AI predictions and the UK sports betting market

Updated October 2026

Licensed

Available in US

Fast payouts

18+ Only

27 October 2026 — Gambling Commission public register, GAMSTOP membership records and each operator’s published terms.

Something has changed about the vocabulary of sports betting. A search for “AI predictions” returns tools that look like a Bloomberg terminal, with match forecasts, probability bars and bankroll curves. The pitch is sharp. So is the question it leaves on the table: when the screen is full of signals, what does a punter actually do with one?

The answer sits inside the same arithmetic the bookmaker uses. A prediction is a number — usually a probability between zero and one — and the price on the slip is another number. The whole conversation is about the gap between the two. AI tools reach for that gap faster than a pencil ever could. They cannot widen it beyond what the data already allows, and the data here is thinner than the marketing admits. Form, line-ups, head-to-head records, expected goals, injury news, even weather — all of it goes into a model. The result is still a probability, not a certainty, and the price the bookie quotes already has the firm’s margin built into it. The fact that a model built the number changes nothing about that.

This page works through what the AI prediction industry actually does, where the value sits and where the trap is. It also lays out ten licensed UK bookmakers side by side, the rules that govern them and the maths that decides whether a price is worth taking. No picks. No “sure” bets. No system that wins.

The bookmakers reviewed here

The comparison below sits in front of the rest on purpose. It names the brands the rest of the page returns to, in the order it returns to them.

A column on AI prediction tooling would have been a column of dashes. The Commission’s public register tells you who holds a licence, not who has built a model. The closest any of these brands comes to a public AI feature is a live stats overlay on a price, and a few of them are explicit about it: bet365’s Match Live, Midnite’s acca builder, Paddy Power’s ACCA stats, BetVictor’s Smart Stats. None of them sells those stats as a prediction. The number the punter bets on comes from the same price the bookie set. A model is what sits in front of the price in the punter’s head, not in the slip.

What “AI predictions” actually means in sports betting

The phrase covers a family of tools. At the simple end: a system that reads form, goals, expected goals, possession and head-to-head data and returns a probability for each outcome. At the more advanced end: a system that processes deep event data, adjusting for squad news and conditions, to calculate a probability. In between are statistical frameworks, ensembles, and tools that synthesise pre-match reports into a confidence score, as well as automated scanners that identify mispriced odds across hundreds of fixtures.

What they share is the output. A number. A probability. The rest is plumbing.

A useful way to think about any of them is as a function from data to probability. The data goes in; a number between zero and one comes out. The number is not a guarantee. The model that produced it has been trained on matches that have already happened, with a margin of error that depends on the depth of the league, the noise in the data and how much of the news cycle the model could read. The price on the slip has been set by a bookmaker with its own data, its own risk team and its own margin.

Where AI genuinely moves the needle for a punter is in the parts of the job a human is bad at. Staking across hundreds of selections without losing track of exposure. Watching a list of 200 pre-match prices for a misprint. Reading every team-news update on a Saturday morning without missing a side. Building a model the punter actually understands, instead of trusting one they don’t. None of those change the fact that a single bet is a single bet.

The data inside the model

Three layers go into any serious model. The first is performance data: results, goals, shots, expected goals, expected assists, xG against, set-piece frequency, distance covered, sprint counts. The second is contextual: line-ups, injuries, suspensions, the manager’s press conference, weather, the gap since the last match, the importance of the fixture, derby records. The third is the market: how the price has moved since it opened, the volume on each side, where the bookies disagree.

The layers interact in ways the model has to learn on its own. A side that scores early changes shape, concedes more shots in the second half, and the model’s first-half probability has to give way to a conditional one. A side missing its first-choice goalkeeper is no longer the same bet as the same side with its number one. The model has to be retrained, or recalibrated, when those changes are large enough.

For a punter, the consequence is practical. The longer the league’s history, the more data the model has to work with, and the lower the variance. A Premier League fixture and a Lithuanian second-tier game are not the same proposition. The model’s accuracy is also conditioned on the market the model is trying to predict. Goals markets are more stable than corners markets. Corners markets are more stable than booking markets. A model trained on goals can be a poor guide to a corner.

