Modern credit risk modelling: the use of machine learning algorithms

Credit scorecards remain a key part of how lenders assess risk. Built using methods like logistic regression, they offer transparency and consistency, and have become standard across retail and commercial lending. But as lenders work with more complex borrower data and aim to make faster, more tailored decisions, traditional scorecards are increasingly being supplemented by machine learning models.

The shift towards machine learning models has been much faster in origination, monitoring and collections than in regulatory capital models: the European Banking Authority (EBA) observed in 2021 that the IRB models that institutions were using to calculate credit risk capital "do not differ materially from the approaches used 15 to 20 years ago when the first Basel Accord was put in place".¹

Why move beyond logistic regression?

Most scorecards are based on logistic regression, a linear model that estimates the probability of default from a set of manually engineered features. Its main benefit is simplicity: each input contributes in a clear, weighted way to the outcome.

Linear models make strong assumptions, though. They don’t handle non-linear relationships well, and they miss interactions between variables unless those interactions are modelled explicitly. This is where tree-based machine learning models offer a real advantage, and where the regulatory constraint begins. Article 171(1)(a) and (b) of the Capital Requirements Regulation (CRR) sets the criteria for assigning an exposure to a grade or pool, and the EBA’s reading is that this "may constrain the use of models where you do not have a clear economic link between the input and the output variables".¹

Key machine learning models in credit scoring

Random forest

A random forest is an ensemble method that builds multiple decision trees from bootstrapped samples of the data. Each tree votes on the outcome, and the final prediction is the average for regression, or the majority vote for classification.

• Strengths: robust to overfitting, a good baseline model, and a strong performer on high-dimensional data.
• Drawbacks: less interpretable than a single decision tree, and relatively slow on large datasets.

Gradient boosting machines

Gradient boosting machines (GBMs) build trees sequentially, each new tree correcting the errors of the ones before it. That makes them flexible and powerful. Two implementations dominate.

XGBoost (extreme gradient boosting) is known for speed and performance. It minimises a loss function using a convex optimisation algorithm and includes a regularisation term that penalises model complexity to reduce overfitting. It’s widely used in production credit risk models.

LightGBM (light gradient boosting machine) is optimised for speed and memory efficiency. It grows trees leaf-wise rather than level-wise, and uses gradient-based one-side sampling to evaluate information gain efficiently, skipping data points with small gradients. That improves training speed and sometimes accuracy, at the risk of overfitting if it isn’t tuned properly.

These boosting models have become the standard for many machine learning credit scoring applications, thanks to their predictive accuracy and their ability to model complex, non-linear relationships.

Benefits of machine learning models

• Predictive performance: tree-based models generally rank borrowers by risk more accurately than linear models, measured on ROC AUC or Gini, the two standard measures of how well a model separates defaulters from non-defaulters. The gain shows up most in segments with thin credit files or limited traditional data. One 2010 study measured what that is worth on a commercial book: applying non-linear, non-parametric models to customer transaction and bureau data, it estimated savings of 6 to 25% of total losses on conservative assumptions about the cost of cutting credit lines.²
• Automatic handling of interactions: GBMs capture interactions between features that would need to be specified by hand in a regression-based scorecard.
• Flexible feature input: machine learning models are more tolerant of skewed distributions, outliers, and missing values. They can use raw or minimally transformed inputs, which speeds up development.
• Feature importance: complex as they are, these models still offer useful insight through feature importance metrics, highlighting which variables most influence predictions.
• Better credit decisions: with sharper segmentation and risk ranking, lenders can approve marginal applications more confidently, price more competitively, and reduce false declines. That means higher profits and wider financial inclusion.

Interpretability

Explainability is often the primary barrier to using machine learning models in regulated lending. Tools such as Shapley additive explanations (SHAP) help. SHAP assigns an importance value to each feature for a given prediction, which makes account-level explanations possible: not just which variables matter across the portfolio, but why this applicant received this score.

SHAP is one of five techniques the EBA names, alongside graphical tools, feature importance measures, local explanations such as LIME (local interpretable model-agnostic explanations) and anchors, and counterfactual explanations.¹ The graphical tools come in two forms: partial dependence plots show the effect on the average prediction, and individual conditional expectations show the effect on a single one.

The EBA’s 2023 follow-up report recorded what 14 consultation respondents said they use: Shapley values led at 40%, ahead of enhanced reporting and documentation of the model methodology at 28%, graphical tools at 20%, and sensitivity analysis at 8%.³

The reason SHAP leads that list is a result rather than a fashion. A 2017 proof established that among methods explaining a prediction as a sum of feature contributions, the Shapley value is the only one satisfying local accuracy, missingness and consistency together.⁴ Consistency is the one that matters to a validator: if a model change increases a feature’s marginal contribution, its attribution cannot decrease, so the direction of an explanation stays trustworthy from one model version to the next.

Model governance and fairness

Machine learning models must meet the same governance standards as traditional scorecards: clear documentation, robust validation, and regular monitoring for performance drift or data shifts.

Fairness is an equally important part of governance. Because these models learn from historical data, they can unintentionally reinforce biases that exist in it. Fairness checks, such as evaluating model outcomes across different groups, should be part of both development and ongoing monitoring. The Prudential Regulation Authority (PRA) gives this a formal foothold: under Principle 1 of SS1/23, 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".⁵ Bias is therefore an input to how the model is tiered, not only a test it has to pass.

What’s next? Neural networks

As lenders collect richer data, neural networks may represent the next step. They can learn highly complex, non-linear relationships without manual feature engineering, and they suit high-dimensional or sequential data such as transaction histories particularly well.

Adoption of neural networks remains thin, though, held back by interpretability challenges, longer training times, and heavier governance requirements. The EBA’s 2023 survey of where machine learning sits in IRB work suggests the realistic near-term role. Use is selective and early-stage, with PD risk differentiation the dominant case. Random forests and gradient boosting trees are used mainly for selecting risk drivers and preparing input data rather than as the primary estimation engine, and in validation the main use is challenger models and robustness analysis.³ There is a data reason for that split: risk differentiation can rest on shorter time series, whereas risk quantification needs the long-run averages the CRR’s five-year minimums require. Advances in explainability tooling and cheaper compute change the calculation, but they do not change the observation period.

Final thoughts

Machine learning isn’t a replacement for traditional credit scoring. It’s an extension.

Models like XGBoost, LightGBM, and random forests uncover patterns and relationships that linear models can’t, which leads to stronger credit decisions and more tailored risk management. For institutions looking to push further, neural networks may offer more predictive power still, especially as explainability tools improve.

The key to successful adoption lies not just in the models themselves, but in governance, explainability, and the alignment of machine learning development with broader risk policies. The EBA offers a test a model owner can apply directly: avoid complexity in the modelling approach that is not justified by a significant improvement in predictive capacity. It names three in particular, and each is a question worth asking of any candidate model.

Does it carry an excessive number of drivers, or drivers with no significant predictive information? Does it use unstructured data where conventional data offers similar predictive power? Are its modelling choices more complex than simpler approaches that yield similar results?

Approached that way, machine learning improves a lender’s ability to manage credit risk. Approached as a performance contest, it buys a rounding point of Gini and a validation problem.

Frequently asked questions