Hospitality · Growing businesses
Rostering a café from the people walking past the door
A week-ahead forecast of street footfall for a café roster, tested on two years of hourly City of Melbourne pedestrian counts on three café streets.
Executive summary
A week-ahead footfall forecast misses a café street’s trading hours by 14.7% against 18.0% for copying last week, and by 14.1% once the weather is added the day before.
Situation
Café rosters are written on Thursday by copying last week. Holidays, school breaks and the university calendar move the street more than the weather, and last week knows none of them.
Approach
Gradient boosting on two years of hourly City of Melbourne pedestrian counts for three café streets, trained on the first year and tested once on the second, every forecast using counts at least a week old.
Result
The forecast beats copying last week in every month of the test year, by 3.3 points on Errol Street. A four-week average gets most of the way; the forecast adds most on public holidays.
The problem
A café or small hospitality group writes next week’s roster on Thursday, mostly by copying last week. Rostering tools such as Deputy and Tanda can show forecast sales beside the roster, and Australian venues use them to keep wages near a quarter of takings, but they need a forecast to start from. Public holidays, school terms and the university calendar move the street more than the weather does.
I wanted to measure how much better a roster could start if it began from a forecast of the people walking past the door, set a week ahead from free data, and how much a day-before weather update adds.
- Team
- Owner-operator or venue manager, 6 to 20 casual staff
- Scale
- One to five venues on a shopping strip or in city laneways
- Systems
- Square, Lightspeed or Toast; Deputy or Tanda for rostering
- Measure
- Staff hours set to the wrong level for the trade; wage cost as a share of takings
What I built
A weekly job reads the City of Melbourne’s hourly pedestrian counts for the sensor nearest the venue, adds the public-holiday calendar, and forecasts each trading hour of the week ahead. The forecast goes into the rostering tool as a demand line beside the roster, so the manager sees which shifts are thin or heavy before publishing. On the day before, the forecast is refreshed with the weather.
Results
I trained on the first year of counts, from November 2024, and tested on the following year to 17 November 2026, every trading hour from 7am to 4pm, on three streets with cafés: Errol Street in North Melbourne, Lygon Street in Carlton and Degraves Street in the city laneways.
Against copying last week the forecast is 3.3 points better on Errol Street, with a 95% interval of 2.5 to 4.3 points from resampling the test days. Against a plain four-week average the gain is under a point, which is the result I would act on: a café without any modelling should roster from the four-week average, and the forecast earns its place on public holidays, where copying last week misses by 42%.
| Street | Passers-by an hour | Last week | Four-week average | Forecast | Day-before update | Daily total: last week | Daily total: forecast |
|---|---|---|---|---|---|---|---|
| Errol Street, North Melbourne | 168 | 18.0% | 15.2% | 14.7% | 14.1% | 11.1% | 9.0% |
| Lygon Street, Carlton | 229 | 24.5% | 19.2% | 18.6% | 16.7% | 18.9% | 13.8% |
| Degraves Street, city laneways | 832 | 14.4% | 12.2% | 12.0% | 11.2% | 10.2% | 8.2% |
| Street | Public-holiday hours: last week | Public-holiday hours: forecast |
|---|---|---|
| Errol Street | 41.9% | 29.6% |
| Lygon Street | 37.9% | 34.3% |
| Degraves Street | 29.1% | 23.8% |
The daily total, which sets how many staff are on, is easier to forecast than each hour: on Errol Street the forecast misses the day by 9.0% against 11.1% for last week.
Data and method
- Data
- City of Melbourne Pedestrian Counting System, hourly counts, 19 November 2024 to 17 November 2026, CC BY 4.0; Open-Meteo historical weather for Melbourne, CC BY 4.0
- Streets
- Sensors Errol23_T, Lyg161_T and FLDegS_T, trading hours 7am to 4pm
- Split
- Train on the first year, test once on the second; every forecast uses counts at least seven days old
- Methods
- Gradient boosting with an absolute-error objective on hour, weekday, month, public holiday, the same hour one to four weeks back and their average; the day-before version adds temperature and rain
- Error
- Weighted absolute percentage error: total hourly miss divided by total passers-by
Limitations
- Passers-by are not customers. The share who buy differs by venue and hour, so the forecast is checked against the venue’s own till data before it sets shifts.
- The day-before update uses observed weather in place of the forecast issued that day, so its gain is an upper bound.
- Sensors cover the City of Melbourne. A venue elsewhere starts from its own till counts, which carry the same weekly pattern.
- Events, roadworks and sensor outages move single days; a known-event list fixes the first, and outages are flagged and skipped.
How I would take it live
- Connect the nearest sensor and twelve months of till transactions by hour
- Check the forecast against the till for a month and set the buy rate by hour
- Show the demand line in Deputy or Tanda for one venue for six weeks
- Compare wage share of takings with the same weeks last year, then extend