Where the tool helps and where it doesn’t

A tool helps when the task is routine and the gain is in volume. A punter with two hours on a Saturday is not going to build 30 separate models for 30 leagues. A model can. It also helps when the model is well calibrated — when, of the matches it rates as 60% home wins, about 60% end in home wins. A poorly calibrated model looks the same on the screen. The work is in the calibration, which is invisible to the punter who only sees the output.

A tool does not help when the fixture is decided by information the model cannot read. A press conference at 6pm where the manager names a weakened side. A pitch that turns to mud in the half-hour of rain. A player playing through illness. The model’s probability for that fixture is built on the side that started, not the side that finished.

A tool also does not help when the punter’s decision is not the probability but the price. A probability of 0.6 on a home win is irrelevant if the price is 1.50. The implied probability is 0.67. There is no value in the bet. A probability of 0.4 on an away win at 3.20 is a different story. The implied probability is 0.31. The model says 0.4. The bookie says 0.31. The edge, if any, lives in the gap, and the gap is what the punter bets on, not the probability itself.

What a value bet is, and why no tip is ever certain

A value bet is a price that is longer than the true probability justifies. The arithmetic is short. If the model puts the home win at 0.6, the true fair decimal odds are 1 / 0.6, or 1.67. Anything bigger than 1.67 on the slip is a value bet, in the model’s view. Anything shorter is a price the model says the punter should leave alone.

The reason no tip is ever certain is in the same sentence. The probability is the model’s, not reality’s. On a sample of 100 matches the model rates at 0.6, some 60 will go as the model says, on average, if the model is well calibrated. The other 40 will not. A punter with a single bet does not have a sample. A punter with 1,000 bets over a year has a sample, and a model with a small edge will produce a small profit, before the bookmaker’s margin, the staking and the variance of real life.

The trap is to confuse the model being right on the average with the model being right on the next bet. The two are different things. A model that wins 55% of the time at prices of 1.90 still loses 45% of bets, and any of them can be the next one. The discipline of value betting is to bet when the edge is there, and not bet when it isn’t, and to let the sample do the work over weeks and months.

A worked example

Take a single match. The model rates Manchester City at home as 0.55 to win, 0.25 to draw and 0.20 to lose. The fair decimal odds are 1.82, 4.00 and 5.00. The bookmaker is quoting 1.70, 3.75 and 5.50. The model’s 0.55 implies a fair price of 1.82. The slip shows 1.70. There is no value.

Look at the same fixture through a different model, one that puts the draw at 0.22 and the away win at 0.23. The fair price on the draw is 4.55, the bookie is at 3.75. Still no value. The home price of 1.70 against a true probability of 0.55 is a 0.04 margin in the bookie’s favour. The away price of 5.50 against a 0.23 probability is a 0.21 gap in the punter’s favour. The bet is the away win, on this model.

The point is not that the away win is “the right answer”. The point is that the value is the only thing the punter can act on. The outcome is a coin with a 0.23 bias, and the bet wins some 23% of the time at a price of 5.50. Over a year of similar bets, the model wins some, loses some, and finishes roughly 5% ahead of a flat-stake punter who took the favourites. Or it does not. The calibration is the whole game.

The difference between the model’s probability and the model’s tip

A tipster who says “I have City to win” is selling a confidence. A model that says “0.55” is selling a probability. The two are different products. The first tells the punter what to do. The second asks the punter to do their own arithmetic against the slip. For a punter who can read the slip, the second is worth more. For a punter who cannot, the first feels more useful and costs more in overconfidence.

The point is not that tips are worthless. The point is that a price the punter can verify is worth more than a price the punter is told to trust. The model’s probability can be back-tested, by the punter, on a year of fixtures. The tipster’s reputation cannot be tested the same way.

How to read the bookmaker’s margin in the odds

Every price the bookie quotes has the firm’s margin built in. The margin is the cut the bookie takes on every unit staked, before the punter and the rest of the market is paid out. The way to see it is to invert the price, add the inverted prices together and look at how much bigger the sum is than 1.

Take a two-way market. The bookie quotes 1.90 on each side, decimal odds. The inverted prices are 1 / 1.90 and 1 / 1.90, or 0.526 and 0.526. The sum is 1.053. The overround is 5.3%. The margin is 5.3%.

