XGBoost stacks spark model-risk debate

The validation team is only half right: weak single-variable predictors are not automatically fatal inside XGBoost, but “the ensemble will average it out” is not a sufficient defense for a credit decision stack. The sharper critique is whether the full system stays stable, calibrated, and explainable when the population shifts.

• –Low IV is not a knockout argument against tree feeders, because boosted trees can exploit interactions and correlated variables in ways linear scorecards cannot.
• –The best audit challenge is out-of-time performance, drift, and retraining stability, not univariate significance alone.
• –The top logistic layer is where multicollinearity among feeder logits can actually bite, making coefficient estimates and reason codes less stable.
• –If weak variables never show durable split gain, stable SHAP rankings, or business justification, they deserve a parsimony and data-quality challenge.
• –Missing SHAP or LIME is less important than missing sensitivity, calibration, and change-control evidence across vintages.