Revenue, orders and average order value answer different questions.
Revenue can rise because more orders were completed, because the average order value increased, or both. Forecasting only revenue makes it harder to diagnose a miss. A useful restaurant sales view forecasts revenue and order count separately, then derives average order value as a consistency check.
Channels add another layer. Dine-in, takeaway, direct delivery and aggregators can have different demand patterns and economics. Channel-level forecasting is worthwhile only when channel attribution is complete and comparable over time.
Define “sales” before modelling it.
A POS may expose gross sales, item subtotal, discounts, service charge, tax, refunds, voids, tips and net settlement. Choose the measure that matches the decision and document it. Daily totals should follow the restaurant’s business-day cutoff, which may extend beyond midnight.
| Field | Decision question | Validation |
|---|---|---|
| Net revenue | How much trading revenue is expected? | Reconcile daily total to the POS close report. |
| Completed orders | How many transactions must operations serve? | Exclude cancellations and define split/merged bills. |
| Item quantity | What will guests buy? | Handle modifiers, combos, returns and complimentary items. |
| Channel | Where will demand arrive? | Reconcile POS and aggregator totals. |
Add outlet and daypart only when they change an action.
An all-outlet total is useful for ownership, but outlet managers need their own forecast. Dayparts can support staffing and batch preparation when order timestamps and trading-hour definitions are reliable. Avoid multiplying sparse series unnecessarily: a low-volume item split across outlet, channel and fifteen-minute interval may be too noisy for a useful forecast.
Use a hierarchy. Forecast or reconcile from item and outlet levels to consolidated totals, and check that grouped values conserve the same bounded transaction population.
Worked example: seven completed days.
Across the latest seven days in the synthetic dataset, the three outlets recorded approximately ₹1,053,603 from 1,768 completed orders, an average of about ₹596 per order.
Jubilee Hills has the highest seven-day revenue, while Gachibowli has more completed orders. That difference is visible because the view keeps order count and average order value alongside revenue.
Evaluate the components, not only the total.
Use MAE or WAPE for interpretable revenue and order error, and inspect bias to see whether the model consistently over- or under-forecasts. Report results by outlet and important calendar segments. Percentage metrics can behave badly on low-volume days, while one large outlet can dominate an aggregate absolute error.
For a daypart model, test whether the extra granularity improves the staffing or preparation decision. Better technical accuracy at a finer grain is not automatically operationally useful.
Use the forecast as a planning envelope.
Revenue informs targets and cash expectations; orders inform staffing context; item demand informs kitchen preparation. Do not convert revenue directly into ingredient purchasing without the menu and recipe layers in between.
When the forecast changes materially, show why: new actuals, calendar changes, override or model version. A manager should not have to infer whether the business changed or the system changed.