The short answer: occasionally, on paper, under narrow conditions — and the people who manage it in practice tend to find their accounts restricted. The longer answer is more interesting, because the research shows why beating betting odds is so hard and why a profitable backtest proves less than it seems. Here are three studies that are often cited, and what each one actually found.
Why the odds are a tough opponent
Betting odds are themselves a forecast. They combine the bookmaker's models with information from the money that flows through the market, and they include a margin. To make money, a model has to be more accurate than that forecast where it matters, by more than the margin, consistently. Being a good model is not enough; it has to be better than one of the best forecasters in sport.
Study 1: the bookmakers' own numbers
Lisandro Kaunitz, Shenjun Zhong and Javier Kreiner (2017) did not build a football model at all. They used the odds published across the market to estimate each outcome's probability, then bet when an individual bookmaker offered odds that looked mispriced against that consensus. They tested the idea in a 10-year historical simulation with closing odds, a 6-month simulation with minute-by-minute odds, and five months of betting with real money. All three were profitable.
The second half of their title explains what happened next: "how the online sports betting market is rigged". The authors report that bookmakers responded to their success with discriminatory practices against successful clients — in plain terms, restricting their accounts.
Key point: The inefficiency was real, but it was found between bookmakers, not by out-predicting football. And a strategy that works can stop being available to the person using it.
Study 2: be different from the market
Ondřej Hubáček, Gustav Šourek and Filip Železný (2019) built a machine-learning system for basketball, including a convolutional neural network fed with a large number of player statistics. Their key idea was not just to maximise accuracy but to reduce the model's correlation with the bookmaker's predictions, so that it disagrees with the odds where the odds are wrong. Combined with a betting strategy borrowed from portfolio theory, it produced positive cumulative profits in experiments on NBA seasons 2007–2014.
Two points matter when reading this result. It comes from experiments on historical data, not from years of live betting. And it shows that a model that simply copies what the market already knows cannot profit, however accurate it is.
Study 3: profit is not proof of accuracy
Fabian Wunderlich and Daniel Memmert (2020) asked a question that should come before any claim about "beating the bookies": are betting returns even a good way to judge a forecast? Using theory, simulations and real data from three sports, they showed that positive returns can appear without a more accurate model, systematically or just by chance. Their conclusion: betting returns measure profitability, but they are not a valid measure of model accuracy.
What this means for "AI tips" you see online
When a service claims its AI beats the bookmakers, these studies suggest four questions:
| Question | Why it matters |
|---|---|
| Was it tested on matches the model had not seen? | Any model can fit the past |
| Was it real money, at real stakes, over a long period? | Backtests ignore limits and changing odds |
| How large was the sample? | Profit over a few hundred bets is mostly noise |
| Is accuracy reported with proper scoring rules? | Profit alone can be luck, as Study 3 shows |
Watch out: A service that sells winning predictions has an obvious question to answer: if the edge were real and repeatable, why sell it? Treat any guaranteed-profit claim as marketing, not evidence.
The honest summary
Models can find occasional mispricings, and research shows how. But the margin, the strength of the market, account restrictions and the noise in short records make consistent profit very rare, and much harder to prove than most marketing suggests.
To see how forecasts are judged without looking at profit at all, play with the Elo rating calculator, and read How AI Predicts Football Matches.
Further reading
- Beating the bookies with their own numbers — and how the online sports betting market is rigged — Kaunitz, Zhong and Kreiner, 2017 (arXiv)
- Exploiting sports-betting market using machine learning — Hubáček, Šourek and Železný, International Journal of Forecasting, 2019
- Are betting returns a useful measure of accuracy in (sports) forecasting? — Wunderlich and Memmert, International Journal of Forecasting, 2020
The AI Sports team writes about data, models and technology in sport.