The general pattern is the same on every market. In a three-way 1X2 market, the three prices invert and add to something larger than 1, and the difference is the margin. The same logic applies in a market with five runners: the prices are inverted, summed, and the surplus above 1 is the margin. The lower the margin, the more of every unit staked the bookie returns to the market on a winning bet. The higher the margin, the more the bookie keeps.

An illustrative two-way market at 1.90 / 1.90 — used here as an example rather than a quoted price — produces a margin of roughly 5%, with most of the return going to the winning side and a small share left as the bookie’s take. The exact percentage is set by the arithmetic, not by the marketing, and a bookie with a lower margin on the same market is offering a better deal in the long run, all else equal.

The practical consequence is that any model the punter builds is competing with a margin that already sits in every price. A model that returns a 5% edge over the true probability is matched, before staking, by a 5% margin in the bookie’s favour. A model that returns a 3% edge is behind from the first bet. A model that returns a 10% edge has a 5% net edge after the margin, which is the kind of edge a flat-stake punter can build a bankroll on if the sample is large enough.

Why the margin varies by market

The margin is not the same on every market at the same bookie. A Premier League 1X2 market is sharper than a Championship 1X2 market, and both are sharper than a League One 1X2 market. The sharper the market, the lower the margin, because more bookies are quoting and more money is moving. A betting exchange, where punters back and lay against each other, runs a margin close to zero. A traditional bookie, which carries the risk, runs a margin to cover it.

The punter’s strategy follows from the structure. The model is best deployed on the markets with the lowest margin, because the model’s edge is not eaten by the bookie’s cut. A model with a 5% edge on a 2% margin is in front. A model with a 5% edge on an 8% margin is behind. The markets the bookie advertises most heavily are usually the markets with the highest margin. A free bet offer is paid for by the margin on the bets the punter places to clear it.

How the return on a bet is worked out

The arithmetic is short. For a single bet at decimal odds O with stake S, the return is S × O if the bet wins. The net profit is S × (O − 1). The return is the total paid back, including the stake; the profit is the part the punter keeps after the stake is returned.

A worked example. A punter places £10 on a home win at 2.50. The return is £10 × 2.50 = £25. The net profit is £10 × (2.50 − 1) = £15. The return includes the £10 stake, which is paid back with the win. The profit is the £15 the punter did not put down.

A combination bet is the same arithmetic, with the legs multiplied. A double at 2.00 and 3.00 is a combined price of 6.00. A £10 stake returns £60 on a win, with a profit of £50. A treble at 1.50, 2.00 and 3.50 is a combined price of 10.50. A £10 stake returns £105 on a win, with a profit of £95.

The point of the multiplication is the same as the point of the model. The combined price is a single probability turned upside down. The punter’s edge, if any, sits in the gap between the combined price and the true combined probability. The bookie’s margin sits inside every leg and is compounded across them, which is why a four-fold acca is a worse proposition, on the same model accuracy, than four singles.

The free bet and the stake

A free bet is the same arithmetic, with one change. The stake is not paid back with the win. A £10 free bet at 2.50 returns £25 on a win, of which £10 is the bonus and £15 is the profit. The punter’s net is £15, not £25. The arithmetic is the same as a single bet at 2.50, but the punter has not put any of their own money on the line, and the return to the punter’s bankroll is the profit only.

The reason this matters is in the offer’s wording. A free bet whose stake is paid back is, in effect, a cash bonus. A free bet whose stake is not paid back is a smaller value at the same price. Most UK free bets are of the second kind. The 10x cap on wagering introduced in 2026 means that a £10 bonus at 50x wagering would have needed £500 of bets; under the cap no more than £100 is required, and the offer’s true value is whatever the punter can clear, at the model’s edge, before the expiry.

The each-way bet

An each-way bet is two bets of equal stake, one on the win and one on the place. The stake doubles. The win leg pays at the full price. The place leg pays at a fraction of the price, with the fraction set by the number of runners and the type of race. A £10 each-way bet at 10.00 with a quarter odds place term is two £10 bets. The win leg pays £100 plus the £10 stake on a win. The place leg pays £25 plus the £10 stake on a place. The total stake is £20 and the maximum return, on a win, is £200 plus the £40 in stakes.

