Churn/Value OLIST · Executive Report

Artifacts served from /static/data/OLIST/** (PNG/CSV/JSON by execution TS).

Recommended model: Gradient Boosting

Best balance of ROC AUC, Average Precision, and calibration (Brier). Its calibrated probabilities enable reliable per-segment thresholds for campaign targeting. Use Lift/Gain charts to select the optimal decile cutoff: targeting the top 20% captures ~65% of repurchases while minimizing contact cost.

Summary

  • Classification (churn): Decision Tree (pruned), Random Forest, Gradient Boosting, MLP.
  • Regression (value): Linear Regression with |coef| + betas & p-values.
  • Data cleaning: ±3σ, anti-leakage, drop |corr|≥0.99 with target.
pandasscikit-learnstatsmodelsmatplotlib

Global artifacts

Churn rate over time Latest log

Notes

Comparison: ROC AUC, Average Precision, F1, Precision, Recall and Brier (↓ better). Regression appears in its own section.

Model comparison (Classification)

Model ROC AUC Average Precision F1 Precision Recall Brier
Best value per metric is highlighted (Brier: lower is better).