Are football prediction sites accurate? Almost any of them can be made to look accurate, or hopeless, depending on what you let it count. Our own record is public: on 11 Nov 2026 it showed a mean calibration error of ±0.1% across 6,597,384 graded predictions (all-time; all markets, player markets included). This page explains the trap in the word "accurate" and what to ask for instead, before trusting anyone’s numbers, ours included.
The Base-Rate Trick
Football's outcomes are not evenly likely, and that asymmetry quietly manufactures impressive statistics. A tipster who predicts "at least one goal in the match" every single time is right in the overwhelming majority of fixtures without knowing anything about football. Call "no draw" in every top-flight match for a season and you land roughly three times in four. Always back the favourite in a division dominated by two clubs and your hit-rate looks excellent, because the league did the work, not the judgement. Any headline accuracy claim therefore answers a question you did not ask: not "how good is this model?" but "how predictable was the thing it chose to count?"
Why "Accuracy" Asks the Wrong Question
There is a deeper problem than cherry-picking. A serious prediction site does not output verdicts; it outputs chances. When a model rates an outcome six-in-ten likely, the other four-in-ten is not an error allowance — it is part of the forecast. Mark that model right-or-wrong match by match and you have thrown away the very information it exists to provide. Over any short stretch, a well-built probabilistic model will look mediocre by binary scoring, and a reckless one can look brilliant. Binary accuracy is simply the wrong ruler for probabilistic output, in football as in weather forecasting.
Calibration Is the Right Question
The correct test asks whether stated chances survive contact with reality. Take all the graded predictions a site has ever made; group them by claimed likelihood; then compare claim with outcome inside each group. Outcomes rated six-in-ten likely should have arrived close to six times in ten; outcomes rated as long shots should have stayed rare. Plot claim against reality and you get a reliability curve; summarise the gaps and you get a mean calibration error. Three conditions make the exercise meaningful: a large sample (hundreds of settled predictions at minimum, ideally thousands), a complete ledger with every losing call still in it, and continuous updating — a record that stops when form dips is not a record.
A Live Example, Not a Claimed One
This is the standard TACTIX volunteers for. The live scoreboard grades each prediction with a result against the published method — wins and losses, by market and by confidence tier, with the underlying record downloadable as CSV or JSON. As of 11 Nov 2026: mean calibration error ±0.1% across 6,597,384 graded predictions (all-time; all markets, player markets included) — and those figures move as the results come in, which is the point of a live record. You can watch the same record accumulate fixture by fixture on the Premier League predictions page.
Three Checks Before Trusting Any Site
First, hunt for the method: a publisher confident in its model explains it. Second, find the ledger and scroll for losses; a history of winners only is a sales page, not evidence. Third, look for the word calibration — or any sign the site compares its stated chances with observed reality at all. Sites pass or fail these checks in minutes, and the checks themselves are the honest answer to this page's title: prediction sites are exactly as accurate as they can prove, and the proof has a specific shape.
Prediction-Site Accuracy — FAQs
Are football prediction sites accurate?
Some are honest about how good their probabilities are; many never show you. Judged by headline hit-rate alone the question cannot be answered, because hit-rates mostly reflect what a site chose to count. Judge instead by calibration — whether stated chances match observed outcomes over a large, complete, published sample.
What is a base-rate artefact?
An impressive-looking score that comes from the underlying frequencies of football rather than from any forecasting skill. Predict "at least one goal in the match" every time and you are right in the overwhelming majority of games; predict "no draw" every time and you are right in roughly three matches out of four. A site can headline either number without possessing any insight at all.
What is calibration in football prediction?
The property that stated chances match reality in aggregate: of all outcomes a model rated six-in-ten likely, close to six in ten should actually have happened. It is measured by grouping settled predictions into confidence bands and comparing claimed frequency with observed frequency — the result is a reliability curve and a mean error.
What accuracy figure should a good site quote?
Ideally none in isolation. A single percentage without its market, league mix, time window and sample size is marketing, not measurement. A trustworthy publisher shows the whole distribution: performance by market and confidence band, sample sizes attached, losing calls included, updating as results settle.
Where can I see a live calibration record?
The TACTIX scoreboard is exactly that: a continuously updated public record of the predictions with a result, each graded win or lose, with reliability figures and downloadable data — the long-form walkthrough lives at how accurate is TACTIX.
More From TACTIX
- How accurate is TACTIX? The graded record
- Live model scoreboard — the full calibration record, losing calls included
- Methodology & editorial standards
- Premier League predictions — the format in practice
- Free football stats hub
18+. TACTIX guides and comparisons are model-informed analysis — not betting advice or guarantees. Please gamble responsibly — support at BeGambleAware.org.
This post is generated from TACTIX’s match data by our software and checked automatically. How it works →
See today’s free slate
Free, every day: the model’s fixed slate of picks, each with its confidence tier and that tier’s settled record — plus yesterday’s slate, graded. Pro adds the full board, the value edge and the written analysis. No payment card.
Create free account