’t been altered.
Recent advances in deep neural networks (a type of machine learning algorithm) research have improved image recognition and classificationi. This type of model allows a lender to process submitted documentation much more quickly, consistently and accurately. Using machine learning, documents can be reviewed and approved in minutes as opposed to days, speeding up the application process and improving the overall customer experience.
Measure transactions against a baseline to identify potential fraud
Fraud is a serious issue at the point of application and point of payment. Fraudsters often apply for credit with no intention of ever repaying. Fraudulent credit card transactions lead to large, system-wide losses from chargebacks, and sometimes systems can’t distinguish between legitimate and fraudulent transactions.
With the proliferation of digital channels, knowing your customers has become more important than ever. Machine learning algorithms use characteristics generated in real-time to establish a baseline against which individual transactions are measured to identify potential fraud. These characteristics can include how long it takes an applicant to read the terms of the loan, and addresses or phone numbers used to perpetrate fraud in the past. Potentially fraudulent applications can be set aside for review, which can contribute to large savings.
According to research by Javelin Strategy, cardholders in the U.S. have $118 billion annually from legitimate credit card transactions wrongly rejected due to suspected fraudii. The dynamic, non-linear nature of machine learning algorithms can improve the classification of transactions. Complex patterns based on historical spending and usage patterns are established as baselines against which future transactions are compared and classified. Improved classification may lead to happier customers and greater revenue.
Identifying at-risk consumers
Machine learning models can be used to identify consumers at risk of moving to another lender, and pre-approved offers can be quickly delivered to retain their business.
Using chatbots to ease problem resolution
Problem resolution is often handled by a representative of the lender. Chatbots based on natural language processing combined with machine learning present an avenue for a more cost-effective yet still personalized experience. They may also integrate easily with existing hardware — for example, a consumer can ask Alexa to transfer money between two accounts or pay a bill without having to open a lender’s web page or app.
Of course, there are many things to be mindful of when rolling out machine learning solutions. Lenders must be able to detect and remove biases in their models, particularly where there’s potential to penalize groups or individuals. And, as much as possible, they need to cost-effectively combine machine learning with existing processes and infrastructure. Those that can implement algorithms across the credit lifecycle may be in a stronger position to succeed in the market.
iSee, for example, https://towardsdatascience.com/deep-learning-for-image-classification-why-its-challenging-where-we-ve-been-and-what-s-next-93b56948fcef
iiSource: Javelin Strategy & Research, 2015