THE GUIDE / PRAEDIXA
Forecasting restaurant sales: method and limitations.
Sales forecasting estimates future activity within a defined scope. For that estimate to inform inventory or staffing, you need to define what is forecast, use information available at the right time and then compare the result with observed sales.
1. Choose the decision and its horizon
Start with the operational question. Preparing lunch ingredients requires a different view from organizing next week’s team. Specify the restaurant, period and quantity being studied: covers, product quantities or revenue.
These measures are not interchangeable. Revenue may rise because prices increase while quantities stay unchanged. Quantities by recipe are particularly useful for ingredient calculations. For organizing service, the timing of demand also matters.
2. Prepare an interpretable history
Gather sales over comparable periods. Identify closures, opening-hour changes, promotions and menu changes. A day with no sales because the restaurant was closed is not the same as an open day with no demand.
Also examine stockouts: observed sales may be limited by unavailable products. Retain information that explains these episodes. The quality of a history is not just its length; it depends on the meaning of its values.
3. Build a simple baseline
Before comparing complex methods, choose an understandable baseline: for example, the same service in the previous week, after reviewing exceptional events. Evaluate more elaborate methods on the same restaurants, dates and forecast horizon.
Weather and events can add context. Use only information known when preparing the forecast. Testing a past service with the weather that actually occurred, when it was uncertain beforehand, can make the result appear too favorable.
4. Measure forecast errors
For each evaluated service, match a dated forecast with observed sales for the same restaurant, product and period. Keep the unit of measurement and the forecast issue date. Absolute error measures the size of the difference without allowing overestimation and underestimation to cancel each other out.
Average the absolute errors over the selected observations. Also examine signed errors, with an explicit convention, to identify a tendency to overestimate or underestimate. A service with missing data must not become a zero-valued observation.
The calculationMean absolute error = sum of absolute differences between forecasts and actuals ÷ number of observations
| Data | Check |
|---|---|
| Dated forecast | Version available before service |
| Observed sales | Same product, period and unit as the forecast |
| Absolute error | Absolute difference between forecast and actual |
| Signed error | Forecast minus actual; sign retained |
5. Test without using future information
Evaluate forecasts on periods not used to build the method. Repeat the exercise by moving forward through time and retaining only information available at each date. This is the principle of chronological evaluation.
Compare horizons and locations separately. An overall average can conceal large errors for particular services. Document the number of observations, covered periods and excluded data before interpreting the reliability of the method.
6. Connect the estimate with a decision
Translate forecast sales into requirements using recipe specifications, then examine available inventory. For scheduling, compare demand with roles, skills and availability. Both decisions start from the same expected activity but involve different constraints.
Keep uncertainty visible: a forecast does not remove unexpected events. Record discrepancies and their context to improve service preparation. Praedixa connects these workflows, while the evaluation method must remain explicit and suited to your data.
REFERENCES