S13 · stress test simulation

Stress-test the booking forecast before its feed migration

People, data and outputs in this scenario are simulated. Each stage is recorded when you commit; the reveal comes after.

4stages
35 minabout
0stages committed

The brief

The synthetic airline's 21-day-ahead booking forecast (a LightGBM model on route, capacity, bookings on hand at 21 days, calendar flags, competitor price and competitor capacity) sets fare-class availability three weeks out. Trained on 2022-2023; clean WAPE on the 2024 flight-days (6,954) is 2.00%. Next month the booking-curve snapshots move to a new events pipeline, owned by another team. You have a half day to stress-test the model; everything below was measured on flight_day and booking_curve (synthetic).

Stage 1 of 4

The model serves 11 routes from one feature table. In an earlier ablation, removing bookings on hand at 21 days raised WAPE from 2.0% to 14.2%, while removing either competitor feature left it at 2.0%.

Which tests belong in the first-pass plan? Select all that apply.

Missing and late values for bookings on hand, including a multi-day outage and scattered nulls.A wrong or stale snapshot: the 30-day count used where the 21-day count is expected.Duplicated events that double-count bookings on a few flights.Add the same 10% random noise to every feature so that all inputs are treated equally.Re-run the clean 2024 holdout with ten more random seeds to confirm the 2.00% is stable.