Logistic Regression for Credit Risk Scorecards: From Risk Drivers to Probability of Default

The controlled modelling chain starts before regression

The target must first say who can default, what default means and within which future horizon it is observed. Entimema's work on default definition establishes the event boundary; observation and performance windows establishes the prediction clock. Regression inherits both. It cannot correct them.

What the model actually receives

Source systems must reproduce a historical snapshot at a declared observation date. Eligibility rules then select the correct unit—borrower or facility—and predictor extraction must use information available at that instant. Multiple facilities require an explicit aggregation or row-selection rule; duplicated borrowers cannot quietly become independent observations. Exclusions, missing values, category definitions and temporal joins are model specifications, not data-cleaning footnotes.

The resulting feature matrix needs lineage back to source fields, extract timestamp, rules and model population. A categorical value not known in development needs an approved fallback. A currency, percentage or day count needs a stable unit. Rebuilding the same snapshot should produce the same eligible rows and features.

The model translates predictors into log-odds, then probability

For borrower i, PDᵢ = P(Yᵢ = 1 | Xᵢ): the conditional probability of default given the controlled predictor vector. Odds express default probability relative to survival probability, PD/(1−PD). Log-odds place that positive ratio on an unbounded additive scale:

The intercept β₀ is baseline log-odds when every encoded predictor equals zero. Each coefficient βⱼ changes log-odds for a one-unit change in predictor Xⱼ while other included predictors remain fixed. Its sign gives the conditional direction; it is not a constant percentage-point change in PD because the logistic curve is nonlinear. The final transform constrains the output to zero through one.

The conditional odds multiplier is e^βⱼ. In a WoE model, that statement applies to a one-unit change in WoE—not directly to one percentage point of utilisation. Interpretation also depends on convention. This article defines WoE = ln(distribution of defaults / distribution of non-defaults), so positive WoE denotes a riskier bin and positive coefficients preserve that direction. Reversing the WoE convention should reverse expected coefficient signs.

Why logistic regression fits traditional credit scorecards

Its binary target, explicit probability link, additive contributions, transparent coefficients and deterministic implementation suit validation, governance and score conversion. Risk committees can challenge driver direction; validators can reproduce contributions; engineers can execute the same arithmetic. That does not make it universally superior. Additivity in log-odds can miss interactions, estimates can be unstable under correlated predictors, extrapolation can be unsafe and changing populations can invalidate relationships. Interpretability is earned by coherent design, not conferred by an algorithm name.

WoE makes structure inspectable—not automatically valid

Binning can represent nonlinear raw relationships, isolate missing values and turn deployment into explicit range-to-value mappings. WoE aligns bins to target separation and makes contributions easier to inspect. Yet WoE transformation does not make a predictor automatically suitable for regression. Stability, redundancy, economic interpretation, leakage, interactions, coefficient behaviour and production availability still need challenge.

One borrower, reproduced from raw values to raw PD

Consider a hypothetical applicant scored by an original four-variable model. No proprietary data is used. Every WoE and coefficient below is illustrative; the declared positive-risk WoE convention applies.

The estimated intercept is β₀ = −3.200000. Individual contributions retain six decimal places:

Using the unrounded linear predictor, e^1.222 ≈ 3.39406, so PD = 1/(1 + 3.39406) = 0.227584, or 22.7584%. A production golden-borrower test should retain the approved coefficient precision and tolerance rather than copy a displayed rounded number.

Recent delinquency contributes +0.910000, the largest upward movement in log-odds. All four positive coefficients mean riskier positive-WoE bins raise conditional odds under this convention. These contributions describe the model's relative structure; they do not prove causality. Nor does 22.7584% automatically represent the portfolio's final PD: sampling and the development-period default level may require a separately approved calibration layer.

Variable selection is controlled judgement, not a leaderboard

Candidate selection should combine univariate discrimination and Information Value with business meaning, missingness, operational availability, leakage review, correlation, coefficient stability, significance and out-of-time behaviour. Automatic stepwise selection can optimise noise and repeatedly test the development sample until chance structure looks intentional. Statistical significance cannot rescue an economically incoherent model.

Multicollinearity destabilises the explanation

Utilisation, revolving balance and minimum-payment burden may encode the same behaviour. Together they can produce unstable coefficients, unexpected signs, inflated standard errors and sensitivity to minor sample changes. Correlation matrices and VIF help, but no mechanical threshold answers whether variables are economically redundant. Compare nested models and coefficient direction across development, validation, time and segments. A sign reversal may reveal correlation or a suppressor effect, not a newly discovered economic law.

Statistical significance is not economic materiality

With enough rows, a tiny effect can produce p < 0.05 while adding negligible separation. A conceptually strong driver can look unstable in a limited default sample. Selection therefore requires Statistics + Economics + Stability + Implementation. Committee-ready evidence should show not just a p-value, but incremental contribution, plausible direction, temporal behaviour and dependable availability at decision time.

