Key findings
- Estimates of in-play’s share of betting range from about 47% to 75%, depending on market, metric and source. The highest figures describe mature European markets.
- Repricing engines are fast, but the delay between the venue and the engine is usually the real bottleneck.
- Machine learning is most visible in risk and liability management, and trading desks have not disappeared.
- Research suggests calibration matters more than raw accuracy, and that bookmakers restrict accounts that beat them.
- Micro-betting is the live regulatory flashpoint, with caps in baseball and unresolved bills in New Jersey.
How big is live betting?
In-play betting, also called live or in-game betting, lets customers wager after an event has started. Its share of the market depends on where you look and what you count.
Sportradar’s own analysis puts the global in-play share of handle at roughly 70 to 75 percent (Sportradar analysis, reported by InGame, February 2026). Sportradar client data cited by Sports Betting Dime suggests that US in-game handle was about 35 to 40 percent over the past year and approached 50 percent near the end of 2025, while mature European markets sit nearer 70 to 75 percent (Sports Betting Dime). DraftKings’ chief executive said live betting passed half of the company’s handle for the first time in the first quarter of 2025 (Covers).
Other estimates are lower. Stats Perform, a data vendor, reports that in-play accounted for over 62 percent of online sports betting worldwide in 2025 and over 80 percent of tennis turnover (Stats Perform). A Bettingranker study reported by European Gaming put in-play at 47 percent of global wagers in 2024 (European Gaming).
The gap comes from definitions: share of money wagered versus share of bets placed, operator figures versus market-wide figures, and mature versus emerging regions. So when an article says “70 percent of all online bets are live,” treat it as the top of the range, not the norm.
The commercial incentive is clear. An analyst note summarising Sportradar management’s view estimated that each percentage point of in-play share growth adds about €6 million of revenue for the company (Macquarie via Next.io). That is why suppliers keep investing in faster data and automated pricing.
How live odds are computed: from the venue to the screen
If you have wondered how sportsbooks calculate live betting odds, the process has four stages: capture the event, estimate probabilities, add margin and risk controls, then publish or suspend the market.
Stage one: capturing the event
Official data suppliers such as Sportradar, Genius Sports and Stats Perform collect event data using cameras, sensors and human scouts. Optical tracking varies by league. The NBA’s former Second Spectrum system recorded the coordinates of every player and the ball 25 times per second (Tech Briefs). From the 2023-24 season the league moved to Hawk-Eye’s three-dimensional pose tracking, which it said delivers data at sub-second latency (SportBusiness), with Sportradar as the league’s exclusive data provider (ESPN).
Many sports still depend on people. An operator-facing vendor guide says raw data typically reaches a provider’s systems within one to five seconds for well-covered events and up to 15 to 30 seconds for lower-tier coverage (Source Code Lab). That is a vendor estimate rather than an audited benchmark, but it matches the structure of the problem: the lower the tier, the slower the feed.
Stage two: estimating probabilities
Most pricing models descend from statistical work on scoring. The foundational football paper by Mark Dixon and Stuart Coles, published in 1997 in the Journal of the Royal Statistical Society, Series C, built a parametric scoring model on English league data specifically to look for inefficiencies in the betting market (Lancaster University repository).
In-play pricing adds the clock and the game state. Robberechts, Van Haaren and Davis showed at the 2021 ACM KDD conference that win-probability models borrowed from other sports struggle in soccer, especially late in games. They proposed a Bayesian model that treats the goals still to come as a temporal stochastic process and reported well-calibrated results (Robberechts et al., KDD 2021). The authors note that time remaining and the current scoreline have the largest influence on the estimate (KU Leuven DTAI blog).
Stage three: margin and risk
A fair probability is not a price. Operators add margin and adjust for exposure. Sportradar’s Alpha Odds tool, an extension of its Managed Trading Services, recalculates a sportsbook’s financial exposure after each bet. Selections that would increase liability are shortened, and those that reduce it are lengthened (SBC News interview with Sportradar’s Darren Small).
Stage four: publishing and suspending
The UK Gambling Commission notes that sportsbooks build in a delay to make sure their own prices are correct, and that the length of delay varies by operator, by event and by the latency of the data source (UK Gambling Commission). An operator guide describes a typical bet-acceptance delay as one to five seconds (Track360 operator guide).
A correction worth making: you will often see claims that odds update “in under 50 milliseconds.” We could not find a primary source for an end-to-end figure like that. What sources support is sub-second latency for some tracking systems and seconds-level delays elsewhere in the chain.
