Generalized Additive Models: Interpretable Credit Scoring
Generalized Additive Models: Interpretable Credit Scoring
An additive model scores an applicant as a baseline plus one learned contribution from each input. When an applicant's predicted default risk rises from 4.74% to 18.2%, the model can say exactly how much came from 85% utilization, four recent inquiries and a two-year credit history, with no approximation. Generalized additive models and explainable boosting capture the nonlinear patterns that make machine learning accurate while staying as readable as a scorecard.
What you'll learn
The trade-off that may not exist
Credit modelers are often told they must choose between accuracy and explainability: a logistic scorecard anyone can read, or a gradient-boosted ensemble that predicts better and explains itself only through approximations. Generalized additive models, introduced by Hastie and Tibshirani, sit between the two. Each input gets its own flexible curve, learned from the data, and the prediction is the sum of those curves. The model can capture the nonlinear shapes that make boosting accurate while remaining exactly as readable as a points-based scorecard.
Modern versions, such as the explainable boosting machines developed from the work of Lou, Caruana and Gehrke, fit each curve with boosted trees and can add a small number of two-variable interactions. On many tabular prediction problems they come close to the accuracy of unrestricted boosting. Rudin has argued that for high-stakes decisions this makes black-box models with after-the-fact explanations hard to justify.
How an additive model scores
The model predicts the log-odds of default as a baseline plus one contribution from each input. Take a baseline of −3.0, which corresponds to a 4.74% default probability, and an applicant with these characteristics.
| Input | Applicant value | Contribution to log-odds |
|---|---|---|
| Revolving utilization | 85% | +0.9 |
| Inquiries in the last six months | 4 | +0.5 |
| Age of oldest account | 2 years | +0.4 |
| Delinquencies in the last two years | 0 | −0.3 |
| Total | −3.0 + 1.5 = −1.5 |
The adverse action reasons fall out directly. Ranked by contribution, they are high utilization, recent inquiries and limited credit history. If the applicant's utilization were 30%, where the curve contributes −0.2, the log-odds would fall by 1.1 to −2.6 and the probability to 6.9%. Per our adverse action analysis, Regulation B requires specific and accurate principal reasons for a denial; an additive model provides them by construction rather than by estimation.
Reading and fixing the shapes
Because each input's effect is a single curve, reviewers can look at it and ask whether it makes sense.
- Expected shapes confirm the model. Risk rising with utilization, falling with file age and rising with recent delinquency are what credit experience predicts.
- Unexpected shapes reveal data problems. Caruana and colleagues described a pneumonia risk model that learned asthma patients were at lower risk, because those patients had been sent straight to intensive care. The additive model made the finding visible, and the curve could be corrected before use. Credit data has the same kind of trap: a segment that looks safe because it was always declined unless exceptionally strong.
- Curves can be edited. A modeler can smooth a jagged curve, impose a monotonic shape where policy requires it, or flatten a variable's effect, and the model remains consistent. Per our monotonic constraints analysis, enforcing known risk directions is one of the most effective protections against overfitting.
- Proxy risk is easier to see. Per our proxy variable analysis, a curve that tracks a protected characteristic, such as one for a geographic input, can be reviewed directly rather than inferred from approximate attributions.
How the approaches compare
| Logistic scorecard | Additive model | Unrestricted boosting with post-hoc explanations | |
|---|---|---|---|
| Nonlinear effects | Only through manual binning | Learned per input | Learned, with interactions |
| Interactions | Only if specified | A few, chosen and visible | Many, implicit |
| Explanation | Exact | Exact | Approximate |
| Illustrative AUC on the same data | 0.74 | 0.77 | 0.78 |
The accuracy figures are illustrative, and the gap varies by problem. The pattern they show is common in published comparisons on tabular data: most of the gain from moving beyond a logistic scorecard comes from capturing nonlinearity in individual inputs, which an additive model does, and a smaller remainder comes from complex interactions, which it partly gives up. Per our model explainability analysis, post-hoc methods such as SHAP values can explain an unrestricted model reasonably well, but their attributions are estimates that can vary with the method and the reference population.
Accuracy you can read
HL Hunt AI Underwriting builds additive models with learned risk curves for each input, enforces monotonic shapes where credit logic requires them, and generates adverse action reasons directly from the model rather than from approximations.
