Two years ago, Foresportia published its first football match predictions.
The project began with a fairly simple question: can data, statistics and artificial intelligence produce useful match probabilities while remaining transparent enough for us to assess the results afterwards?
Two years later, Foresportia has changed a great deal.
The models have evolved, new competitions have arrived, the analysis has grown richer, and new tools have gradually taken shape around the prediction engine.
For this second anniversary, we wanted to look back at how far we have come. And, above all, at a few numbers.
Nearly 19,000 matches with archived probabilities
The first number probably gives the clearest sense of the project’s scale.
Foresportia’s data now includes 18,812 matches for which 1X2 probabilities have been archived.
That spans several football seasons, 59 competitions and around 1,700 different teams in the data.
And 1X2 is only part of what we calculate.
Over time, Foresportia has added other markets and scenarios: double chance, Draw No Bet, BTTS, total goals, likely scores and various selections derived from the probabilities produced by the models.
The archive now contains more than 103,000 recorded picks across multiple markets.
These records do more than preserve the past. They also let us compare what the models forecast with what actually happened on the pitch.
Keeping the successes and the mistakes
One idea has guided Foresportia from the start: a prediction is only truly useful if it can be checked afterwards.
In sports predictions, it is easy to show only what worked.
A screenshot of a successful pick tells us almost nothing about a model’s quality. What matters is how it behaves over hundreds, then thousands, of matches.
That is why Foresportia’s archives also retain incorrect predictions.
A badly predicted match remains a badly predicted match.
This depth of history now lets us study performance by stated probability, league and match stability level.
In the 1X2 results archive, matches rated Stable currently exceed 80% correct predictions, while those rated Very stable are around 88%.
These categories deliberately cover only a portion of the matches analyzed. The clearer the statistical scenario has to be, the more selective the model becomes.
We have never aimed to present football as perfectly predictable. It is not.
The aim is to build probabilities whose reliability can be measured over time, and gradually understand where the model performs well and where it should remain cautious.
That approach to calibration and historical checks is now an integral part of how Foresportia works.
86,000 active users recorded by Google Analytics
Foresportia has also grown considerably in audience.
Since launch, Google Analytics has recorded around 86,000 active users, nearly 750,000 page views and more than 4.4 million events on the site.
These are figures recorded by Google Analytics, so they do not necessarily match the exact number of individual people who have used Foresportia. Some visitors may not be measured, while some automated traffic may occasionally appear in analytics tools.
Even so, they give a useful sense of how much the project has grown.
Today, several dozen people regularly use Foresportia at the same time; during especially busy periods, simultaneous visitors can reach several hundred.
That is a long way from the project’s first few weeks.
More than 350 accounts created in just over a month
Another figure is much more recent.
Foresportia accounts have only existed since August 2026.
In just over a month, more than 350 users have already created an account.
The account system should gradually make the experience more personal and support new services built around how people use the site.
It is still an early figure, but that is precisely what makes it interesting: unlike the traffic statistics covering two years, these first few hundred registrations came in only a few weeks.
From an initial automated model to a probabilistic engine
Today’s Foresportia looks quite different from the version launched in September 2024.
The first engine, now called P0, was primarily an initial automated foundation: gathering data, running an early statistical model and regularly publishing predictions.
Several successive generations followed.
P1 strengthened the probabilistic framework and calibration. P2 and then P3 gradually added more context, additional markets and more advanced ways to track the model’s reliability.
Among the added markets, likely scores show several possible outcomes; our exact score analysis explains how to read them without treating any one score as certain.
The current engine monitors calibration, data quality and statistical changes over time, among other things. Historical results also help identify situations in which predictions prove more or less reliable.
That is probably one of the biggest changes in two years: the system no longer simply calculates a prediction. It also observes its own results over time.
Regular reassessment, self-training and monitoring gradually help adjust parameters, detect changes in how leagues behave and avoid assuming that a model that performed well yesterday will automatically do so tomorrow.
And now, live matches
One of the most recent developments concerns probabilities during matches.
For supported matches, Foresportia can now recalculate the probabilities of a home win, draw or away win as the game unfolds.
The live model starts from the pre-match analysis, then incorporates factors including the score and elapsed time to reassess the different scenarios.
This feature is still in beta and will keep evolving in the coming months.
Notifications have also become part of the experience, making it easier to follow a match without keeping its page open all the time.
That would have been hard to imagine when Foresportia offered only a handful of predictions calculated before kick-off.
Foresportia has changed visually, too
Not every change over these two years has been about the models.
The site itself has been extensively redesigned.
The goal has gradually shifted from a mostly technical tool to an interface that makes a match easy to understand at a glance: probabilities, stability level, history, team form, possible scenarios and now live information.
The presentation of probability, stability and confidence has been reworked in particular to clearly separate what the model estimates from what it has actually demonstrated in the past.
This evolution also brought the Foresportia Android app, making predictions easier to access on mobile.
Over two years, Foresportia has grown from a prediction engine with a few web pages into a full set of tools built around match analysis.
What about year three?
There is still a great deal to improve.
And that is probably a good thing.
Improving the probability model itself will remain a priority: calibration, understanding mistakes, behavior across leagues and handling situations where little data is available.
We also want to explore more kinds of predictions over time.
Player markets, along with analyses and probabilities for different periods of a match, are among the ideas we want to explore. Several users have asked for first-half data in particular, rather than only full-time results.
The aim would be to offer a more detailed view of a match while maintaining the same standards of calibration, archiving and transparency.
The model used for live probabilities will also need to keep evolving. In particular, we want to move it gradually closer to the probabilistic approaches found in academic football modeling, while preserving one of Foresportia’s essential requirements: results must be explainable and verifiable.
An iOS app is also among the developments we would like to offer eventually.
This is not a fixed roadmap, however.
Foresportia’s first two years have shown that the best improvements are not always the ones planned months in advance.
Some came from the data.
Others came from problems encountered in production.
And many simply came from your requests and feedback.
That is the approach to continuous improvement we want to carry forward.
Two years in numbers
Today, Foresportia amounts to around 86,000 active users recorded by Google Analytics, 750,000 page views, 4.4 million events, more than 350 accounts created since August 2026, 18,812 matches with archived 1X2 probabilities, more than 103,000 multi-market picks, 59 competitions and around 1,700 teams in the data.
Behind those numbers are two years of revised models, analyzed data, new features, fixed bugs, experiments that worked and others that were abandoned.
And, most of all, a growing number of people using Foresportia.
To everyone who checks the predictions, comes back to see the results, uses the app, reports a problem or simply takes a few minutes to send us a suggestion:
thank you.
Foresportia is two years old.
And we still have a lot to build. ⚽🤖