Traditional credit scoring systematically disadvantages people new to credit (students, recent immigrants, young adults) and those whose credit history was disrupted. A UK challenger bank targeting underserved segments discovered that adding alternative data to their LightGBM scoring model improved AUC from 0.81 to 0.94, reduced non-performing loans by 34%, and enabled automated approval for 71% of applications previously requiring manual review.
Why Traditional Credit Data Fails Thin-File Applicants
Roughly 5.8 million UK adults have thin or no credit files (Experian, 2024). The alternative data thesis: creditworthiness correlates with behaviours leaving digital traces outside the traditional credit system.
Key alternative data sources: (1) Open Banking transaction data -- income regularity, spending patterns, balance volatility, overdraft frequency; (2) Rental payment history -- paying rent on time monthly is arguably stronger than any credit card signal, yet invisible to bureaux; (3) Utility payment patterns -- regular on-time payment of energy, broadband, and mobile is highly predictive; (4) Payroll patterns in Open Banking confirming income reliability more accurately than a PDF payslip.
All require explicit customer consent under GDPR and UK FSMA. Open Banking requires authorisation through a regulated AISP. Consent architecture must be built before data can be used.
Feature Engineering on Open Banking Data
Raw transactions are aggregated into signals over 30, 60, and 90-day windows. Income signals: count and regularity of income credits, coefficient of variation, declared vs actual income ratio. Balance behaviour: average end-of-day balance, standard deviation, days-in-month with negative balance, overdraft frequency and depth. Spending discipline: essential vs discretionary ratio, gambling transaction presence (risk signal), subscription commitment level. Payment discipline: recurring outgoing payment count, missed recurring payments, days-late on identified recurring payments. Stress indicators: emergency cash withdrawals, payday loan transactions, balance trajectory direction.
Combined with bureau data and application data, the full feature set reaches 180+ variables per applicant.
LightGBM + SHAP Architecture
LightGBM handles missing values natively (critical since not all applicants have all data sources), runs fast on high-dimensional data, and integrates cleanly with SHAP.
Training setup: 18 months of applications with 12-month performance labels (default = 90+ days past due). Class imbalance: SMOTE oversampling plus LightGBM scale_pos_weight. Cross-validation: 5-fold temporal split preventing data leakage. Hyperparameter tuning: Optuna Bayesian optimisation. Final model AUC: 0.94 on out-of-time holdout (3 most recent months).
SHAP is integrated at inference time. Every application produces both a credit score and a SHAP vector. When declined or approved at a premium tier, the top 3 SHAP drivers generate the customer-facing Consumer Duty explanation.
Results and the Fairness Outcome
After 12 months: AUC 0.81 to 0.94 (+16 percentage points). Non-performing loans (90+ DPD at 12 months): -34%. Automated approval rate: 41% to 71%. Average time to decision: 48 hours to 4 minutes. Approval rate for thin-file applicants: 12% to 38%.
The fairness monitoring outcome is particularly important: the alternative data model showed significantly less demographic disparity. The bureau-only model had a 31 percentage point gap between most- and least-approved demographic postcodes; the alternative data model reduced this to 14 percentage points. SHAP analysis confirmed the improvement came from Open Banking income signals -- more democratically distributed than historical credit behaviour.
For Ukrainian FinTech firms expanding to the UK, the advantage is structural: Ukrainian applicants often have no UK credit history but strong digital data trails -- exactly where alternative data outperforms bureaux.
Related Codeworth Services
Need Help with Implementation?
The Codeworth team is ready to implement these ideas in your project. Free consultation.
Discuss the ProjectCo-founder of Codeworth with 7 years of UK fintech ML experience. Specialises in credit scoring with FCA compliance, ensemble methods including XGBoost and LightGBM, and ML experiment management via Python and MLflow. Has built production systems for credit decisioning processing millions of applications annually under FCA oversight. Holds deep knowledge of FCA model explainability requirements and Consumer Duty documentation standards.