Choosing a Football Data Analysis Site: Seven Revealing Questions

Why These Sites Are Hard to Compare

Search for a football data analysis site and the results all look approximately the same. Every one of them has charts. Every one has a win rate somewhere on the homepage. Every one describes its method as advanced, proprietary or AI-driven. Read three of them back to back and the only thing that reliably differs is the colour scheme.

That is not because the field is uniformly good. It is because the marketing surface of a strong service and a weak one converge: both are built from screenshots, percentages and confident sentences, and none of those things can be checked from outside.

So comparison has to move off the marketing surface. The seven questions below do that. None of them ask whether a site is right — being right is a property of a long sample and you cannot assess it in an afternoon. They ask whether being wrong would be visible. Most services fail on that axis, and they fail in the same few ways.

Run these on whichever site you are considering. Then run them on us; the second half of this article does exactly that, including the questions we do not pass cleanly.

The Seven Questions

1. Does it publish the losses?

The first question, and the one that eliminates the most candidates. Scroll back through a service's public feed and look specifically for losing calls. Not a rough week acknowledged inside a summary post — individual losing calls, in the same format and the same place as the winning ones.

If the losers are absent, the feed is a highlight reel. Nothing in it needs to be false for the whole to mislead: every surviving post can be genuine while the sample is manufactured. Quiet deletion is the cheapest trick available here, and it leaves almost no trace.

2. Is the record verifiable, or only viewable?

A record you can look at is not a record you can check. Verifiable means an outside reader can reconstruct it: named fixtures, stated sides, stated prices, and a settled outcome for each one against a real final score.

The specific failure to look for is the unstated denominator. A stated accuracy figure with no accompanying count is not a measurement — over how many calls, in what window? And if the count exists but the individual items do not, you are being asked to check the publisher's arithmetic against the publisher's own summary of the publisher's own data.

3. Is there a timestamp on every call?

This one works almost as a single-question filter for the whole category. A prediction with no publication time cannot be distinguished from a note written after full time.

Look for two things. First, that calls appear on a public surface that is append-only in practice — one where an edit or a deletion is itself visible, and where readers who saw the original are witnesses the publisher does not control. Second, that in-play calls carry the match minute. A live call with no minute on it is unfalsifiable by construction: nobody can tell whether it went out before the moment that made it look smart, or after.

4. Is the settlement rule written down?

Football markets produce partial results. Quarter-ball Asian handicap lines split, and pushes return the position with no outcome at all. A service that has never stated how it treats halves and voids is holding a free parameter, and free parameters drift toward whoever holds them.

Look for a formula, in one line, that names the buckets and says which of them sit in the denominator. If that sentence appears nowhere on the site, the win rate on the homepage has no definition — which means it has no meaning either.

5. Is it published before the event, or explained after it?

Retroactive publication is the most common trick and the hardest to see, because the content of the post can be entirely truthful. A correct analysis of a match that has already finished is still a correct analysis. It simply is not a prediction.

The tell is structural rather than textual. Post-match material tends to be longer, better organised and more certain than anything written before kickoff, because writing uncertainty well is genuinely difficult. A feed where the pre-match content is thin and the post-match content is rich has already told you where the effort goes.

6. Is the access logic transparent?

You do not need a service to be free. You need to know what you are buying before you buy it, and what changes when you do.

Two things to look for. First: is the tier structure stated plainly, or does finding it require a conversation with someone in a chat window? Second, and more important: does paying change the analysis, or only the coverage? A service whose higher tiers claim better signals is making a claim that cannot be verified from outside, and that none of the six other questions can reach.

7. Will it put its name on a position?

The last question is about exposure. Anyone can produce commentary. The question with teeth is whether the service commits to specific, settleable positions on a clock — a named fixture, a named side, a stated price, published before the event — or whether it stays in the safe register of "one to keep an eye on" and "value looks to be with the home side".

Commentary is never wrong. That is precisely the problem with it.

Running the Seven on Ourselves

Applying a checklist to yourself is only useful if you are willing to fail parts of it in public. Here is where we actually stand.

1. Losses. Yes. Every published signal is settled against the real final score, and the settlement is posted daily with the losses included, on X at @Oddsflow_Nat and Threads at @oddsflow.ai.

2. Verifiable. Partly. Individual signals are all there with fixtures, sides and prices, so the record can be reconstructed by hand. But there is no third-party auditor, and we are not going to pretend that self-publication and independent audit are the same thing. What we can offer is that the checking is possible. Doing it is on you.

3. Timestamps. Yes. Publication timestamps on public channels, plus the elapsed match minute on in-play cards. The field-by-field layout is in How to Read an OddsFlow Signal Card: Every Field Explained.

4. Settlement rule. Yes, in one line: win rate = won ÷ (won + lost + half), with void excluded from the denominator and half counted in full. Half staying in the denominator is the choice that works against us, and it is deliberate. The reasoning is in How We Count Our Record: The Formula and the Timestamps.

5. Before the event. Yes. Signals publish before kickoff, or before the in-match moment they refer to, and never afterwards.

6. Access logic. Mostly. New users get a three-day full trial, there is a free daily pick after it, and continued access runs on tokens earned through daily check-in and /follow tasks. VIP changes coverage, not analysis: the model, the filter and the settlement rules are identical across tiers. The part we cannot verify for you is that last sentence — you would have to watch the free tier and the full feed settle over time and see whether they behave the same way.

7. Exposure. Partly, and this is the honest gap. The agent commits to specific positions with fixtures, sides, prices and timestamps, and every one settles in public. But a published position is not a wagered one, and no bankroll ledger is published. Anyone telling you that publishing and staking are the same commitment is overselling.

There is also an eighth question nobody asks, which we fail outright: sample length. A selective filter — roughly one candidate in twenty-two clears — produces long streaks in both directions, and a few weeks of watching cannot separate a good process from a lucky one. That is a limit of the arithmetic, not a matter of trust.

What the Checklist Cannot Do

These seven questions detect the absence of accountability. They do not detect skill. A service can publish its losses, timestamp everything, define its settlement precisely, and still be a losing operation with excellent hygiene.

Passing means something narrower and more useful: the service is built so that being wrong shows. That is the precondition for evaluating anything at all. Fail these questions and there is nothing to evaluate — only material to admire.

For the version of this aimed specifically at machine-learning claims, AI Football Prediction Tools: Four Types and How to Verify Them takes the same posture toward anything calling itself AI. And for what sits behind the answers above, What Is OddsFlow? From Model to Telegram Agent is the map.

Take these seven to whichever site you evaluate next. If it answers them, you have learned something real. If it will not, you have also learned something real, and rather faster.