How efficient is the betting market?

Can the betting market be beaten?

Few questions move as much money around sports betting as this single sentence. Tipster sites, "proven systems" and paid groups are all built on the same promise: that some knowledge exists which lets a bettor come out ahead of the bookmaker in the long run. We did not start from a promise but from a measurement: we built our own probabilistic models and then systematically tested whether there is any point where the market’s pricing is exploitably wrong.

The short answer, which this article backs with numbers: on our data, with our methods, nowhere. On the main football markets prices are so accurate that the errors left in them are smaller than the transaction cost of betting, that is, the bookmaker margin. This null result sounds disappointing at first; in reality it is the most useful thing a data-analytics platform can say, because it marks out precisely what public data can do and what it cannot.

The article walks through the question, the methodology and the results of our seven experiments, then shows a live league benchmark table computed from the platform’s historical corpus: the average margin per league, how accurately the market prices matches, and how often the favourite wins. The closing section lists the limitations, because a measurement is worth exactly as much as its honesty about its own boundaries.

What does an efficient market mean?

The betting version of the efficient-market idea known from financial markets goes like this: the odds incorporate all publicly available information, so no systematically better probability estimate can be produced from public data. This does not claim that the price is always correct. It claims that its errors cannot be found in advance, reliably, with enough confidence to still leave a profit after the margin is paid.

The key actor in this idea is the closing line, the last odds in force at kickoff. By then the market has absorbed everything from injury news to the money of professional syndicates, so the closing price can be treated as the best publicly available probability estimate. Someone who consistently bets at better prices than the close over the long run probably holds a genuine informational edge; where that is not the case, short-term profit and loss are the work of variance. The trade calls this ratio CLV (closing line value).

Efficiency is not magic but the product of feedback. A mispriced line is immediately bet by well-informed money, and the bookmaker responds by moving the price. By the end of the process the price sits where neither side is worth taking. The more liquid a market, the faster this correction, which is why the main markets of the big leagues are the most accurate, and why any offer promising easy edge exactly there deserves extra suspicion.

What does the published research say?

The question has decades of literature behind it, and the lessons are surprisingly consistent once you read the details. Dixon and Coles’ 1997 paper, by now a classic, still found exploitable mispricing in the English market of the time using a goals-based model. Hvattum and Arntzen’s 2010 ELO study, however, already showed that estimates derived from the odds were more accurate than every rating model they examined. The more recent results keep pointing in one direction: the market keeps getting sharper:

• 1Kaunitz, Zhong and Kreiner (2017): a strategy betting prices that deviated from the consensus showed +3.5% over 479 thousand matches across ten years, and +8.5% in a five-month real-money test. The experiment was stopped not by the market but by the bookmakers, who limited the winning accounts to minimal stakes.
• 2"Not feeling the buzz" (2023): a recomputation of a strategy previously published as profitable. After fixing the data errors the profit essentially vanished (most of it came from a single outlier bet), and none of the examined strategies produced returns after 2020. According to the authors, the market has grown measurably sharper over the past few years.
• 3Hubáček, Šourek and Železný (2019): a machine-learning model deliberately decorrelated from the bookmaker’s estimate produced positive returns on NBA games in a 2007 to 2014 backtest. Important caveats: it is backtested on the prices of one bookmaker only, ignores account limiting, and was never validated on a live market.
• 4Walsh and Joshi (2024): when selecting models, calibration rather than hit-rate accuracy is the right criterion; the model chosen by calibration produced better returns in their experiment. This is partly why we too measure Brier score and calibration on every one of our models.

Our own measurement program

Behind the platform stands a closed historical data corpus that is not extended after the fact: more than 50,000 matches across 11 seasons. The first, smaller-sample phase of the work was published in a peer-reviewed journal article (Pál and Bíró, 2025); there, over 539 matches, the best model-strategy pairing still looked profitable. Our walk-forward evaluation on the full corpus did not confirm that as a durable edge: our models are calibrated, but none of them produced a more accurate estimate than the closing price.

We then designed the profit-hunting program around seven structurally different methods, and ran every one of them with the same discipline: walk-forward evaluation, in which the model may only see data known before the match at every decision; nearly 49 thousand matches with opening and closing odds; decision rules fixed in advance, with no after-the-fact selection. Walk-forward is non-negotiable because the most common failure of backtests is looking into the future: a single crumb of information leaking back from the future is enough for a beautiful paper return to evaporate in reality.

It is also important to state what we did not do. We did not cherry-pick among the experiments afterwards, and we do not hide the failures: the list below reports the results in the form the measurement produced them. For a research program, a refutation is an outcome worth just as much as a confirmation, even if it looks worse in an advertisement.