The each-way is a bet on the horse to place, not just to win. The place leg is the part that flatters the price. A punter who wants the place should bet the place. A punter who wants the win should bet the win. The each-way is for the punter who wants the place more than the win and is willing to pay the second stake for it.

The accumulator and the bookie’s margin

An accumulator is a combination bet with three or more legs. The price is the product of the legs. A four-fold at 1.50, 1.80, 2.20 and 3.00 is a combined price of 17.82. A £10 stake returns £178.20 on a win. The probability is roughly 0.056, on a model that has each leg at the true probability. The book’s price implies a probability of 0.054, and the difference is the margin compounded across the four legs.

The compounding is the reason a 10-fold acca is a worse bet than ten singles, on the same model accuracy. The book’s margin on each leg is multiplied, and the punter’s edge is divided. An accumulator insurance offer, which refunds the stake as a free bet if one leg lets the punter down, softens the loss but does not change the arithmetic. A free bet, as above, returns the profit only. A £10 stake on a 10-fold acca at a combined price of 1,000 returns £10,000 on a win. A refund of £10 as a free bet returns the profit on a free bet at the same price, which is £9.99, not £10.

What the law actually says about AI sports betting in the UK

The regulator is the Gambling Commission, a non-departmental public body sponsored by the Department for Culture, Media and Sport. The framework is the Gambling Act 2005, which covers Great Britain — England, Scotland and Wales. Northern Ireland is outside the Act and runs on its own older law. The Act licenses the whole commercial market. There is no state monopoly, no cap on the number of licences, and no statutory list of permitted bets. An operator holds a licence for a category — remote general betting standard (real events) for the brands reviewed here — and offers markets under its own rules and the Commission’s Licence Conditions and Codes of Practice.

Since the Gambling (Licensing and Advertising) Act 2014, an operator whose remote facilities are used in Great Britain needs a Commission licence wherever it is based. A Curaçao, Maltese or Gibraltar licence is not a substitute. The Commission’s public register is the whole test of who holds one. It can be searched by business name, trading name, domain or account number, and a licence that is not on the register is not in force. The brands in the table above are all on the register, and the account numbers in the table are the ones to check.

The legal test for a punter in Great Britain is whether the operator is licensed at the point of consumption. Using an unlicensed site is not, in itself, a criminal offence aimed at the customer. The customer’s loss is the protections the licence confers: no GAMSTOP coverage, no Commission complaints route, no approved dispute-resolution body, no British consumer redress if funds are withheld, and no application of the 10x bonus cap or the £150 vulnerability check. The enforcement is against the operator, through cease-and-desist notices, search-engine referrals and work with registrars, hosts and payment providers. The Commission has been given, or is about to receive, court-order DNS blocking powers. ISP blocking is not in operation.

The minimum age and the identity check

The minimum age to bet in Great Britain is 18. Since 7 May 2019, name, address and date of birth are verified before a customer can deposit or gamble, including with a free bet or bonus. There is no anonymous play at a licensed site. The verification is a check of public records, not a check of income. The financial vulnerability check applies when a customer’s deposits minus withdrawals exceed £150 in a rolling 30 days; it is a public-record check (bankruptcy, county court judgments), not a credit reference check, and it has been in force since 28 February 2025. The financial risk assessment — what the press calls “affordability checks” — is announced in a staged plan with no start date.

Credit cards

Credit cards are banned for gambling in Great Britain since 14 April 2020, including credit-card-funded payments through an e-wallet. Debit cards and bank-based payments are not covered. A British named-account rule, a state list of permitted payment methods and any rule for or against cryptocurrency deposits at licensed operators are not established. Payments are described by what each operator’s terms say.

Advertising and bonuses

Advertising is not licensed by the Commission. It runs through the CAP Code (non-broadcast) and BCAP Code (broadcast), enforced by the ASA, alongside the LCCP marketing conditions. The CAP Code section 16 requires that marketing be socially responsible, not be of strong appeal to under-18s, and not feature anyone who is or seems to be under 25 in a significant role, with an exception for marketing placed where a bet can be made directly. That rule concerns who may appear in the advertisement, not who may bet.

