How accurate were the estimates?

The model is frozen on one set of hypothetical joiners and scored on joiners it never saw, each at their own horizon, with inverse-probability-of-censoring weights. Every number here is copied from the backtest reports in the repository.

What could not be tested on Fall 2026

• The first week of instruction. Berkeleytime recorded nothing for any section from Aug 19 to Sep 1, 2026, which covers the end of the adjustment period and the first week of class, so no joiner's outcome at the first-day-of-class or last-run horizon is observable. Fall 2026 validates clearing during Phase 1 and Phase 2 only; the 14- and 28-day horizons are the ones scored.
• Reserved seats. A section can hold seats for a category the joiner is not in; the counts do not say who was eligible, so the estimates treat the queue as first-in-first-out.
• The Cox model across the Phase 1 to Phase 2 shift. Fit on Phase 1 joiners, it under-predicts Phase 2 clearing; its scores are shown for honesty, but the pages use the Kaplan-Meier curves, not the model.