Guide

Variance in Football: Why Good Models Still Lose

Why a well-built football model still loses: variance, luck and small samples explained with a simple worked example.

5 min read20 November 2026by AI Football Predictor

Your team dominates on Saturday. They create chance after chance, the stats show about 2.3 expected goals against 0.4, and then a deflected shot in the 78th minute decides the match 1-0. On the way home you read that the model gave your side the best chances. So was the model wrong?

Usually, no. What you just watched is variance, the gap between what was likely to happen and what actually happened. Understanding it is the single most useful step towards reading football numbers sensibly.

What variance actually means

Variance describes how much real outcomes scatter around the average outcome. If you flip a fair coin ten times, you expect five heads, but getting three or seven is perfectly normal. The coin is not broken. Ten flips is simply a small sample.

Football works the same way. A model can say a team has a 60% chance of winning. That also means it has a 40% chance of not winning, and that outcome will happen regularly. A probability describes how often something happens over many similar situations, not what will happen in this one.

Why football is especially noisy

Some sports let the better side win almost every time, because there are many scoring events. Football is the opposite. Several features make it unusually exposed to luck:

  • Few goals. Top leagues average roughly two and a half to three goals per match. When the final score is built from so few events, one lucky bounce changes the result.
  • Rare, high-impact moments. Deflections, goalkeeping errors, penalties, red cards and offside calls are uncommon but can swing a match entirely.
  • Draws are common. Roughly a quarter of matches in many leagues end level, so the outcome space is never a simple two-way split.
  • Finishing is uneven. Even excellent strikers go through spells where good chances do not go in, and average ones go through spells where everything does.
  • Short seasons. A league campaign is around 38 matches per team, which is not much when each match is so noisy.
AIFP
These numbers, for every match todayProbabilities, expected goals and checked reliability, free, on the site.
See the probabilities

A probability is not a promise

The right question about a model is not 'Did it call this match correctly?' but 'Do its probabilities match reality over many matches?' This property is called calibration.

Imagine a model assigns 70% to a thousand different home wins. If the model is well calibrated, about 700 of those teams win and about 300 do not. Those 300 are not failures of the model. They are exactly the share the model said would happen.

A model that never lost would actually be suspicious. It would mean it was claiming certainty in a sport that does not offer any.

A worked example: a 60% favourite, five matches

Let us make this concrete. The numbers below are a simple illustration, not real data. Suppose a team has a 60% chance of winning each of five separate matches, and the matches are independent. How many wins should we expect, and how much can reality drift from that?

Wins out of 5Approximate chance
01.0%
17.7%
223.0%
334.6%
425.9%
57.8%

The average is three wins, and three is also the single most common result. Yet look at the spread:

  • The team wins two or fewer of the five in about 32% of cases, almost one run in three.
  • It wins all five in only about 8% of cases.
  • Even though each match favoured the team, a poor-looking run is completely normal.

Nobody would say the 60% figure was wrong after seeing two wins from five. The sample is simply too small to tell. This is why judging a model on a weekend, or even a month, says very little.

Variance inside a single match: xG versus the scoreline

Expected goals (xG) estimates the quality of each chance, based on factors such as shot location and type. Adding up the chances gives a measure of how many goals a team 'should' have scored on average.

xG does not remove variance. It describes the average, and the real scoreline is one draw from a wide range of possibilities. A team with 2.0 xG will score fewer than 2 goals in plenty of matches and more than 2 in others. Over a few matches, the gap between goals and xG can be large. Over a full season, it usually narrows, although some finishing skill and goalkeeper quality do persist.

The key reading rule: xG tells you about performance, not about the result you were owed. Football does not keep a ledger that pays back missed chances.

How to judge a model fairly

If single results are a poor test, what is a good one? A few principles help:

  • Use large samples. Hundreds of matches say far more than a handful.
  • Check calibration. Group matches by predicted probability, such as all those near 30%, 50% and 70%, and see whether the real frequencies line up.
  • Use proper scoring measures. Metrics such as the Brier score or log loss reward honest probabilities and punish overconfidence, unlike a simple count of correct calls.
  • Compare with a baseline. A model should beat simple references, such as league-wide home, draw and away rates, to prove it adds information.
  • Watch for hindsight. After a surprise, every explanation sounds obvious. Judge the numbers as they stood before kick-off.

Common mistakes when reading football numbers

  • Treating the most likely result as a certainty. Even the most likely outcome often has a probability below 60%, so it fails a lot of the time.
  • Ignoring the draw. Many fans mentally split the match between two teams, but the draw usually carries a meaningful share.
  • Judging by one upset. A 15% event happens about once every seven matches of that kind, so it should happen often across a season.
  • Confusing a big win with a big probability. A 4-0 scoreline does not mean the winner had a 95% chance beforehand.
  • Expecting luck to even out quickly. It evens out over large samples, not over a weekend.

Using probabilities wisely

The healthiest way to use match probabilities is as a way to understand how the match is likely to unfold, how balanced it is and how much uncertainty surrounds it. AI Football Predictor publishes these numbers, including result probabilities and xG, for every match, so you can compare what was expected with what happened and see variance in action over time.

If you ever connect statistics with betting, remember that no number is a guarantee, that models can and do lose, and that gambling is for adults only and carries real financial and personal risks. Play within your limits, or better, enjoy the numbers purely as a way to understand the game.

A good model is not one that never loses. It is one that is honest about uncertainty and holds up over a long run. In football, being right most of the time is not the goal, because the sport itself guarantees you will be surprised.

faq

FAQ

Does a model losing a match mean it is bad?

No. A single result tells you very little. If a model gives a team a 65% chance, the other 35% still happens regularly. Quality shows over hundreds of matches through calibration and scoring measures, not through individual outcomes.

Does xG remove luck from football?

No. xG estimates the average number of goals chances should produce, but the real scoreline is still one random draw around that average. Over a few matches goals and xG can differ a lot, and the gap usually shrinks over a longer period.

How many matches do you need to judge a football model?

There is no exact number, but hundreds of matches is a sensible minimum, and more is better. Within that sample, check whether events predicted at around 30%, 50% or 70% really happen at those rates. Statistics are never guarantees, and gambling is for adults and carries risks.

The probability of every match, free

Today's most likely matches, live, team sheets and proof of reliability, free on the site.