Calibrating Explanation Stability: Input Noise Moves SHAP Rankings Far Less Than Retraining Does

Studies of explanation robustness usually report how much feature attributions change under small input perturbations, but rarely say how large that change is compared with the variability already present between equally valid retrained models, and they often compare explanation stability and prediction stability on different scales. This study calibrates perturbation-induced SHAP instability against a retraining noise floor. Across 13 real tabular datasets and three model classes (Random Forest, XGBoost, logistic regression), independent and correlation-preserving Gaussian noise of 1–10% of each feature’s standard deviation is added, and the resulting change in exact SHAP explanations is compared with the change caused by retraining on bootstrap resamples. At 5% noise, global SHAP rankings change only 5–7% as much as under retraining for tree ensembles (16–17% at 10%), and top-5 feature sets are essentially unchanged. On this matched scale, global rankings are significantly more stable than predictions (Holm 𝑝 ≤ 0.0012). Local explanations behave differently: at 10% noise they change about as much as retraining does and track prediction flips (exploratory 𝜌 ≈ 0.6). A pre-specified predictor of rank fragility based on baseline importance gaps was not supported. A naive comparison of unweighted Kendall’s 𝜏 with flip rate would suggest a mixed or opposite picture. The analysis plan, code and per-run results accompany the paper.