A 70% favourite in a football match wins 7 times out of 10 on average. In any single match the same favourite loses 3 times out of 10, and the 70% number is unaffected. Probability is not a prediction. It is a long-run average that, on any given match, has a wide range of possible outcomes.
What “70%” actually means
Calibrated analyses give a probability per outcome. The favourite wins 70% of the time across a large sample of similar fixtures, not in any single fixture. Read as a distribution, not as a vote.
Across 10 matches at 70% favourite, the expected wins are 7. In any given 10-match stretch, you might get 5, 6, 7, 8, or 9 wins. The 70% is the long-run expected value, not a per-match guarantee. The same logic applies on a per-match basis: a single 70/20/10 forecast splits 70-20-10 only across thousands of similar fixtures.
The single-match distribution is wide. Even a “70%” headline leaves a 30% chance the favourite does not win. That is the correct shape of the forecast. It is not a calibration failure.
The 1-in-3 problem
Across the Big Five from 2020-21 through 2024-25, the public favourite in a heavy-mismatch fixture won 68.9% of the time. That is 1 win short of 70%. The other 31% are draws or losses. A heavy favourite loses or draws nearly one-third of the matches they should win.
When you see a 70/20/10 distribution on a heavy mismatch, the 10% is not a rounding error. It is roughly 1 in 10 matches. The 20% is roughly 1 in 5. Together they account for nearly 1 in 3 of those matches not going as the public favourite expected.
That is the correct shape of the forecast. It is not a calibration failure.
How to read a forecast
Three frames actually matter:
Compare it against your own expectation. If your read is “this team is 65% to win” and the model says 70%, the 5 percentage-point gap is what the model adds on top of your intuition. Both can be calibrated to different priors.
Compare it against baseline. A 60% home-favourite win rate is meaningful only because the baseline rate (across all fixtures) is around 43% for slight favourites and around 50% in the aggregate. Without a baseline, the 60% is just a number.
Compare it across distributions, not headlines. A 60-25-15 call is meaningfully different from 60-30-10. The first concentrates confidence on the favourite. The second distributes it across draw and underdog. Both are “60% wins”; only the first tells you the favourite is much more likely to win than to draw or lose.
A model that always says 60-25-15 is committing fraud in the noisier matches. A model that lands at 40-30-30 is being honest about uncertainty. The distribution shape is the forecast. The headline number is one of three figures in that shape.
What good reading looks like
A model that calls 100 heavy-favourite matches and lands 69 out of 100 wins is doing its job. Whether the next single match goes the right way is noise. Whether the model lands close to the long-run distribution across many matches is signal.
The mistake readers make is treating the forecast as a vote. The forecast is a calibrated probability. Reading it as a vote means expecting 100% of “favourite” calls to land. Reading it as a probability means expecting ~70% of them, and understanding that the other 30% is the model saying “this match can go the other way and we are not surprised when it does.”
What this means for an analysis
The right read on a 60-25-15 call is not “60% likely to win.” The right read is “60% likely to win across many such fixtures; in any single match this could be a win, a draw, or a loss with the model’s stated distribution.” That is the calibration.
For a reader comparing per-league pre-match reads, the more useful question is not “did the model pick the winner” but “is the model’s distribution close to the long-run outcome across many fixtures.” A 60-25-15 model that hits 58% wins, 26% draws, 16% losses over 100 matches is well-calibrated. A 60-25-15 model that hits 70% wins, 18% draws, 12% losses over 100 matches is mis-calibrated toward the favourite.
A read is a lens for separating signal from noise. The forecast’s job is to land close to the long-run distribution, not to be right on every match. The reader’s job is to read the distribution, not to expect a vote.
Continue with the methodology
A full walkthrough of how the Goalsforge model reads a football fixture, what it weighs, and how published calibration is computed.