CALIBRATION LAB / MODEL RECEIPTS
The model took a
closed-book exam.
We selected the model using 2022, 2023, 2024, then graded it against an untouched 2025 season. Bad outcomes stay in the test. No victory laps around cherry-picked players.
Scoring
Test window
POSITION RESULTS / Half PPR
Where it earns trust. Where it doesn’t.
Rank correlation measures ordering quality. MAE is average absolute points error. “Prior year” is the deliberately simple baseline the model must beat.
QB47 players
Rank correlation0.625
Prior year: 0.601
- Mean error
- 80.6 pts
- Vs. prior year
- 5.1 better
- Top-player hit rate
- 66.7%
- 80% residual band
- -155.3 to +119.2
RB99 players
Rank correlation0.692
Prior year: 0.678
- Mean error
- 47.3 pts
- Vs. prior year
- 2.6 better
- Top-player hit rate
- 66.7%
- 80% residual band
- -83.5 to +60.1
WR126 players
Rank correlation0.630
Prior year: 0.638
- Mean error
- 47.1 pts
- Vs. prior year
- 0.6 better
- Top-player hit rate
- 50.0%
- 80% residual band
- -89.6 to +39.9
TE49 players
Rank correlation0.564
Prior year: 0.501
- Mean error
- 38.3 pts
- Vs. prior year
- 4.2 better
- Top-player hit rate
- 41.7%
- 80% residual band
- -61.5 to +57.4
SLEEPER EDGE / 2025 HOLDOUT
A useful signal. Not magic.
The static score was built on 2022–2023–2024and checked against the unopened 2025 season. The live next-turn and source-conviction adjustments remain provisional.
The static core was evaluated on 2022-2024 and then checked once on 2025. MFL ADP supplies market cost; prior nflverse production supplies projection, opportunity, and availability inputs; same-season half-PPR points above positional replacement supply the outcome.
SELECTED CONFIGURATION
70 / 20 / 10
The winning model assigns 70% of its historical signal to last season, 20% to two seasons ago, and 10% to three seasons ago, with position-specific age curves. It beat eight alternatives on the development seasons before the holdout was opened.
Last season70%
2 seasons back20%
3 seasons back10%
2025 HALF-PPR / BIGGEST MISSES
The ugly table.
A credible model shows its bruises. Positive error means the player outscored the forecast; negative means the model was too optimistic.
PlayerProjectedActualError
Christian McCaffreyRB88365.6+277.6
Jayden DanielsQB355.8114.3-241.5
Joe MixonRB194.10-194.1
Joe BurrowQB317.5134.5-183.1
Kyler MurrayQB258.777.8-180.9
Jacoby BrissettQB53.1227.4+174.3
Drake MayeQB177.1351.5+174.3
Lamar JacksonQB388.4214.9-173.6
TEST DESIGN
No time travel.
Walk-forward validation. Model configuration was selected using 2022-2024 only; 2025 was held out from selection. Predictions use only the three seasons preceding each target season.
Players at QB, RB, WR, or TE with at least four games and a minimum opportunity threshold in the immediately preceding season. Eligibility uses no target-season information.
WHAT THIS DOES NOT PROVE
- The cohort evaluates returning NFL players and does not validate rookie or no-history baselines.
- Prior-season eligibility includes players who subsequently retired, were released, or missed the target season; those players remain in the results with zero actual points.
- The test validates historical production forecasts, not licensed expert consensus ranks, market ADP, or causality.
- Intervals are empirical 10th-to-90th percentile residual bands, not guarantees for an individual player.