Hourly Demand + Cross-Sell Mining — Florida Bakery Group

Flat staffing and production plans across all dayparts despite a 5x intraday demand swing.

• Production decisions used weekly averages, missing the structural shape of demand within a day
• Morning rush regularly stocked out of high-velocity items while afternoon hours wasted prep
• No system to detect non-obvious item-to-item purchase associations across 3 years of order history
• Operator intuition could not scan 685,000+ line items for cross-sell patterns at scale

Built an hourly-granularity demand forecasting model on 36 months of POS data and ran association-rule analysis to surface previously-invisible cross-sell relationships. The pilot concluded when the client sold the business; the methodology — hourly demand forecasting + market-basket lift mining + service-level/waste frontier modeling — is preserved as a reusable template for future retail/food engagements.

• Cleaned and normalized 685K+ line items across 176,000+ orders covering 36 months of activity
• Trained a per-hour demand forecast for top-volume items (top savory item: ~15 units/hr AM peak vs ~3 units/hr PM trough)
• Recommended the Conservative strategy variant (65% service-level target) — the production-recommended posture from a service-vs-waste trade-off frontier
• Ran market-basket / lift-ratio analysis on 685K line items to surface co-occurrence patterns invisible at the operator level
• 147-day backtest window with holdout validation for forecast accuracy reporting

The demand model captured intraday shape that flat human planning had missed for years. The market-basket analysis surfaced a 2.6x co-occurrence lift between eggs and the classic latte — 666 corroborating co-purchases across 176,000 orders, strong enough to redesign the morning bundle. The numbers below are from the backtest deliverable; the methodology and code are now templated for future retail/food clients.

“The eggs-to-coffee association lived in 685,000 line items as 666 corroborating co-purchases. No human review of monthly summaries would have found it. That is the gap between dashboards and models — dashboards report what happened; models surface the patterns nobody was looking for.”

— DMG engagement notes — Florida QSR bakery pilot

Timeline

4 weeks data ingestion and cleaning, 6 weeks model training and backtest, 2 weeks delivery and handoff. 147-day holdout window for forecast validation.