Credit-Scoring Algorithms | Logistic Regression and Model Risk

Quick read. A credit score is a prediction, not a verdict about a person. A transparent educational model can take measured inputs, transform them into a weighted sum, convert that sum into a probability, and then test whether the predicted probabilities match what later happens. The hard part is not only fitting the equation. The hard part is deciding whether the data, target, calibration, threshold and use of the model remain valid.

This article explains general mathematics used in credit-risk modelling. It does not disclose or attempt to reconstruct any proprietary scoring formula, does not describe any individual lender’s underwriting rules, and is not financial or legal advice.

The one-sentence answer

A common interpretable credit-risk model estimates a probability such as “probability of default within a stated horizon” from observed features, then validates whether that probability is discriminative, calibrated, stable, explainable and suitable for the decision in which it is used.

Why this belongs in mathematics education

Credit modelling joins algebra, functions, logarithms, probability, statistics, classification, optimisation and decision theory. It also teaches a deeper lesson: a numerical output is only meaningful when the target and conditions are explicit.

The Consumer Financial Protection Bureau describes a credit score as a prediction of credit behaviour based on information from credit reports and notes that people can have multiple credit scores depending on the data, model and purpose. MoneySense Singapore likewise explains that credit scores are used alongside other considerations such as income when financial institutions assess creditworthiness.

1. Start by defining the prediction target

Before choosing an algorithm, define the event. “Risk” is too vague. A model might instead predict an event such as default within 12 months, serious delinquency within a defined period, or another precisely stated outcome. Change the target and you change the meaning of every coefficient and probability.

Observation time: when are features measured?
Outcome window: over what future period is the event observed?
Positive class: exactly what counts as the event?
Eligible population: which applications or accounts belong in the model’s scope?
Decision use: is the model estimating risk, ranking cases, setting a review priority, or supporting another process?

2. A transparent teaching model: logistic regression

Logistic regression is useful for teaching because it converts a linear combination of features into a probability between 0 and 1. Let x_1, x_2, …, x_k be measured features. The model first forms:

z = beta_0 + beta_1 x_1 + beta_2 x_2 + ... + beta_k x_k

Then it applies the logistic function:

p = 1 / (1 + exp(-z))

The result p can be interpreted as an estimated probability only if the model, target, training process and calibration justify that interpretation. A score that merely ranks cases well is not automatically a well-calibrated probability.

3. Why log-odds make the equation linear

Probability is bounded between 0 and 1, but the linear predictor z can range over the real numbers. Logistic regression solves this by modelling log-odds:

log(p/(1-p)) = z

This gives students a useful connection between exponential and logarithmic functions. A one-unit change in a feature adds beta_i to log-odds, holding other modelled features constant. Exponentiating beta_i gives a multiplicative change in odds under the model.

4. Never confuse a teaching model with a commercial credit score

Commercial credit scores can use proprietary data transformations, segmentation, interactions, score scaling and other methods. Their exact formulas may be confidential intellectual property. The mathematics in this article is therefore a public conceptual model: it teaches how a probability-scoring system can work without claiming that any named bureau or lender uses these exact equations or coefficients.

5. From probability to score

Once a model produces log-odds or a probability, a system can map that quantity to a score scale. A generic linear transformation might be:

Score = A + B × log(odds)

The constants A and B determine the displayed scale and direction. This is another important lesson: two scoring systems can encode similar risk ordering on different numerical scales. The number itself has no universal meaning outside its model definition.

6. Discrimination asks whether the model separates outcomes

A model that assigns higher risk to cases that later default more often has useful discrimination. Common tools include the receiver operating characteristic curve and area under that curve. But discrimination is only one property. A model can rank cases well while systematically predicting probabilities that are too high or too low.

This is why a single metric such as accuracy or AUC should never become the entire audit.

7. Calibration asks whether predicted probabilities are numerically believable

Suppose a model places many comparable cases near predicted risk 0.10. Across a sufficiently appropriate sample and horizon, roughly 10% of those cases should experience the defined event if the model is well calibrated at that level. Calibration plots compare predicted probabilities with observed frequencies.

A Brier score can also measure the mean squared difference between predicted probabilities and binary outcomes:

Brier = mean((p_i-y_i)^2)

Lower is better for the same prediction task, but context still matters. Class prevalence and the intended use affect interpretation.

8. Thresholds turn probability into action

A probability model does not itself decide what to do. A separate decision rule may compare the score with one or more thresholds. For example, a low-risk band might proceed automatically, an uncertain band might receive human review, and a high-risk band might require another process. The exact decision policy is an institutional choice subject to law, regulation, risk appetite and product design.

