Forecasts: Read the Range Around the Number

Ask what the central number represents

A forecast of 100 covers looks straightforward until two people use it differently. One treats it as the most likely outcome; another assumes there is an equal chance of finishing above or below it. Those interpretations can differ. Ask whether the output is a mean, a median or another quantile, and confirm the service window and target. Precision in the display should not exceed the practical precision of the information.

A central estimate is useful for coordination, but it is not a promise. Some variation reflects factors that were not observable at the planning deadline, such as late cancellations. Other uncertainty comes from limited history, measurement problems or an unfamiliar operating regime. A model cannot reliably eliminate a gap in knowledge merely by producing more decimal places.

Distinguish scenarios from prediction intervals

A lower, central and upper scenario can structure a briefing even when no probability has been estimated. Label it as a scenario exercise. A prediction interval makes a stronger claim about future outcomes. For example, an interval intended to cover 80% of comparable services should contain the observed outcome on roughly that proportion of services over a sufficiently informative evaluation sample, if well calibrated.

That is not a guarantee for the next dinner, nor does it mean every outcome inside the interval is equally likely. A confidence interval for an estimated average is also not the same as a prediction interval for a particular service. Ask a supplier which quantity is being described and how its coverage was tested. Do not borrow probability language for an untested best-case and worst-case pair.

Compare against a credible simple baseline

Start with a transparent reference, such as recent comparable weekdays adjusted for known closures. A seasonal naive forecast, which repeats an appropriate previous seasonal observation, is another useful benchmark. Statistical models, gradient-boosted trees and neural networks all need to demonstrate improvement against that reference under the same conditions. Complexity is not evidence of operational value.

Evaluate in time order. Train on earlier services, forecast a later period, then move the cutoff forward. This rolling-origin approach reflects repeated real decisions more faithfully than randomly mixing past and future records. Fit preprocessing steps using training information only. If you test many model variants, keep a final holdout period separate from the tuning process.

Match uncertainty to the horizon

A forecast made several weeks before a service usually has different information from one made that morning. Report results separately by horizon rather than blending them into one flattering accuracy figure. Booking pickup can narrow uncertainty as the date approaches, but a late large-party cancellation may still create a material change. More information does not guarantee that every update is closer to the final outcome.

Forecasting at shorter intervals introduces another trade-off. Half-hour arrival estimates may support role coverage, but they are noisier than service totals. Independently estimated quantiles cannot simply be added to obtain a valid quantile of the total. If totals and intervals disagree, discuss aggregation and dependence with the analyst rather than quietly choosing whichever number suits the plan.

Watch calibration and usefulness together

Coverage alone is insufficient. An interval spanning nearly every possible outcome could achieve high coverage while providing little planning help. Review interval width alongside coverage and use an interval scoring method that penalizes both excessive width and missed outcomes. Inspect performance on relevant groups such as weekend dinners, terrace services and known event days, while acknowledging the small sample size of rare categories.

Also review signed errors. A model that repeatedly underestimates busy services may look acceptable on average yet put pressure on the same teams. Avoid using percentage error as the sole metric when actual covers can be zero or very small. Mean absolute error, bias and operational examples usually provide a more interpretable starting conversation.

Turn the range into a conditional plan

Do not simply staff to the top of every range or force the forecast upward to justify a preferred rota. Keep the demand estimate separate from the decision. Discuss what the lower, central and higher cases would mean for seating pace, preparation, breaks and role support. Identify a trigger for revisiting the plan and who has authority to act.

For example, a late increase in confirmed large-party arrivals may justify a briefing change, but any staffing adjustment still needs to respect availability, contractual terms and notice requirements. Where flexibility is limited, change the operating plan rather than assuming workers can absorb uncertainty without cost. The purpose of a range is to make trade-offs explicit, not to shift risk invisibly onto the team.