The promotion rules in force since 19 January 2026 cap wagering requirements on bonuses at 10x. A £10 bonus at 50x needed £500 of bets; under the cap no more than £100 is required. Mixed-product promotions — an offer that requires gambling on two or more product types such as betting, casino, bingo and lottery, for example “bet on football to get free spins” — are banned. Whether the cap applies to free bets with no wagering condition, and any cap on the money value of a bonus, are not established.

“Whistle-to-whistle” is an industry code, not law. Since 1 August 2019, betting advertisements are not shown on TV from five minutes before a live sporting event to five minutes after it, before the 9pm watershed, with horse and greyhound racing excluded. The Premier League clubs’ withdrawal of gambling sponsorship from the front of matchday shirts, from the 2026/27 season, is voluntary. Sleeves and hoardings remain.

Tax

The player pays no tax on betting winnings in Great Britain. General Betting Duty is the operator’s tax on gross profits, currently 15%, with a new remote betting rate of 25% from 1 April 2027. Remote bets on UK horse racing, bets on self-service betting terminals in licensed premises, pool betting and in-person betting stay at 15%. Until 1 April 2027 the remote rate stays at 15%, and whether operators pass duty on through odds is not established. Spread betting has its own rates, 3% on financial and 10% on other. The Commission’s statutory levy is a separate charge, not a tax.

What responsible gambling looks like in practice

GAMSTOP is the national multi-operator online self-exclusion scheme and a mandatory licence condition for online operators since 31 March 2020. The periods are six months, one year, five years, or five years with auto-renewal, and after the minimum period the exclusion stays in place for a further seven years unless the customer asks GAMSTOP to remove it. It covers online sites and apps. Betting shops use a separate service. The brands reviewed here are all GAMSTOP members, and a self-exclusion is honoured across the whole market.

The safer-gambling tools the bookmakers offer are a separate layer. Unibet, for example, lists deposit limits, a Player Profile test, product restrictions and self-exclusion, alongside live chat support and links to BeGambleAware, GamCare, GAMSTOP, Gambling Therapy and the BGC. Betwright states it is licensed in the UK and provides responsible gaming tools, and its customers must be aged 18 or over. London Bet is published by Thistle Bet Limited, with a 17+ App Store age rating that reflects the same floor. The tools are operator-side: the punter sets a deposit limit, a stake limit or a time-out, and the bookie enforces it. The model is no help on its own, and a punter who bets more than they planned has lost before the price is set.

Help is available outside the bookmakers. GamCare runs the National Gambling Helpline. GambleAware funds treatment and education. The NHS runs gambling clinics in several cities. Gamblers Anonymous runs meetings. The Commission’s register is the first place to check a brand, the safer-gambling page on the bookie’s site is the second, and a self-exclusion is the third if the punter has decided the question.

The deposit limit and the punter’s budget

A deposit limit is a ceiling on the punter’s deposits over a period. The first changes under RTS 12 have applied since 31 October 2025, with a second phase postponed to 30 September 2026. There is no state-set ceiling: the limit is the customer’s choice. A punter who sets a limit and sticks to it has put a floor under the worst case. A punter who raises the limit after a loss has put the floor in the wrong place.

The model’s role in this is small. A model that says the home win is 0.55 does not say anything about the punter’s budget. The stake is a function of the bankroll, the price and the edge, not of the probability. A flat 2% of bankroll per bet is a common starting point. A Kelly fraction, where the stake scales with the edge, is the mathematical optimum and the practical route to ruin for a punter who overestimates the edge. The size of the bet is a separate question from the bet itself.

Mobile and app: what changes on a handset

The betting app is a different surface from the website, and the same bookie can be a different experience in each. The differences are not in the markets — the same Premier League fixtures are priced — but in the way the punter reaches them. The bet365 app, for example, offers in-play betting and live streaming of events worldwide, with Bet Track, Match Live and a Live Golf Tracker for following bets in play. The Paddy Power app adds Bet Builder+, Popular Bet Builders, Super Sub where a bet rolls over to the replacement if a selected player is substituted, expanded Win More markets and player and ACCA stats. The Midnite app surfaces an acca builder, popular accas and price boosts with maximum-stake limits, alongside in-play events. The Lottoland app combines casino, lottery betting, sports betting and scratchcards, with a stated Gambling Commission licence for customers in Great Britain.