Seven methods, seven results

The seven avenues examined and the measured outcome of each:

• 1Recreational bookmakers’ closing prices against the fair closing price of a sharp book: the excess return computed on above-threshold deviations has a confidence interval that includes zero. No exploitable gap can be demonstrated.
• 2Beating the close at the opening price ("beat the close"): +4.7% return, but only in oracle mode, that is, selecting with knowledge of the future closing price. Built from information available in real time, the result is around zero or negative.
• 3Predicting the closing price with machine learning: the model reproduced the closing price almost exactly, yet betting on its estimates returned -11%. The cause is the optimizer’s curse: value selection harvests exactly those cases where the model overestimates.
• 4Lower divisions: the tempting +30% on paper came from just 132 matches, with a confidence interval running from -7% to +67%. That is noise, not signal, and in an illiquid market there is no sharp reference price against which value could even be measured.
• 5Lay-off, that is, hedging a position taken at the opening price on an exchange before kickoff: the price movement is real, but the net result is +0.6%, which commission, stake limits and the lack of liquidity consume.
• 6Cross-market pricing, the 1X2 recomputed from the goals market: the two markets are consistent with each other, and the goals market carries no extra information over the 1X2.
• 7BTTS and related side markets: adequate historical data is missing, so no meaningful test can be run on them, and their prices derive from the same goal distribution on which the earlier avenues already failed.

The program’s conclusion: pricing on the main markets is efficient within the band of transaction costs. None of the tested methods earned back the margin, and it is more honest to say so than to sell a system that merely looks like it works.

League benchmark from the historical corpus

Each row is one league. The average margin is the overround built into the consensus prices (computed from the average of real bookmakers, excluding panel aggregates). The Brier score is the squared error of the margin-free market probabilities against the actual result: a lower value means more accurate pricing. The favourite hit rate shows how often the outcome priced as most likely by the market actually won.

Implied and observed frequency

Outcomes are grouped into bands by their margin-free market probability. In a well-calibrated market the observed frequency stays close to the implied probability band by band; a large, systematic deviation would indicate exploitable mispricing.

The aggregates above are computed from the platform’s historical corpus: 60,210 completed, priced matches between 2015-08-08 and 2026-09-20. Only leagues with at least 100 matches appear in the table; the page publishes no match-level data and no bookmaker names.

What does this mean from a bettor’s perspective?

The first consequence is a change of perspective: betting has a price, and that price is the margin. The table above shows per league what it costs to be in the market; the pattern is usually the same, with the big liquid leagues cheaper and more accurately priced, and smaller competitions more expensive. As orders of magnitude: sharp bookmakers work with around 2-3% on the main markets, recreational bookmakers typically with 5-8%, and on side markets and small leagues higher values are not rare.

The second consequence is the primacy of measurement. If the market is this accurate, short-term profit proves nothing on its own: a sample of a few dozen bets is mathematically too small to separate luck from skill. What can actually be measured is the relation of the price taken to the closing price, and the calibration of one’s own estimates. That is exactly why the platform’s tools are built on measurement, not on tips: a betting journal, calibration curves, Brier scores.

The third consequence is the economics of suspicion. If a research program dedicated to the task could not demonstrate a durable edge from public data, then any offer promising the same for a small fee and with a guarantee carries the burden of proof. The way to verify a tipster record is the same as in science: tips announced and fixed in advance, a complete rather than curated history, and a comparison against the closing price.

Limitations

No single measurement is universally valid, and this study is no exception. The results should be read within the following limits:

• 1Coverage: the table only shows leagues with at least the threshold number of completed, priced matches in the corpus; the threshold, the covered period and the total sample size are stated in the line below the table.
• 2Market: the measurement concerns the main 1X2 market, at consensus prices that are averages of real bookmaker prices. The consensus is not the same as the best price actually available at a given moment.
• 3The seven-method program ran on our own closed corpus, on the opening and closing odds of nearly 49 thousand matches. On other markets, other periods or other data the result could in principle differ.
• 4Measuring the past is not a forecast: a historical aggregate says what was, not what will be. The structure of the market changes, according to the 2023 correction study typically in the direction of greater efficiency.
• 5The null result is no proof of impossibility: it shows that on the avenues we tested, the margin could not be earned back from public data. It claims nothing about actors working with non-public information or about the markets of other sports.

Conclusion

For us, market efficiency is not an article of faith but a measurement result, one we tried to overturn from seven different directions and failed. That failure is the foundation of the platform: since we found no edge, we sell no edge. What we offer is an understanding of how the market works and an honest measurement of one’s own results, which are the two things our measurement shows public data is actually good for.

This article is educational and research content, not betting advice. Gambling involves risk, and no tool of the platform increases the chance of winning.

Sources

The studies referenced in the article. The open-access items are directly available at the links: