Chronology, not a sealed holdout
2025 is a chronology-controlled retrospective replay, not a sealed or untouched test set.
Methodology · Week 4 2026
How the numbers
are actually made.
This page is the part most forecasting sites leave out. It describes what the model does, what it is allowed to price, and what the historical record does and does not prove. Nothing here is behind a subscription.
- Players forecast
- 1,688
- Game environments
- 16
- Pick-enabled markets
- 5
- Serving snapshot
- 289c2edf
The forecast
One joint model, not seventeen separate ones
01
Shared game futures
Every matchup is modelled as one joint distribution over score, pace, efficiency, and pass/rush exposure, weighted across game scripts. Player outcomes are conditioned on that shared game rather than forecast in isolation, so a receiver's ceiling and his quarterback's volume cannot disagree with each other.
02
Regime selection
Each matchup resolves to between three and six game-script regimes, selected by a contract rather than chosen by hand. The regimes are what the site exposes as game paths: pick one and every player forecast in the room re-conditions on it.
03
Role and participation
Availability, snap share, teammate competition, and route or carry competition are modelled explicitly. Bench and zero-role players are represented with explicit zero forecasts rather than omitted, so coverage gaps cannot masquerade as confidence.
04
Recentering against marginals
Scenario-conditioned player means are integrated back to guarded specialist marginals, so the conditional views and the headline distribution stay arithmetically consistent. The launch audit bounds this error at 0.00004.
05
Pricing, only where earned
A distribution becomes a fair probability only for markets whose specialist has passed the approval gate. Everything else is served as a forecast and is never priced against a book.
The gate
What it takes for a market to be priced
A market is pick-enabled when its forecast and execution rule are timestamp-valid, its result is stable across seasons or a frozen cross-season development/holdout split, and the surviving effect has a coherent football or market-structure mechanism. Research exposure is disclosed explicitly; it does not disqualify historical evidence, but it prevents us from describing that evidence as prospective.
5 markets have passed: QB pass attempts, RB rushing yards, QB passing yards, Player sacks, Solo tackles. Each has a frozen price rule, and a pick publishes only when the live price still clears it; when no price does, the board says so rather than manufacturing a selection. The rules themselves are not published. Every other market is served as a forecast and is never priced against a book.
The honest part
What the 2025 record does not prove
Selection policy
Candidates fit on 2024 Weeks 1–12 and are selected on Weeks 13–18 using MAE and CRPS. A challenger that does not improve both remains an exact no-op.
Replay discipline
The configuration was selected on 2024 and every 2025 week uses only information available before that week. Because development continued after 2025 was observed, results are retrospective rather than sealed holdout evidence.
Deployment policy
Current models may be refit through 2025 for 2026, but they remain prospective and cannot inherit the retrospective model's evidence.
Known limits
Conditional scenario sensitivities are a serving overlay rather than separately selected conditional heads on a chronological holdout. Pregame availability does not identify the timing of an in-game injury. Route, package, and read-progression labels stay partial where public sources do not preserve them.
Flagship calibration on the 2025 replay: predicted 62.3% against a realized 61.9%. See the full ledger.
Reproducibility
Every published forecast has a receipt
Each publication run writes a content-addressed snapshot that is never overwritten, records input hashes and model versions in the serving manifest, and fails the build if player coverage, regime probabilities, marginal recentering, or fair-probability bounds fall outside contract. The forecasts on this site come from snapshot 289c2edf56c1d356a8b97ce1e58a0cf004981376a8e6beb986c6fc674130fc9d, frozen September 29, 2026 at 3:27 AM UTC.