Checked in public, before the games.
Accuracy record
Forecasts, frozen before kickoff
Predictive accuracy
Collecting results. Forecasts have been frozen every morning since Oct 27, 2026 (3 so far). The first week to be scored is week 8, after its last game (Nov 2, 2026). Results appear here then; until 4 weeks are in, treat them as early.
- What's frozen
- Every morning, for every player, what Who should I start shows for the coming week in PPR, half PPR and standard; his season form; and the start/sit chance a gap between two players carries. Each freeze is stamped with its time, the model's fingerprint and the data date. A player counts only with a freeze made before his own kickoff.
- Whose numbers
- The weekly projections are Sleeper's, scored by Swindlr under each scoring setting (Swindlr's season form stands in where Sleeper has none). Swindlr's own are the scoring, season form (also the rest-of-season rate) and the start/sit call with its chance, calibrated on the 2025 season. A result for the weekly projection is mostly a result for Sleeper's.
- Baselines
- For projections, each player's season-to-date average (points a game so far, with two or more games). For the chance, a coin flip: always 50%.
- Never rewritten
- Frozen forecasts are kept as written; each file's SHA-256 fingerprint is below, so a later change would show. A new model gets a new fingerprint, and its results are reported apart, not merged into the old ones. Nothing frozen is used to tune the model.
Freeze log
| Frozen | Week | Model | Players · PPR / half / std | Data as of | SHA-256 |
|---|---|---|---|---|---|
| Oct 29, 2026, 10:01 a.m. ET | 2026 wk 8 | ymqmg8 | 765 / 763 / 756 | 2026-10-29 | 34924d31ceb5… |
| Oct 28, 2026, 6:00 a.m. ET | 2026 wk 8 | ymqmg8 | 766 / 764 / 757 | 2026-10-28 | 46d1d8ceb598… |
| Oct 27, 2026, 10:15 p.m. ET | 2026 wk 8 | ymqmg8 | 766 / 764 / 757 | 2026-10-27 | a4d141eee325… |
What the record can't show
- One season's weeks aren't independent: the same players, teams and weather. Early weeks give wide intervals.
- Projection misses leave out players who didn't play (an injury isn't a projection miss); the decisions count them as zero.
- Decisions are scored on points, not lineups or matchups won; a pair is one choice a manager might face, not every manager's.
- No comparison with other sites' projections or values is made, and none is claimed: Swindlr doesn't say it's more accurate than anyone until forward results like these support it.
Agreement with accepted trades
Not a forecast
Values are tuned on 18,849 accepted dynasty trades from about 2,000 public Sleeper leagues, 562 redraft and 129 keeper trades, and checked on halves they weren't fit on. On 4,698 held-out dynasty trades, a typical trade lands 22% from even under Swindlr's values, and 46% land within 20%. Those halves were reused over many rounds of tuning, so they aren't untouched.
The clean test: trades accepted from Oct 28, 2026 on are kept out of every tuning step and scored on their own. None in the sample yet (the trade sample refreshes weekly); collecting until there are enough to say something.
Agreement means the values price trades the way managers did. It doesn't say those trades were good, or forecast anything.
Points added up right
Not a forecast
99.99% of player-games scored exactly as Sleeper scored them: 244,311 player-games, every rostered player in 148 real Sleeper leagues under their own settings, weeks 1–7 of 2026. That's arithmetic on games already played, checked so a league's scoring settings give the points its platform gives; it says nothing about forecasts.
Record updated Oct 29, 2026, 11:38 a.m. ET by the daily job (scripts/forward-scorecard.ts). Methodology: How Swindlr works. The tool these forecasts come from: Who should I start?