Credit assessment — determining whether to lend money to a borrower, and at what interest rate — is one of the oldest and most consequential applications of quantitative analysis. For decades, banks have used statistical scorecards based on credit bureau data: payment history, outstanding debt, credit utilisation, and account age. AI is disrupting this model, expanding both the data sources available for assessment and the sophistication of the modelling approaches, while raising important questions about fairness and regulatory compliance.
Limitations of Traditional Credit Scoring
Traditional credit scores, such as FICO in the US and the equivalent bureau scores in the UK, serve their purpose well for borrowers with established credit histories. However, they systematically exclude or disadvantage large segments of the population: young people applying for their first loan, recent immigrants, individuals who have paid exclusively in cash, and those recovering from past financial difficulties. An estimated 1.7 billion adults globally are "credit invisible" — lacking the credit history required for a conventional assessment.
For included borrowers, traditional scores use a limited feature set that may not capture the full picture of creditworthiness. A borrower with good payment history but deteriorating recent cash flow — observable from bank transaction data — may receive a credit offer that does not reflect current ability to repay. Classical models also assume linear, additive relationships between features and default probability, potentially missing complex interactions that ML models can detect.
Machine Learning Models in Credit Underwriting
Gradient-boosted decision trees — XGBoost and LightGBM in particular — have become dominant in credit underwriting for their accuracy on tabular financial data, built-in handling of missing values, and relatively straightforward interpretability through feature importance and SHAP values. Compared to logistic regression scorecards, ML models typically achieve 5–15% reduction in credit losses at equal approval rates, or equivalent loss rates at substantially higher approval rates.
Deep neural networks are used for processing non-traditional data sources: raw transaction sequences, document images (bank statements, payslips), and text data. LSTM models applied to sequences of bank transactions can identify spending and income patterns that predict creditworthiness independently of bureau data — enabling assessment of thin-file borrowers who would otherwise be excluded.
Alternative Data Sources
The most transformative development in AI credit assessment is the incorporation of alternative data — information beyond the traditional credit bureau file. Mobile phone usage patterns (frequency of use, social network characteristics, geographic mobility) have been used successfully for credit assessment in developing markets where bureau coverage is limited. Rental payment history — not traditionally reported to credit bureaus — is a strong predictor of creditworthiness and is now being incorporated into assessment models by several UK lenders.
Open banking data, enabled by PSD2 regulation in Europe and similar initiatives in the UK, provides lenders with consented access to applicants' transactional bank account data. This rich data source — typically months of day-by-day income and expenditure — allows far more granular assessment of financial health than static bureau data. Income verification, affordability assessment, and detection of financial stress signals all improve substantially with transactional data.
Regulatory Landscape
Credit risk models are among the most heavily regulated AI applications. In the UK, lenders must comply with the Consumer Duty, GDPR's right to explanation, and the Equality Act's prohibition on indirect discrimination. The Financial Conduct Authority (FCA) expects firms to be able to explain credit decisions to customers and to demonstrate that models do not produce discriminatory outcomes for protected groups.
Model risk management requirements — derived from guidance such as SS1/23 and the EBA's guidelines on internal governance — mandate independent model validation, robust backtesting, and ongoing performance monitoring. The increasing use of ML models in credit decisions has prompted the FCA and other regulators to develop more detailed guidance on validation expectations for algorithmic underwriting, reflecting the unique challenges that ML interpretability poses in a regulatory context.