E-Commerce RTO Prediction with Random Forest

This project builds a machine learning pipeline to predict Return-to-Origin (RTO) risk for e-commerce orders before delivery.
The focus is on real-world ML practices such as leakage detection, threshold optimization, and explainability, rather than inflated accuracy.

---

Problem Statement

Given order-level attributes (payment method, discount, category, location, time), predict whether an order will result in RTO.

Reducing RTO helps lower:

• reverse logistics costs
• inventory blocking
• delivery inefficiencies

---

Key Insights (Directional Patterns)

• Cash-on-Delivery (COD) orders show higher RTO risk, likely due to lower delivery commitment.
• High discounts increase RTO likelihood non-linearly, especially when combined with COD.
• Certain product categories exhibit structural RTO risk, driven by size/fit or quality expectations.
• Geographic context matters more than order value alone, highlighting operational challenges.

These insights were validated using SHAP explainability.

---

Model & Approach

• Model: RandomForestClassifier
• Encoding: One-hot encoding
• Tuning: RandomizedSearchCV
• Decision Rule: Custom probability threshold (not default 0.5)
• Explainability: SHAP (global + local)

The model outputs probabilities, which are converted into final predictions using a business-aligned threshold.

---

Evaluation Results (Test Set)

After removing leakage and tuning the decision threshold:

• ROC-AUC: ~0.65
• Accuracy: ~0.68
• F1 (Returned / RTO): ~0.60
• Recall (Returned / RTO): ~0.60

These results reflect a challenging, low-signal problem and demonstrate honest model behavior rather than overfitting.

---

Key Learnings

• Feature engineering matters more than hyperparameter tuning
• Threshold tuning exposes weak models early
• Static order-level features alone provide limited signal
• Interpretability is critical for operational decision-making

---

Future Improvements

• Add customer-level and location-level historical aggregates
• Introduce time-based behavioral features
• Compare with gradient boosting models (XGBoost / LightGBM)
• Deploy as an internal logistics risk-scoring tool

---

Disclaimer

This project demonstrates methodology and decision logic using anonymized or public data and is intended for learning and portfolio purposes.