Why draws are the hardest result to predict in football: the view from a settled 192-fixture record

Why draws are the hardest result to predict in football

Draws are hard to predict because they sit in an awkward middle: common enough that ignoring them is wrong, roughly one match in four, yet almost never the single most likely result of any one fixture. In our settled 192-fixture record, 26% ended level. Among the fixtures our model scored at 7.0 or higher, one in twenty-two did.

How often do lower league matches actually end in a draw?

The long-run figure quoted for the English leagues sits around a quarter of matches, and our own settled record, 192 fixtures across four divisions from 14 April to 25 May 2026, landed almost exactly there: 50 of 192, or 26%. The division-by-division split is where it gets more interesting.

League One was the draw-heaviest division in our window, and it is no coincidence that it also produced two of the model's three biggest misses. The National League drew least, and it was also the division where the model separated fixtures most cleanly, going 8 of 8 on its 7.0+ calls. Fewer draws means more fixtures where one side's case is genuinely stronger, which is exactly the condition a factor model works best in.

Why do prediction models struggle with draws?

Because a draw is rarely the modal result. Take a fixture between two evenly matched sides: the home win might be 35% likely, the away win 33%, the draw 32%. The draw is a huge slice of reality, but it is still the third most likely single outcome, so any system built to name the most probable result will almost never name it. The sport's most common phrase for this is that the draw "splits the difference", and models are built to find differences, not splits.

That is not a flaw unique to models; pundits do not predict many 1-1s either. The difference is that a model can at least tell you where the draws will cluster. They cluster where nothing separates the sides, and "nothing separates the sides" is precisely what a low conviction score means.

What our conviction bands say about draw risk

This is the table we have not published before. Take every settled fixture in the record, including the ones where the model named nobody, bin them by conviction score, and count the draws.

The base rate across the whole record is 26%. Above 6.0 conviction, draws nearly vanish: 3 in 50 fixtures. Below 5.0, the draw rate climbs past a third, well above the base rate. In other words, the conviction score is not only grading how strong the named side's case is; it is, in practice, measuring the fixture's draw risk. A 7.0+ score has meant "this match will very probably have a winner, and the data says which one". A sub-5.0 score has meant "this might be anyone's, including nobody's".

The one 7.0+ fixture that did end level deserves its name on the record: Southampton 2-2 Bristol City, scored 7.1. Standard fence applies: this is one spring window of 192 fixtures, an observation and not a law, and we will re-run this exact table in public once the new season's record builds.

When the model misses, is it a draw or a defeat?

Of the 184 fixtures where the model named a side last spring, 80 calls were correct and 104 were not. The anatomy of those 104 misses: 47 draws and 57 outright defeats. Nearly half of everything the model got wrong was not the other side winning; it was football declining to pick a winner at all.

At the top of the scale the picture tightens. The model missed only 3 of its 22 calls at 7.0+, and just one was a draw. The other two were proper defeats, and both are worth remembering when a 7.0+ score reads like certainty: Mansfield 5-4 Cardiff, against a 7.2 on Cardiff, and Stockport 1-2 Port Vale, against a 7.0 on Stockport.

Mansfield, in fact, are the record's designated agent of chaos. Three of the model's fourteen highest-conviction misses were Mansfield wins: 5-4 at home to Cardiff, 1-0 away at Stockport against a 6.6, and 4-1 away at Huddersfield against a 6.2. The single time the model flagged Mansfield themselves, at 4.3, they drew. There is no deeper lesson in that beyond the one this whole page is about: the sport keeps a share of results, roughly a quarter of them, that no scoring system will ever own.

Does the model ever predict a draw?

No, by design. The model names the side with the stronger 8-factor case and says how strong that case is, or it stays quiet: in eight fixtures last spring, all in League One, it named nobody, and three of those eight ended level. A draw prediction would be a different product making a different promise. What the conviction score offers instead is the honest version of the same information: when the score is low, draw risk is high, and the model says so by refusing to be confident. How the 8 factors produce that score is set out in how the conviction score works, and the full settled record, misses included, lives on the accuracy page. Every new fixture is scored in the tool once team news and rest profiles are in.

All figures from Lower League Analyst's settled-fixture record, 14 April to 25 May 2026: 192 settled fixtures across the Championship, League One, League Two and National League, of which 184 carried a named predicted winner. Lower League Analyst is not a bookmaker and does not take bets. It is the research: the hard work done for you, the working shown, and the record published. No model can guarantee a result, and past performance is not a guide to future outcomes. 18+. Please gamble responsibly. begambleaware.org