A football match is 0–0 after five minutes. Nothing strange there. It is still 0–0 after 25 minutes. Fine. Then 45 minutes arrive without a goal. To a casual viewer, the same thing has happened for almost an hour: nobody scored.
To an algorithm, however, quite a lot has happened.
Every minute without a goal is new information. The clock has moved, the amount of football remaining has shrunk and another opportunity for the expected goals to arrive has disappeared. In live betting, “nothing” is not really nothing. It is data.
The clock is quietly changing the match
Imagine a model expects a match to produce roughly three goals. Before kick-off, there are more than 90 minutes for those goals to appear. If the game reaches half-time at 0–0, those three goals have stubbornly refused to cooperate with the forecast.
That does not mean the model simply says: “Well, three goals are still coming.”
It updates.
Traditional football modelling has often used Poisson-type processes to describe goal scoring. The basic idea is relatively intuitive: goals are relatively rare events occurring over time, and probabilities can be assigned to different numbers of goals. More sophisticated models make those scoring intensities dynamic rather than fixed. They can respond to the score, red cards, team strength and other match events.
So a 0–0 at minute 10 is statistically a different object from a 0–0 at minute 70.
At minute 10, there is plenty of time for 2–1, 3–0 or 2–2. At minute 70, those outcomes require several goals squeezed into a much smaller window. They are still possible, obviously. Football enjoys embarrassing probability models occasionally. But their likelihood has changed dramatically.
Not scoring is an event too
This is where things become more interesting.
Suppose a strong attacking team was expected to dominate. Twenty minutes pass without a goal. That alone does not necessarily say much. But what if 60 minutes pass? And what if the favourite has created very little?
The model now has two kinds of information. First, time has disappeared. Second, the match itself may be producing evidence that the pre-match assumptions were too optimistic.
Modern systems can incorporate live variables rather than watching only the scoreboard. A 2026 study by Lawrence Clegg, Zixing Song and John Cartlidge combined pre-match betting-market information with post-shot expected goals as a changing in-play variable. Tested on 140 Premier League matches, its model reached 70.2% classification accuracy, compared with 70.6% for Betfair Exchange prices.
That is a useful reminder of what the algorithm is actually doing. It is not staring hypnotically at “0–0.” It is repeatedly asking: given everything we knew before the match, plus everything we have observed so far, what is likely to happen during the time that remains?
A quiet match can therefore become very noisy
This continuous updating is visible on live betting platforms. On 22Bet, for example, in-play football odds move as the match progresses, while live markets react to the changing game state. The funny part is that the score does not need to change for the probabilities behind those prices to move.
A goal is spectacular information. A red card is spectacular information. But ten additional goalless minutes are information too, just delivered without fireworks.
And time does not necessarily affect scoring probability in a perfectly straight line.
A 2025 study examining 3,433 matches across 21 leagues and competitions found that goals were more likely as matches progressed, with fewer goals than expected during the early minutes of each half. That makes intuitive football sense too. Teams tire. Tactical risks change. A side needing three points may eventually decide that defending politely is no longer enough.
Score state also influences behaviour. Research using Brazilian league data found that a team’s goal rate increased when it was behind, while receiving a red card substantially reduced its scoring intensity. In other words, models benefit from treating a match as a moving system rather than a 90-minute lottery ticket printed at kick-off.
The strangest information is sometimes an absence
This explains why 0–0 can become fascinating mathematically while becoming slightly painful for the neutral viewer.
By minute 75, the algorithm knows much more than it knew before kick-off. It knows the original expectations did not materialise for 75 minutes. It knows how the teams have actually performed. It knows the score state, disciplinary events and, depending on the model, shots or expected-goal information. Most importantly, it knows that only a small slice of the match remains.
Probability has not been standing still just because the scoreboard has.
That is perhaps the cleverest thing about live football modelling. Algorithms do not only learn from events. They learn from events that failed to happen.
Sometimes the loudest piece of data in a football match is still a zero.
