Demand Forecasting and Data Analytics
Demand Forecasting and Data Analytics tests whether you can build a forecast at the level where decisions are made, distinguish constrained from unconstrained demand, choose an appropriate method, and judge forecast quality. Expect questions on pickup, booking curves, accuracy metrics, denials and regrets, and data sources such as PMS, CRS, RMS, and competitive benchmarking.
Forecast at the decision level
A property-level forecast cannot set a rate fence. Forecast by segment, rate code, arrival date, day of week, and length of stay, then aggregate. Use a 30/60/90-day rolling horizon plus a weekly short-term forecast. Forecast unconstrained demand first, then apply controls; otherwise the forecast only describes the inventory you chose to sell.
Methods and when they fit
- Pickup and booking-curve models: use historical pickup factors from the same lead time and day type.
- Moving average and exponential smoothing: stable, low-volume series.
- Regression: day of week, seasonality, events, price, comp-set occupancy, airlift.
- Machine learning: large, rich datasets; risk of overfitting and poor explainability.
Accuracy metrics: MAPE, MAE, RMSE, and bias. Bias matters as much as absolute error because a consistently low forecast under-protects inventory.
Constrained versus unconstrained demand
Denials ('no availability'), regrets (rate or fence rejected), turnaways, closed-to-arrival dates, and minimum-LOS blocks all hide demand. Add logged denials and regrets back to sold rooms to estimate unconstrained demand, and record why a request was refused.
Worked example
Ninety days before a Tuesday, 60 rooms are on the books. The historical 90-day pickup factor for that day type is 1.6, giving 96 rooms. Capacity is 120. Actual demand materializes at 110 rooms, but the hotel logged 8 regrets at the rate fence. Unconstrained demand was therefore 118, close to capacity. Forecast error against the 96 pickup forecast is 14 rooms (12.7% low); against 118 it is 22 rooms, but that larger gap is a signal about constrained data, not a reason to abandon the method. A three-period MAPE check: actuals 100, 120, 80 against forecasts 110, 110, 90 gives absolute percentage errors of 10.0%, 8.3%, and 12.5%; MAPE = 10.3%.
Traps
- Forecasting constrained demand and then wondering why the hotel is full too early.
- Confusing pace (versus the same time last year) with pickup (change from now to arrival).
- Using MAPE on low-volume or zero-demand days; it explodes.
- Ignoring forecast bias and managing only average error.
- Overfitting an event model to one year of history.
- One forecast for all segments, then blaming sales for the mix.