Problem
Netflix launch preparation needs the final media file, the IMF, or an earlier non-final version, the Locked Cut. Teams must choose when to start, using delivery dates that content partners enter by hand in production schedules. These dates are often missing, and they are least accurate far from delivery. Titles whose schedules were less accurate more often had a launch miss, meaning a delayed title.
Example from the article: waiting for the IMF risks a compressed timeline if it arrives late, while starting on the Locked Cut risks extra conformance work if the final files differ.
Solution in brief
Train a boosted-tree regression that predicts the days left until each asset (IMF, Locked Cut) is delivered, using daily snapshots of production signals. This fills the gaps where no manual date exists and corrects inaccurate ones. Compare predicted and scheduled dates with an error metric accumulated over time, and serve whichever is more reliable per segment.
Questions this case answers
- How to predict the time until a milestone for in-progress projects when human estimates are sparse and drift from reality
- How to compare model forecasts with human estimates at every horizon before the event, not at a single point in time
- How to predict when a production asset will be delivered when manual schedules are incomplete and inaccurate?
- How to measure schedule inaccuracy over time and link it to missed launches?
Tools and methods
- Boosted tree regression — predicts days until delivery for each in-progress production
- Daily feature snapshots — each feature stored with its value on every day of the production
- Accumulated Error Days (AED) — schedule-accuracy metric tied to launch misses
- Backtest by horizon and segment — compares predicted with scheduled dates
Google or a one-time email code. No password.