Why it matters for governance
- Validation is easier. Per our model validation analysis, validators can review every effect directly, test stability curve by curve, and confirm that reasons match the score.
- Monitoring is clearer. When performance drifts, each curve can be refit or compared with its earlier version, which localizes the change. Per our population stability analysis, knowing which input moved is the first step in deciding whether to recalibrate or rebuild.
- Policy and model can be reconciled. Per our rules and models analysis, credit policy is written in terms of individual factors, and an additive model expresses risk in the same terms.
- Fair lending review is more direct. Per our AI lending governance analysis, searching for less discriminatory alternatives is simpler when each input's effect can be adjusted and the accuracy cost measured.
Built for validators and regulators
HL Hunt AI Underwriting documents every risk curve, tracks shape stability over time and supports fair lending testing input by input, so model governance reviews the model itself rather than an explanation of it.
The case for unrestricted models
Additive models give up something. When important risk depends on combinations of inputs, such as income volatility mattering more for borrowers with little savings, an additive model captures it only if the interaction is included explicitly, and only a few can be included before readability suffers. In large, data-rich portfolios, the remaining accuracy gap can be worth real money: per our rare events analysis, small improvements in ranking compound at scale. Post-hoc explanation methods have also matured, and many lenders use them successfully under regulatory review. The practical test is empirical. Fit both on the same data, measure the out-of-time accuracy difference, translate it into expected losses and approvals, and compare that with the cost of approximate explanations in validation, adverse action and fair lending review. Per our model sourcing analysis, many lenders that run this comparison find the additive model close enough that its transparency decides the question.
Frequently asked questions
A model that sums one learned curve per input, capturing nonlinear effects while staying exactly interpretable.
The largest positive contributions are the reasons, exactly. In the example: high utilization, recent inquiries and short credit history.
Often close, because most gains come from per-input nonlinearity. The remaining interaction gap varies and should be measured.
Every effect is directly reviewable, drift is easier to localize, and fair lending adjustments can be made input by input.
Key takeaways
- Additive models predict log-odds as a baseline plus one learned curve per input.
- They capture nonlinear effects like boosting while remaining exactly interpretable.
- In the example, an applicant moves from 4.74% to 18.2% default risk, fully attributed to specific inputs.
- Adverse action reasons come directly from the largest contributions, not from approximations.
- Reviewing each curve exposes data and selection problems, as in the asthma pneumonia case.
- Curves can be smoothed, constrained to monotonic shapes or flattened without breaking the model.
- Most of the accuracy gain over logistic scorecards comes from per-input nonlinearity.
- Measure the out-of-time gap against unrestricted models before giving up transparency.
References
- Hastie, T., & Tibshirani, R. (1986). Generalized Additive Models. Statistical Science, 1(3), 297–310.
- Lou, Y., Caruana, R., & Gehrke, J. (2012). Intelligible Models for Classification and Regression. Proceedings of the 18th ACM SIGKDD International Conference on Knowledge Discovery and Data Mining, 150–158.
- Lou, Y., Caruana, R., Gehrke, J., & Hooker, G. (2013). Accurate Intelligible Models with Pairwise Interactions. Proceedings of the 19th ACM SIGKDD International Conference on Knowledge Discovery and Data Mining, 623–631.
- Caruana, R., Lou, Y., Gehrke, J., Koch, P., Sturm, M., & Elhadad, N. (2015). Intelligible Models for HealthCare: Predicting Pneumonia Risk and Hospital 30-day Readmission. Proceedings of the 21st ACM SIGKDD International Conference on Knowledge Discovery and Data Mining, 1721–1730.
- Rudin, C. (2019). Stop Explaining Black Box Machine Learning Models for High Stakes Decisions and Use Interpretable Models Instead. Nature Machine Intelligence, 1(5), 206–215.
- Regulation B (Equal Credit Opportunity), Notifications, 12 C.F.R. § 1002.9(b)(2).
This guide is educational and does not constitute model risk, legal or regulatory advice. Baselines, contributions, probabilities and accuracy figures are stylized illustrations and do not describe any particular lender or model. Model performance comparisons vary by data and problem; validate any model against out-of-time outcomes before use.