Engineering guide
How to Build an Explainable AI Credit Scoring Model
A practical sequence for an underwriting model that is accurate, explainable to each applicant, and acceptable to compliance from the first version.
- Back-test accuracy in our build
- 92%
- Recommendations overridden
- Under 12%
- From kickoff to full cutover
- 16 weeks
In lending, an accurate model that cannot explain itself does not ship. Compliance needs reason codes for every decision, underwriters need to see why a case was flagged, and applicants who are declined are often entitled to know the main reasons.
This guide follows the approach we used for a consumer lender with about 200,000 active borrowers: decisions in 48 hours instead of 14 days, 92% accuracy on a 12-month back-test against real repayments, and underwriters overriding fewer than 12% of recommendations.
The steps
- 1. Fix the data first
- Most errors start in document extraction. Clean, verified income, employment and bank data paid back several times over in our project.
- 2. Define the target and the window
- Agree what counts as a bad loan and over what period, then build a back-test on held-out months, not random rows.
- 3. Start with a baseline
- A logistic regression scorecard shows how much the more complex model really adds.
- 4. Train an explainable model
- Gradient boosting with monotonic constraints where the direction is known (for example, more income should never raise risk).
- 5. Generate reason codes
- Map the largest SHAP contributions per applicant to approved, plain-language reasons.
- 6. Test fairness
- Compare approval rates and error rates across groups, investigate proxies, and document the results for compliance.
- 7. Put people in the loop
- Underwriters see the recommendation and reasons, and record overrides with a structured reason.
- 8. Monitor and retrain
- Track drift, approval mix and actual repayments, and retrain with versioned models and an audit trail.
Model choices compared
Common model types for credit riskScroll horizontally to compare all columns.
| Model | Explainability | When it fits |
|---|---|---|
| Logistic regression scorecard | Very high, coefficients are the explanation | Baseline and small datasets |
| Gradient boosting with SHAP | High per decision, with care | Most production underwriting with rich data |
| Neural networks | Harder to defend in review | Rarely the first choice for regulated credit decisions |
Frequently asked questions
- Is SHAP enough for regulators?
- SHAP gives per-decision attributions, but regulators look at the whole system: data, model choice, testing, documentation, monitoring and human oversight. Treat SHAP as one part of the evidence.
- How much data do we need?
- Enough repaid and defaulted loans to back-test over a meaningful period. Where history is thin, start with a scorecard and rules, and add machine learning as outcomes accumulate.
- Will underwriters trust it?
- They trust what they can see and correct. Clear reasons, a single workstation and structured overrides kept the override rate under 12% in our project.
- Can you build this for us?
- Yes. See our AI loan underwriting software service, or book a free consultation for a written plan and a fixed quote.
Sources and scope
This is an engineering planning guide. Requirements depend on the product and deployment. Consult the relevant authorities and qualified advisers before launch.
- SHAP documentation
Reference for computing per-prediction feature attributions.
- Explainable AI in Credit Risk Management (arXiv 2103.00949)
Research on applying LIME and SHAP to credit risk models.
Planning your AI loan underwriting software development project?
Bring your workflow, integration requirements, and deployment constraints. We will identify the discovery work, specialist reviews, and dependencies needed for a realistic delivery plan.
