What does a 70% win probability actually mean?

When a model says a team has a 70% chance to win, most people hear "this team will win." That is the wrong reading, and it leads to a predictable cycle: the favorite loses, someone declares the model broken, and the next number is ignored. The correct reading is less exciting and far more useful.

The frequency reading

A 70% win probability means: across many games that look like this one, the favored team should win about seven of every ten. It is a claim about a group of similar situations, applied to the single game in front of you because that is the best information available.

That framing immediately explains why an upset is not a failure. If a model gives ten different teams a 70% chance and all ten win, the model was actually wrong — it should have said something closer to 100%. If seven win and three lose, the model did exactly what it claimed.

Why you cannot judge one game

A single outcome tells you almost nothing about whether a 70% estimate was good. The team either won or lost; both results are fully consistent with a 70% forecast. You need a sample. With 20 games you can start to see whether 70% favorites are winning closer to 50% or closer to 90%. With a few hundred, the picture gets reasonably sharp.

This is the main reason EdgeIQ suppresses accuracy figures when a sample is too small. Showing "100% accurate" after three games would be technically true and completely misleading.

How to check whether 70% really means 70%

The check is called calibration. Group every past forecast into bands — say 60–65%, 65–70%, 70–80% — then compare the average stated probability in each band with how often those teams actually won. In a well-calibrated model the two numbers track each other.

Here is what that looked like for EdgeIQ's NFL engine (version 3.1.0) on its 768-game chronological backtest:

Those bands landed close to what they claimed. That does not guarantee future seasons will look the same, but it is the kind of evidence that gives a 70% label meaning.

Probability versus confidence labels

EdgeIQ also shows a confidence label (for example low, medium or high). The label is a summary of the probability plus data quality — how many prior games each team has played, whether key inputs are missing. It is a reading aid, not a separate prediction. Two games can both be 68% and carry different labels if one of them rests on much thinner data.

Common misreadings

• "70% means they win by a lot." Not necessarily. Probability describes how often, not by how much. A projected margin answers the second question.
• "30% means no chance." Three in ten is common. Over a weekend with a dozen games, several 30% underdogs should win.
• "The model flipped from 70% to 64%, so it was wrong before." Probabilities should move when new information arrives. Stubbornness is not accuracy.
• "70% here is the same as 70% in another sport." The meaning is the same, but how often 70% forecasts come up, and how reliable they are, differs by league.

Converting a probability into expectations

A practical habit: when you see a probability, translate it into an expected record. If you follow ten games where the favorite is listed around 65%, expect those favorites to go roughly 6–4 or 7–3. If they go 10–0 or 3–7, that is worth noticing; anything in between is normal variation.

Read a win probability as "how often, across many similar games," never as "what will happen tonight."

Why this matters for trust

A source that publishes probabilities and then grades them openly is accountable in a way that a source publishing only picks is not. Picks can only be right or wrong. Probabilities can be checked for honesty. That is why EdgeIQ publishes probabilities, confidence labels and a calibration record together — and why we think every prediction source should.