This page does not rank apps
We have not downloaded, used or compared any football prediction app. That means we will not name a best one. A list that ranks apps without checking their predictions against a full season of results is showing you affiliate arrangements, not evidence about the apps.
Instead, here is a cheaper and more useful approach: six tests that help you tell a product built around a model from one built around a slogan. Use them for apps, websites and this site too.
Six checks that matter
A rough order from strongest signal down.
Probability first, not just a selection
A pick cannot really be judged from one match. It was either correct or “unlucky”, and you cannot know which. A probability makes a testable promise: when something is priced at 60%, it should happen about six times in ten over a large set of matches. If a product only gives selections, it has chosen the format that is hardest to score.
The number should have clear ingredients
Ignore phrases such as “advanced algorithms” or “AI-powered” unless the real inputs are explained. You should be able to see which matches are used, how far back the data goes, how it is weighted and what data is included. If the same description could sit on any similar product, it has not described anything useful.
A backtest needs a date and a sample size
Look for a record from a season that was not used to fit the model. The useful details are the number of predictions scored, plus the number left out and the reason for leaving them out. If a record has no skipped games and no wrong calls, it is more likely a marketing choice than a serious test.
Calibration matters more than hit rate alone
This check removes many weak products. Picking the favourite every time can create a high hit rate, but it does not prove much. A model is useful when the size of its probability is close to reality, not when it simply picks the obvious side. If the only proof offered is a percentage of correct picks, that is the easiest figure to make look flattering.
An accuracy claim must be able to fail
Ask what result would prove the claim wrong. “85% accuracy” without a period, market or sample size cannot be tested, which is why it is often written that way. A claim worth your time states the season, the market and the count.
Price and result should be kept separate
“Value” means the price is considered wrong. It does not mean the outcome is certain. Even a real edge on a 25% chance still loses three times in four. When a product presents an edge as a likely win, it either does not understand its own number or expects you not to understand it.
Apply the same checks to us
We show probabilities for every fixture, not picks. The inputs and constants are set out on the mathematics page. The accuracy page has a dated backtest on a hold-out season, with skipped fixtures counted and calibration error published league by league, including the league that missed our own tolerance. Check number four is the one we would start with.
There is no app. Alin is a website, and everything it does works in a browser.