Red cards in football predictions: what a sending-off really costs
By FootInsights · Published · 6 min read
A defender mistimes a challenge, the referee reaches for his pocket, and every probability attached to the match becomes obsolete at once. Red cards in football predictions occupy a strange position: they are among the largest single shocks a match can absorb, and they are also the shock no pre-match model attempts to forecast. Both of those facts are true at the same time, and understanding why is more useful than any rule of thumb about backing the eleven.
What follows is the arithmetic of a sending-off: what it does to goal expectancy, why the minute on the clock matters far more than the card itself, and what it should change about how you judge a prediction that ended up on the wrong side of one.
A sending-off is a goal-rate shock, in both directions
Football probabilities are built from goal expectancies. A model estimates how many goals each side should score over ninety minutes, then converts that pair of numbers into 1X2, Over/Under and both-teams-to-score prices (the mechanics of that conversion). Anything that changes a match probability has to do it by changing a goal rate. A red card does this unusually violently, and it does it twice over.
The sanctioned side loses roughly a tenth of its outfield personnel, but the damage exceeds one-eleventh because the team also loses shape. Ten players cannot press, hold a defensive line and support an attack at once, so almost every side abandons one of those jobs — usually the attack. Their scoring rate falls further than a headcount would suggest.
Meanwhile the opposition’s rate rises. They have more of the ball, more time on it, and more space to attack into against a defence that has stopped stepping out. Published studies of large match samples — including work on international tournament data — consistently find both effects: the scoring rate of the punished team drops while the scoring rate of their opponents climbs, and the two moves compound in the same direction on the match result.
That compounding is why a red card outweighs almost any pre-match adjustment. A significant injury absence might be worth a tenth or two of a goal in expectancy; a sending-off can be worth several times that per remaining minute of play.
Why the minute matters far more than the card
Here is the part that most commentary skips: a red card is not a fixed quantity. It is a rate change multiplied by the time left to apply it.
A dismissal in the eighty-eighth minute changes the state of the match for two minutes. A dismissal in the eighth changes it for eighty-two. The rate shift may be identical; the effect on the result differs by more than an order of magnitude. Any statement of the form “a red card is worth X% to the opposition” is incomplete unless it names a minute.
Work it through with deliberately hypothetical numbers. Take an evenly matched fixture: both sides at 1.35 expected goals across ninety minutes, which puts each team’s scoring at roughly 0.015 goals per minute. Suppose a sending-off cuts the punished side’s rate by 40% and lifts their opponents’ by 30% for the rest of the match.
- Card in the 80th minute. Ten minutes remain, so about 0.15 goals of expectancy sit on each side beforehand. Afterwards the ten-man side holds roughly 0.09 and their opponents roughly 0.19. The whole match has been re-priced by about a tenth of a goal in total. On a scoreline grid, that is a handful of percentage points — real, but it rarely reorders the favourite.
- Card in the 20th minute. Seventy minutes remain, carrying about 1.05 goals of expectancy each. Afterwards it is roughly 0.63 against 1.37. Combined with whatever was already on the board, that is a swing of well over half a goal — enough to turn a coin-flip into a clear favourite and to move the draw and Over/Under lines substantially.
Same card, same teams, two entirely different matches. Timing is not a refinement here; it is the dominant term.
Two caveats keep the arithmetic honest. The rate shift is not constant across scorelines — a ten-man side already ahead defends deeper and suppresses scoring more than one chasing a deficit. And a red card is not randomly assigned: teams under pressure commit more desperate fouls, so part of the observed post-card gap reflects the match state that produced the card rather than the card itself. Naive comparisons of “matches with a red card” against “matches without” overstate the effect for exactly this reason.
Red cards in football predictions: why pre-match models leave them out
Given how large the effect is, it would be reasonable to ask why a model does not forecast sendings-off. The answer is that it already does, in the only way that is statistically defensible.
A pre-match probability is an average over every way the match could unfold. Some of those paths contain a red card. The historical base rate of dismissals is baked into the goal expectancies the model learned from — those expectancies were estimated from real matches, a small fraction of which featured a sending-off. The model is not ignoring red cards; it is pricing them at their long-run frequency because it has no basis for doing anything else.
Forecasting a specific dismissal would require predicting a rare, referee-dependent, largely momentary event. The signal available before kick-off is weak, and a model that tried to sharpen it would add far more noise than accuracy. Predicting slightly more cards for a physical fixture with a strict referee is defensible at the margin; predicting this red card in this match is not.
This is the same logic that governs any unpriceable in-play event, and it is worth separating from information that genuinely is available beforehand. Team news is known before kick-off and can be quantified in advance; a red card is neither. Confusing the two leads people to demand that models account for things no honest model can see.
Wrong prediction, or unlucky one?
The most useful consequence is what this means for evaluating predictions after the fact.
When a 65% favourite goes down to ten men in the twenty-fifth minute and loses, the natural reaction is that the prediction was wrong. It was not — at least not on the evidence of that match. The prediction described the fixture as it stood at kick-off; a low-probability event then removed a player and re-priced everything. Forecast and outcome disagree, but the disagreement carries almost no information about the model’s quality.
That is not an excuse, and it must not become one. It is the reason single results cannot judge a forecaster, and the reason honest evaluation runs over hundreds of predictions with a proper scoring rule rather than anecdotes. Across a large sample, red cards land on both sides of a model’s calls in rough proportion to how often they occur, and their contribution to the score averages out. Across one match, they are simply noise — the same noise that puts a firm ceiling on how accurate any football forecast can be (why that ceiling sits lower than people expect).
The practical test is the same one that applies to every claim about prediction quality: look at the full record, not the memorable matches. Every prediction we publish stays on our public track record, including the ones a sending-off destroyed, because a record that quietly excluded the unlucky results would be measuring something other than skill.
What to actually do with this
Three things follow, none of them a betting tip.
Read a red card as a rate change with a clock attached. An early card and a late one are different events wearing the same name.
Do not adjust a pre-match forecast for the possibility of a red card. It is already there at its base rate, and no reliable way exists to sharpen it. Adjusting twice is worse than not adjusting at all.
Do not update your view of a model on a match that turned on one. The information content of that result is close to zero. Save your scepticism for the calibration curve and the long-run scoring, where it can tell you something.
A red card is football’s clearest demonstration that a probability is a statement about a distribution of futures, not a promise about one afternoon. Models that pretend otherwise are easier to sell, not more accurate.