Preserve what was known before service
A fair review begins with the forecast that actually informed the decision. Store the target, issue time, horizon, model version, input snapshot and any human adjustment. Keep the unadjusted estimate too. If the forecast is updated during the day, retain each version rather than overwriting the previous one. Otherwise, a late forecast can accidentally receive credit for an earlier staffing choice.
Use a stable definition of the actual outcome. If the target was seated covers at dinner, compare it with seated covers at dinner, not all-day transactions. Reconcile corrections through a visible process. A change in how the host records children or walk-ins can look like model drift even when guest behaviour has not changed.
Calculate a small set of complementary measures
Define signed error consistently. In this guide, error means actual covers minus forecast covers, so a positive value indicates underforecasting. Absolute error removes the sign and describes the size of the miss. Across services, mean absolute error gives an interpretable average magnitude, while mean signed error indicates directional bias. Neither alone describes every operational concern.
Suppose a purely illustrative forecast was 100 covers and actual seated covers were 118. The signed error is +18 covers and the absolute error is 18 covers. That single observation cannot establish model quality. Review repeated services and show the number of observations. Percentage errors are problematic when actual covers are zero or very small, so do not make them the only headline measure.
Compare like with like
Use the same services, issue times and target definitions when comparing a model with a baseline. A morning forecast with near-final bookings should not compete against a forecast issued several days ahead. Report separate results for the horizons that support purchasing, scheduling and pre-service decisions. Avoid quietly excluding difficult services after seeing the errors.
Break results into operationally meaningful groups such as weekday lunch, weekend dinner, private events or terrace services. Small groups produce unstable summaries, so retain the sample count and avoid strong claims based on a handful of nights. Review both aggregate totals and the peak intervals that create pressure. Errors in opposite directions can cancel at service level while still causing a difficult evening.
Evaluate ranges as well as central estimates
If a model supplies prediction intervals, record whether actual outcomes fall inside them and how wide the intervals are. An interval targeting 80% coverage should be assessed over enough comparable forecasts to make the observed proportion informative. A miss outside the interval is not automatically a failure; occasional misses are expected. Persistent undercoverage, however, suggests understated uncertainty.
Also check whether the range is useful. A very wide interval may cover outcomes but offer little guidance. Interval scores can balance width against misses. Do not judge scenario ranges by probabilistic coverage unless they were explicitly designed and calibrated as prediction intervals. Keep those labels consistent in reports and briefings.
Separate forecast error from operating changes
An equipment failure, an early closure or a decision to stop accepting walk-ins changes the relationship between demand and served covers. Record the event without silently removing it from the main review. Consider a secondary analysis of comparable unrestricted services, with the exclusion rule stated in advance. This preserves a view of real-world performance while helping diagnose the cause.
Discuss staffing outcomes separately. A busy service that ran well may reflect an effective contingency, not an accurate forecast. A correct cover total can still conceal an arrival spike or unexpected menu mix. Ask hosts, kitchen leads and floor teams where pressure appeared, which assumptions failed and whether the planned breaks and close-down work remained feasible.
Make one testable improvement
Choose a specific change linked to evidence: repair booking snapshots, add a reliable event flag, revise an opening-hours field or test a different interval method. State the expected benefit and the evaluation window before making the change. Retain the simple baseline and compare on future services. Resist replacing the whole process after one conspicuous miss.
Watch for structural change after a menu redesign, refurbishment, pricing change or new service channel. Old patterns may no longer represent the restaurant. Review input distributions and errors together, and decide whether recent data should carry more weight. A drift alert is a prompt for investigation, not an instruction to retrain automatically.
Close the learning loop with the team
A short review can contain the original estimate, actual result, signed error, key service events and one agreed action. Assign an owner and a date to check the action. Keep guest and employee identifiers out of the note unless they are genuinely necessary and lawfully handled. The goal is an explainable operating record, not a surveillance archive.
Over time, the review should improve both the forecast and the decision made around it. Better data may narrow uncertainty; clearer briefings may improve responses even when accuracy barely changes. Count those as distinct improvements. A forecast earns trust when its limits are visible and its mistakes lead to a disciplined next step.