AI sales forecasts: plan your restaurant with useful data

Start with the decision the forecast supports

Are you planning next week's staffing, tomorrow's preparation or the monthly finances? These are different tasks. A daily revenue forecast does not automatically tell the kitchen how many portions of a particular dish the lunch rush will require.

Choose a measure, a period and the deadline for producing the forecast. Sales excluding VAT provide a clear financial measure. Dish quantities by service may be more useful to the kitchen. Guest counts require reliable cover records; they are not necessarily the same as receipt counts.

Build an understandable comparison

Start with sales from comparable weekdays and opening hours. Flag anything that makes the history misleading: refurbishment closures, price changes, a new terrace or a different menu.

Add known bookings and events as explicit assumptions. Use information available when the forecast was created. Feeding in the final booking list afterwards gives an unfair impression of how well the model would have supported planning.

Weather may matter for a terrace, but the relationship needs local assessment. Plan weaker- and stronger-demand scenarios when uncertainty is high. Such a planning range is not automatically a statistically validated prediction interval.

Test on sales the model has not already seen

A model can fit historical data well and still forecast poorly. Assess its performance on new periods that were not used to fit the model. Hyndman and Athanasopoulos on forecast evaluation.

Save the forecast before the period begins. Compare both the AI method and the simple historical estimate with the same actual results. Do not change the saved forecast after seeing the answer.

Hypothetical example: forecast sales excluding VAT were SEK 20,000 and actual sales were SEK 18,000. The absolute difference is SEK 2,000, about 11.1 percent of actual sales. One day tells you little; track several comparable periods and check whether forecasts consistently overestimate or underestimate demand. Percentage errors become difficult to interpret when sales are very small or zero.

Turn the forecast into practical decisions

Consider capacity as well as demand. Staff skills, working hours, delivery lead times and ingredient shelf life affect what you can change. Purchasing also requires information about actual stock and consumption; POS sales are not a stocktake.

Decide who can adjust the plan and when it needs to be finalised. Record significant manual changes so you can understand the outcome afterwards.

Use Vendion to support planning

Vendion Analytics helps you follow sales, while the AI assistant supports questions about operational data. Scheduling suggestions can use staffing history alongside sales and bookings; the responsible manager reviews the proposal.

Ask to see which inputs and functions your setup uses. Then assess whether the support improves planning in your own restaurant, without assuming a universal accuracy guarantee.