Home
Draft
Draft War RoomPreseason RankingsMock DraftRookie BoardHandcuff FinderStrength of ScheduleBye PlannerCheat SheetValues & ReachesScarcity CurveStack Finder
In-Season
Weekly ProjectionsSeason Budget PaceWaiver WireInjury WatchScenariosLineup OptimizerTrendingScores
Tools
CompareTrade AnalyzerWhat-IfMy LeagueConnect LeagueTrade FinderTeamsRoster Report CardExport
Research
Divergence WatchLine MoversSignal LibraryArmchair QBAsk NovaFloor / CeilingOpportunityConsistency
BlogPricing
About
About NovaPredictMethodologyAccountabilityChallenge the ModelLeaderboard
← Glossary
Glossary

Calibration ledger

The calibration ledger is NovaPredict's public check that start/sit win probabilities mean what they say: 72% calls should hit about 72% of the time.

The calibration ledger is NovaPredict's public record showing that its start/sit win probabilities mean what they say. When Nova says a start wins 72% of the time, calls like that one have to hit about 72% of the time, and the ledger is where that gets checked.

Calibration is the difference between a number and a vibe. Plenty of tools will tell you "start Player A, 72% confidence" and never look back. NovaPredict groups its start/sit calls by stated probability, then tracks how each group actually performed. The 60% calls should win around 60% of the time, the 80% calls around 80%, all the way up the scale. Every call is locked before kickoff, so there is no quiet rewriting after the games.

How to Read It

The ledger tells you how literally to take a start/sit probability.

  • Stated matches actual: the probabilities are trustworthy. A 55% edge really is a coin flip with a lean, and you should treat it that way, not as a lock.
  • Stated runs hot: if calls labeled 75% are only winning 60% of the time, the model is overconfident, and you should mentally discount its strong calls.
  • Stated runs cold: if 60% calls keep winning at 70%, the model is underselling itself.

The practical payoff is decision-sizing. A calibrated 52% is a "start him, but do not agonize" call. A calibrated 80% is a "do not overthink this" call. Without calibration, both are just tones of voice.

What It Is Not

The ledger is not a win-loss record for Nova, and a losing 70% call is not evidence of a broken model; 70% calls are supposed to lose three times in ten. What the ledger judges is whether the percentages, taken together, are honest. The live ledger begins with Week 1 of the 2026 season, so early entries deserve patience before conclusions.

For how Nova's boom and bust probabilities are graded, see Brier score and skill.

The ledger lives, in public, at /accountability.