What we are estimating
Our model estimates future hockey outcomes and the FSPI paths that could follow from them. FutureSports sets the official scoring rules, and prices come from the market. Within NHLFutures, the “canonical” label simply identifies the model we use as the published reference.
What is included
The documented reference combines a decayed team-rate core, player availability and injury-impact estimates, expected-goals (xG) evidence, five-on-five regularized adjusted plus-minus (RAPM), bounded market priors where permitted, and team-rate uncertainty. League re-anchoring is enabled. Market inputs mean the reference should not be described as wholly independent of market information.
What each part contributes
Each part of the model contributes a different kind of evidence. We combine them carefully because the same information can otherwise appear twice—for example, once in a team’s recent results and again in a market price shaped by those results.
| Layer | Question it addresses | Important limit |
|---|---|---|
| Decayed team-rate core | What does prior team performance suggest about future rates? | Recent games are informative but noisy. |
| Availability and injury impact | Who may contribute, and how could missing contribution affect the team? | A reported absence and a forecast of return are different kinds of evidence. |
| Expected goals | How dangerous were the opportunities created and allowed? | An xG estimate is not an observed goal or an official FSPI scoring entry. |
| Five-on-five RAPM | What player-level signal remains after accounting for shared ice time and context? | Regularized attribution is not a causal rating or a guaranteed trade effect. |
| Bounded market priors | What information is already reflected in eligible prices? | Quotes need a valid timestamp; market-informed forecasts are not independent of those prices. |
| League re-anchoring | Has the league-wide scoring environment shifted at the forecast origin? | The adjustment must not learn from later realized games. |
| Team-rate uncertainty | How wide should the plausible future paths be? | More simulated paths cannot remove uncertainty in the model itself. |
How the pieces become a forecast
The process normalizes hockey observations, reconstructs scoring history, fits pre-origin team and player evidence, and simulates the league jointly. Joint paths keep opponents, standings, playoff progression, and index outcomes consistent within each simulation. Reported probabilities are proportions of simulated outcomes; intervals describe a range rather than a guarantee.
How a player update reaches the forecast
Imagine a hypothetical important forward whose availability outlook changes from several missed games to a possible early return. The observation is the new report; availability over future games is still uncertain. A forecast must translate that uncertainty into plausible contributions rather than pretending the player is definitely back at full impact.
The effect then passes through future games and their shared league consequences. A changed result for one team affects an opponent’s result in the same simulated game. Standings and playoff paths follow from those games, while the FSPI scoring rules translate their statistics and milestones into index paths.
The result is a range of possible outcomes rather than one fixed “injury adjustment.” If you reduce a player’s impact in a scenario, you are exploring an assumption about that player’s future contribution. The scenario should not be read as an injury report or as a proven estimate of what his absence would cause.
What a probability or interval means
A hypothetical 20% playoff probability means the modeled team qualifies in roughly one fifth of the simulated outcomes under that run’s assumptions. It does not mean the model is “20% confident” in a deterministic answer. An index interval describes the spread of simulated index values; it is not a guaranteed boundary.
Random simulation variation can be reduced by running more paths. Data uncertainty, parameter uncertainty, missing information, and structural model error remain. Producing 100,000 paths is therefore not evidence that a forecast is calibrated. That requires checking distributions against outcomes over appropriately held-out cases.
What has been validated so far
Historical evaluation uses forecast cutoffs and held-out outcomes. The September 4, 2026 review preserves the general reference model but establishes no new next-game or horizon-specific champion. Prospective 2026–27 validation and Cup-probability calibration remain unestablished. Historical performance must not be advertised as a demonstrated live trading edge.
How we read the evidence
We use three kinds of checks. First, we ask whether an individual component predicts its own target, such as shot outcomes. Then we test whether adding that component improves the complete hockey or index forecast against a fixed reference. Finally, before publishing a forecast, we check that its inputs, history, and output are complete. A component can look useful on its own without improving the finished model, and a well-tested model can still produce an unusable publication if its inputs are stale.
Historical tests supported the xG and RAPM composition in the current reference, although the review records limitations in recent seasons and at the tails of the distribution. We have not established that it leads at every horizon, that every ingredient adds value independently, or that its differences from market prices remain profitable after costs.
The model and the latest forecast have separate status
The same review found a publication gap: the source forecast lacked explicit next-game probabilities and complete live injury/odds freshness evidence. Rebuilding that publication is blocked pending a valid forecast. This page documents model design as reviewed on September 4; it does not certify that the latest website output is fresh or production-approved. Read each result’s own status and timestamp.
What we have left out
The trade-deadline overlay is disabled. Named replacement/deployment modeling and future-injury incidence remain research directions. Schedule-aware injury returns are enabled in the composite, but their incremental benefit needs a fresh comparison under the current validation contract. Fitted weights, calibration settings, protected feature recipes, raw licensed feeds, and proprietary simulation mechanics remain private. We publish the model’s purpose, evidence families, limitations, and evaluation principles.
Based on the internal model catalog, current model state, and September 4, 2026 model/preset review. This public summary intentionally omits protected implementation details.
Independent educational research. Examples are hypothetical unless stated otherwise. Read current exchange rules and fund documents for product terms.
Inspect the model workspace →