Time series and forecasting · insight-0233

Use bookings already on hand as a known-at-origin input for departure-level forecasts

Situation
A revenue-management team forecasts final bookings for a departure 21 days out using only past departures.
Action
Benchmark pickup-style models (final bookings equal current bookings plus historical average pickup from this point on the curve) against time-series and regression models, all evaluated at the same days-to-departure snapshots.
Reason
Bookings on hand at the snapshot are known at the origin, and pickup and regression methods performed well in published hotel revenue-management comparisons.
Boundary
Airlines, hotels, events, cruises, any advance-booking business.

Symptom or risk

Ignoring the bookings already received throws away the strongest signal about this specific departure.

Bookings on hand are censored by capacity and fare controls; unconstrain first.

Multiplicative pickup is unstable when early bookings are tiny; additive pickup is more robust there.

Building the snapshot from the final booking table instead of archived booking-curve snapshots.

The synthetic booking_curve table stores cumulative bookings at 13 snapshots from 90 down to 0 days before departure.

In a hotel revenue-management comparison, pickup and regression methods had the lowest error on one dataset, and exponential smoothing, pickup and moving average were most robust on another.
A comparison of forecasting methods for hotel revenue management