UPS Location Model

A parcel network wants to grow its drop-box footprint, but “where next” is usually decided by intuition. With almost 30,000 ZIP codes to weigh, the real question is which underserved areas have the demand to justify a new location. We framed it as a binary suitability problem and let the data rank the candidates.

We merged US Census data — income, employment, internet access, housing, age — with a UPS facility dataset, then cleaned 29,943 ZIP codes: standardised fields, converted placeholder tokens to missing, removed ghost-town ZIPs with zero households, and imputed the rest. The features carried business meaning rather than raw counts: a tech-wealth index (income times internet penetration), a total tech volume that separates small rich enclaves from large high-value markets, and an employment-density proxy for daytime foot traffic.

I led the initial UPS research and shaped the analytical flow, then guided data collection, variable selection, and the exploratory analysis, with hands-on coding support. I had a say in model selection — and I owned the part that makes a model matter: translating the outcomes into plain business language, the ethics analysis, and the implications for where and why UPS should expand.

Across Logistic Regression, Random Forest, and XGBoost, XGBoost won at ROC-AUC 0.78 with balanced precision and recall — chosen on ROC-AUC because two thirds of areas are the negative class. Household count, employment density, and the tech-wealth signal drove predictions, so connectivity and wealth mattered more than raw population. Calibration held: predicted probabilities matched actual frequencies across confidence bands.

The model surfaced “missed opportunity” ZIP codes — areas with no current drop box but high predicted suitability. That’s a direct expansion lead list. We paired it with a threshold tuned as a business trade-off and a phased deployment plan (pilot the top candidates, monitor, then scale) rather than treating the score as a verdict. At heart it’s network design: deciding where physical nodes should sit to serve demand efficiently.