No rankings here
We have not installed or measured any football prediction app, so we will not name a winner. A “best app” list is only meaningful if its author scored every app’s predictions against real results over a full season — without that, it is a list of commercial relationships.
Instead, here are six checks that reveal whether a product has a real model behind it or just a tagline. They work on apps, on websites, and on TipsterGrid.
The six checks
Ordered roughly by how much each one reveals.
Probability or pick?
A single pick can never be shown wrong — it either won or was unlucky. A probability makes a testable commitment: if it says 60%, the outcome must happen around six times in ten over many matches. A product that only ever names a selection has picked the format nobody can score.
Does it explain the inputs?
“Smart algorithms” and “AI-powered” explain nothing. Look for specifics: which matches, how far back, how they are weighted, which data. A description that would fit every competitor tells you nothing about this one.
Is there a dated backtest with a sample size?
Look for results on a season the model was not trained on, with the number of predictions scored — and the number left out, with the reason. A record with no exclusions and no losses reflects a marketing choice, not a model.
Calibration, or just a hit rate?
Most products fail here. Backing the favourite every time gives a respectable hit rate and proves nothing. A model shows its value by getting the size of each probability right, not by naming the obvious winner. A product that only quotes its percentage of correct picks is quoting the easiest figure to flatter.
Could the accuracy claim turn out false?
Ask what result would disprove it. “85% accuracy” with no time period, no market and no sample size can never be disproved — that is exactly why it is phrased that way. A claim worth trusting names its period, its market and its sample.
Does it treat price and outcome separately?
Value is a verdict on the odds, not on who wins. Find a true edge on a 25% shot and you will still lose that bet roughly 75% of the time. A product that presents an edge as a probable win has either misread its own number or hopes you will.
Apply the checks to TipsterGrid
We publish probabilities for every fixture instead of picks, list every input and constant on the maths page, and report a dated backtest on an unseen season — excluded fixtures counted, calibration error shown per league — on the accuracy page, where the league that broke our tolerance is named rather than hidden. If you only run one check on us, make it number four.
TipsterGrid has no app. It is a website, and all of it works in your browser.