📍
Situation
Two cross-border risks owned by two different teams
🎯
Task
Two models, two scores, each with a reason
⚡
Action
Leakage checks · XGBoost & RF · threshold calibration · checked SHAP
📈
Result
Working prototype with a live demo on real test data
Companies buying from abroad carry two separate risks: a supplier failing, which sits with credit and procurement, and goods arriving late, which sits with logistics. Commercial supplier scores exist but are proprietary and rarely explain themselves.
With a four-person graduate team, build two independent classifiers on public data (Taiwan bankruptcy, DataCo supply chain, NY Fed GSCPI), judge them on metrics that suit their class balance, and make every score explainable to a non-specialist.
A working prototype, not a production system. XGBoost reached F1 0.529 on a 3.2% positive class. Calibrating the shipment threshold to 0.40 lifted recall to 0.825 and caught 3,705 more late shipments than the default, at the cost of 4,200 more false alarms. Tuning did not beat the default settings on either layer, and the site says so. Team: Ankit Kumar, Divya Gopal, Aswin Shriram Thiagarajan.
F1 0.529 on a 3.2% positive class (bankruptcy)
Actual · held-out test set, 1,364 firms
Recall 0.825 on late shipments after threshold calibration
Actual · held-out test set, 34,393 orders
+3,705 late shipments caught vs the 0.50 default
Actual · with 4,200 more false alarms