Not tips and not prices. Each row is the model's expected count for one player in this match — how many fouls, tackles or shots it thinks he records — plus the full spread of outcomes around it. Nothing in the projections has been compared to a bookmaker; that comparison is what the Value tab is.
Step 1
His own record
Every completed match in the dataset, as a rate per 90 minutes. The n column is how many matches that rests on.
Step 2
Pulled toward his position
A thin record gets pulled hard toward other players in his position, in his league — a goalkeeper toward goalkeepers, not toward the outfielders around him. How hard is fitted per market, not chosen: tackles pull about four times harder than shots on target.
Step 3
This fixture's conditions
How much of the stat the opponent concedes, what the referee’s matches look like, and home or away. This is where a projection stops being a season average.
Step 4
Scaled to expected minutes
A rate becomes a count only once you say how long he plays. 60 minutes carries two-thirds the exposure of 90, and the distribution is built at that exposure.
- Model /90
- The model’s rate for this match with the minutes taken back out — steps 1 to 3 already applied. It moves with the opponent and the referee, so it is not his raw career rate; the builder shows career and model rates side by side. Open any row to see it.
- Exp. mins
- Minutes the model expects. Confirmed means the lineup is out; provisional means it is guessing from his start rate, and is the widest of the three.
- Projected
- The expected count in this match, at those minutes. This is the number a line is set against.
- Over 1.5, etc.
- Read straight off the distribution beside it, not off how often he has beaten that line before. It answers what the model thinks, not what has happened.
- n
- Matches behind his own rate. Under 30 it turns amber: the projection is mostly his position group speaking, and §6.3 will not publish an edge on it.
- Distribution
- The whole spread, not just the average. The dashed amber cut is the line; jade bars beat it. An expected 1.8 fouls made of 2 every week and an expected 1.8 made of a 0 and a 5 are different bets, and this is where you see which one you have.
Counts are modelled as negative binomial rather than Poisson, because fouls and tackles are overdispersed and a Poisson would systematically underprice the tails — which is exactly where over bets live. Every market on this page cleared its calibration gate before it was allowed to produce a projection: when the model says 60%, it lands within a point or so of 60% out of sample.
The full write-up — every gate, every fitted parameter and what the model refuses to price — is on the methodology page.