Where AI actually sits: risk, liability and personalised pricing
The most documented use of machine learning is in trading and risk. Sportradar launched Alpha Odds in 2022 (iGaming Business), and by April 2025 the company said it had been rolled out across 80 percent of events covered by its Managed Trading Services. It also reported that the tool lifted client profits by an average of 11 percent in soccer, basketball and tennis in 2024 compared with conventional odds services (AGBrief). Those are company-reported figures, not independent audits. In the second quarter of 2025, Sportradar’s chief executive said the company was trading close to €45 billion in turnover through the service (Legal Sports Report).
Have human oddsmakers been replaced?
Not entirely. A Genius Sports executive told iGaming Business that automation reduces manual work, especially in risk management and customer segmentation, while trading desks retain subjective input. Sportradar’s Darren Small said many operators are still working out how automation changes the desk, and an industry commentator quoted in the same piece said that “solving” sports betting the way high-frequency traders believe they have solved financial markets remains a long way off (iGaming Business). The claim that manual trading is dead overstates the evidence.
Do books just balance the money?
The textbook view says a bookmaker sets prices to attract equal money on both sides. Steven Levitt tested this on NFL data in a paper published in The Economic Journal in 2004 and found that bookmakers are better forecasters than bettors and deliberately choose prices that deviate from the market-clearing price in order to exploit bettor biases (Levitt, 2004, via IDEAS/RePEc). The implication for AI is that a modern risk engine is better understood as a profit optimiser than as a simple book balancer.
Integrity monitoring and personalisation
Sportradar says its AI-based Universal Fraud Detection System monitors global markets, and that when a match is flagged its trading service is notified and can suspend markets or reduce liability limits (AGBrief, March 2026). Small has also said the depth of Sportradar’s data could eventually support odds tailored to individual customers (SBC News). Our view is that personalised pricing is the area most likely to attract regulatory questions about fairness, though no source we reviewed says regulators have acted on it yet.
Latency, bet delays and courtsiding
Because the data pipeline is imperfect, anyone who learns about an event first has a potential edge. The UK Gambling Commission defines courtsiding as using or transmitting information from a live event for gambling, taking advantage of the gap between live action and television or other feeds (UK Gambling Commission).
Tennis shows the problem clearly. Points are short, matches run in parallel, and official scoring has historically relied on a human entering results. A data-provider executive speaking at a 2018 industry conference called latency an “ongoing challenge” for the sport (SBC News), and a 2026 industry write-up notes that odds can still be exposed to delay when the official score has not yet been entered, for example during clay-court mark inspections (G3 Newswire). A Guardian investigation reported that some umpires had been banned or suspended for deliberately delaying score entry on a data tablet, allegedly after bribes from betting syndicates (Gambling Insider).
The sport’s governing bodies responded by banning spectators from transmitting live scores for commercial purposes, and jointly invested more than $10 million in their own live-data system (FiveThirtyEight). Whether courtsiding is a crime is another matter. The same report describes a British courtsider charged under a sporting-integrity law whose charges were later dropped, and Huck magazine describes the legal position as a grey area (Huck). Anyone considering this should check local law and event rules first.
Can bettors beat automated odds?
The honest answer is that it is hard, it is rarely sustainable, and the best-known evidence is not about live betting.
The odds-comparison study
In a 2017 preprint, Lisandro Kaunitz, Shenjun Zhong and Javier Kreiner analysed historical odds from 32 bookmakers across 818 leagues and 479,440 football matches over the decade to 2015. Rather than building a better forecasting model, they looked for disagreement between bookmakers. Their simulated strategy placed just over 50,000 bets and reported a profit of $98,865. When they tried real money, bookmakers began severely limiting their accounts within a few months (Cosmos Magazine; MIT Technology Review). Caveats: it is a pre-match football study, it relies on historical data, and the paper was a preprint.
Calibration beats accuracy
Walsh and Joshi, writing in Machine Learning with Applications in 2024, trained NBA models over several seasons and tested them on one season of published odds. Selecting models by calibration produced an average return on investment of +34.69 percent, against −35.17 percent when selecting for accuracy (University of Bath research portal). A December 2025 preprint on the English Premier League, which has not been peer reviewed, reported accuracy of roughly 51 percent but no consistent profit, and attributed the gap to calibration error (Qeios preprint). Both results point the same way: a model that is confidently wrong loses money faster than one that is modestly right.
The “contextual blind spot” idea
A common claim is that algorithms miss weather shifts, body language or blowout motivation. These are plausible hypotheses, and it is reasonable for a person watching a match to notice things a data feed has not yet registered. However, we found no peer-reviewed study that quantifies such an edge in live markets, so we treat it as an open question rather than a strategy.
Micro-betting and regulation
Micro-betting means wagering on the next discrete event, such as whether the next pitch is a strike or whether the next play is a run or a pass. Sportradar has expanded micro-markets across NBA, MLB and WNBA products, adding eight extra micro-markets for the NBA playoffs compared with the regular season (Legal Sports Report).
