Logistic Regression for Credit Risk: From Log-Odds to a Production Scorecard

A binary outcome needs a probability architecture that survives production

The target inherits the institution's Default Definition, observation point, performance horizon and maturity rules. Observation and Performance Windows determine which features and outcomes are legitimate. A technically correct estimator cannot rescue a contaminated target.

A linear probability model, P(Y=1|X)=β₀+β₁X, can predict below zero or above one, imposes a raw linear probability effect and has heteroskedastic errors. Its limitation motivates a bounded link rather than a detour into ordinary regression.

Odds make the convention explicit

At PD=10%, default-to-survival odds are 0.10/0.90=0.1111, or approximately 1:9 Bad:Good. Credit scorecards often use the reciprocal Good:Bad odds. Every implementation must name the convention and preserve it through scaling and interpretation.

The probability is nonlinear and bounded, while effects remain additive in log-odds. Operationally, that means variable contributions can be reconciled before conversion to PD and points.

Coefficients describe conditional odds—not percentage-point probability changes

β_j is the change in log default odds for a one-unit increase in X_j, holding other model variables constant. Exponentiating gives the odds ratio:

The example means approximately 35% higher default odds per unit under the specified scale and model—not a 35% increase in PD. The probability effect depends on the borrower's starting risk and all other contributions.

WoE sign conventions must agree with the target

Under WoE=ln(DistGood/DistBad), higher WoE indicates better relative credit quality. With Y=1=Bad, a standalone coefficient should generally be negative: increasing “good” evidence should lower bad log-odds. Multivariate correlation can complicate signs, so investigate rather than mechanically constrain them. Reversing the target, reversing WoE or reversing only the interpretation creates a production error.

Weight of Evidence & Information Value develops binning, stability and univariate evidence. WoE controls raw non-linearity and extremes; it does not validate the multivariate specification.

A single WoE predictor makes every transformation visible

A positive +0.75 coefficient under this target and WoE convention would predict higher bad odds for better-quality bins. That is not impossible in a correlated model, but it is a sign anomaly demanding investigation.

A four-variable model decomposes borrower risk

For any borrower, Contribution_ij=β_jX_ij and z_i=β₀+ΣContribution_ij. That additive audit trail becomes points, diagnostics and candidate reason-code evidence.

Maximum likelihood estimates probabilities—not accuracy at an arbitrary threshold

Estimation chooses coefficients that make the observed pattern of Goods and Bads most probable under the assumed functional form. It does not maximise classification accuracy. Credit risk normally needs ranking, probability estimation and strategy simulation—not one binary classification threshold.

Rare defaults expose the weakness of accuracy

At a 2% bad rate, predicting Good for every account achieves 98% accuracy and zero useful discrimination. AUC/Gini, KS, calibration and proper scoring rules address more relevant properties.

Oversampling changes the probability level

Bad oversampling can support coefficient estimation and ranking, but raw model probabilities reflect altered prevalence. Use a documented intercept correction or recalibration to the intended population; do not report development-sample PDs as production probabilities.

Apparent simplicity can conceal specification failure

A raw continuous X assumes a linear effect on log-odds unless transformed. Binning/WoE, splines, polynomial terms and domain transformations can represent non-linearity. Interactions such as Utilisation × Tenure may capture genuine conditional risk, but an interaction zoo built to optimise in-sample metrics weakens stability and explainability.

A model can rank well while individual coefficients are unstable and explanations are economically incoherent. Random development/validation splitting tests sample generalisation; out-of-time evidence challenges temporal robustness.

Regularisation and ML belong in the challenger set

Shrinkage can control correlated predictors and overfitting, but changes coefficient and selection interpretation. Machine learning can add nonlinear and interaction power. Compare all challengers through incremental decision value, stability, explainability, operational cost and governance—not novelty.

Ranking and calibration answer different model-risk questions

A logistic model can discriminate well and still be badly calibrated. If predicted PD is 5% for a coherent group, approximately 5% should realise the governed default event over the matching horizon under the intended interpretation. PD Ranking & Calibration develops these cases in depth.

