Retail & Restaurant
Flat production plans miss a 5x intraday demand swing. Operator intuition misses cross-sell signals buried in 685,000 line items. Our retail/QSR methodology — built and validated on a prior pilot with a Florida bakery group — is preserved as a reusable template for future retail and food clients.
Three Signal Categories We Mine
Demand shape, cross-sell associations, and the service-level / waste trade-off frontier. Each one is a distinct decision frame — and each one is invisible without the right granularity.
Hourly demand forecasting
Most production plans use weekly averages. But retail demand has a structural shape inside each day — morning rushes, mid-afternoon troughs, end-of-day spikes. Our models forecast at hourly granularity for top-velocity items, so production matches the curve.
Real example
Top savory item hourly model: peak ~15 units/hr in the morning rush, trough ~3 units/hr in the afternoon. The 5x swing was invisible to flat human planning across 147 holdout days.
Association-rule cross-sell mining
Which items reliably move together? Operator intuition catches the obvious pairs and misses the high-lift ones. Market-basket / lift-ratio analysis on 6-12+ months of POS data surfaces the non-obvious associations that justify menu redesigns and bundle pricing.
Real example
Eggs ↔ classic latte: 2.6x lift, 666 corroborating co-purchases across 176,000+ orders. Strong enough to redesign the morning bundle. Invisible at the daily-summary level.
Service-level / waste trade-off frontier
Higher service level = less stockouts = more waste. Lower service level = less waste = more stockouts. The trade-off is a frontier, not a number. We model the frontier per-item and let you pick the operating point — Conservative (65% service level) is the production-recommended default.
Real example
On a 36-month POS dataset, the Conservative posture cut waste materially without exceeding the operator-tolerable stockout floor on top-velocity items.
The Engineering Underneath
Modeling is the visible part. The infrastructure that makes it survive real POS data — schema mapping, item-name canonicalization, holdout validation — is what separates a usable forecast from a dashboard nobody trusts.
Modeling at the right granularity
Daily aggregates hide hourly demand shape. Monthly aggregates hide weekly seasonality. We model at the granularity that exposes the operating decisions you actually make — hourly for production / staffing decisions, weekly for purchasing, monthly for menu planning.
POS-format-agnostic ingestion
Toast, Square, Clover, Lightspeed, and the long tail of restaurant POS systems all export differently. We accept order-level transactional data in whatever format your provider gives us and handle the schema mapping internally — typically 1-3 days from raw export to first model run.
Holdout-validated forecast accuracy
Every forecast model ships with a holdout-window evaluation: MAPE, RMSE, and confidence intervals against data the model never saw during training. We will not deploy a model that fails the holdout test, and we will tell you if your data does not support the accuracy you need.
Item-name normalization that survives operator drift
POS systems accumulate item-name variants over time: the same savory item can appear as its full menu name, an abbreviation, and an all-caps register alias. Without normalization, the same product looks like five different items in the analysis. We canonicalize before any modeling runs.
Florida QSR bakery group — 36 months of POS data
176,000+ orders and 685K+ line items processed. Hourly demand model captured a 5x intraday swing on top-velocity items. Market-basket analysis surfaced an eggs ↔ classic-latte association at 2.6x lift with 666 corroborating co-purchases — strong enough to drive a morning-bundle redesign. The engagement concluded; the methodology + pipeline are preserved as a reusable template for future retail/food clients.
Line items processed
685K+
Cross-sell lift discovered
2.6x
Frequently Asked Questions
What POS systems do you support?
We have ingested data from Toast, Square, Lightspeed, Clover, and custom in-house systems. Anything that can export order-level transactional data in CSV / JSON / Parquet is supportable. If your POS supports API access, that is preferred over manual exports for ongoing pipelines.
How much POS history do you need?
Hourly demand forecasting needs 12+ months minimum to capture seasonal patterns. Market-basket / cross-sell analysis can produce useful signal with 6+ months of moderate-volume data. More history improves accuracy and surfaces longer-cycle patterns (e.g., yearly seasonality) but at diminishing returns past 24-36 months.
Do you handle multi-location operations?
Yes. Multi-location operations benefit from cross-store comparison — surfacing which item-pairing patterns generalize and which are location-specific. The same modeling infrastructure handles a single-store pilot and a 50-store rollout; the difference is data volume and the addition of location-as-a-feature.
How do we put predictions into production?
Forecast output ships as either (a) a scheduled report your kitchen / floor team can act on each morning, or (b) a real-time scoring API your POS / production system can call. The reporting path is faster to deploy; the API path enables tighter operational integration. We will recommend the right architecture for your team's technical capacity.
Curious what is hiding in your POS history?
A discovery call takes 30 minutes. We will look at what your data shape will support and tell you honestly whether modeling is the right next move — or whether you would get more value from dashboard cleanup first.
Engagement numbers refer to DMG's anonymized prior pilot with a Florida QSR bakery group (engagement concluded; methodology preserved as reusable template for future retail/food clients). Cross-sell lift and demand-shape figures are validated against held-out POS data across a 147-day window. Full methodology.