Explainable AI for Credit Scoring

Traditional logistic regression credit models have done this job well, providing high levels of transparency and easy-to-explain results. Yet as the financial services sector grapples with managing credit risk through uncertain times, lenders are looking for more stable models, using a variety of data to get more accurate lending decisions whilst maintaining the just as important explainability. And this is where artificial intelligence (AI) and machine learning techniques like neural networks have stepped in, delivering superior performance and prediction compared to logistic regression models. 

However, the lack of understanding of AI neural networks has hampered their widespread adoption in risk analysis. The black-box nature of machine learning algorithms means that outcomes are difficult to interpret and explain. Being unable to articulate the reason for a credit decision is less than ideal, especially during times of disruption. Robust risk analysis should not only accurately assess risk in a relevant population, but also have the capability to provide logical and actionable explanations. 

Clear, transparent and highly predictive credit decisions are crucial as financial institutions face changing consumer behaviour and increasingly complex open banking data relationships.

Best practice risk modelling

The launch of the new generation of credit score – Equifax One Score - shows that it is possible to deliver a highly predictable, explainable score by taking the transparency associated with conventional models and combining this with high-performing AI neural networks. 

This new generation credit score provides financial institutions with an enhanced ability to approve more consumers for credit without taking on additional risk, as well as better explain their lending decisions to customers. This transparency into the factors that contributed to an individual’s score supports a better customer experience by providing an incentivised path to improved financial health.

Powered by a neural network and designed to reflect COVID-19 induced changes to consumer credit behaviour, the predictive power of the new Equifax One Score remains robust under coronavirus-simulated testing conditionsⁱ. Unlike logistic regression models, neural networks have the unique ability to automatically find pockets of non-linear data and interaction relationships – without the need to use manual techniques like segmentation or added interactions to optimise for each population.

Within a population of consumers who have adverse payment data in their bureau file, for example, there will be a range of variables for why their bills went unpaid. Perhaps they were going through a rough patch, and their financial circumstances have now improved? Maybe the defaults were from years ago? So rather than drawing a linear correlation between adverse data and high credit risk, the model automatically segments and subsets the data, acknowledging that within this ‘adverse’ population, there may be a proportion of consumers who are now creditworthy. 

Using this example, with the new Equifax One Score, you are likely to see greater spread in the score distribution. Rather than the entire ‘adverse’ population receiving a low score, the model will find cohorts of better risk for a more in-depth and accurate prediction of creditworthiness. With elevated default and hardships rates expected in the coming year, this deeper understanding of prospects and customers will be invaluable to more accurately predicting opportunity and risk.

Explainability and transparency

One of the greatest limitations of AI systems is a lack of visibility around how decisions are made. In contrast, explainable AI (xAI) seeks to produce transparent explanations of how conclusions are reached.  

A key feature of the new generation Equifax One Score is an explainable AI (xAI) modelling technique known as NeuroDecision™ Technology (NDT). Explainability goals are built-in at the design stage, enabling the development of a high-performing, explainable neuro-network model. This groundbreaking explainable AI patent in the credit industry is among the first machine learning scoring methodologies to provide explainable reason codes to consumers. 

Unlike other xAI techniques that rely on proxy or ‘average’ explanations, the explanations (reason codes) provided by Equifax One Score are personalised to an individual consumer. NDT optimally constrains the neural network to be monotonic, so the process of generating the score and extracting the rationale for why a consumer would be denied credit comes from exactly the same model. This monotonicity allows the NDT algorithm to reward positive behaviour (score increases) and penalise negative behaviour (score decreases). When a borrower re-pays debt every month on time, NDT identifies the behaviour as positive and increases the score. 

The reward of this tool of explainability is a common, standard appreciation for what drives credit risk. With a greater connect between risk managers, customers and regulators into why an individual scores a certain way, meaningful progress can be made towards championing the behaviours that improve financial wellness for all. 

Equifax One Score gives lenders a step up in risk prediction. To find out more about how it delivers a more complete and consistent picture of a customer’s credit risk, email a One Score Sales Specialist or we recommend reading: The Next Big Thing in Credit Scores.

i The performance of Equifax One Score compared to our previous generation score was analysed under the COVID-19 scenario that a borrower had lost 30% of their repayment history from deferrals and had no recent defaults loaded to their credit report.