Why almost every accuracy claim you see is unverifiable

Accuracy claims in this industry fail for one of three reasons. Sites grade only the picks that won and quietly delete the rest. Sites count a draw as a half-win, or a void, instead of a loss. Or sites publish so many picks across so many markets that some subset always looks brilliant in hindsight. None of that is measurement. It is marketing arithmetic.

Real accuracy requires four rules, applied before the matches kick off: every pick published in advance, every pick graded against the official full-time result, draws counted as losses when a winner was picked, and unverifiable fixtures voided rather than claimed. Those are the rules we grade ourselves by, and the entire log, losses included, sits on our homepage where nobody can edit it after the fact.

Our actual numbers, band by band

Since 4 August 2026, every pick on our board has carried a probability from our model. Here is how each confidence band has performed against reality so far:

Band	Model probability	Record	Hit rate	What calibration predicts
Strong picks	70%+	3W–0L	100%	75–85% over time
Good picks	55–70%	3W–1L	75%	55–70% over time
Close calls	Under 55%	1W–3L	25%	Under 55% over time
All graded picks	—	6W–4L	60%	—

Two honest observations about that table. First, the sample is ten graded picks, which is small; the strong band will not stay at 100%, and we say so on every page that mentions it. Second, and more importantly, the bands are behaving exactly as calibrated probabilities should: high-probability picks win far more often than coin-flips. That ordering, not any single percentage, is the real test of whether a prediction model knows anything.

What a probability model can and cannot see

Modern football prediction models, ours included, are built on inputs a machine handles better than a pundit: long-run team strength, home advantage, scoring rates, and the brutal three-outcome maths of football. Where models are structurally blind is everything that happens between data updates: a dressing-room falling out, a keeper injured in the warm-up, a manager resting nine players for a European tie four days later. This is why our published probabilities move less than betting markets do on matchday morning, and why we tell readers on the tomorrow predictions page to re-check picks once line-ups drop.

It is also why the ceiling exists. Football has three outcomes and the most common winning margin is a single goal. Even the best-resourced models in the public literature sit near 55% on match results across full seasons of top-flight football. A site claiming to beat that by thirty points is claiming to have solved a problem the entire quantitative industry has not.

How to judge any prediction site’s accuracy in 60 seconds

Ask three questions. Can you find a pick that lost, with its final score, published on the site? If every visible pick won, the record is curated. Are picks timestamped or clearly published before kick-off? Hindsight picks are worthless. Do they tell you draws count as losses? If draws vanish from the denominator, the headline number is inflated by roughly a quarter in a typical league season. Our full checklist for this is in what makes a good football prediction site.

FAQ

Is 60% accuracy actually good?

For one-outcome-per-match picks across a mixed slate of leagues, yes. Random guessing across three outcomes lands near 33%; always picking the favourite lands near 45–50% depending on the league. 60% with honest grading beats both. Whether it makes money depends entirely on the odds you take, which is a different question from accuracy.

Do AI predictions beat human tipsters?

On volume and consistency, yes: a model never gets bored, never chases a narrative, and prices every match the same way. On information that emerges hours before kick-off, an attentive human still has an edge. The strongest setup is the one we run: model probabilities first, human judgement on top, and public grading to keep both honest.

Where can I check your record myself?

The Results board lists every graded pick with its competition, final score, our stated confidence, and the outcome. It updates automatically after each round, and voided picks are labelled as voids, never counted as wins.

Predictions are information, not financial advice. Never stake money you cannot afford to lose. 18+ · Responsible gambling resources