How Kuponcu selects its tips: the method, explained

In short

Each Kuponcu tip is a single bet on a single match: the match result (home win, draw or away win), over or under 2.5 goals, or both teams to score. Every other day at 08:00 UTC, the system checks prices for matches taking place over the next three days across eleven European leagues and the three UEFA club competitions, using data from about twenty bookmakers.

It estimates the fair value of each outcome, keeps the selections where the available price is high enough, then asks two AI analysts to make the case both for and against each one. The tips that pass those checks are published, usually 15 to 30 per run. Every tip stays on the record and is tracked through to the final whistle, whether it wins or loses.

The data

An odds feed supplies pre-match prices from around twenty bookmakers and betting exchanges, including some of the sharpest markets. The system reads those prices at the start of each run and ignores any price that is more than an hour old.

Match results come from football-data.co.uk. They are imported every morning at 06:00 UTC, two hours before the run, and are used to fit the goal model. Final scores used to settle tips come from football-data.org and the odds feed, with both checked every hour.

For team news, the system checks recent headlines about each club: up to six per team, published within the previous three days. Headlines offer clues, not confirmed line-ups. If an important injury has not been reported during that window, the system will not know about it.

The fair price

Most bookmakers take their prices from a small group of "sharp" firms. These accept large bets from professional bettors and adjust their odds quickly when the market shows they are wrong. Research over many years, along with our own backtests, points to the sharp closing price as the most accurate public forecast of a football match. Kuponcu therefore uses the sharpest bookmaker in its feed to set the fair price. If that bookmaker has no price, it uses the median from at least six others.

Bookmaker prices include a margin, which needs to be removed before we can estimate a fair price. Take a match priced at 2.10 for the home win, 3.40 for the draw and 3.60 for the away win. Dividing one by each price gives implied probabilities of 47.6%, 29.4% and 27.8%, adding up to 104.8%. That leaves a margin of 4.8%. Kuponcu removes it using the "power" method. This takes slightly more probability from long shots than from favourites, reflecting where bookmakers tend to build in more of their margin. In this example, the adjusted probabilities are 46.0% home, 27.8% draw and 26.2% away, giving fair odds of 2.17, 3.60 and 3.82.

The goal model

For every league match, Kuponcu also runs its own goal model: Dixon-Coles, a long-established statistical model of football scores. It estimates each team’s attacking and defensive strength, as well as home advantage, using the previous two years of league results. Recent matches carry more weight: a game from a year ago counts half as much as one played last week.

Those ratings produce an expected-goals figure for each side, say 1.6 for the home team and 1.1 for the visitors. The model then estimates the probability of every score from 0–0 to 10–10 and groups those scores by market. In this example, it gives the home win a 49% chance, the draw 25%, the away win 26%, 51% for over 2.5 goals, and 53% for both teams to score.

The goal model does not set the fair price. We tested adding its estimates to the fair price over two seasons, but the extra bets lost money. A model based only on goals has less information than a market that also reflects injuries, line-ups and betting activity. We use it as a second opinion instead: it helps rank selections, contributes to each tip’s strength score, and gives the AI reviewers another view to compare with the market. There is no goal model for UEFA competitions because teams from different domestic leagues rarely play each other.

Value and the minimum odds

A bet’s edge is its expected return per unit staked, assuming the fair probability is correct. Calculate it by multiplying the fair probability by the odds, then subtracting one. For example, a 46% chance at odds of 2.30 gives an edge of 0.46 × 2.30 − 1 = 5.8%.

A tip is a value bet when its edge is between 2% and 15% (above 15% it is almost always a stale or incorrect price), its odds are between 1.40 and 3.00, and its chance of winning is at least 20%. Only one tip is allowed per match. Value bets carry the "Value" label on the site.

On many days, the system finds fewer than 25 value bets. When that happens, it fills the remaining places with the strongest other selections: their odds must be at least fair, but their edge is below 2%. A price on its own is not enough to make one of these picks worthwhile; the goal model and the review must also support it from a football point of view.

Odds can change as soon as a tip is published, so each one comes with a minimum price at which it is still worth considering. For a value bet, that is the price where its edge falls to 2%. For a top-up pick, it is the fair price. If the odds available to you are lower, leave the tip alone: the value is no longer there.

The AI review

Each selection is first reviewed by an AI analyst, currently Anthropic’s Claude Opus. The analyst receives the figures — prices, fair probability, edge, the goal model’s assessment and each team’s recent matches — along with the team-news headlines. It then sets out the arguments for and against the bet, identifies the main factors and risks, and gives it a score from 0 to 100. It can reject a selection outright or veto it when a headline raises a serious concern, such as a key striker being ruled out or a heavily rotated squad.

