Analysis
Why 83% accuracy claims are false
Search for football predictions and you will find sites claiming 80%, 85%, even 93% accuracy. None of them are measuring 1X2 outcomes over a public log. Here is the arithmetic that shows why.
Start with what the market already knows
Bookmakers price thousands of matches with far more data than any public model, and they compete on accuracy because being wrong costs them money. If 80% 1X2 accuracy were achievable, their odds would reflect it and there would be no market left to bet into.
A simple benchmark: always pick the home team. Across major European leagues that wins roughly 47% of the time, for free, with no model at all. Any honest claim must be measured against that floor, not against zero.
Three outcomes, not two
Football's 1X2 market has a draw. Draws occur in roughly a quarter of matches and are close to unpredictable — the whole point of a draw is that two sides cancelled out. A model that never tips a draw forfeits about 25% of matches before it starts; one that tips draws often is wrong most of the times it does.
That structural constraint is why serious published work on 1X2 forecasting clusters in the low 50s. Our own held-out validation landed at +7.1 percentage points over baseline, which puts the honest ceiling around 51%.
How the big numbers are manufactured
The claims are not usually lies so much as a different measurement wearing the word "accuracy":
| The trick | Why the number inflates |
|---|---|
| Double chance counted as a hit | "1X" wins if the home side wins or draws. Two outcomes out of three. Around 70% before any skill is involved. |
| Over/Under 1.5 goals | Roughly 75% of matches go over 1.5. Predicting it every time looks like a strong model. |
| Counting only "confident" picks | Publish 200 predictions, report accuracy on the 20 you liked most. Nothing constrains which 20. |
| Results recorded after the match | Without a timestamped pre-kickoff log, nothing stops a losing pick being quietly revised. |
| No sample, no denominator | "85% accuracy" with no count and no date range is not a measurement. |
The test to apply to anyone, including us
- What is the denominator? 85% of how many, over what dates?
- What counts as a hit? Straight 1X2, or double chance and over/unders folded in?
- Were the picks published before kick-off? If they cannot show you that, the number means nothing.
- Are the losses shown? A record with no bad weeks is a marketing page.
- Is it above the always-home baseline? Below roughly 47% on 1X2, a model is losing to a coin with a preference.
Apply it here. 157 graded 1X2 picks, logged before kick-off, 43.3% accuracy, value bets 19 of 73 at −21.4% ROI. That is currently below the always-home baseline, and it is on the record page with the reasons and the caveats.
We would rather publish a number you can check than one you would prefer.
What to do with this
Treat any 1X2 accuracy claim above roughly 55% as a measurement question, not a skill claim. Ask what is being counted. The gap between 51% and 85% is not talent; it is almost always a different denominator.
Figures current as of 27 August 2026. The baseline and ceiling figures come from our own backtesting, described in how the model works.