The short version
- Every morning we pull the officials assigned to each game once the league posts them.
- Each official has a career profile built from play-by-play: how often they call fouls or penalties, when, and on whom.
- We combine the crew, compare it with what the two teams normally do, and write the report.
- Where the whistle should move a line, we say how much, with a confidence level. Where it should not, we say that too.
Data sources
All of it is public. NBA assignments and play-by-play come from the league’s own stats service, with ESPN’s public feeds as a backup. NHL data comes from the league’s public game centre feed, which lists referees and linesmen. MLB will use Retrosheet and Statcast; NFL will use nflverse. Closing lines come from public odds feeds.
The model
A hierarchical model per official. Every official starts at the league average and moves away from it only as the games pile up. Crew-level tendencies take roughly 30 to 50 games to settle, so our minimum sample for a read is 47 games; below that the official is shrunk hard toward the league prior and we say so.
We always show the expected value next to the actual. A crew that called 50 fouls in a game with 60 borderline contacts is not the same as a crew that called 50 in a clean one.
What we will not claim
- Nothing about intent. We measure how a game was called, not why. We make no claim about any official’s bias or integrity.
- No single-game verdicts. One night is mostly noise. Reads come from career priors updated by games.
- No “this ref hates this team”. Splits by team or player stay noisy for years. We do not publish them.
- Totals effects are weak. In public NFL research, crew effects on final scores mostly vanish after adjusting for era. The stronger story is how the game gets called: minutes lost to foul trouble, power-play time, pace.
Backtesting
Every read is tested against closing lines over the 2021-22 to 2025-26 seasons before it goes live. We will publish the backtest, including where the model is wrong, before we put a probability on a live game. Until then, numbers on this site are labelled as samples.
Tiers
- Best bet: the crew moves the number past the line with room to spare.
- Value play: a smaller edge, usually a player prop priced without the whistle.
- Stay away: the whistle is already in the price. Often the most useful read on the card.
No locks. No guarantees. A 56% read loses 44% of the time.
Corrections
When we get a number wrong, we fix it on the page and note the change underneath. Send corrections to corrections@whistlebets.com.