How to choose a football prediction site
Every site in this category claims to be accurate, so the claim itself carries no information. What separates a model worth reading from a marketing page is whether its claims can be checked at all - and six questions settle that, none of which require statistical training.
1. Is there a graded record, and is it complete?
The first question is not how accurate a site is but whether it publishes the outcome of every prediction it made, including the ones that failed. A record that only contains selections the site chose to keep is not a record - it is a portfolio.
What to look for: a page listing individual past predictions with their outcomes, reachable without an account, covering a continuous period rather than a curated selection. What to be wary of: a single headline percentage with no underlying list, results that only appear as screenshots or testimonials, or a history that begins conveniently recently.
This is the check that most sites fail, and it fails cheaply - publishing a complete record costs nothing technically. A site that has not done it has usually decided not to.
2. How large is the sample, and over what dates?
A percentage without a denominator is not a statistic. Forty predictions can produce almost any hit rate through chance alone, and a figure drawn from one hot month tells you about that month rather than about the model.
Any published accuracy figure should state how many graded predictions it covers and the date range it spans. If a site reports by market, each market should carry its own count, because a model can be genuinely good at one and poor at another - and a blended figure across markets is where that difference goes to hide.
Two-outcome markets like both teams to score sit structurally higher than three-outcome markets like the match result. Any site quoting one number across all of them is, intentionally or not, publishing a figure shaped by which markets it covers more of.
3. Is calibration reported at all?
Calibration is whether stated probabilities match reality: of everything a model called at 70%, close to 70% should happen. It is the property that makes a probability useful rather than decorative, and it is the one thing a site cannot fake by picking favourable examples.
The reporting takes one table - predictions grouped by stated probability, each group's claim set against what actually happened, with a sample size per group. A site that shows this is handing you the tools to audit it. A site that talks about accuracy but never about calibration may simply never have measured whether its percentages mean anything.
This is the strongest single signal available, because it is both easy to verify and rare. If a site publishes nothing else useful, a calibration table would still tell you more than a homepage full of figures.
4. Is the method described in checkable terms?
"Advanced AI" and "proprietary algorithm" describe nothing. A site that has built something can say what kind of model, what data goes in, how often it retrains and how probabilities are produced - and none of that gives away anything a competitor could copy.
What a real description sounds like: named model families rather than adjectives, a specific count of input features rather than "millions of data points", and an explanation of how raw model output becomes a stated percentage. That last step is the one most often skipped, and it is where a number either acquires meaning or does not.
Vagueness here is not always dishonest - some sites are genuinely thin and have nothing to describe. Either way it tells you what you are reading.
5. How does the site make money?
A prediction site that is paid for sending visitors to a particular third-party operator has an interest in which predictions it shows you and where those predictions point. That is not an accusation against any specific site; it is an ordinary incentive that exists whether or not anyone acts on it, and knowing it is present changes how you read the output.
The question is answerable in a minute. Are outbound links tracked affiliate or referral links? Does the site name commercial partners? Is it funded by subscriptions, advertising, referral fees, or some combination? A site that states its revenue model plainly has removed a reason for you to wonder.
Subscription-funded is not automatically better - it creates its own pressure to look impressive. But it is at least a pressure aimed at the quality of the predictions rather than at their direction.
6. What does it say about being wrong?
Football is uncertain in a way that caps how much anything can know. A red card in the ninth minute is not a signal a model missed; it is noise the sport produces. Any method built on the game inherits that ceiling.
So the way a site discusses its own limits is diagnostic. Does it name the markets where it performs worse? Does it state the sample size, so a thin figure cannot pass for a solid one? Does it explain that a 75% forecast failing is expected roughly a quarter of the time rather than treating it as an anomaly?
The inverse is the clearest warning sign in the category. Language that promises outcomes or certainty is describing something the sport does not contain, and a site willing to say it about football is telling you how carefully it says everything else.
Where ScorePredicts stands on all six
Applying a framework to everyone except yourself would be worth nothing, so: every forecast is graded against the final result and is not changed afterwards, and the record is public and requires no account - aggregate figures on the accuracy page, individual graded predictions on every league and team page. Headline accuracy is reported separately for each prediction type, each with its own sample count, and the mix of markets those predictions came from is published beside it - so the figure can be taken apart rather than taken on trust.
Calibration is published for the match-result market as a bin-by-bin table with sample sizes and 95% intervals, alongside a comparison against a naive home-win baseline on that market. Where the data cannot support a claim, the page says so: calibration for other markets is not yet published, and for both teams to score it is withheld deliberately because the forecasts are too concentrated to report honestly. Every hit rate is shown with the number of graded predictions behind it, so a small sample can be seen for what it is, a calibration bin needs at least 50 fixtures before it is plotted, and fixtures that were never completed are excluded, with the count shown. The same numbers are downloadable as JSON under a CC BY 4.0 licence.
The method is described by component - how many models, how they are grouped by competition, how many input features, and which calibration method converts model output into a stated percentage. Revenue comes from subscriptions only, billed through Paddle: no advertising, no affiliate or referral links, no commercial partnerships with third-party operators, no commission from anyone, and no links sending visitors to any third-party operator. And nothing here is presented as certain, because a probability is not a promise and a site claiming otherwise would be failing its own sixth test.