Predict how many guests come each hour
A five-step method for one independent restaurant, using the checks your register already saves and a spreadsheet. Each step has a worked example from an invented New York bistro.
The five steps
From your register, export the number of checks (or covers) for each hour of each day. Leave out nights you were closed or ran a private event. They skew the average.
Example: A 40-seat bistro in Williamsburg exports ten weeks. Its Tuesday dinner reads 6 PM, 7 PM and 8 PM, one row per Tuesday.
checks-by-hour.csv
Checks for each hour of each Tuesday Tuesday 6 PM 7 PM 8 PM Sep 2 15 18 14 Sep 9 17 20 15 Sep 16 16 21 16 Sep 23 18 22 17 Sep 30 Closed, left out One row per Tuesday, ten weeks in all. Example figures.
Compare Tuesday with Tuesday, never with Saturday. The median of the last six to eight same weekdays ignores one odd night that an average would chase.
Example: Its last eight Tuesdays at 7 PM had 18, 20, 21, 22, 22, 23, 25 and 41 checks (the 41 was a birthday party). The median is 22; the average would say 24.
Eight Tuesdays at 7 PM
The median is 22. The party night does not move it.
18, 20, 21, 22, 22, 23, 25 and 41 checks. Example figures.
Find how rain changed your own past nights. Divide the median of rainy nights by the median of dry nights for the same hour. Use your own ratio.
Example: Rainy Tuesdays at 7 PM ran at 0.85 of dry ones. Rain is forecast, so 22 becomes 19.
Tuesday, 7 PM
Median
22
× 0.85
Plan for
19
Rainy Tuesdays ran at 0.85 of dry ones. Use your own ratio. Example figures.
Mark US holidays, school breaks and anything that fills or empties your block: a game, a parade, a street fair, a subway line closed for the weekend. Check what the same night did the last time it happened.
Example: A street fair on the block last June doubled the 6 PM hour and emptied 9 PM. That night the bistro plans from last June’s hours instead of the median.
The night of the street fair
- 95 PM
- 326 PM
- 227 PM
- 198 PM
- 09 PM
- 610 PM
Last June’s fair doubled 6 PM and emptied 9 PM. Example figures.
Write a low and a high number for each hour. After each shift, write the real count beside it. If more than one night in ten falls outside the range, widen it. Track WAPE, the total misses divided by the total guests, week by week.
Example: It plans 7 PM as 16 to 23 checks and staffs for the top of the range. The real count was 21.
Tuesday, 7 PM, scored
- Planned
- 16 to 23
- Staffed for
- 23
- Came
- 21
Inside the range. Widen it if more than one night in ten falls outside. Example figures.
checks-by-hour.csv
| Tuesday | 6 PM | 7 PM | 8 PM |
|---|---|---|---|
| Sep 2 | 15 | 18 | 14 |
| Sep 9 | 17 | 20 | 15 |
| Sep 16 | 16 | 21 | 16 |
| Sep 23 | 18 | 22 | 17 |
| Sep 30 | Closed, left out | ||
One row per Tuesday, ten weeks in all. Example figures.
Eight Tuesdays at 7 PM
The median is 22. The party night does not move it.
18, 20, 21, 22, 22, 23, 25 and 41 checks. Example figures.
Tuesday, 7 PM
Median
22
× 0.85
Plan for
19
Rainy Tuesdays ran at 0.85 of dry ones. Use your own ratio. Example figures.
The night of the street fair
- 95 PM
- 326 PM
- 227 PM
- 198 PM
- 09 PM
- 610 PM
Last June’s fair doubled 6 PM and emptied 9 PM. Example figures.
Tuesday, 7 PM, scored
- Planned
- 16 to 23
- Staffed for
- 23
- Came
- 21
Inside the range. Widen it if more than one night in ten falls outside. Example figures.
Where Jevan fits
- Step 1JevanReads every check by hour
- Step 2JevanTakes the same-weekday median
- Step 3JevanAdds the weather forecast
- Step 4YouEvents, parades, subway closures
- Step 5JevanScores each night's 90% range
Jevan runs steps 1, 2, 3 and 5 every night from the checks your register saved. It has orders for each day and hour, with a 90% range from its own past misses, ready by 5 AM.
It adds the weather forecast and US holidays. It does not read local events or subway service, so step 4 stays yours. It needs 28 consecutive saved days, and it imports past sales when you connect the register.
Questions owners ask about hourly forecasts
- How many weeks of sales do I need to predict guests per hour?
- About eight weeks. That gives six to eight nights of the same weekday for each hour, enough for a median that one odd night does not move. Jevan's own forecast starts after 28 consecutive saved days.
- Should a restaurant forecast covers or checks?
- Use the number your register counts reliably. Every register counts checks. Covers depend on servers entering the party size. Pick one and keep it, or the history mixes two different measures.
- How do I know if my hourly forecast is good?
- Add up the misses for a week and divide by the total guests. That is WAPE, and 100 minus WAPE is demand accuracy. Also count how often the real number fell inside your range. A 90% range should hold about 9 nights in 10.
Jevan does this every night, before you unlock the door