Best Paper Winner: Can Credit Scoring Be More Accurate, Adaptive, and Safer for High-Risk Lending?

What is it about?

A good online credit-scoring model should not only predict default accurately—it should minimize costly false negatives and adapt when borrower behaviour changes. Online lenders and financial institutions increasingly use machine learning to assess whether borrowers are likely to default. However, in high-risk lending, a model that incorrectly classifies a risky borrower as safe can be especially costly. This study proposes a new non-parametric credit-scoring approach based on the Kruskal–Wallis statistic. The method automatically selects and weights predictive features, works with both small and large datasets, and is designed for continuous online credit assessment. It also includes a warning mechanism for detecting data drift and triggering model renewal when borrower behaviour changes. The results show a very low false-negative rate, lower computational cost, and approximately 18% improvement in Recall/Sensitivity, making the approach particularly relevant for unstable or rapidly changing environments.

Featured Image

Photo by Markus Winkler on Unsplash

Photo by Markus Winkler on Unsplash

Why is it important?

Credit-scoring models are often judged by overall accuracy, but for lenders the type of error matters. Failing to identify a genuinely risky borrower can create greater financial losses than incorrectly rejecting a low-risk applicant. This research introduces a more conservative credit-scoring strategy that prioritizes the detection of high-risk customers while remaining computationally efficient. It can also adapt to behavioural change by monitoring data drift and automatically reselecting relevant features. This makes the approach useful for online lending, open banking, and periods of economic instability, when borrower behaviour and risk patterns may change quickly.

Perspectives