Held between 1 and 3 July at Celtic Manor near Newport, this year’s Credit Week put credit modelling centre stage. Jaywing were asked to take part in the headline panel, The Future of Credit Risk Modelling: Embracing AI and Machine Learning, given the role we have played in pioneering these approaches in the credit risk marketplace.
With interest in machine learning and AI at an all-time high, but barriers to adoption still real for many lenders, the panel struck a practical tone. Here, we capture the standout themes from the event, client response, and what risk leaders should be watching closely over the coming year.
The future of credit risk modelling: 5 key takeaways from the panel
Chaired by Stephen Holmes (Novuna), the panel brought together leaders from Evlo, Jaja Finance, Equifax and Jaywing to discuss where credit risk modelling is headed and what’s getting in the way.
Here’s what we took from it:
1. AI and machine learning are in use, sometimes in combination with traditional models
Panel members shared examples of practical deployment. At Jaja Finance, machine learning techniques are integrated into traditional scorecard models to increase predictive power without compromising transparency. Our own experience at Jaywing was welcomed by the panel, where long-run data from 30 implementations showed that true non-linear models (neural networks and gradient boosted methods) had been proven to deliver 15-20% reduction in bad debt for the same accept rate over traditional linear methods.
2. Progress is slower for larger banks due to regulatory hurdles
While challenger lenders are making faster use of new techniques, larger Advanced IRB firms face a more complex regulatory environment. The panel acknowledged this gap and linked it to resource and knowledge constraints at the regulator. There was a shared interest in whether future regulatory guidance might facilitate greater adoption of AI and ML models within the IRB framework to unlock benefits for the lending industry and consumers themselves.
3. Increased governance is expected and worthwhile
There was broad agreement that governance requirements should increase in line with model complexity. Each panellist had a slightly different perspective, but all supported the principle that additional oversight is necessary and worthwhile for what it enables. Stronger governance was not seen as a barrier, but rather as a prerequisite for progress.
4. Ethical considerations are front of mind
Bias, explainability and control were raised as key themes. There was no appetite for pushing boundaries without clarity on how those models would be governed. Transparency remains a non-negotiable requirement. However, there was also recognition that bias is not unique to AI/ML models, but rather an important consideration independent of methodology.
Where the conversation led next: Governance
The panel made a strong case for the careful and safe use of machine learning in credit risk modelling. There was alignment that these methods can deliver clear benefits, such as enabling more people to access credit, and maintaining low bad debt levels across large portfolios. However, that potential depends on how the technology is utilised and whether it maintains broad acceptance among customers, regulators, and the lending industry.
On implementation and governance, the views were consistent. While each panellist had their own approach, there was a consensus that as models and data become more complex, governance must be strengthened in parallel. The additional oversight was not seen as a trade-off; instead, it was considered necessary for the wider benefits it brings to the market.
Looking ahead: Where change could come next
Panellists were asked to name the biggest opportunity ahead. Self-updating models, also known as self-improving or self-evolving models, are suggested as the next big opportunity and are something that is currently possible, potentially significantly improving responsiveness and accuracy. These approaches are already being explored, but the regulatory environment is unlikely to be ready for them in the credit risk space soon.
These takeaways reflect a clear appetite for safe, practical progress. Judging by the reaction in the room, it’s a direction many are ready to take.
Credit Awards 2025: Recognition and market signals
Later that evening, attention turned to the Credit Awards. Jaywing sponsored the Alternative Lender of the Year category. This is an area where we’ve done extensive work, particularly in partnership with Fair4All Finance.
Since last year’s fraud success story, our proposition has developed further. We’ve added more experience, refined our approaches, and are now seeing strong interest in our fraud health checks, especially from lenders who want a second opinion before adjusting systems or processes. Fraud often sits unseen in parts of the process that haven’t been reviewed in years. An external check can make all the difference.
This ties closely to the work we’re doing with non-linear models that can help avoid blunt decisioning, offering a fairer and more accurate way to serve the whole market. Following several studies with traditional and challenger banks we have added Gradient Boosted Methods (GBMs) to Archetype to give our clients a one-stop shop to explore the two leading approaches to credit risk modelling, GBMs and Neural Networks (NNs).
Closing thoughts from Credit Week
Credit Week 2025 showed how regulatory and governance requirements have historically constrained the use of machine learning models in credit risk, but also that the appetite was now there to fully explore the transformative capabilities that Jaywing has helped prove beyond question over 7 years of implementations.
For lenders large and small, there is plenty now within reach. Model performance can be significantly improved, without compromising explainability, and with proper governance and controls in place. Reach out now to see how Jaywing can help you on a path to generate with domain-specific AI modelling expertise grounded in 7 years of real-life data.
Learn more about Archetype.