Logistic regression is linear in log-odds, not raw economics

Binning and WoE can represent nonlinear raw effects. Explicit interaction terms or segmentation can represent conditional effects—for example, high utilisation may be materially more dangerous when recent delinquency is present. Interactions need prior rationale, sufficient observations, stability assessment and executable definitions. Uncontrolled interaction mining simply enlarges the space in which development noise can win.

Defaults are scarce—and sampling changes prevalence

Oversampling defaults, undersampling non-defaults or weighting observations can support estimation and learning from scarce events. The estimation population then differs from the calibration population. A sampling strategy can help model estimation while simultaneously changing observed default frequency. Unless the method explicitly corrects that shift, especially its effect on baseline odds, raw logistic output must not be labelled portfolio-level calibrated PD.

Validation needs time, not only a random split

Use distinct development, validation and out-of-time samples. Random splitting tests repeatability within a mixed historical population; it can distribute the same macro regime, policy, channels and products across both sides. Credit portfolios evolve through economic conditions, underwriting, product design, channel mix, customer mix and data collection. Out-of-time evidence asks whether relationships travel. Credit Vintage Analysis helps identify when cohort behaviour reflects changing origination conditions rather than a stable model relationship.

A model can rank correctly and estimate the wrong probability

ROC, AUC, Gini and KS are useful only insofar as they answer an economic question: does the model consistently assign higher estimated risk to borrowers who subsequently default? They assess ordering or separation, not whether 5% predictions default at approximately 5%.

This is the central transition from discrimination to calibration. Pricing, expected loss, provisioning, limits and risk-adjusted return require meaningful probability levels. Entimema's analysis of PD Model Ranking vs Calibration develops that distinction; here it becomes an implementation requirement: Borrower characteristics → contributions → log-odds → raw probability → calibration → final PD.

Score scaling changes representation, not borrower risk

A traditional scorecard may express the same model on a points scale. If odds mean non-default/default, a generic increasing-goodness convention is:

The base score is assigned at declared base odds. Points to double the odds (PDO) defines how far score moves when good-to-bad odds double. With base score 500 at good:bad odds of 20:1 and PDO 40, Factor = 40/ln(2) = 57.7078 and Offset = 500 − 57.7078×ln(20) = 327.1229. At 40:1, score is 540. Reversing the odds definition or desired score direction changes signs; it does not change PD.

Score and PD are different representations. Production documentation must state odds convention, offset, factor, rounding and whether calibration occurs before or after a score-to-PD mapping. An incorrectly signed scaling formula can make the riskiest borrower appear best even while the underlying regression is correct.

The analytical model becomes executable through explicit artefacts

The deployed chain is Source data → feature pipeline → transformation rules → WoE mapping → coefficients → linear predictor → logistic transform → calibration layer → PD → score / risk grade → decision engine. Each stage needs deterministic transformations, schema validation, explicit missing and unknown-category behaviour, timestamp consistency, feature availability checks, audit logs and fallback logic.

A production model version must bind target definition, population rules, feature definitions, bin maps, coefficients, calibration, score scaling and decision interface. Changing any one can change the effective model. Version identity should appear in every scored record with input timestamp, feature values, mapped bins, contributions, raw PD, calibrated PD and decision response sufficient for reconstruction.

Development–production parity is a model-risk control

Equivalent input must yield equivalent bins, WoE values, contributions, raw PD, calibrated PD and score within approved numerical tolerances. Different missing rules, rounding, category maps, data types, units, coefficients, bins or versions break parity. A statistically perfect model can therefore fail operationally.

This is not ordinary software QA alone. A mapping error changes the statistical model being used and therefore belongs to model-risk control, validation evidence and release approval.

Failure propagates from statistics to real decisions

Other recurring failures include coefficient sign instability, overfitting, incorrect probability code, wrong score scaling and stale coefficients. Controls should fail visibly: a silent default value is not resilient fallback unless its statistical treatment, allowed use and escalation were approved.

Monitor the whole chain after deployment

A stable final score can conceal a broken feature whose effect is offset elsewhere. Monitoring must connect layer signals and retain model version and policy chronology. Entimema's PD Model Monitoring develops this diagnostic architecture from input integrity through realised portfolio and decision effects.

The PD model supplies risk; it does not make the credit decision

The next layer is PD model → decision strategy → decision engine → automated credit decision. The engine may combine calibrated PD with affordability, policy rules, exposure, price, limits, fraud indicators and manual-review rules. Keeping those layers separate preserves a crucial distinction: changing a cut-off or affordability rule changes strategy; changing a bin map or calibration changes the effective model.

A credible future workflow is a PD Model Implementation & Parity Monitor. It could repeatedly compare development and production outputs, run golden borrowers, validate WoE maps, detect version or feature-pipeline changes, monitor missingness and score distributions, flag unexpected PD movement and prepare diagnostics for human model owners. This is an automation opportunity, not a claim that such an Entimema agent currently exists.