UK High Street Bank

Credit scoring has been a cornerstone of UK retail banking for over 25 years, but the model has a fundamental flaw: the same rule-based criteria are applied uniformly to every applicant regardless of their individual financial behaviour. For a bank serving approximately 20 million customers, this blunt instrument produces two costly outcomes simultaneously — creditworthy borrowers are denied loans, shrinking the addressable lending market, while genuinely risky customers who happen to score well slip through, creating undetected default exposure. The bank recognised that its legacy scoring system was leaving significant value on the table in both directions, but lacked a method to identify the behavioural signals its linear models could not see.

The bank engaged Kortical to test whether machine learning could outperform traditional credit scoring on default prediction. To establish credibility before accessing proprietary data, Kortical first demonstrated a market-leading solution by topping a public Kaggle credit scoring competition against 924 teams. With access secured, Kortical's AutoML platform ingested over 220 million rows of anonymised transaction history and CRM data drawn from a representative sample of the bank's customer base. The platform automated data cleaning, transformation, and feature engineering while training thousands of ML models in parallel. Critically, the platform's explainability layer was used not just to validate the model but to surface distinct customer segments with fundamentally different behavioural risk drivers — a level of granularity that rule-based systems cannot reach. The model was validated and production-ready within four weeks.

The ML model delivered results that exceeded expectations across both dimensions of the credit risk problem:

• 83% of bad debt missed by traditional credit scoring was correctly identified, substantially reducing the bank's default exposure at identical loan approval rates
• 77% more customers could be offered loans if the bank chose to maintain its existing default rate rather than reduce risk
• 4 weeks from data ingestion to a production-ready model

The explainability analysis also produced actionable segmentation insights: for customers aged 45–65, late-night spending (midnight to 6am) was a significant default predictor, while unarranged overdraft use — a strong signal for younger customers — carried no predictive weight for this cohort. The top 15 risk drivers were entirely different across age segments.

• Segment-aware modelling is not optional in lending: uniform scoring logic systematically misreads risk for distinct demographic groups; ML's ability to isolate segment-specific drivers is a structural advantage, not an incremental one.
• The credit risk trade-off is a false constraint: the same model can simultaneously reduce default exposure and expand the eligible borrower pool — the right framing is risk-adjusted growth, not risk versus volume.
• Data access is the real bottleneck: proving model quality on public benchmarks before requesting proprietary data was the unlock; organisations should anticipate this negotiation and prepare evidence in advance.
• 220M+ rows can be operationalised in weeks with the right AutoML tooling: the constraint is rarely data volume, it is data preparation and iteration speed.