Model comparison — out-of-fold predictions
4 models evaluated on identical stratified folds. LightGBM leads on every metric; the WOE logistic regression scorecard stays close behind despite being fully interpretable.
Reading this: KS statistic and Gini are the two metrics credit risk teams report internally
more than plain AUC. A KS above 40 is generally considered strong for retail credit scoring — LightGBM
and XGBoost both clear that bar here.
Top features by Information Value (IV)
WOE/IV ranking computed during EDA — used both to shortlist scorecard variables and to validate against SHAP importance later.
Note on EXT_SOURCE_MEAN: its IV of 0.61 is flagged "suspiciously strong" by design —
it's an average of three external bureau risk scores, i.e. partly a restatement of another institution's
own risk model rather than a fresh signal. The individual EXT_SOURCE_1/2/3 features (IV 0.15–0.33) are
the more defensible, non-redundant predictors.
Scorecard points by bin — EXT_SOURCE_3
Points-based scorecard derived from the WOE logistic regression (PDO convention). Higher external bureau score → more points → lower risk.
Reading this: this is the same logic a real credit scorecard uses — each bin of a
feature contributes an additive number of points, and an applicant's total score is just the sum
across all their features' bins. Monotonically increasing points here confirm the scorecard behaves
exactly as intended: safer applicants score higher.