Methodology.

01What we calculate

For every match in the competitions we cover, BetPlanners calculates the probability of different outcomes: who wins, how many goals are scored, whether both teams score, corners, cards and other markets you will see on the match page. Next to each probability we show the fair odds derived from it, which is simply its inverse, with no margin. Where market odds are available, we compare them with ours.

When we publish a selection of the day, nobody picks it by hand. It is a rule: of every forecast in the window of matches we cover, those that do not clear our probability, odds and value thresholds are discarded, and of those left we publish the one with the highest probability. The weekly balance measures that selection at a flat stake of one unit per pick: a unit is one bet, not an amount of money, and the balance is published whether it wins or loses.

02Where the data comes from

We work with two kinds of data, and it is worth telling them apart:

• Sports data: fixtures, results, line-ups and match statistics, collected from public third-party sources. We normalise them into our own corpus, which today holds more than 318,000 matches with a final score.
• Odds: we read them periodically from a comparison site that aggregates Spanish-licensed bookmakers. They are not our odds and we have no agreement with any bookmaker.

We do not independently verify each sports datum against a second source. If a result or a statistic arrives wrong from origin, our calculations will inherit it until the source corrects it.

03How the model is trained

The model learns from matches already played: from each team history, the competition context and accumulated statistics, it estimates how often similar outcomes have occurred in similar situations.

It is trained and validated with a time separation: to evaluate how well it does on a match, it may only use information from before that match. It is never trained on data later than what it predicts. That is the only way for measured quality to resemble the quality obtained in production.

The model changes over time: it is retrained and adjusted. The probability we give a match today may not be the one we would have given a month ago, and that is expected.

04Coherence across markets

The probabilities for one match are produced by several specialised models, which can lead them to contradict one another. Before anything is shown, they all pass through a single reconciliation point that corrects those contradictions, so that what you see in one part of the page is not denied by another. It is also the point that the probabilities pass through when you record a prediction, so that it settles at the same figure you saw.

05How several selections are combined

The parlay calculator multiplies the probabilities of the selections you tick and returns the joint probability and its fair odds. Multiplying is only correct if the selections are independent of each other, so that assumption is not taken on trust: it was measured.

Across different matches, independence holds. Measured over 264,486 matches spread across 7,661 days, comparing the real joint probability against the product of the individual ones, the deviations stay below 1%. And the half that makes the result trustworthy is the control: repeating the measurement with the days shuffled at random, where a day effect cannot exist by construction, produces deviations of the same size. Which means even those tiny deviations are not a phenomenon: they are the noise of the method. Multiplying across different matches is legitimate, including matches from the same round of the same competition.

Within a single match, it is not. There the selections condition each other and multiplying yields a number that corresponds to no real bet: combining "under 2.5 goals" with "both teams to score" gives a product 54.4% above the true joint probability, which would sell that combination at nearly twice what it is worth. That is why the calculator rejects two selections from the same match instead of computing them with a warning.

Each selection also carries the measured calibration deviation for its market, where we have one. Derived markets (a double chance is not a model of its own, it is the sum of two parts of the 1X2) also carry the sum of their two errors, which puts them around one point above directly trained markets. A market for which we lack that measurement is declared unmeasured and the calculator says so: it does not inherit the neighbouring market's figure.

The fair odds of the combination are the inverse of the joint probability, with no margin added. It is not a recommendation to bet it: it is the price at which that combination would be neutral according to our model, so you can compare it against what the market offers you.

06The odds

Odds are read periodically, not continuously. Every card and every comparison carries the time of the reading it is based on, and that stamp is part of the data: an odd without a time is a claim with no expiry. Between two readings bookmakers can move their prices, and they do as kick-off approaches.

When we show the best odds for a market, it is the best among the bookmakers we name, not necessarily the best that exists.

We call value the gap between what a bookmaker pays and our fair odds, in percentage points. A 6% value means that price pays 6% above what our calculation considers fair. It is not expected profitability nor a forecast of gain: if our probability falls short, the value we show does not exist.

07Which markets we cover

Each match page shows the markets available for that fixture. Not every match has every market: it depends on the competition, on how far back the data for those teams goes and on there being enough history for the calculation to mean something. When there is not, the market does not appear, rather than appearing with a poorly founded figure.

08What this does not do

Stated expressly, because it defines the object of the service:

• We do not predict results. We give probabilities, and a probability is not a promise: an outcome we put at 60% not happening is not a fault of the service.
• We do not give advice tailored to you. When we publish our own selection, it is a general opinion based on these calculations, the same for everyone and without knowing your circumstances.
• We do not guarantee the accuracy of third-party sports data, nor do we verify each datum against a second source.
• We do not guarantee continuous availability or any particular update frequency.
• Past results do not guarantee future results.

09Changes to this page

This page describes how the service works at a given moment, and the service evolves. When we change something substantial of what is described here, we will update the date above. If the change affects what you contracted as a paying subscriber, the amendments section of the Terms of Use applies.