Guide

Modelling player cards: the referee is half the bet

By Adam · Updated 16 July 2026 · 7 min read

A card bet is the only player market where the biggest multiplier in the model belongs to the one man on the pitch who can't be booked. Bookmakers call the market "to be shown a card"; the model treats it as three questions — how often does this player get booked, who is he playing, and who's holding the whistle — and the third question moves the price more than most people who watch every match would believe. This is the walk from raw booking rates to a probability you could defend, companion to the shots on target guide and a plain-English disclosure of how this site's own cards model works.

The spine is the same, the rates are smaller

The machinery starts exactly where the shots model starts: per-90 rates with minutes as the denominator, recent form weighted on a half-life so the whole season still counts, and small samples shrunk toward a sensible prior. The difference is scale. The average Premier League player picks up a card about once every five full matches, so a raw booking rate is even noisier than a raw shot rate, and the shrinkage earns its keep even harder. Two early-season bookings tell you almost nothing; the model knows that even when the market's memory doesn't.

Position is the prior

What a thin sample gets shrunk toward matters, and for cards the positional differences are big. Midfielders are the most booked players in the league — the tackle-heavy middle third, roughly 40% more cards per 90 than forwards — with defenders just behind and goalkeepers barely on the chart. So a new signing with no Premier League history gets priced as his position first and earns his own number match by match, at exactly the speed the minutes justify.

The referee is half the bet

Now the centrepiece. The referee's appointment is public before kick-off, his history is measurable, and his effect is enormous: among officials with a meaningful sample, card rates run from about 30% below the league's average of four a game to a shade over 10% above it. The same lunging challenge is a booking under one whistle and a talking-to under another. On a player whose baseline chance is 15%, the referee alone stretches the true probability between roughly one-in-ten and one-in-six — a spread wider than most of the value you'll ever find in the player himself.

The market's weakness here has a mechanical cause. A referee moves every player's number a little rather than one player's number a lot, so no single price ever looks obviously wrong — and diffuse effects are exactly what a price-setter under deadline smooths over. The model has no deadline. One discipline, though: a referee factor is only as good as its sample, so the confidence grade on every card pick discounts referees with only a short record in the data rather than trusting a dozen matches to describe a career.

Some opponents drag you onto a card

Bookings aren't taken, they're induced. Quick, tricky sides that invite the desperate lunge; sides that draw fouls and surround the referee; sides that slow the game down until frustration does the rest. The model measures this directly — how many cards teams actually pick up against this opponent, relative to the league norm — and the spread runs from about 0.7× to 1.2×. It's the same idea as the shots model's opponent factor, pointed at temperament instead of defence.

Derbies count, and the model refuses to get excited

Derbies reliably raise card counts, so the model applies an uplift for a fixed list of genuine rivalries — modest ones, 15 to 30% depending on the fixture. Deliberately modest: the market prices derby heat too, and juicing a multiplier because a game feels spicy is how a model starts manufacturing edge that was never there. What stays out entirely is the subjective layer — stakes, must-win nights, a manager under pressure. Real factors, unquantifiable ones; they belong in your judgement, weighed knowingly, not smuggled into the arithmetic.

One card is the whole bet

Multiply it out — booking rate, share of the 90 he's expected to play, referee, opponent, derby — and you have the expected cards for this player in this match. The step to a probability is deliberately plain: cards are rare events, and a simple at-least-one calculation, recalibrated against seasons of history, beats cleverness here. Note what doesn't exist: a 2+ cards line. A second yellow is a red, and send-offs are a separate, much longer price. One card is the whole bet, which is why the market is a clean yes/no rather than a ladder of lines.

The price you can't fully de-vig

Here's a wrinkle the other player markets don't have: the card market is one-way. There's a price for the booking and no price against it, because a market in players avoiding cards would be all margin and no punters. With shots you can strip the bookmaker's margin out properly because both sides exist; with a one-way price the margin stays hidden inside the number. The tool's answer is to say so — card legs are flagged un-de-vigged rather than dressed up with a guess — and to remember that margin bites hardest on longshots, which is precisely what most booking prices are.

What the record says

Last season the model made 9,521 card predictions out of sample: it said a booking 13.9% of the time, and bookings happened 14.3% of the time. Calibration is the strong suit. The honest caveat is that the margin over a naive baseline is thin — in some leagues it doesn't clear it at all — because most players, most games, don't get booked, so a lazy "nobody gets carded" guess scores deceptively well. The paying edge lives in the confluences: a combative midfielder, a strict referee, a rival in the wrong mood. The model's advantage isn't seeing something you can't see. It's running the same checklist nine and a half thousand times without once getting bored.

Frequently asked

Keep going