Questions from the pass
FAQ: Answers Before the Next Shift
Short answers for the briefing. Follow the planning-cycle articles when you need the assumptions behind them.
Is AI necessary to start forecasting?
No. Begin with a transparent comparable-service baseline. A more complex model should improve forward-looking evaluation at the same planning horizon.
How much history do we need?
There is no universal minimum. You need comparable services for the patterns being estimated, reliable definitions and a separate evaluation period.
A full calendar year may expose annual events but still contain too few examples of a rare service type. Changed opening hours or menus can make older records less comparable.
Can bookings stand in for covers?
Not without qualification. Bookings can include cancellations and no-shows and omit walk-ins. Keep booking snapshots and realized seated covers separate.
Should we include weather and local events?
Only when the information was available at the planning deadline and improves a fair forward-looking test. Relevance varies by restaurant and service.
Use the weather forecast available then, not the later observed weather. An event near the restaurant is not automatically evidence of additional covers.
What does an 80% prediction interval mean?
If well calibrated for comparable services, it is intended to contain future outcomes about 80% of the time over repeated forecasts. It is not a guarantee for one service.
A scenario range with no estimated probabilities should not be given that label. Read the uncertainty guide →
Can the model write our rota?
A demand estimate does not establish a lawful or workable rota. Human review must account for tasks, skills, breaks, availability, contracts and applicable requirements.
Which accuracy measure should we use?
Use absolute error for magnitude, signed error for bias and interval coverage with width for uncertainty. Review operational consequences and comparable horizons too.
Percentage error alone is unreliable around zero covers. Build a review routine →
What if the restaurant is new?
Use explicit scenarios and provisional assumptions, record outcomes consistently and avoid presenting borrowed patterns as locally validated accuracy.
Related venues can inform a starting hypothesis, but differences in catchment, service format and pricing still need local review.
Does a quiet service prove we overstaffed?
No. Preparation, cleaning, training, supervision and essential coverage still matter. Judge the plan against information available before service.
Review the quality of the decision and its assumptions rather than treating hindsight as information the manager should have had.
Can I upload my restaurant data here?
No. This is an information website. Its contact tool creates a local plain-text draft; it does not upload records, run forecasts or send messages.
Please do not send guest lists, employee records or other unnecessary personal information. Draft a general question →