The role of data in credit modelling

The hottest topic in credit risk modelling often centres on which algorithm to use. The choice of algorithm matters, but it isn’t where the biggest improvements in performance come from. Logistic and linear regression are valued for their simplicity and transparency. More sophisticated machine learning models, like random forests and gradient boosting machines, hold more predictive power but are harder to interpret. Our guide to machine learning in credit risk modelling weighs that trade-off model by model.

The real progress is in the data

The biggest improvements in credit model performance come from the data itself: how it’s used, and how different sources are brought together to reflect real-world behaviour.

Lenders now have access to far more than traditional application and credit bureau data. Open banking, transactional patterns, behavioural trends, internal customer relationship management (CRM) systems, and even unstructured notes from customer conversations all hold valuable signals.

The next challenge isn’t collecting more data. It’s integrating what we already have.

The data scientists making the biggest impact aren’t just tuning models. They’re connecting messy datasets, aligning definitions, resolving inconsistencies, and engineering features that reflect how customers actually behave. Often that means aggregating transactions into meaningful summaries, building variables that represent business logic, or cleaning records so that models don’t learn from noise. These steps can drive larger gains than tuning an algorithm. One 2010 study put a number on it: joining a commercial bank’s own customer transaction records to credit bureau data improved delinquency forecasting enough to save an estimated 6 to 25% of total losses, on conservative assumptions about the cost of cutting credit lines.¹

The importance of collaboration

One of the most underrated aspects of credit modelling is communication. The best models aren’t built in isolation. They’re made by talking to underwriters, front-line teams, and the people who use them day to day. Those conversations clarify what variables mean in practice, how they’re used in decisions, and what really drives credit risk.

Without that input, it’s easy to build a model that is technically sound but misaligned with the business. Local knowledge is not a nicety here. One study looked at account-level credit card data from six major commercial banks between January 2009 and December 2013. It found that risk factors, their sensitivities and the predictability of delinquency all varied significantly between institutions holding revolving credit card portfolios.² A feature set that works at one lender is a hypothesis at the next. It might rely on unstable or unnecessary data, miss important signals, or use variables in ways that make no sense to decision makers. Good data scientists bridge that gap, making sure features are grounded in real-world understanding and that models genuinely support business needs.

What lenders should do next

It’s easy to chase marginal gains through model tuning or the latest algorithm. The less glamorous work is cleaning the data, enriching it, and engineering better features, and it usually carries the larger lift in discriminatory power.

Integrate open banking data

Traditional bureau data is useful but limited. It gives a slightly delayed view of a customer and often misses key segments, such as thin-file or new borrowers. Open banking data gives a real-time view of income, spending habits, and account activity, and it exists in usable form because regulation made banks expose it. In the UK that was the CMA’s retail banking order and in the EU it was PSD2. The Basel Committee’s May 2024 paper on the digitalisation of finance notes that open banking regimes differ worldwide in whether they are mandatory or voluntary and regulatory led or market driven, and in whether they mandate the use of APIs at all.³ Lenders should prioritise integrating open banking data, and engineer powerful features from it such as income volatility and discretionary spending patterns.

Capture first-party behavioural data

Application forms, call centre notes, CRM data, and product usage logs all provide signals that bureau files can’t. Missed direct debits on other products, the frequency of service interactions, or how quickly applicants respond to document requests can all predict future risk. These sources are often underused, but hold real value when cleaned and engineered properly.

Use AI for conversational data

Call notes, emails, and chat logs often contain early warning signs, such as complaints or income issues, but they’re hard to use at scale. Large language models can now scan and categorise these conversations automatically, flagging risk-relevant events like mentions of financial hardship, broken arrangements, or repeated contact. That turns unstructured communications into structured risk signals a model can learn from, bringing previously untapped data into the credit decision. Reading call notes with a large language model also changes the model’s governance profile.

The Prudential Regulation Authority (PRA) sets out its model risk expectations in SS1/23. Under Principle 1, a firm’s assessment of model complexity may consider the use of unstructured data alongside interpretability, explainability and transparency, and the potential for designer or data bias.⁴ Expect a model that reads call notes to sit higher in the firm’s own risk tiering than one that does not.

Strengthen cross-team communication

Involve underwriters, product managers, and credit policy owners early. If a model uses "number of accounts opened in the last three months", ask what it’s really capturing: fraud, opportunism, or normal customer onboarding? That context shapes whether a feature should be included, excluded, or transformed. The best models reflect not just what is predictive, but what is usable in a decision, and the way to get there is a shared understanding between model owners, data scientists, data engineering, operations, and product.

Prioritise data monitoring

As alternative data grows, so does the risk of drift. Lenders need granular, ongoing monitoring of data inputs: tracking changes in open banking connection success rates, for example, or distribution shifts in derived income features.

The usual measure of a distribution shift is the population stability index. It compares the share of the population falling in each bucket of a feature today against the share that fell there in a reference period, and returns zero when the two are identical. The industry reads it against conventional thresholds: below 0.10 for stability, 0.10 to 0.25 for modest drift, and above 0.25 for drift material enough to investigate.⁵ The index works one feature at a time, so it will not catch the case where every feature looks unchanged on its own but the way they move together has shifted.

Build for explainability

Even complex, high-feature-count models need to be explainable. Use tools like SHAP early in development to spot misleading variables, or features that can’t be easily communicated to the business or the regulator. SHAP, short for SHapley Additive exPlanations, attributes a single prediction to the contribution each feature made to it. The method’s standing rests on a 2017 result: among the ways of explaining a prediction as a sum of feature contributions, the Shapley value is the only one that satisfies local accuracy, missingness and consistency together.⁶

Final thoughts

Predicting credit risk is at the heart of any lending business. Models keep evolving, but the most meaningful improvements now come from getting the data right: pulling in the right sources and creating features that reflect real behaviour, not just what’s easy to model. It also means working closely with the people who use these models every day, so the outputs are trusted, explainable, and aligned with how lending decisions are actually made.

The most successful lenders won’t necessarily be the ones with the most advanced algorithms. They’ll be the ones who know how to turn messy, fragmented data into features a model can learn from and an underwriter can defend.

Frequently asked questions