Any selection that passes the first review goes to a second AI, called the "opposing bettor". Its task is to look for every credible reason the bet might be wrong: why the market may have it right and our estimate may be off. Both arguments appear on the site, so you can weigh them up for yourself.

The AI is instructed to work only with the information supplied to it, rather than relying on its own memory, which may be months out of date. It cannot watch matches, know a line-up before it is announced or see news that has not yet been published. It is a reviewer, not an oracle. Research on AI forecasting (covered in our guide) shows that language models on their own do not beat the betting market. If the AI is unavailable, tips are published using prices alone; the site says so, and those tips are never labelled "strong".

Tip strength

Every tip receives a final score from 0 to 100, calculated as a weighted average of five parts: price (35%, based on how far it beats fair, with a 10% edge earning full marks), the goal model (15%), the AI analyst’s score (20%), team news (10%), and the weakness of the opposing case (20%).

A score of 70 or more makes a tip strong, provided the analyst’s uncertainty is not high. Scores from 60 to 69 are good tips; anything below 60 is a small tip. Value bets scoring under 55 are removed. Strength measures how convincing the overall case is, not the chance of winning. Even a strong tip at odds of 2.80 will lose more often than it wins. Check the win probability shown with each tip for that.

Settling and tracking every tip

Every published tip stays in the record. Nothing is deleted or left out. We track each one as a one-unit stake at the odds shown when it was published: a win adds the odds minus one, while a loss costs one unit. Final scores are checked every hour. If a match is postponed or abandoned and still has no result after ten days, it is voided: the stake is returned and the tip is recorded as neither a win nor a loss.

Short-term results can be noisy, so we also measure closing-line value (CLV): how the odds on our tip compare with the fair price at the final check before kick-off. If a tip was taken at 2.30 and the fair price closed at 2.10, it beat the closing price by about 9.5%. A method that regularly beats the closing price is much more likely to be profitable over time, even with losing runs along the way. A method that fails to beat it is more likely to lose over time, even if it enjoys a few lucky streaks. The Results page shows every tip, its outcome and its closing-line value.

Premium betslips

Premium betslips are made from the tips produced in each run. Each slip combines two to four selections from different matches, all available with the same bookmaker; a single slip cannot be split across sites. Each leg must be priced close to fair — at least 97% of the fair price — and the combined slip must be at or above fair, with total odds between 2 and 15.

Premium slips are grouped into three levels according to their chance of winning: safe (30% or more), balanced (15–30%) and long shot (under 15%). Each run can include up to three slips at each level. Every slip shows the lowest total odds worth taking and is tracked and settled in the same way as an individual tip.

Limits and honest numbers

The method can only work with the information in its data. It may miss late injuries, confirmed line-ups, weather changes or problems inside a dressing room. Its edge depends on finding a better price than the sharp market, which is only possible when other bookmakers are slower to adjust. It also cannot prevent bookmakers from limiting accounts that win.

The record is still new. Kuponcu began publishing on 13 October 2026. By 4 November 2026, 176 tips had been settled: 88 won and 88 lost, for +6.3 units (+3.6% per tip) at average odds of 2.27. The fair prices implied a 49% win rate; the actual figure was 50%. In addition, 85% of the tips beat the closing price, by +1.7% on average. That is an encouraging start, but 176 bets are nowhere near enough to prove that the method works. At these odds, a run of fifty tips that loses money overall would not be surprising. Nor would a lucky run.

Not every area has done well so far. Picks with a win chance under 35% won only 4 of 29 (around 8 were expected), while the Championship, Bundesliga and Eredivisie are also behind. Before launch, backtesting led us to change two rules: the goal model no longer affects the fair price, and top-up picks must be priced at or above fair. Top-ups at lower odds lost 4.3% per bet. When the results show that a rule is not working, we change it and record the change here.

Who builds Kuponcu

Kuponcu’s entire research department is one independent software developer, a lot of coffee, and two AIs that are paid to disagree with each other. I wrote the system on this page, from the odds feed and the goal model to the AI review and the tracking, and I check its results every week. No lab coats, but every tip is on the record.

Kuponcu takes no bets, is not paid by any bookmaker and does not link to any. It earns from advertising and Premium subscriptions, never from your losses. Questions, corrections and ideas are welcome at the contact address on the About page.