Brier score evaluates probabilistic error while combining discrimination and calibration characteristics. Log loss penalises confident wrong predictions heavily. Neither is a universal decision rule; interpret them with ranking, calibration, stability and portfolio use.

Score scaling turns odds into a controlled operating language

Using Good:Bad odds so higher score means lower risk:

For base score 600, base Good:Bad odds 20:1 and 50 Points to Double the Odds:

Because log Good:Bad odds = −log Bad:Good odds = −z, the equivalent model mapping is Score=Offset−Factor×z.

The intercept becomes base points: Offset−Factor×β₀. It can be held once or distributed across variables, but development and production must use identical architecture. A score is a monotonic transformation of odds, not PD without its scale and calibration mapping.

Recover PD from a score

At S=600, odds are exp((600−383.904)/72.135)=20, so PD=1/(1+20)=4.76%.

An original end-to-end borrower from raw data to production score

Consider a fictional scorecard developed on 200,000 consumer-loan applications. One applicant maps to these governed bins and Good/Bad WoE values:

Base points are 383.904−72.135(−3.20)=614.736. Adding the rounded variable points gives approximately 520. Rounding explains small display differences; the governed engine should calculate with full precision and round only at the specified stage.

This chain—Borrower data → Bin → WoE → Contribution → Logit → PD → Odds → Score—makes statistical and operational reconciliation possible. Largest adverse contributions can inform explanations, but a mathematical contribution is not automatically a legally valid customer-facing reason code.

The logistic model produces risk information, not the lending decision

Credit Cut-Off Strategy connects PD and score to expected loss, revenue, affordability, policy, risk appetite and marginal economics. The same score can support different actions under different products or constraints.

A model fitted only on historical accepts estimates relationships within a selected population. If policy expands beyond that region, Reject Inference explains the overlap and identifiability problem. Logistic regression does not remove selection bias.

Production replication is part of model validity

Annual versus monthly income, percentage versus decimal, or one currency versus another can destroy model meaning without any software exception. New categorical values need a governed Other bucket, fallback or model exception—never arbitrary WoE.

Rounding each variable's points can differ from rounding only the final total, especially at a cut-off. Test boundary records, missing values, unseen categories and exact bin edges at contribution, PD and score level.

Monitor who arrives and whether risk relationships still hold

Track score and input distributions, bin populations, missing rates, discrimination, calibration, realised bad rates, approval rates and overrides. A stable applicant population does not guarantee stable coefficients or PD relationships. PD Model Monitoring connects these signals to investigation and action.

If ranking remains useful while risk level moves, recalibration may suffice. If coefficients, variable relationships or ordering materially deteriorate, re-estimation or broader redevelopment is required. Mathematical transparency does not protect an obsolete model.

Logistic scorecards remain practical for high-volume non-bank lending

Consumer finance, fintech, digital instalments and short-tenor portfolios often need lightweight deployment, high-volume decisions and explainable logic with limited modelling resources. Logistic models provide a compact deterministic path from features to score.

Rapid population drift, strategy changes, thin history and higher-risk mix increase the need for OOT testing, frequent calibration review, vintage analysis and transformation monitoring. Fast outcomes shorten feedback; they do not remove target, sample or selection risk.

Twenty failure modes between equation and decision

A Credit Scorecard Development Agent can automate evidence work—not borrower decisions

A future agent could fit candidate logistic models, test coefficient stability, flag sign anomalies, diagnose multicollinearity and separation, compare regularised challengers, calculate discrimination, perform OOT validation, generate score scaling, reconcile score-to-PD mapping, test implementation parity and prepare development evidence.

Its role is model-development automation + diagnostics + validation support. It must not autonomously approve or reject borrowers. Human owners govern model purpose, lawful variables, target meaning, risk appetite and adverse decisions.

The broader Credit Scorecard Development pillar connects this estimator to population, target, binning, scaling, calibration, strategy and monitoring. Entimema's Credit Risk practice supports that lifecycle; Decision Automation turns approved model and policy logic into controlled execution.