Industry Insights
How bookmakers set odds using automated systems and human traders. Discover AI-powered sports betting risk management, in-play technology, and the future hybrid model.
By Eleanor Caldwell
Picture a bookmaker's trading room in the early 2000s: traders calling across the desks, prices tweaked by hand, a roomful of people reacting to every twist in the market. Now picture the same job today. Real-time data feeds, automated risk systems and pricing engines are crunching millions of data points in milliseconds, and most of the room is looking at screens rather than shouting.
So are the traders finished? Not a chance. The modern sportsbook is a hybrid. Algorithms handle routine pricing, risk checks and market management at scale, while people provide oversight, deal with the awkward cases and make the calls a model can't. Kambi, one of the industry's biggest platform providers, reports that AI-powered pricing and trading now covers a large and growing share of the bets on its network.
We'll look at how odds were set before automation, what the algorithms do now, where traders still earn their keep, and what the next few years might bring. Along the way, there's plenty for bettors to use.
Algorithms win on speed, scale and consistency. Traders win on context, psychology and the messy situations data can't explain. We think the sharpest bettors can find their edge where the machines have the least to work with: thin markets, lower leagues and teams in flux.
To appreciate what algorithms changed, start with the people who came first. For decades, traders and odds compilers were the backbone of the industry, setting and nudging prices with a mix of mathematics, deep football knowledge and instinct.
Their toolkit still underpins betting maths today. Supremacy was their view of which team was better and by how many goals. Expectancy captured the goals each side was expected to score, based on history and form. The Poisson distribution, a statistical model of goal probabilities, turned those inputs into the likelihood of every scoreline from 0-0 upwards. Our guide to how football betting markets are created takes you through the process step by step.
From there, compilers built a "tissue price", the trade term for an initial, margin-free price, and then added the margin, known as the over-round, which gives the bookmaker its long-term edge whatever the result.
Just as important was the human side that a spreadsheet struggles to capture: team morale, dressing-room tension, motivation. Traders knew England games would draw strong patriotic money whoever the opponent was, and could tell when a team's reputation was running ahead of its form. That judgement still matters. As Sky Bet's football desk manager Paul Lowery puts it, being a bookmaker is about offering the right bet at the right price to the right people.
The over-round is the total of the implied probabilities in a market, minus 100%. Take a two-way market priced at 1.90 and 1.90. Each price implies a probability of about 52.6% (1 divided by 1.90), so the total is about 105.3%. That extra 5.3% is the bookmaker's built-in margin. The exact figure varies by market and competition, but the principle never changes: the implied probabilities always add up to more than 100%.
SVP Trading at Kambi
“The market-leading betting product of the future will be so complex that it must be automated to produce it in a scalable and profitable way.”
— Simon Noy, SVP Trading at Kambi
The move from human to algorithmic bookmaking was gradual, pushed along by three things: an explosion in the number of betting markets, the rise of in-play betting and the need to manage risk across thousands of simultaneous events.
The maths underneath hasn't really changed. Modern models still use Poisson distributions, still work out supremacy and still add margin. What's different is scale and speed. A human trader might spend hours pricing one match comprehensively; an algorithm turns out prices across many markets in milliseconds, updates them within moments of an event, runs thousands of markets at once and never gets tired or lets emotion creep in. That consistency also cuts down calculation errors.
That power has let operators widen their menus dramatically. Player props and bet builders have grown strongly in recent years, and combination bets like those need automated pricing that can handle the links between outcomes. A major match now carries dozens of markets with many outcomes each, and most are priced without a human touching them.
The toolbox is varied. Machine-learning models, including neural networks and random forests, learn patterns from historical data. Markov chain models treat a match as a sequence of states, with each pass, foul or shot shifting the odds. Metrics such as expected goals (xG), expected threat (xT) and pressing intensity (PPDA) give pricing teams a sharper view of how a game is flowing.
On a modern trading floor such as Sky Bet's, automated alerts flag risky bets and live profit-and-loss views track the day's activity. Routine pricing is largely automatic, and people concentrate on oversight and exceptions.
The modern bookmaker does not run on humans or on algorithms. It runs on both, and the smart ones know which to trust for what.
| Task | Algorithms are strongest at | Traders are strongest at |
|---|---|---|
| Pre-match pricing, major leagues | Generating and updating prices across many markets | Checking prices against news and team information |
| Pre-match pricing, lower leagues | Little, where data is thin | Using experience and judgement on limited data |
| In-play betting | Updating prices within moments of events | Handling unusual events and exceptions |
| Risk management | Monitoring every open bet and exposure figure | Deciding how to respond to flagged risk |
| Fraud and syndicate detection | Spotting patterns across accounts and markets | Investigating intent and deciding on action |
| Market creation | Producing prices for new bet types quickly | Choosing which markets to offer and testing them |
Machines win on volume and neutrality: they crunch huge datasets, ignore a team's reputation, reprice live odds at speed, watch for suspicious patterns and run thousands of markets without a bigger team. People win wherever context matters: motivation and morale, news that has not reached the data yet, telling genuine market moves from coordinated betting, weather and venue conditions, and thin leagues with little history.
The lines are blurring, too. Trading desks are now a mix of traders and compilers, two roles that used to be separate and increasingly overlap, alongside mathematicians trained in probability and pricing.
The easiest way to see the split is to watch it work. Here are six situations and who comes out on top.
Take a standard Premier League game between two well-known sides in normal conditions. The model chews through historical data, builds a Poisson distribution across scorelines, adds the margin and spits out prices in moments. A human could take hours to reach the same depth, and might still miss a pattern the model finds.
