Time series and forecasting · insight-1249
Treat a closed fare class as censoring even when the cabin is not full
- Evidence type
- practitioner heuristic
- Difficulty
- hard
- Roles
- Data scientist
- Industries
- aviation
Summary: situation, action, reason, boundary
Situation
A hypothetical flight with 180 seats has sold 118 when its lowest fare class closes, and that class's booking curve goes flat from that day.
Action
Record class availability at every booking snapshot, mark each closed class's bookings as censored from the closing time, unconstrain per class using comparable flights where it stayed open or a censored-likelihood model, and then aggregate.
Reason
Booking requests are turned away not only at physical capacity but also at booking-control limits, so observed bookings in a closed class measure the control, not the demand.
Boundary
Revenue-managed products with nested fare classes or inventory limits, such as airline, hotel and rental-car.
Symptom or risk
The flat segment is read as a demand plateau, so the low-fare forecast for similar flights is too low and the system opens fewer low fares than demand justified.
Exceptions
If the closed class would have been closed by the same rule anyway in the decision being supported, censoring matters less.
Trade-offs
Per-class unconstraining needs availability history that many systems do not archive.
Pitfalls
Unconstraining only sold-out flights, which leaves the much larger set of flights with a closed class untouched.
Lessons and concepts
Claims and their sources
Weatherford and Pölt compare methods for unconstraining capacity-limited airline booking data, because observed bookings understate demand when flights close (as read from the registry record; full text not opened).
The revenue-management reference by Talluri covers demand forecasting, unconstraining, capacity control and overbooking.