Why a 70% prediction is wrong three times in ten

Here is the sentence that causes the most trouble in sports analytics: a 70% prediction that loses has not malfunctioned.

If it never lost, it was not a 70% prediction. It was a 100% one, and no honest model produces those.

What the number is promising

A probability is a rate, not a verdict. "70%" is a claim about what happens across many similar situations: out of every ten fixtures the model rates this way, it expects roughly seven to land and three to miss.

The three misses are not the model breaking. They are the model being correct about a world that does not always cooperate.

Losing runs are ordinary

People underestimate how streaky an accurate model looks over short stretches. At a genuine 60% strike rate, runs of three and four consecutive losses arrive regularly and mean nothing at all. At the point you have enough results for the pattern to be informative, you have far more results than most people ever look at.

This is why OddUni publishes the sample size next to the record rather than the percentage alone. A win rate with no denominator is not a claim you can evaluate.

How you would actually spot a bad model

Not by whether it lost yesterday. By whether it is calibrated — whether the things it calls 70% land about 70% of the time, across a large number of them.

A model can be badly wrong in two directions:

1. Overconfident — its 70% calls land 50% of the time. The probabilities are inflated and the confidence figure is decoration.
2. Underconfident — its 55% calls land 75% of the time. Less harmful, but still not describing reality.

Both are only visible in aggregate, over a long run, in a record that includes the losses. Which is the entire argument for publishing one.

What this means for reading OddUni

• A losing prediction is evidence of almost nothing on its own.
• A published record with a large sample is evidence of something.
• Past performance still does not guarantee future results — the competitions, the data and the sport itself all move.

The full record, by sport, with every sample size, is on the results page. It contains the losses because a record without them is marketing.

And the part that matters more

None of this is a reason to stake money you need. A probability is a description of uncertainty, not a way of removing it. If gambling has stopped being entertainment, help is available and free.