Research note · November 14, 2026
How the Rebuild Lab weights were validated
The window index estimates how a franchise may score in the current and coming seasons. This note shows how the roster/result blend was compared with roster-only and results-only baselines. It is a forecast model, not a guarantee or a trade recommendation.
Sample and method
- The widened study covered 534 Sleeper dynasty leagues and 6,223 returning-manager teams; the two-season-out test covered 214 leagues and 2,279 teams. The leagues were reached through Da League managers, so this is not a random sample of all fantasy leagues.
- The outcome was each team's future points-per-game rank. Rank correlation was weighted by team count. Four thousand bootstrap samples of leagues estimated 95% intervals for the shipped blend's improvement over roster-only and results-only baselines.
- A separate out-of-sample check split leagues into random halves 200 times. The central 95% of weights selected on the training half fell between 0.5 and 0.6 for preseason, 0.4 and 0.5 for next season, and 0.6 and 0.8 for two seasons out. The held-out comparison below evaluates the fixed shipped blend against the better single-source baseline.
- The shipped weights are 50% roster projection before the season and next season, and 70% roster projection two seasons out. The rest of each blend comes from team results.
Improvement in rank correlation
Each comparison shows the shipped blend minus a single-source baseline. Parentheses give the 95% league-bootstrap interval. Positive values favor the blend.
| Forecast horizon | Roster weight | Vs roster only | Vs results only | Held-out median gain |
|---|---|---|---|---|
| Preseason | 50% | +0.079 (0.068 to 0.096) | +0.038 (0.024 to 0.047) | +0.035 |
| Next season | 50% | +0.091 (0.079 to 0.107) | +0.018 (0.010 to 0.025) | +0.015 |
| Two seasons out | 70% | +0.028 (0.009 to 0.047) | +0.050 (0.027 to 0.078) | +0.029 |
All six bootstrap intervals are above zero. The shipped blend beat the better single-source baseline in all 200 held-out splits; the displayed held-out values are median rank-correlation gains.
Limits and source record
The sample is connected to Da League managers rather than randomly drawn. The model cannot anticipate injuries, NBA trades, role changes, or a breakout after the latest sync. The study used league-season totals and validated weekly results because Sleeper can recalculate old weekly player lines. The source research retains the scripts and input snapshots; raw league-level records are not published here.
Research artifact: docs/research/rebuild-lab-2026-11-14/README.md in the source repository.