Credit Scoring (Scorecards, WoE/IV)
Also known as: Credit Scorecard, Application Scoring, Behavioural Scoring, Kredi Skorlama
Credit scoring is a statistical technique that estimates the probability that a borrower will default on a financial obligation. Using Weight of Evidence (WoE) binning, Information Value (IV) variable selection, and logistic regression, it converts raw applicant data into a single integer score. Formalized by Hand and Henley (1997) and elaborated by Thomas, Edelman, and Crook, the scorecard framework has become the regulatory standard for retail credit risk assessment in banking, lending, and insurance.
Read the full method
Sign in with a free account to read this section.
Method map
The neighbourhood of related methods — select a node to explore.
When to use it
Use credit scoring when the outcome is binary (default vs. non-default), historical labelled data with at least several hundred bad accounts are available, and regulatory interpretability is required. The WoE/IV framework assumes that binned predictors capture the risk relationship adequately and that the log-odds is approximately linear in the WoE scores. It is less suited for very sparse populations, rapidly shifting risk profiles, or settings where complex non-linear interactions dominate. Alternatives include gradient-boosted trees (XGBoost) for predictive accuracy or Altman Z-Score for corporate distress prediction.
Strengths & limitations
- Produces fully interpretable, auditable scorecards that satisfy Basel II/III and GDPR explainability requirements.
- WoE transformation handles missing values, outliers, and non-monotonic relationships without manual imputation.
- Integer scorecard format enables rapid manual override and direct communication of risk to non-technical stakeholders.
- Well-studied statistical properties; Gini coefficient, KS statistic, and PSI provide standardized performance monitoring.
- Assumes the log-odds relationship is linear in the WoE-transformed predictors, which may miss complex interactions.
- Requires a sufficiently large bad-account sample; performance degrades sharply with fewer than a few hundred defaults.
- Static scorecards become miscalibrated when population distributions shift (population stability must be monitored via PSI).
- Manual binning and IV thresholds introduce analyst subjectivity that can affect reproducibility.
Frequently asked
What is the difference between a scorecard and a logistic regression model?
A scorecard is a logistic regression model whose log-odds output has been scaled to an integer score using target-odds and PDO calibration parameters. Every scorecard is backed by logistic regression, but the integer points format makes it auditable and easy to apply manually. The underlying probability of default is recoverable by inverting the scaling transformation.
How do I choose the number of bins when computing WoE?
Common practice is to start with 10–20 quantile-based bins, then merge adjacent bins that violate monotonicity or contain fewer than 5% of the sample. The goal is smooth, monotone WoE values that reflect genuine risk separation without overfitting. Automated binning algorithms (e.g., optimal binning via dynamic programming) can assist but still require expert review of the resulting risk segments.
What IV threshold indicates a variable is worth including in the model?
The conventional rule is: IV below 0.02 is useless, 0.02–0.1 is weak, 0.1–0.3 is medium, and above 0.3 is strong. Variables with IV above 0.5 should be treated with caution as they may be proxies for the target or data leakage. These thresholds originate from industry practice documented in Hand and Henley (1997) and subsequent scorecard literature.
Sources
- Hand, D. J., & Henley, W. E. (1997). Statistical classification methods in consumer credit scoring: a review. Journal of the Royal Statistical Society: Series A, 160(3), 523–541. DOI: 10.1111/j.1467-985X.1997.00078.x ↗
How to cite this page
ScholarGate. (2026, June 2). Credit Scoring (Scorecards, WoE/IV). ScholarGate. https://scholargate.app/en/finance/credit-scoring
Which method?
Set this method beside its closest kin and read them side by side — the library lays the books on the table; the choice is yours.
- Altman Z-ScoreFinance↔ compare
- Logistic RegressionResearch Statistics↔ compare
- XGBoostMachine learning↔ compare