The touch screen changes the workflow. The bet is placed with three taps from the home screen; the price is set in the time it takes to confirm. In-play markets refresh in seconds. Cash out, when it is offered, settles the bet on the current price. The temptation to bet more often, on more markets, on shorter odds, is built into the surface. A punter who planned three bets for the afternoon and placed nine, because the app made each one easy, has lost the discipline the bankroll requires.

The browser is a different medium. The same bookie, on a desktop, is slower to refresh and harder to bet from. The punters who bet on a desktop tend to be the punters who pre-plan. The punters who bet on the app tend to be the punters who react. Both are valid. The model’s role in each is the same: produce a probability, compare to the price, place the bet at a stake the bankroll supports.

The model on the phone

A model on the phone is the same model on the desktop, with a different interface. Some apps offer a price comparison across bookies. Some offer a stats overlay on the fixture. Some offer an alert when a price moves through a threshold. None of them offers a free lunch. The alert is a way to be told when the price is wrong; it is not a way to be told when the model is right. The two are different conditions, and the difference is the gap the punter is paid to find.

A model run in the cloud, with a Python notebook and a paid data feed, is the same model run on a phone with a stats app. The output is the same: a probability, or a price, or a signal. The punter’s job is the same in each: read the slip, set the stake, place the bet. The model is a tool, not a tipster.

How the brands stack up against the question

BetGoodwin

Goodwin Racing Limited, Commission account 2927. A racing-led bookie with Best Odds Guaranteed on all UK and Irish horse and greyhound racing, up to £1,000. The offer is specific to two sports, the ceiling is high, and the rest of the book is a standard fixed-odds line on the same events. For a punter whose model is built on racing form, the BOG is the headline: a starting price that ends up longer than the bet is paid at the longer price, with no cap that the model has to clear. The brand is one of the ten the comparison routes, and it sits at the top of the list because the racing product is the most specific.

Betway

Betway Limited, Commission account 39372. The welcome offer is £40 in free bet tokens, credited as four £10 instalments, on settlement of a first bet of at least £10 at minimum odds of 2.0, with a 7-day expiry and debit card deposits only. The arithmetic on the offer is the standard 10x cap era. A £40 free bet at four instalments, each expiring in 7 days, gives the punter four shots at a price of 2.0 or longer. The model is no help on the offer itself — the offer is fixed — but the model is the reason the punter picks the price. A punter using a model to identify Premier League value at odds of 2.0 or greater can use the four instalments to systematically test their edge, treating each as a distinct opportunity rather than a single lump sum.

bet365

Hillside (UK Sports) ENC, Commission account 55148. The biggest bookie in the world, with in-play betting and live streaming of events worldwide, plus Bet Track, Match Live and a Live Golf Tracker. The interface is a model of its own: the in-play page refreshes in seconds, the live data is layered over the price, and the cash out is on the slip before the next goal. For a punter whose model is built on in-play data — the next goal, the next corner, the next booking — bet365 is the venue with the most data to feed the model and the most markets to bet on. The risk is the same as the rest: the price has the margin in it, the model’s edge is the gap, and the slip is settled in seconds either way.

William Hill

WHG (International) Limited, Commission account 39225. Best Odds Guaranteed on UK and Irish horse racing from 8 AM on race day, on single and multiple bets, with WHTV streaming that requires a funded account or a bet in the last 24 hours and live racing subject to a £1 qualifying bet. The BOG is the racing headline, and the streaming condition is the price of admission. The BOG ensures that if the starting price outperforms the early price, the model’s returns are bolstered on both win and place legs, while the streaming service provides the real-time visual confirmation for in-play adjustments.

Lottoland

EU Lotto Limited, Commission account 38991. A combination of casino, lottery betting, sports betting and scratchcards, with a stated Gambling Commission licence for customers in Great Britain. The sportsbook is one of four products on the same wallet, and the punter who wants lottery and sports on the same site has the option. The model is irrelevant to the lottery side — lottery betting is a different market with a different structure — and the sportsbook is a smaller offering than the dedicated brands. For a punter whose model is sports-only, the rest of the wallet is a distraction. For a punter who wants both, the wallet is the draw.

