Know Thursday will be brutal, before you build Tuesday’s roster.
Hotel demand forecasting software predicts arrivals, departures, occupancy and covers, per department, per hour, from the live pickup in your PMS (OPERA Cloud via OHIP, Mews, Cloudbeds). Manual 90-day forecasts run roughly 28% error; PMS-driven AI forecasting cuts that by about 20 points and refreshes as pickup changes, so Tuesday’s roster is built on what Thursday will actually look like.
How wrong is a spreadsheet forecast? About 28% at 90 days.
Only 28% of hotels run a dedicated revenue management system (Skift Research), the rest forecast in Excel, and the error grows with every week of horizon. Staffing budgets, casual-pool sizing and rosters all inherit that error: 60% of shifts end up overstaffed (Unifocus, 2024) while the real peaks queue. Every point of forecast error is paid twice, once in idle hours, once in review scores. Run your own rooms and wages through the hotel labor savings calculator to see the size of the leak.
The data behind this chart
| Forecast horizon | Manual forecast error (MAPE) |
|---|---|
| 7 days | 8% |
| 14 days | 11% |
| 30 days | 15% |
| 60 days | 22% |
| 90 days | 28% |
One forecast per department, per hour, rebuilt as pickup changes
The forecast is built on your actual PMS data, not a template: internal historical data (up to two years of reservations, occupancy and covers) plus the real-time bookings and pickup flowing in today. On top of that it blends external signals, weather, day of week, public holidays and local events, because a rainy Songkran Tuesday does not book like a dry one. The result is demand per department, per hour: check-ins at the desk, room turns for housekeeping, covers at breakfast. When a tour block books 80 rooms on Thursday night, Friday morning’s forecast already shows it. And where you load a budget, actuals and forecast are shown against it, so variance is visible per department before the month closes, not after.
The same forecast feeds AI hotel staff scheduling: rosters are drafted from it, and published shifts get flagged the moment pickup drifts away from them. Connection details per system: Oracle OPERA Cloud OHIP integration · Mews integration · Cloudbeds integration.
The data behind this chart
| Day | Actual covers | HotelCadence forecast | Manual forecast |
|---|---|---|---|
| D1 | 230 | 244 | 310 |
| D2 | 195 | 210 | 310 |
| D3 | 410 | 385 | 310 |
| D4 | 380 | 357 | 310 |
| D5 | 240 | 254 | 310 |
| D6 | 210 | 198 | 310 |
| D7 | 285 | 268 | 310 |
| D8 | 420 | 395 | 310 |
| D9 | 250 | 265 | 310 |
| D10 | 215 | 203 | 310 |
| D11 | 235 | 249 | 310 |
| D12 | 395 | 372 | 310 |
| D13 | 415 | 390 | 310 |
| D14 | 300 | 282 | 310 |
Worked example, figures indicative, 250-room Bangkok property in tour-group season.
Can your managers trust a forecast they didn’t build?
Every number opens into the pickup behind it: which segments, which group blocks, which day-of-week pattern. F&B managers see covers by arrival mix; housekeeping sees departures and stay-overs. The conversation changes from “the system says 340” to “two tour groups check out Tuesday.”
Accuracy is measured the way savings are measured: tracked against your actuals every week and visible in the same report you see, not claimed once in a sales deck. Guest service stays the constraint, not the casualty: the forecast exists so peaks are staffed, not just so quiet hours are cut.
Tokyo, business hotel: Golden Week housekeeping finally tracked departures
A 30-day-out manual forecast (~15% error) missed the Golden Week departure wave two years running: 34 overtime hours, rooms returned late, arrivals queuing in the lobby. A departure-driven forecast per day rebuilt the housekeeping plan as pickup firmed: overtime −18%, every room back by 15:00, review scores unchanged. Read how it was measured →
Measured result, property anonymized at client’s request.
Forecasting questions revenue and ops teams ask
How accurate are hotel demand forecasts?
A manual spreadsheet forecast typically runs 8% error (MAPE) 7 days out, 15% at 30 days and 28% at 90 days. PMS-driven AI forecasting cuts that error by roughly 20 points at longer horizons. In HotelCadence worked examples, department-level accuracy improves from about 71% to 94%, enough to staff to the curve instead of the average.
What data does forecasting need from Opera, Mews or Cloudbeds?
Read-only access to your actual PMS data: reservations and pickup, arrivals and departures, occupancy, rate and segment mix, and F&B covers where available. OPERA Cloud connects via OHIP; Mews and Cloudbeds connect via their open APIs. The model combines this internal data (history plus real-time bookings) with external signals such as weather, day of week and public holidays. No guest payment data is needed, and setup takes 1-2 weeks.
Can it compare the forecast and actuals against our budget?
Yes. If you load a budget or an existing forecast, HotelCadence shows actuals and the AI forecast against it per department, so variance is visible while there is still time to act on it. Where no budget exists, the agreed baseline plays that role.
How is this different from a revenue management system?
An RMS forecasts rooms to set prices. HotelCadence forecasts operations to set staffing: per department, per hour, check-ins at the desk, room turns for housekeeping, covers at breakfast. The two run side by side; only 28% of hotels use a dedicated RMS anyway (Skift Research), and neither replaces the other.
How often does the forecast update?
Continuously, as pickup changes in your PMS, a tour block booked Thursday night shows up in Friday morning’s forecast. Published rosters are flagged when they no longer match the updated forecast, so managers can adjust before the shift, not after.
Can it forecast tour groups, public holidays and local events?
Yes. Group blocks and nationality mix come straight from the PMS; holiday calendars (Golden Week, Chinese New Year, Songkran) and local event patterns are configured per property during the pilot. These are exactly the days manual forecasts miss by the widest margin.
How much history do you need to start?
Ideally 12-24 months of PMS history for seasonal calibration; the pilot can work with less. History is pulled through the same read-only connection during the 90-day pilot (from US$7,500), and the forecast is validated against your own recent months before any roster relies on it.
30 minutes · your own PMS data if you share access · no obligation.