Retail Stockout Prediction

Over half of the transactions in the dataset — 51.86% — ended in a stockout, proving existing replenishment logic couldn't anticipate demand spikes or supply delays. The goal: predict high-risk store-product combinations before they stock out, then allocate a fixed restocking budget against them. A 4-person team project; I was the EDA & Visualization Lead.

• Led the exploratory analysis across 5 stores × 8 products, surfacing the risk patterns that shaped the model — including the counterintuitive ones.
• Found that rainy (54.0%) and cloudy (53.2%) days drive more stockouts than sunny ones (49.1%) as customers move indoors, that Saturday peaks at 54.2% (mid-week replenishment misses weekend traffic), and a 'buffer illusion' — stockout vs non-stockout transactions had nearly identical inventory buffers (151 vs 155), proving simple threshold rules fail.
• Built the visualizations and insight summaries that justified tree-based modeling over linear baselines and a reconstructed, operationally-meaningful target.

• The team's tuned XGBoost hit 0.77 AUC and 76.6% recall — recall-prioritized, because missing a real stockout costs more than an early reorder.
• A Gurobi optimizer allocated 1,685 restocking units across 11 priority stores within a $32,836 budget, turning risk scores into a concrete reorder plan.
• Final recommendation: a Score-Monday → Optimize-Tuesday → Order-Wednesday workflow that converts reactive restocking into predictive inventory management.