PredictZ accuracy: the ten-minute verifiability test for any tips site
By FootInsights · Published · 4 min read
“PredictZ accuracy” is one of the most-searched phrases in football prediction, and what it returns is a familiar swamp: third-party sites guessing, review pages with an affiliate link where a method should be, and no number anyone can check. That is not a PredictZ problem — to be explicit, nothing here claims their predictions are poor. It is a structural problem shared by every prediction site on the internet, ours included: an accuracy figure is only worth what the record behind it can prove.
We’ve already published the slow instrument for this question — the 30-day audit you run yourself on any site you rely on. This piece is the fast one: five properties you can check from your chair in about ten minutes, which together decide whether a site’s record is even capable of supporting an accuracy claim. Score any site out of five — PredictZ, us, whoever — before you spend a month auditing it.
The five properties of a checkable record
1. Predictions are frozen before kick-off, provably. A prediction only counts if it demonstrably existed before the result did. Look for timestamps on the prediction itself, and check whether old prediction pages survive unchanged — try a few in the Wayback Machine. A record that lives in an editable database, displayed after full-time, proves intentions the way a diary written in December proves January’s plans.
2. The denominator is complete. Find the site’s results page and count backwards: does every fixture that got a prediction appear in the settled record, including the bad weekends? The fastest tell is coverage asymmetry — hundreds of daily predictions but a results archive that only surfaces highlights. Whatever isn’t in the denominator is doing silent work for the numerator.
3. “Accuracy” has a stated definition. A match-outcome hit rate, a correct-score hit rate and a both-teams-to-score hit rate are wildly different numbers — correct scores land far less often than 1X2 picks do, through no fault of any model. If a percentage is advertised without saying which market, over which period, on which fixtures, it isn’t a measurement; it’s a slogan.
4. Probabilities, not just picks. A bare pick (“home win”) cannot be scored for calibration — there is no way to ask whether the site’s 70% claims land 70% of the time, because no such claims exist. Only published probabilities can be held to a proper scoring rule like the Brier score. A site that publishes picks alone has chosen a format that cannot be judged on honesty, and that choice is itself information.
5. The record survives independent reconstruction. The strongest version of all of the above: could a stranger, using archived snapshots and public results, rebuild the site’s record and get the same number? This is the property that makes the other four enforceable — and it is the one almost nobody, anywhere in this industry, fully provides.
What a picks-only record can still tell you
Suppose a site fails on probabilities but passes on freezing and completeness. Its hit rate is then real but nearly meaningless on its own, because hit rate is mostly a function of which fixtures get picked: backing the home team in every match of our nine-league dataset lands around 44% by itself, and favourites-only selections beat it comfortably while still losing money at the prices. The only honest use of a picks-only record is comparative — the same fixtures, the same period, against a no-skill baseline, exactly as step four of the 30-day audit sets up. If you can’t run that comparison from the site’s own archive, property 2 has already failed.
And before any number impresses you, remember the sample: thirty picks separate nothing from nothing. The arithmetic of why is in sample size in betting records, and the honest ceiling for the whole exercise is in how accurate football predictions can actually be.
Scoring ourselves, since we brought it up
Fair is fair. FootInsights against the five: predictions are frozen before kick-off with a model version attached (1); every prediction is settled on the track record, misses included, per league and per market (2); the metric is named — Brier score, not a headline win rate (3); we publish full probabilities for every priced market (4). Property 5 — independent reconstruction — is where we are honest about the limit: like everyone else, we host our own record, so we’d encourage exactly the scepticism this article teaches. Archive our pages. Check us later against what you saved. How our predictions work explains what you’d be checking.
That is also why you will find no accuracy percentage typed into this or any other article here — a number in static text can’t update, can’t be audited, and can’t go stale honestly. The live track record is the claim.
Run the ten-minute test on whichever site you were about to trust. If it scores well, graduate to the 30-day audit. And whatever the spreadsheet eventually says, treat betting as entertainment with a cost, never as income — an edge, even a real one, is small, and variance is patient.