Can AI predict football? What research and our results actually show

The short answer

No method, including AI, can reliably pick winners. Three decades of research point to the betting market — especially the closing prices of the sharpest bookmakers — as the best public forecast of a match. AI models rarely beat that market on their own. A careful system can still look for prices that are better than the market’s fair price, then use AI to check whether there are good reasons not to take those selections. That is the approach behind Kuponcu, and this page explains the evidence for it.

The market is a strong forecaster

In 2009, Martin Spann and Bernd Skiera compared three ways to forecast German Bundesliga matches. Betting odds predicted the correct result in 53.7% of matches, almost matching a prediction market. Newspaper tipsters managed 42.6%, below even a simple rule (50.9%). Many studies since have reached a similar conclusion: published odds are difficult to beat as forecasts.

The market is not infallible. In 1997, Mark Dixon and Stuart Coles developed a statistical model of football scores and showed that it could identify bets with positive returns against bookmakers’ prices at the time. Their model remains a foundation for many football models, including Kuponcu’s. But betting markets have become much sharper since the 1990s.

Machine learning against the bookmaker

Later research showed that how a model is used matters more than how often it predicts the winner. Ondřej Hubáček and colleagues (2019) found that a model makes money only when it differs from the bookmaker’s view in the right places. A model that closely copies the odds may be accurate, but it offers no useful edge for betting.

Conor Walsh and Alok Joshi (2024) tested this idea using basketball bets. Models selected for well-calibrated probabilities — where a 60% forecast wins about 60% of the time — returned +34.7% on average. Models selected for accuracy, or picking the most winners, returned −35.2%. In betting, reliable probabilities matter more than a high strike rate.

Lisandro Kaunitz, Shenjun Zhong and Javier Kreiner (2017) took another approach. Rather than predict matches, they treated the average odds across bookmakers as the fair price and bet only when one bookmaker offered clearly more. The strategy made a profit across ten years of historical data and with real money, until bookmakers restricted their accounts. Kuponcu uses a similar idea, but takes the sharpest bookmakers rather than the average as its reference.

What about ChatGPT and Claude?

Large language models have become much better at forecasting real-world events than they were a few years ago. In 2024, Danny Halawi and colleagues at UC Berkeley built a system that searched news and reasoned through questions. Its forecasts came close to those of the human forecaster communities on competitive prediction sites — a demanding standard to meet.

Football is a tougher test because the betting market is already a strong forecaster. The 2026 World Cup provided a clean test using matches that no model could have seen during training. One study used four leading models — Claude, ChatGPT, Gemini and Grok — as research agents across all 104 matches. They chose the same favourite in 92% of matches, none forecast better than the betting market, and backing the market favourite alone earned more than any of the four models. Another benchmark, WorldCupArena, found that the best AI systems were only slightly better than the betting market at predicting results and exact scores.

How Kuponcu uses AI, and why

The research and our own tests point in the same direction. In every league we cover, our goal model is less accurate than the sharp market. When we added its estimates to the fair price, the extra bets lost money. Kuponcu therefore does not ask an AI or a model to beat the market. The market sets the fair price; the system looks for bookmakers offering more than that. The AI then does a different job: it reviews form, the model’s view and team news, sets out arguments on both sides, and filters out selections with a clear reason to avoid them.

A second AI makes the case against every selection. Both systems are told to use only the supplied information, not their own memory, and both sides of the argument are shown with each tip.

Our own record so far

A method should be judged on every selection it publishes, not just its best results. By 4 November 2026, three weeks after launch, 176 Kuponcu tips had been settled:

• Overall: 88 won and 88 lost; +6.3 units at average odds of 2.27 (+3.6% per tip). Fair prices implied that 49% would win; 50% did.
• High win probability (55% or more): 60 tips, expected win rate 61%, actual win rate 60%.
• Medium win probability (35–55%): 87 tips, expected win rate 46%, actual win rate 55% — luckier than expected.
• Low win probability (under 35%): 29 tips, expected win rate 29%, actual win rate 14% — unluckier than expected.
• Closing-line value: 85% of tips beat the closing price; the average was +1.7%.
• The sample is small. Results over a few weeks can swing either way through chance; closing-line value is a more useful early signal. The latest figures are on the Results page.

How to judge any AI tipster

• Are all selections published and kept on record, including losses, with the odds shown when each one was posted?
• Is performance reported in units at flat stakes over hundreds of bets, rather than built around a few big wins?
• Do the stated probabilities match how often the selections actually win (calibration)?
• Does the method beat the closing price? That is one of the hardest figures to fake.
• Does it promise sure wins, fixed matches or "guaranteed" profit? Walk away. No honest method can make those promises.

Sources

The research discussed on this page:

• Spann & Skiera (2009), Sports forecasting: a comparison of the forecast accuracy of prediction markets, betting odds and tipsters, Journal of Forecasting 28(1)
• Dixon & Coles (1997), Modelling association football scores and inefficiencies in the football betting market, Journal of the Royal Statistical Society C 46(2)
• Hubáček, Šourek & Železný (2019), Exploiting sports-betting market using machine learning, International Journal of Forecasting 35(2)
• Walsh & Joshi (2024), Machine learning for sports betting: should model selection be based on accuracy or calibration?, Machine Learning with Applications 16
• Kaunitz, Zhong & Kreiner (2017), Beating the bookies with their own numbers, and how the online sports betting market is rigged
• Halawi, Zhang, Chen & Steinhardt (2024), Approaching human-level forecasting with language models
• FIFA World Cup 2026 as a contamination-free benchmark for LLM forecasting agents: four models, a bookmaker and 104 matches (2026)
• WorldCupArena: fine-grained evaluation of language models and deep-research agents on football forecasting (2026)