Forecasting, Scheduling, and Ordering in Hospitality

The coupled loop: demand → labor → purchasing

In restaurants and hospitality, three decisions are tightly linked:

1. How many guests/orders will arrive? (forecast)
2. How many labor hours should be scheduled? (staffing)
3. How much inventory should be prepped/ordered? (purchasing + production)

When these decisions are managed in separate tools (or by separate people), the organization usually pays twice:

• Over-forecasting drives overstaffing and overproduction (waste).
• Under-forecasting drives understaffing (service failures), emergency purchasing (higher unit costs), and burnout (churn).

The waste signal is large enough to matter

ReFED estimates that restaurants and foodservice generated 12.5M tons of surplus food in 2024, with nearly 70% attributed to plate waste (food served/taken but not eaten). [1]

You cannot eliminate plate waste purely via forecasting — but forecasting, portioning, and prep planning materially influence the “overproduction” slice and the frequency of stockouts that lead to reactive cooking patterns.

Labor churn is an operational constraint

In BLS JOLTS data, the quits rate for Leisure and Hospitality remains materially higher than many industries; for example, the quits rate in Dec 2025 is listed at 4.5% (seasonally adjusted) versus 2.0% overall. [2]

This has two direct operations implications:

• You are continuously training new staff; skill distributions shift week-to-week.
• Schedule instability has a compounding cost: it degrades retention and forces more last-minute coverage.

Scheduling systems fail when inputs are low quality

Teams often jump to “optimization” (a solver, an AI scheduler) before the basics are true:

• Employee availability is correct and up to date
• Constraints are explicit (labor laws, minor rules, max hours, rest periods)
• Roles/skills are mapped (who can run expo, who can close, who can bartend)
• Demand signals are aligned to work content (orders ≠ labor hours unless you model prep/service mix)

Academic work on human-computer interactions in labor scheduling highlights that managers frequently override AI-generated schedules, spending substantial time doing so and potentially reducing schedule consistency — a reminder that tools don’t remove judgment; they shift where judgment is applied. [3]

Forecasting: what “good” measurement looks like

Forecast accuracy is not “how well the model fits the past.” Hyndman & Athanasopoulos emphasize that you must evaluate on new data not used during fitting, using training/test splits or time-series cross-validation (rolling origin). [4]

Practical accuracy measures (choose 1–2 and standardize)

Common measures include:

• MAE (mean absolute error): interpretable in the unit you forecast (covers/transactions)
• RMSE (root mean squared error): punishes large misses more heavily
• MAPE/sMAPE: unit-free but unstable near zero (use carefully)
• MASE: scaled, robust across series (recommended as an alternative to MAPE) [4]

The key is to use one measure consistently and to report it per daypart and channel (dine-in vs takeout vs delivery), not only in aggregate.

Exception-based management (the highest ROI control loop)

Most managers don’t need a better dashboard; they need a smaller list of decisions that matter today.

Define exceptions like:

• Forecast miss > X% for a daypart
• Labor-to-sales deviates beyond a band
• Waste events exceed a threshold (by item family)
• Out-of-stock events exceed a threshold

Then route each exception to an owner with a short playbook (“if this happens, do that”).

Manual vs. system-driven operations

Implementation sequence (what to do in the first 60 days)

Weeks 1–2: instrumentation baseline

• Standardize definitions: sales, covers, orders, labor hours, waste events
• Create a daily extract (even CSV) with:
  • per store/daypart demand
  • scheduled hours by role
  • realized hours (timeclock)
  • waste events (count + cost proxy)

Weeks 3–6: forecasting + measurement discipline

• Select a baseline forecast (seasonal naive / last-week same-daypart)
• Evaluate out-of-sample per Hyndman & Athanasopoulos’ guidance [4]
• Track forecast error by channel and daypart

Weeks 7–8: scheduling hygiene + controlled overrides

• Implement availability refresh rules
• Introduce role/skill tagging
• Log overrides as events (who, when, why) so you can learn where the tool is wrong [3]

Weeks 9–12: ordering + waste control

• Tie prep/ordering to forecast distributions, not point estimates
• Use ReFED’s category framing to target waste causes (plate waste vs overproduction) [1]
• Implement a “top 10 waste items” weekly review with one intervention per week

Next steps

If you want a system that improves margins without burning out managers, start by instrumenting the coupled loop, measuring forecasts correctly, and treating scheduling inputs as master data.

References

1. ReFED: Restaurants and Foodservice — 2024 surplus and causes (12.5M tons surplus; plate waste share)
2. BLS JOLTS Table 4: Quits levels and rates by industry (2025 M12) (Leisure & hospitality quits rate context)
3. Kwon, Raman, Tamayo (HBS Working Paper, 2024): Human-Computer Interactions in Demand Forecasting and Labor Scheduling Decisions (PDF) (Overrides, manager time cost, schedule consistency concerns)
4. Hyndman & Athanasopoulos: Evaluating forecast accuracy (Forecasting: Principles and Practice) (Out-of-sample evaluation; MAE/RMSE/MAPE/MASE)
5. ReFED: 2024 Food Waste Report (PDF) (System-wide context; definitions; sector framing)
6. Google SRE: The Art of SLOs (handbook PDF) (SLO-style thresholds and error-budget thinking for exceptions)