METHOD

Why are "87% accuracy" claims not realistic?

What is the realistic ceiling for predicting match results? How do "87% accuracy" numbers come about when even the betting market stays at 57%, and how do you test a prediction claim?

LigLab Editorial Team · · 6 min read

Anyone who searches the internet for football predictions meets the same promise: "87% accuracy", "90% success", "9 out of 10". The numbers are big and the delivery is confident. We, on the other hand, put 54.5% on LigLab's home page. At first glance the gap is against us.

This article explains where that gap comes from. The short answer: 87% is not possible in match-result prediction. The long answer contains everything you need to test a prediction claim yourself.

Why is a match result hard to know?

A football match has three results: home win, draw, away win. Someone who picks at random without knowing anything gets right about a third of matches over the long run, around 33%. That is the baseline every prediction is compared against.

The ceiling is set by football itself. It is a low-scoring game; a single deflection, a single red card changes the result. Two things make it particularly hard:

  • The draw. About a quarter of matches end level. But a draw is almost never the most likely result in any match. A model that predicts "who wins" loses most of those matches from the start.
  • The favourite doesn't always win. If a team's probability of winning is really 60%, it won't win 4 matches out of 10. That's not the model's mistake; it's what probability means.

That is why even a perfect model that knew the true probability of every match would stay around 60% over the long run. That's the ceiling.

The realistic ceiling: the betting market

The best way to see where the ceiling is, is to look at the strongest prediction in the world. That prediction is the closing odds of the betting market: injury news, line-ups, weather and the information of big money are all gathered in those numbers just before the match begins.

LigLab uses no betting odds and shows no odds. But to measure our model honestly, we took the market as a benchmark and put the two side by side on the same matches:

Match result accuracyAccuracy in the most confident matches
Random guess33%–
LigLab (Consensus)54.5%72.2%
Betting market (closing)57.2%76.8%

The measurement was made on 1,028 matches from three Süper Lig seasons (2023/24–2025/26) that the model had never seen. "The most confident matches" are the favourites given at least 60% probability.

Two conclusions come out of the table. First, we are behind the market and we don't hide it. Second, and important for this article: even the world's strongest prediction stayed at 57% in these matches. Someone who says 87% is saying they beat the market by 30 points. Someone who could really do that wouldn't be selling predictions.

The picture is the same in Europe's four big leagues: in the locked test our accuracy is between 53% and 54%. The current numbers are on the Record page, and the detail of the comparison is on the Comparison with the market page.

So where does "87%" come from?

Most of these numbers aren't made up; they just measure something else. Usually one or more of these routes is used:

  1. Counting the easy question. The answer to a two-option question like "will there be at least one goal in the match?" is "yes" in most matches anyway. Accuracy above 80% is ordinary on these predictions and should not be confused with match-result accuracy.
  2. Choosing a short period. "87% in the last 7 days" can be true. A good week happens to every model; the bad weeks don't make the headline.
  3. Counting only the big favourites. Accuracy in the matches it is most confident about is naturally high. Ours is too: 72%. But that isn't the accuracy on all matches.
  4. Changing or deleting the prediction afterwards. Nobody can test whether a prediction that wasn't recorded before the match landed.
  5. Not saying how many matches. Anything is possible in a sample of 20 matches. In 1,000 matches the room for chance shrinks.

How do you test a prediction claim?

When you see a site or app's accuracy rate, ask five questions:

  • What is being predicted? The match result, or a two-option question?
  • How many matches? If the number is under a few hundred, the rate says little.
  • Which period? Matches the model hadn't seen before, or a period chosen by looking at the past?
  • Are the predictions recorded before the match? Can all past predictions be seen, including the ones that missed?
  • What is the benchmark? Is the rate given next to random guessing and the market?

An "87%" that can't answer these questions clearly is not a measurement; it's an advertisement.

Accuracy doesn't say everything

There is one more detail. Two models can name the same favourite in the same match, one saying 55% and the other 85%. If the favourite wins, both count as "correct". But if it loses, the one that said 85% has made a much bigger error.

That's why, besides accuracy, we also measure probability error (a measure called RPS; lower is better). On the same 1,028 matches, the market's error is 0.1834 and ours is 0.1905. A good prediction says not only who will win but how sure to be. If about 70% of the matches where we say "70% to win" are really won by that team, you can trust the number.

What do we do?

LigLab's promise isn't a high number but a testable one:

  • Every prediction is locked before the match; it is never changed or deleted.
  • The predictions that miss are on the Record page too.
  • Our rate was measured on seasons the model had never seen, and the whole method is written on the Method page.
  • Predictions use no betting data; they work from match results, shots and squad data.

Each league's current predictions and this season's record are on its own page: Süper Lig, Premier League, La Liga, Serie A, Bundesliga.

In short

  • In match results, a random guess gets about 33% right; the market, the world's strongest prediction, stayed at 57% in the 1,028 Süper Lig matches we measured.
  • "87%" usually counts easy questions, short periods or only the big favourites.
  • Five questions are enough to test a claim: what, how many matches, which period, recorded in advance, what benchmark?
  • A good prediction says not only who will win but how sure to be.

Sources

Statistical model prediction; for entertainment only, not betting advice. 18+ How this article is produced and where the numbers come from: How we work. You can add LigLab as a preferred source on Google.

Related articles

Guide
How to set up a prediction league with friends

The steps to start a friends' league in LigLab, the scoring rules, and a few tips for competing on the same table as the bots. No money, no prizes, no betting; just predictions.

16 November 2026 · 4 min read

Guide
Connect LigLab to Claude and ChatGPT: ask for the locked predictions directly

How to connect LigLab's MCP server to Claude, ChatGPT or Claude Code, what you can ask, and where the answers come from. Free, read-only, no betting data.

16 November 2026 · 3 min read

Method
How to predict a football match: the seven bots' methods

Goals, shots, form, ratings? Each of LigLab's seven prediction bots looks at a match from a different place. What does each rely on, and why are there seven bots?

16 November 2026 · 6 min read

← All articles