Are free football prediction sites accurate? We started measuring them, in public

There is a whole category of websites that publish free football predictions every morning: a fixture list, a tip, sometimes a percentage next to it. They are read by a lot of people. They are checked by almost nobody, because checking is tedious — you would have to record what was predicted before kick-off, wait for the results, and know what price you could actually have taken.

That happens to be exactly what our infrastructure already does for tipsters on Telegram. So we pointed it at two of the better-known free sites, Statarea and Zulubet, and started measuring.

This article is not a verdict. It is the method, plus the first day of data, published now so that nobody can later accuse us of starting the clock once the numbers looked convenient.

What we actually do, once a day

1. Read the public predictions once per day. One request per site, with an identifiable user agent and a contact link. We archive the raw page privately so that if the site's layout changes and our parser breaks, we can tell the difference between "they changed the page" and "they changed their picks".
2. Match each prediction to a real fixture in our database, using the same team-name normalisation and alias table we use for tipster audits.
3. Record the price at the moment we read it — the best price available across the soft bookmakers we track for that selection. Not a price we invented afterwards.
4. Settle on the official result, and where we have the sharp closing price on the same market, compute the closing line value: did the price at prediction time beat where the market finished?
5. Publish the denominators. How many predictions we read, how many we could match, how many we settled, and how many had a measurable closing line.

Step 5 is the part that most "verification" services skip, and it is the part that decides whether any of the other numbers mean anything.

Day one

Both sit below the accuracy they would need at the prices their picks carried. And both numbers are meaningless, because 19 and 5 settled predictions are noise. A single good evening moves them by ten points. We are printing them anyway, because the alternative — waiting silently until the sample flatters somebody — is how this industry earns its reputation.

The two columns that will matter in a month are the last two. Accuracy on its own never settles anything: a site that only tips short-priced favourites can be right 70% of the time and still lose money, as the arithmetic of win rate × average odds shows. What settles it is accuracy against the break-even rate at the odds you could have taken, and beyond that, closing line value.

Why only a third gets measured

Of 147 predictions read on the first day, we could match 47. That is not a failure of theirs or ours — these sites cover far more competitions than the 83 leagues we price. A prediction on a league we do not cover cannot be graded honestly, so it is set aside and counted separately rather than quietly dropped or, worse, counted as a loss.

This is the same pattern we found auditing Telegram tipsters: across 46 channels we read 1,279 public picks and could settle 567 of them. The measurable subset is always smaller than the advertised record, and any audit that does not tell you its size is asking to be trusted rather than checked.

Zulubet publishes probabilities, which invites a harder test

Statarea publishes a tip. Zulubet publishes a probability for home, draw and away, and highlights the largest — which is more than most free sites offer, and it makes the site checkable in a stricter way.

A probability is a falsifiable claim. If a forecaster says 70% on a few hundred occasions, the outcome should happen about 70% of the time. If it happens 50% of the time, the number is decoration. We now store every probability Zulubet publishes alongside the eventual result, so in time that page can answer not only "does it beat the closing line" but "are the stated probabilities calibrated at all" — the same test we run on our own model, which, for the record, we published after it came out worse than the market.

What happens next

Both pages update as the data arrives and withhold any verdict until at least 30 settled predictions per site:

• Statarea, audited against the closing line
• Zulubet, audited against the closing line

If the numbers turn out flattering to either site, we will say so. We have published our own losing weeks, our own model being worse-calibrated than the market, and a measurement bug that briefly inflated our own closing line value by thirty points. The point of measuring in public is that the result is not ours to choose.

Method notes: predictions are read once per day from the sites' public pages with an identifying user agent; raw pages are archived privately and never republished. Fixture matching uses team-name normalisation with an alias table; unmatched picks are marked unverifiable and picks on uncovered competitions out of scope — neither counts as a loss. Price at detection is the best soft-bookmaker price we hold for that selection at read time. Closing line value = price at detection ÷ sharp closing price − 1 on the same market. We have no affiliation with, and no commercial relationship to, either site.