MODEL ACCOUNTABILITY · 2026/27

Model Performance

What Fantasy Hockey Intel predicted, when the forecast was recorded, and what happened afterward. Published outcomes include misses, incomplete samples and unresolved calls.

Data sourcesActual scoring and games played, plus observed deployment: NHL.com. Advanced shot-process and expected-goal inputs: MoneyPuck. Fantasy Hockey Intel forecasts, Expected P/G and discovery signals are model outputs, not statistics published by either source.

Deployment Forecasts

Explicit role predictions compared with observed 5v5 lines, defensive pairs and power-play assignments.

0 frozen predictions
Forecast vs observationDeployment Score is a model value, not a prediction of PP1 or a specific line. Only a role forecast saved before the relevant games is evaluated here. Outcomes cover the next 10 or 20 player games beginning on the day after the forecast.
No deployment forecasts have been recorded yet. Evaluation begins with forecasts saved before the relevant games.

A forecast is marked Met when the predicted role is observed in at least 60% of valid role-evidence games after a full 10/20-game horizon, with usable evidence in at least 60% of those games (minimum six). A PP game without a team power play is excluded from the evidence denominator. This rubric is published in advance and applies equally to favorable and unfavorable outcomes.

Methodology

Preseason rate ranges are fixed historical statements supplied in an archive; Radar discoveries and explicit role predictions are timestamped prospectively. The existing model Forecast Ledger retains immutable versions of Expected P/G, Driver, Deployment and Confidence. We do not backdate new predictions, revise recorded starting values, or classify a few early games as final evidence.