The accuracy scoreboard
Nobody in fantasy analytics publishes how wrong they were last year — a public error metric can only hurt a free product bundled with a league host. We publish it. That is the moat, and it only works if it is genuinely honest, including the parts where we lose.
Generated 2026-08-21 from 2016–2025 of nflverse + FFC ADP data, 12-team PPR. Walk-forward and out-of-sample throughout — every number below is regenerated by pipeline/backtest/export_scoreboard.py, not copied into this page.
We do not beat the market at ranking
Five approaches have been tested against consensus ADP, walk-forward, on realized points-above-replacement. All five lost, or were killed before the comparison was even worth running.
How much of realized season value ADP alone already explains. Every ranking attempt below is measured against this.
TrueValue's essential structure, tested in its simplest form: last season's rate alone, no situation, no QB context.
2018–2025 · 32 position-seasons · its own local ADP bar: 0.460
Opportunity per game, yards per game, receptions per game and durability, regressed on the market anchor's residual. Four collinear features split credit unstably.
2018–2025 · 29 position-seasons · its own local ADP bar: 0.477
The diagnosis of the full model's regression: one feature (non-TD points per game), raw-points residual, ridge-shrunk. Beat the full model. Still short of ADP.
2018–2025 · 29 position-seasons · its own local ADP bar: 0.477
The best overall score of the five, and still inside the noise of its own local bar.
2018–2025 · 31 position-seasons · its own local ADP bar: 0.468
Killed at its own pre-registered sanity check (XFP-SCOPE.md kill criterion 1) before it was ever scored against ADP: prior-season xFP predicted next season's rate WORSE than prior-season actual points did, at every position (actual 0.721 vs xFP 0.635, overall FAIL).
None of these five feed the board’s ranking. TrueValue does not claim a rank edge over the market anywhere on this site.
Where it does add value: the calibrated range
The isotonic anchor behind the board’s season outlook is monotone by construction — it cannot change a ranking. What it can do is tell the truth about spread. Checked out-of-sample, its declared confidence bands land almost exactly where they claim to:
| Quantile | p10 | p25 | p50 | p75 | p90 |
|---|---|---|---|---|---|
| Target | 0.10 | 0.25 | 0.50 | 0.75 | 0.90 |
| Measured | 0.104 | 0.245 | 0.485 | 0.709 | 0.855 |
Not a ranking — the isotonic anchor cannot change Spearman by construction. Its value is here: the empirical coverage of its declared confidence bands, checked out-of-sample.
The one narrow positive: availability at TE
Adding measured durability — how many games a player at that cost actually plays — beat the market at tight end, over 2019–2025 (7 seasons). The only position where adding measured durability (games missed, by draft cost) beat the market. Everywhere else it was flat or worse — see the pooled L1 x L2 row above.
See how TrueSplits works for the full model. MODEL-SPEC.md in the repo has every rejected approach in full, including the ones not shown here.