Mathematically, the threshold should reflect asymmetric error costs. A false positive and a false negative rarely have equal consequences. This is why “maximum accuracy” is often not the right optimisation target.

9. Selection bias: the labels you observe may depend on earlier decisions

Credit data contains a difficult structural problem. Outcomes are easier to observe for applicants who were approved than for applicants who were rejected. If historical approval decisions depended on earlier models or human judgement, the training data is already filtered by past decisions. This can create selection bias.

Any model should therefore ask: “Who is missing from the labelled sample, and why?” A large dataset does not automatically remove this bias.

10. Data leakage can make a weak model look brilliant

Leakage occurs when a training feature contains information that would not genuinely be available at prediction time, or indirectly includes knowledge of the future outcome. A leaked model can produce impressive test metrics and fail immediately in production.

Was every feature available at the real decision time?
Were transformations learned only from training data?
Were later account states accidentally included?
Were multiple records from the same customer split across training and test sets in a way that leaks identity?

11. Model drift means yesterday’s relationship may not hold tomorrow

Economic conditions, borrower behaviour, products, policy, reporting practices and application channels change. Feature distributions can shift; outcome rates can shift; relationships between features and outcomes can shift. A stable equation can therefore become an unstable model of the world.

Monitoring should separate at least three questions: are the inputs changing, are the outcome frequencies changing, and is predictive performance changing? A change in one does not prove a change in all three.

12. Fairness cannot be reduced to deleting one sensitive column

A model can still produce unequal effects even when a protected attribute is excluded, because other variables can be correlated with it. Fairness analysis therefore requires outcome testing, feature review, legal interpretation and attention to the actual decision process. Different fairness metrics can also conflict mathematically when base rates differ, so there may be no single scalar that proves a system fair.

In the United States, the CFPB has emphasised that lenders using complex models still need to provide accurate and specific reasons for adverse credit actions. The broader lesson is computational: explainability obligations apply to the real decision pathway, not merely to a decorative explanation generated after the fact.

13. Model risk is the risk of being wrong in a systematic way

Model risk arises when the model is incorrect, misused, insufficiently validated, poorly monitored or applied outside its intended scope. In April 2026, U.S. banking agencies issued revised model-risk guidance; the OCC described the updated framework as addressing development and use, validation and monitoring, governance and third-party models. This is useful context for students because it shows that a formula is only one component of a dependable modelling system.

14. A practical validation ladder

Definition check: is the target precise and useful?
Data check: is the sample representative of the intended population?
Leakage check: was every feature genuinely available at prediction time?
Functional check: do signs and relationships make mathematical sense?
Discrimination check: does the model separate higher- and lower-risk cases?
Calibration check: do predicted probabilities correspond to observed frequencies?
Stability check: do inputs and performance remain stable over time?
Fairness check: are outcomes and errors examined across relevant groups?
Explanation check: can the real drivers of a decision be communicated accurately?
Use check: is the model still being used for the task for which it was validated?

15. Falsifiers: evidence that should make us distrust the model

Predicted 10% risk groups repeatedly show outcomes far from 10%.
Performance collapses on a later time period.
A feature only exists after the decision being predicted.
A small operational change radically changes scores without a plausible causal reason.
Applicants outside the training population are scored as though they were in-scope.
Explanations do not match the actual features driving the score.
Error rates shift sharply for a particular group or channel.

16. Connection to school mathematics

Linear functions: weighted combinations of features.
Exponential and logarithmic functions: logistic probability and log-odds.
Probability: predicted likelihood is not certainty.
Statistics: sampling, bias, calibration and validation.
Graphs: ROC curves and calibration plots.
Decision mathematics: thresholds trade one kind of error against another.

Bukit Timah Tutor’s public Probability and Statistics HELP Tasks provides a school-mathematics route into several of these ideas.

17. What should trigger an update?

Revalidation should be triggered by material changes in applicant population, product rules, data sources, economic conditions, outcome definitions, reporting systems, policy, model code or the decision process that consumes the score. A model should not be treated as current merely because its software still runs.

Sources and further reading

Final idea. A credit model is not trustworthy because it produces a precise number. It becomes useful only when the number has a clear target, sound data, tested calibration, appropriate decision rules, known failure modes and a process for discovering when the world has changed.

Small Group Tutorials

How Credit-Scoring Algorithms Turn Data Into Probabilities: Logistic Regression, Calibration and Model Risk