Canonical data
Provider observations are normalized, deduplicated and provenance-tracked before they are allowed into model inputs.
HOW PROBHOOPS WORKS
Methodology
Pregame features
Team state is frozen before the game being predicted. Current-game outcomes cannot leak backward into the pregame row.
Model governance
Champions are promoted from temporal out-of-sample evidence. Better-looking in-sample or aggregate results are not enough.
League simulation
Pregame win probabilities propagate through NBA standings, tiebreaks, Play-In and fixed playoff-bracket rules.
The promoted bridge combines opponent-adjusted team efficiency with a governed Elo component.
A separately governed score model supplies projected margin, total and score means.
Formal evaluation is time-ordered and reported on the Accuracy page.
AF-SCORE NBA
Measured performance stays separate from prediction
AF-Score describes the basketball a team produced on court. It is era-adjusted, uses pace-aware rate metrics and is never interpreted as a probability. The v1.1 dimensions remain Scoring Efficiency 30%, Defensive Resistance 30%, Possession Control 20%, Creation 12% and Discipline 8%.
Era adjustment
Each component is expressed relative to the contemporaneous NBA environment before fixed 0–100 normalization, reducing distortion from league-wide scoring changes.
Early-season shrinkage
New-season performance is shrunk toward league average using eight equivalent games so one exceptional opening night cannot manufacture an elite ranking.
Same-games comparison
Historical evolution is compared after 5, 10, 20, 41, 60 and final games, rather than relying on nominal calendar dates.
Full-observed history
AF-Score v1.1 now uses 100% observed regular-season detailed team boxes for all 30 teams across all seven completed seasons. Historical normalization and imputation are both zero.
PUBLICATION GATES
A number is not public just because the model can calculate it
Operational facts
Schedule, score and box data must pass canonical source and freshness checks. Last-good data may be retained; missing facts are never fabricated.
Player availability
Availability must be timestamped and NBA-official before it can become canonical. DNP or zero minutes are not retroactively re-labeled as injuries.
Forecast activation
The 2026-27 structural prior is not a public forecast. Publication requires an independently governed roster-adjusted candidate tied to current operational facts.
What If and leverage
Scenario intelligence uses deterministic paired simulation. Game Leverage is derived from conditioned outcomes, not assigned as a subjective score.
Movement causes
A probability delta is movement. A cause is published only when deterministic counterfactual components reconcile to that observed delta.
AI
AI may narrate validated evidence. It cannot create canonical sports facts, probabilities, model promotions, injury status or movement causes.
READ THE EVIDENCE
Separate accuracy from trust
Accuracy answers how the promoted models performed on frozen historical predictions. Trust answers whether the current data, publication and governance gates are satisfied. ProbHoops keeps those questions separate.