← back to tools

Return Risk Predictor

A tiny, pre-fit logistic-regression scorer for "will this order be returned?" — built from a simulated labelled dataset of 2,000 orders. Score a single order, tune the decision threshold, and see the ROC curve live. The real model at work used XGBoost and hit ROC-AUC 0.92; this is its simpler cousin.

Score an order

Return probability
—
Decision
—

Model performance

ROC-AUC
—
Precision
—
Recall
—
F1
—
ROC curve (AUC shown above) Random classifier Current threshold
Under the hood: logistic regression σ(w·x + b) with weights hand-picked to mimic the signs you'd expect (higher discounts and higher customer prior return rates ↑ risk; longer tenure and higher order value ↓ risk, slightly). I evaluate it on a held-out simulated test set of 2,000 labelled orders. The ROC curve and metrics re-compute against your chosen threshold. In a real system the weights come from sklearn.linear_model.LogisticRegression, and you'd sanity-check with calibration plots before shipping.