With a new boss, the tactics are a mystery and recent history can mislead, so a model may overweight poor form. A trader understands the psychology and can anticipate changes from the manager's past work. They also know the new manager bounce is a short-term lift that is far from guaranteed. Context beats data here.
When the ball hits the net, every related price has to move at once. The system rebuilds them in milliseconds, factoring in time decay and the new game state. Do it by hand and you leave a window sharp bettors would happily climb through. This is pure arithmetic, so speed is everything.
Unusual volume on an obscure market from fresh accounts sets off a joint response. The algorithm flags the pattern, links the accounts and totals the exposure. Traders then work out whether it is syndicate activity or genuine interest and decide what to do.
In lower divisions with little history, models can produce shaky numbers. Traders lean on experience of similar leagues, judge team quality and motivation, and may widen the margin to cover the uncertainty. If you like hunting for value in the less glamorous leagues, this is where your chance lies.
Big finals bring huge volume, heavy media attention and odd betting patterns. Algorithms cope with the load and flag anything strange, while traders add the colour: the emotional stakes, the narratives that drive public money and the patriotic punting that arrives with international football.
Nowhere is the shift clearer than in risk. Old-school risk management was reactive: it relied on the attention of the people watching, concentrated on the big markets and spotted trouble slowly. Liability spreadsheets, hard-won experience, a feel for customers' betting histories and basic position limits did the job, until volumes outgrew them.
Today's systems watch every market around the clock. They correlate related bets to spot coordinated activity, calculate exposure in real time and flag anything that strays from normal behaviour: a sudden burst of bets on an obscure market, a cluster of linked accounts, or a run of unusually large stakes on outcomes that look safe. Velocity checks reveal automated betting, dynamic limits shift with each customer's risk profile, short bet delays blunt latency tricks and markets are suspended the instant uncertainty creeps in.
The cleverest touch is the granular suspension. Instead of shutting a whole market, the system pauses only the affected part, which protects the bookmaker without frustrating everyone else.
The best operators run both layers together. The algorithms patrol every market and deal with clear threats, while people investigate the odd ones, make the strategic risk calls and rewrite the rules as experience teaches them something new.
In-play betting is where automation stopped being a nice-to-have. In the early 2000s, traders nudged live prices by hand, very few markets were available and the lag between something happening on the pitch and the price reflecting it could be long. Today in-play is a huge slice of turnover on major games, and almost all of it is priced by machine.
It all starts with real-time data: goals and events, player tracking, ball position, shot location and quality, possession and territory. Predictive models then treat a match as a chain of states and estimate how likely each move is, say a corner, then a header, then a goal. They apply time decay as the clock runs down and try to sniff out momentum swings before the scoreboard changes. Expected goals (xG) captures shot quality, expected threat (xT) shows how dangerous a position is, and pressing intensity hints at fatigue or aggression.
The newest frontier is micro-betting: narrow events such as the next corner, the next free kick, who scores next, or the next point in a tennis game. Each one settles in moments, so the price has to be rebuilt almost as soon as the event ends. That needs sub-second data, instant updates and automated settlement, which is why these markets only exist because of the machines.
Live betting is far from solved, though. Operators still wrestle with four problems:
Expect evolution rather than revolution. Machine learning will keep creeping further into pricing, offers will get more personal and new markets will be created automatically as soon as the data allows.
Personalisation is the big one: limits and offers tailored to the individual customer, which would be unthinkable without algorithms. Richer models are coming too, using tracking and video data to read tactics, and perhaps fatigue and injury signals. Risk tools will get sharper at spotting exposure building up, syndicate behaviour and the early signs of problem gambling. Talk of artificial general intelligence running the trading floor, or quantum computers rewriting pricing, is still firmly in the realm of speculation.
The trader's role will keep shifting from operator to overseer. The work will centre on building and checking models, handling exceptions, designing new markets and looking after customers. The valuable skills are data and machine-learning literacy, a feel for where a model breaks, commercial judgement and a good grip on regulation. There will be fewer traditional pricing jobs, but plenty of demand for people who can translate between the technical teams and the business.
For all the processing power, some parts of bookmaking stubbornly resist automation, and they are the reason trading teams are still hiring.
Psychology is the big one. A relegation-threatened side's final home game, a local derby, a manager's farewell: the emotional temperature of a match rarely shows up in a dataset, but a seasoned trader feels it straight away.
Curveballs are another. A scandal in the squad, a sudden storm or a tactical surprise leaves historical data with little to say. Models catch up once the change appears in the numbers, whereas people can anticipate it from context.
Syndicates need a human eye. An algorithm can wave a flag at an odd pattern, but deciding whether it is genuine interest or coordinated manipulation means judging intent, and sharp bettors work hard to disguise theirs.
Thin data is the fourth. Lower divisions, youth competitions and emerging sports often lack the history for a reliable model, so experienced traders price them with feel and comparison.
Finally, customers remain a human call. Spotting a sharp, looking after a VIP and deciding whether to restrict an account all involve judgement that profit calculations alone cannot make.
The Unit
“AI isn't replacing the human touch in betting and gaming. It's refining it. It's helping us build fairer, safer, faster, and more personalized experiences.”
— David McKenny, The Unit
Bookmaking has gone from pencil and paper to systems that crunch millions of data points in milliseconds, and yet the human trader is still in the building. What has changed is the job: less typing in prices, more steering the machine, spotting what it has missed and deciding what to do when the unexpected turns up.
For bettors that means more markets, quicker price moves and ever more elaborate products. It also means the big, heavily traded markets are priced very sharply indeed. The softer spots are where the models have the least to go on, which is exactly where patient, well-researched punters can still find an edge.
Our angles for bettors:
18+. Please gamble responsibly. For free, confidential support visit BeGambleAware.org.