Baseball’s $200 cap
On November 10, 2025, Major League Baseball announced that its authorised sportsbook partners would cap pitch-level wagers at $200 and exclude them from parlays. The league said these markets present heightened integrity risk because they focus on one-off events that a single player can determine and that may be inconsequential to the game’s outcome (MLB press release). The announcement came a day after two Cleveland Guardians pitchers were indicted over allegations they rigged pitches for gamblers (CBS News). Those are allegations, not convictions. Ohio’s regulator then dropped its plan for a player-specific micro-bet ban (Legal Sports Report).
New Jersey and the wider picture
New Jersey’s Senate Bill 2160 would ban micro-bets outright. It cleared a Senate committee on March 23, 2026, while the Assembly’s companion bill, A3258, was narrowed in committee to cover online micro-bets only (Gambling Insider). A law-firm briefing from the period noted that no state or federal law then explicitly banned micro-bets in professional sports (Fox Rothschild). Separately, at least 17 US states restrict college player props (iGaming Business). Legislation moves quickly, so check the current status before relying on any of this.
In-play betting and gambling harm
Researchers have long worried that in-play betting’s structure, with a short gap between bet and outcome and continuous opportunities to bet, could contribute to harm (Gainsbury et al., PubMed Central). In a 2020 survey of 501 Australian online sports bettors, published in Frontiers in Psychiatry, people who bet in-play had higher problem-gambling severity scores, and the authors concluded in-play betting warrants specific regulatory attention (Gainsbury, Abarbanel and Blaszczynski, 2020). The same paper acknowledges that there is limited research showing in-play betting causes harm, so the finding is an association, not proof of cause.
The evidence base is also contested. When the UK Gambling Commission classed in-play betting as an indicator of harm in draft guidance, law firm Harris Hagan and Regulus Partners argued that the cited evidence did not support the classification (Harris Hagan). Both positions are worth knowing before you form a view.
What comes next
This section is our analysis. It extrapolates from reported trends and is not a forecast.
Broader automation. Sportradar said it would make Alpha Odds available for every sport by the end of 2025 and extended it to cricket with more than 400 in-play markets per match (AGBrief). Expect automated liability management to become the default rather than a premium option.
AI-driven content. Sportradar’s Foresight streaming product is deployed across ATP events, and a company case study claimed a 30 percent turnover lift on covered events (Legal Sports Report). That is a vendor claim, but it shows why suppliers see video and pricing as one product.
Guardrails on micro-markets. Baseball’s cap-and-no-parlay template is now a reference point for regulators and other leagues. Whether it spreads by league agreement or by statute is unresolved.
Better evaluation. The research on calibration suggests that the next competitive gap will come from probability quality, not model complexity.
Scrutiny of personalisation. If operators move toward customer-specific pricing and limits, fairness and transparency rules are a plausible next regulatory topic.
Frequently asked questions
Do sportsbooks use AI to set live betting odds?
Largely, yes, especially for risk and liability. Data suppliers sell machine-learning pricing tools to operators, but executives at Sportradar and Genius Sports still describe a role for trading desks (iGaming Business).
How fast do live odds update?
It depends on the stage. Some tracking systems report sub-second latency (SportBusiness), while scout-based data can take seconds to reach a provider and operators often add a deliberate acceptance delay (UK Gambling Commission).
Why do sportsbooks limit winning bettors?
Levitt’s research indicates bookmakers price to maximise profit (IDEAS/RePEc), and the Kaunitz team reported being severely limited after a few months of real-money betting (Cosmos Magazine).
What is latency arbitrage, and is courtsiding legal?
Latency arbitrage means exploiting the gap between live action and the feed an operator prices from. Courtsiding is the venue-based version of it. Legality varies, and tennis authorities ban it (FiveThirtyEight).
Is micro-betting banned?
Not by any enacted US state law in the reporting we reviewed. Baseball caps pitch-level bets at $200 (MLB), and New Jersey bills remain pending (Gambling Insider).
Editorial standards and disclosure
Authorship. This article was prepared by the Adzvelo Editorial Team and last updated on October 23, 2026. It is educational and is not betting, legal or financial advice.
How we sourced it. We prioritised peer-reviewed papers, regulator pages and league press releases, then reputable trade press, and treated vendor and operator material as claims to be attributed rather than facts. Each statistic is linked to its source in the section where it appears. Preprints and company-reported figures are labelled as such.
Limits. Betting markets, data products and laws change quickly. Verify current rules in your jurisdiction.
Disclosure. Adzvelo publishes operator reviews and may earn referral commissions elsewhere on the site. This article does not rank or recommend any operator.