VBet

SCGO Limited, Commission account 44662. Bet Builder, pre-match and in-play betting, live streaming and a live calendar with results and stats. The bet builder is the headline: the same fixture, priced across combinations, with the margin on each leg compounded. For a punter whose model is a same-game multi model, VBet is a venue with the surface to support it. For a punter who bets singles, the bet builder is a feature they will not use.

Midnite

Midnite Limited, Commission account 101839. In-play events, an acca builder, popular accas and price boosts with maximum-stake limits. The welcome offer is £25 in free bets plus 50 free spins for new customers who bet £5 or more on pre-match odds of EVS (2.0) or greater within 14 days of sign-up, valid for 7 days. The price boosts are the interesting surface: a boosted price is a price the bookie has raised, and the punter’s edge is the gap between the boosted price and the true probability. The cap on the maximum stake is the cost of the boost, and the model’s job is to find the boost that is long enough to clear the cap.

Sky Bet

SBG Sports Limited, Commission account 67370. The welcome offer is £30 in free bets, credited as three £10 tokens after a deposit of at least £5 and a first single or each-way bet of at least 5p at odds of 1/1 or greater. The 5p minimum stake is the unusual feature: it lets the punter place the qualifying bet at the smallest stake the site allows, leaving the model to find a price of 1/1 or longer without committing more than 5p. With the model identifying potential value near 1/1, this 5p qualifying threshold allows the bettor to unlock the tokens while maintaining a minimal exposure on the initial account activity.

Tote

TDCO Limited, Commission account 54779. Pool betting on UK and Irish horse racing, in a separate licence class from fixed-odds bookmakers. The dividend is the pool, divided among the winning tickets after the take. The model’s role is the same as on a fixed-odds book: pick the horse, set the stake, accept the dividend at the end. The difference is the dividend is not known at the time of the bet. A punter whose model is a probabilities model has to price the dividend against the implied probability of the field, and the pool’s overround is harder to read than a fixed-odds book. For a punter who likes the pool’s structure, the Tote is the venue. For a punter who likes the fixed price, the rest of the list is the venue.

Paddy Power

PPB Counterparty Services Limited, Commission account 39439. Bet Builder+, Popular Bet Builders, Super Sub where a bet rolls over to the replacement if a selected player is substituted, expanded Win More markets and player and ACCA stats. The Super Sub is the unusual feature: a bet on a player who is taken off before the bet is settled rolls over to the replacement, and the model’s job is to price the replacement, not the named player. For a punter with a model that includes substitution rates by position and by manager, the Super Sub is a market the model can read. For a punter whose model is a price model, the Super Sub is a feature that changes the bet’s structure.

What the numbers mean in the round

A bookmaker’s margin sits inside every price on the slip. A model that returns a 5% edge over the true probability is matched by a 5% margin in the bookie’s favour, before staking. A model that returns a 3% edge is behind from the first bet. A model that returns a 10% edge is in front by 5%, net of margin, which is the kind of edge a flat-stake punter can build a bankroll on if the sample is large enough.

The brand-by-brand differences are in the surface, not the math. BetGoodwin’s BOG is paid on racing only. William Hill’s BOG is paid on racing, with a streaming condition. bet365’s live data is the in-play surface. Midnite’s price boosts have a maximum stake. Sky Bet’s qualifying bet is 5p. Paddy Power’s Super Sub rolls the bet over. The Tote pays a pool dividend. None of those differences is an edge. The edge sits in the model, and the model is the punter’s, not the bookie’s.

The bonus economy is the same arithmetic in a different costume. A free bet is a stake the bookie has paid for, with a price floor and an expiry. A welcome offer is a free bet on a first bet. An acca insurance is a free bet on a near-miss. The punter’s job in each is to find a price that is long enough to clear the floor and the expiry, and short enough that the model says the price is wrong. The 10x cap is the Commission’s contribution to the conversation: a £10 bonus at 50x needed £500 of bets; under the cap, no more than £100. The cap does not make a bad offer good. It makes a 50x offer a 10x offer, and the punter’s job is unchanged.

