What changes when the clock is running
In-play betting is now the core of the sports betting product: Sportradar's annual report for 2025 says it accounts for the majority of gross gaming revenue in the more developed European markets and is expected to grow significantly in the United States. Not every regulator allows it online: Australia's Interactive Gambling Act 2001 bans online in-play sports betting. Pricing it is a different discipline. Pre-match, the model sees the full event as one draw from a distribution. In-play, the event is partly observed: the score, the time, who has been sent off, who is serving. The model has to condition on the current state and price the remainder, and it has to do so faster than the information arrives.
The mathematical structure is the same for every sport: a state, a process that moves the state forward, and a terminal payoff for each market. A football match is a state (score, minute, cards, possibly momentum) and a scoring process (goal hazard rates for each team that depend on the state). A tennis match is a state (sets, games, points, server) and a point-winning probability on serve for each player. The market prices are expectations of the terminal payoff under the process, computed from the current state.
Hazard rates and time decay
In football the natural in-play model is that each team scores according to a Poisson process with an intensity that depends on the state. The pre-match model gives each team's expected goals for the whole match; the in-play model spreads that over the ninety minutes, typically with a rising intensity toward the end, and adjusts it for the current score. In Bundesliga matches from 1968/69 to 2010/11, the total goal rate rose roughly linearly through the match, with a further increase after the 87th minute, which the authors read as more attacking, or weaker defending, as the match closes.
The score adjustment is the important part and it is not symmetric. A team a goal down attacks more and concedes more; a team a goal up sits deeper. The same Bundesliga study found that the total goal rate falls when the score is level late on: at 0-0 it ran about 20% below average in the last five minutes even after three points for a win were introduced, so it is correspondingly higher when the score is unbalanced. The extra goals do not simply go to the trailing team. When the away team led by one goal in the last minutes, the home team's defence weakened so much that the away team became more likely to score the next goal than the home team was to equalise. Models that ignore this will price the total goals market too low late in a one-goal game, and that is a market where sharp in-play bettors live.
Red cards, penalties awarded and substitutions are discrete events that jump the intensities, and Betfair's exchange records a goal, a penalty or a red card as the reason when it suspends a football market. The model needs a rule for each, and the rules should be fitted, not guessed: a study of World Cup 2006 and Euro 2008 betting data found that a red card cuts the penalised team's scoring intensity significantly and raises the opponent's slightly, so that total goals usually fall when the stronger team is penalised but can rise when the weaker team is. Because a changed intensity applies only to the time that remains, the same card moves the totals market further the earlier it comes.
With the intensities in hand, the remainder of the match is a pair of counting processes, one for each team (inhomogeneous Poisson processes if the intensities depend only on time), and the probabilities of every scoreline from here are computable, either analytically for simple models or by Monte Carlo for anything with state-dependent intensities. Every market is then a sum over scorelines. That is what makes the approach coherent: one state, one process, every price consistent.
Point-by-point sports
Tennis, volleyball, table tennis, darts and snooker share a structure: a match is a hierarchy of discrete units (points inside games inside sets) and each unit is won with a probability that depends mostly on who is serving or throwing. The classic tennis model takes two numbers, each player's probability of winning a point on their own serve, and computes the match-winning probability by recursion through the scoring hierarchy. Given the current score the same recursion prices the match, the set, the game and the next point.
Two refinements matter in practice. The serve probabilities are not constant: they vary with surface, with the specific opponent's return, and, measurably, with pressure points, though the pressure effect is small. And the recursion assumes points are independent and identically distributed, which is close enough for pricing but not exact: a study of almost 90,000 Wimbledon points found that winning the previous point raises the chance of winning the next, that servers find it harder to win important points, and that both effects are small and stronger for weaker players, leaving the assumption a good approximation. Models that ignore these effects are slightly miscalibrated after long runs of points.
The commercial feature of these sports is that the state changes with every point and every change moves the price; the Rules of Tennis allow a maximum of 25 seconds between points. That puts the pricing engine, the feed and the suspension logic under continuous load. It also means the value of latency is enormous, which is the next problem.
Latency, feeds and the suspension problem
An in-play price is only as good as the book's knowledge of the state. The state arrives through a data feed: a scout at the venue, a broadcast, or an official data partner, and some jurisdictions require operators to use official data. Every feed has a latency, and any bettor with a faster source of the same information has a near-riskless edge. Courtsiding, the practice of transmitting from the venue ahead of the feed, exists purely to exploit this. In tennis, for anyone bound by the Tennis Anti-Corruption Program, transmitting live results from the venue of an event for betting without consent is a corruption offence, as is deliberately delaying the entry of scoring data.
The defences are structural:
- Suspension on event. When the feed signals a goal, a point, a serve fault, the market suspends before the price is recomputed. The suspension window is a trade-off: too short and late bets get through at stale prices; too long and turnover, or handle, is lost.
- Bet delay. In-play bets are accepted only after a delay of a few seconds during which the price can change and the bet is rejected or re-offered; on an exchange the order is held for a set number of seconds before it is submitted to the market. This is the single most effective control. It has a cost, because customers dislike rejected bets, and data suppliers market their trading services on minimal bet delays.
- Feed redundancy. Two feeds, with suspension when they disagree. Disagreement is itself a signal.
- Margin. Setting in-play margins wider than pre-match for the same market prices the latency risk explicitly.
- Customer-level controls. Accounts that consistently bet in the seconds before a suspension are doing something, and they can be delayed further or limited.
None of these removes the risk; they price and bound it. A desk should measure the cost of stale-price bets explicitly (bets accepted at a price that moved against the book within the delay window) and treat it as a line in the P&L that the controls are managing.
Resulting risk
In-play markets are settled against an official result that can differ from the feed. A goal disallowed on review, a point replayed, a match abandoned: the state the book priced was not the state that counted. Models cannot prevent this; operational rules deal with it. The rules should be published (which source of truth settles each market, what happens on abandonment, how a VAR reversal is handled) and applied without discretion, because discretion in settlement is where disputes and regulatory complaints come from. Some regulators require it: in Pennsylvania a sportsbook's house rules must be immediately available to customers and describe how incorrectly posted events, odds, wagers or results are handled and the effect of schedule changes.
The model does have a role here: it should know which states are provisional. A goal flagged as under review should hold the market suspended rather than repricing on a state that might be reversed.
Model against feed against market
Three things can disagree in-play: the model's price given the feed state, the feed itself (if there are two), and the market (other books, exchanges). The desk needs rules for each disagreement:
- Model versus market, feed agreed: the model may be wrong or the market may be reacting to something visible on the broadcast that the feed has not encoded (a player limping). Widen the margin and cut limits until they converge; escalate to a trader if they do not.
- Feed versus feed: suspend.
- Model versus its own pre-match self: the in-play model at kick-off should match the pre-match price. If it does not, one of them is wrong and the desk should know before the match starts.
The next lesson takes the coherence requirement one step further: markets that are not independent of each other, and the parlay products built on them.