The Next Evolution of Machine Learning

Few can predict what the economic fallout of the COVID-19 crisis will look like long-term, but even before this pandemic, there was a growing need for lenders to get better at predicting credit risk.  With a greater influx of data and rapid changes in consumer behaviour, traditional risk scoring models like linear regression, are often not up to the task of predicting default risk for complex applications involving multiple data sources and relationships.

Newer models of machine learning, such as neural networks, tend to be significantly better at estimating the probability of default, but the results can be difficult to interpret and explain.

The fundamental challenge has been for machine learning to generate not only better scores but also explain the basis for an outcome. This challenge is now met with NeuroDecision^® Technology. Equifax has recently obtained the US patent for this new machine learning model, which combines the speed and accuracy of a neural network with the transparency and explainability of traditional linear models. The result is a unique credit risk model that helps lenders understand how credit scoring decisions are determined, so they can stay compliant and make informed decisions when assessing credit risk.

Unlike many neural networks, the Equifax model has been designed to be seamlessly deployed in the lender’s existing environment, providing opportunities for greater automation and faster decision making. This can not only reduce costs, but provide a more predictive credit scorecard that allows lenders to maintain business growth and reduce credit risk.

Recently, a leading bank in the Asia Pacific region approached Equifax to see if our machine learning model was better at predicting credit risk than their own internal scorecard, while also being easy to explain. The bank used the Proof of Concept to determine its level of investment in machine learning technology. 

Equifax curated a global team of analytics, modelling and project experts to assess the client’s mortgage application scorecard and analyse 12 months of data and behavioural variables. We also applied our bureau data to understand the performance of declined applications, helping the lender improve the accuracy of predicting the likelihood of customers defaulting on payments, or becoming a credit-worthy customer in the future.

Combining our bureau data with the neural network model provided the bank with the ability to monitor and improve scorecard performance over time. Specifically, results showed that implementing the Proof of Concept would allow the bank to:

• either reduce declines by 15% or bad debt by 4.1%, for the same debt profile
• increase accepted credit applications by 6.6%
• identified 22% more high-risk applicants in the lowest 5% of scores than the benchmark
• enjoy significant gains in sub-segments, including high LVR (26%) and rural (13%) customers.

Accuracy, transparency and explainability are vital qualities for credit assessment models as lenders navigate the precarious path left in the wake of COVID-19 and the Banking Royal Commission. As the need to combine data residing in different sources increases, so too will credit models have to handle this complexity with performance and reliability.  The use of new machine learning models like NeuroDecision^® Technology is likely to deliver much-needed operational and financial resilience, providing a key differentiator between those who have the technology, and those that do not.

Equifax is using NeuroDecision^® Technology in custom and configurable solutions we are building for our customers right now using our Ignite^® platform.

See the power of Equifax Ignite^® for yourself, book a demo.

Request a demo

Disclaimer:  The information, materials and opinions contained on this article are for general information purposes only, are not intended to constitute legal, financial or other professional advice and should not be relied on or treated as a substitute for specific advice relevant to particular circumstances. We make no warranties, representations or undertakings about NeuroDecision^® Technology (including, without limitation, any as to the quality, accuracy, completeness or fitness for any particular purpose of that product in the context of your requirements or needs).  The performance of NeuroDecision^® Technology in the referenced case study is not necessarily a reliable indicator of future results in a different context.