The shape of the bettor who benefits is the same. A model, a bankroll, a discipline. A bookmaker with a low margin, a price floor the model can clear, and an expiry the bankroll can survive. A safer-gambling tool the punter sets and the bookie enforces. A self-exclusion the punter can trigger if the tools stop working. The brands in the table are the venues. The arithmetic is the work. The rest is noise.

What an AI tool is, and is not, worth

An AI tool is worth what it does to the punter’s process. A model that automates a 30-minute Saturday routine is worth an hour. A model that watches 200 pre-match prices for a misprint is worth a Saturday afternoon. A model that reads every team-news update without missing one is worth a morning. A model that the punter understands, and can back-test, and can calibrate, is worth the year it takes to build.

An AI tool is not worth a tipster’s reputation. A tipster is paid to be confident. A model is paid to be accurate. The two are different products, and a punter who confuses them is overpaying for confidence. The model’s edge is the gap between its probability and the slip’s price. The tipster’s edge is the punter’s willingness to follow. The first can be back-tested. The second cannot.

A tool that promises a winning system is the wrong tool. A model that says “this is the price, this is the probability, this is the edge” is the right one. The difference is the model’s honesty about its own uncertainty. A well-calibrated model says “0.55” and means “60% of these bets win, on average”. A poorly-calibrated model says “0.55” and means “the punter is welcome to trust me”. The first is a tool. The second is a tipster in a model’s clothing.

The brands in the comparison above are the venues a UK punter can bet at, with a model, a price and a stake. The model is the punter’s. The price is the bookie’s. The stake is the bankroll’s. The maths is the same in every venue, and the maths is the work.

Frequently asked questions

Can AI help with betting?

It can automate parts of the job a human is bad at: scanning large numbers of prices, watching team-news feeds, building a probability from a large dataset, and keeping a staking plan on track. It cannot widen the edge beyond what the data allows, and it cannot know the outcome of a fixture the data has not yet seen. The result is a probability, not a tip, and the price the punter bets on is the bookie’s, with the margin already in it.

Can AI help with sports betting?

Yes, in the same way it helps in any prediction task. A model trained on form, expected goals, line-ups and contextual data can return a probability for each outcome of a fixture. The probability is then compared to the price on the slip, and the gap, if any, is the bet. The work is in calibrating the model, in choosing the markets where the margin is lowest, and in keeping the staking plan on track across a large sample of bets.

Can AI predict sports betting?

A model can return a probability. A probability is not a prediction in the sense of a guaranteed outcome. On a sample of 100 matches the model rates at 0.6, some 60 will go as the model says, on average, if the model is well calibrated. The other 40 will not. The punter with a single bet does not have a sample, and the punter with a year of bets does. The edge, if any, is in the gap, and the gap is the work.

What is a value bet and why is no tip ever certain?

A value bet is a price that is longer than the true probability justifies. If the model rates the home win at 0.6, the fair decimal odds are 1.67. Anything longer is a value bet on the model’s view. No tip is ever certain because the probability is the model’s, not reality’s, and any single bet can lose. A model that wins 55% of the time at prices of 1.90 still loses 45% of bets, and any of them can be the next one.

What is the bookmaker’s margin and how can it be seen in the odds?

The margin is the cut the bookie takes on every unit staked, before the market is paid out. To see it, invert each price, add the inverted prices together and look at the over the result over 1. On a two-way market at 1.90 and 1.90, the inverted prices are 0.526 and 0.526, summing to 1.053. The overround is 5.3%, which is the margin. The lower the margin, the more of every unit staked the bookie returns to the market on a winning bet.

How is the return on a bet worked out from the stake and the odds?

For a single bet at decimal odds O with stake S, the return is S × O on a win, including the stake. The net profit is S × (O − 1). A combination bet multiplies the legs: a double at 2.00 and 3.00 is a combined price of 6.00, and a £10 stake returns £60 on a win, with a profit of £50. A free bet does not pay the stake back, so a £10 free bet at 2.50 returns £25 on a win, of which £15 is the profit.

Content created by the «Free bets» team