Time series and forecasting · insight-0213
Unconstrain sold-out flights before training a demand forecast
- Evidence type
- documented fact
- Difficulty
- hard
- Roles
- Data scientist
- Industries
- aviation
Summary: situation, action, reason, boundary
Situation
A route forecast is trained on passengers flown, and many peak departures hit capacity.
Action
Flag observations at or near capacity as censored, estimate unconstrained demand (for example from booking-curve pickup on comparable unconstrained departures or a censored-likelihood model), and evaluate forecasts on unconstrained or uncensored targets.
Reason
Sales are bounded by available inventory, so they measure demand only when the constraint does not bind.
Boundary
Airlines, hotels, events, any capacity-limited sales.
Symptom or risk
Observed passengers are capped at capacity, so the model learns demand that is too low exactly when demand is highest, and capacity or pricing decisions are under-ambitious.
Exceptions
If the decision only concerns realised passengers under fixed capacity, a constrained forecast may be what you need.
Trade-offs
Unconstraining needs assumptions that cannot be fully checked; report sensitivity to them.
Pitfalls
Dropping sold-out departures from training, which biases demand downward even more.
Example
In the synthetic airline data, final bookings are capped at 1.06 times capacity (about 19% of flight-days sit at that limit) while latent demand is not, so sold-out departures understate demand.
Lessons and concepts
Claims and their sources
Revenue-management practice treats demand forecasting and unconstraining of capacity-limited sales as a core step.