Which football stats actually predict a result? What 182 settled fixtures say about 8 factors

Which football stats actually predict a result? Across 182 settled lower league fixtures, league position separated the correct calls from the misses by the widest margin: it averaged 5.8 on calls that landed and 3.6 on calls that did not. Head-to-head history separated them by 0.15, which is close to nothing at all.

Ask which numbers predict football results and the answer is fairly settled: recent form, expected goals, home advantage, head-to-head history, injuries. That list is not wrong. Our own model uses a version of it. What the list never tells you is which of those things is doing the work once you check it against what actually happened.

How do you test whether a football stat predicts anything?

Our model scores every fixture in the Championship, League One, League Two and the National League across 8 factors, each on a 0 to 10 scale, then combines them into a single conviction score. The scores are stored per fixture, so once the result arrives we can go back and read the factors that produced the call.

The window here is our settled record from the 2025/26 run-in: 192 fixtures between 14 April and 25 May 2026. The model named a side in 184 of them. Of those, 182 carry a full stored factor breakdown. 78 of the calls landed, 104 did not.

That last sentence is the whole method. Looking only at the calls that landed tells you nothing, because a trait shared by the misses is not a winning trait. The comparison is the test.

Which factor separates a correct call from a miss?

Average factor score across the 182 fixtures, split by whether the call landed:

Two caveats sit inside that table. Shot data and xG was present in only 110 of the 182 fixtures, and injuries and suspensions in 113, so those two rows compare a smaller and slightly different set of games. Away record is left out entirely: the model favoured an away side just 8 times in the window, which is too few to read anything into.

What does a league position score of 8 or more tell you?

Enough to sort the record on its own. Taking the league position factor by itself and banding every one of the 182 fixtures by that score alone:

• League position 8.0 or higher: 23 of 27 calls landed (85%)
• 6.0 to 7.9: 14 of 22 (64%)
• 4.0 to 5.9: 19 of 46 (41%)
• Below 4.0: 22 of 87 (25%)

That is a clean staircase from one factor. The gap holds in all four divisions too, at 2.0 points in the Championship, 3.5 in League One, 1.0 in League Two and 2.4 in the National League.

Does head-to-head history predict anything?

Head-to-head is the number supporters reach for first, and in this window it was the one that told us least. The same banding exercise on the head-to-head factor alone:

• Head-to-head 8.0 or higher: 7 of 14 (50%)
• 6.0 to 7.9: 21 of 44 (48%)
• 4.0 to 5.9: 20 of 55 (36%)
• Below 4.0: 26 of 62 (42%)

No ladder. The bottom band outperformed the band above it. Whatever information is in two clubs recent meetings, this record could not find it.

Do the model weights match what the record says?

Partly, and the part that does not match is worth saying out loud.

The 8 factors do not carry equal weight. Recent form, fixture difficulty and home record sit around 0.17 each, shot data and xG around 0.15, goals and clean sheets and league position around 0.13, head-to-head around 0.09 and injuries around 0.08. Weights renormalise when a factor is missing, which is why a fixture scored on five factors is still scored on a 0 to 10 scale.

On the vindicated side: the two lightest weights in the model are on head-to-head and injuries, and those are exactly the two factors that separated correct calls from misses least. The model was already discounting them, and the record says it was right to.

On the other side: league position carries the sixth-heaviest weight of eight and separated the most, by a distance. On this window it looks under-weighted. That is not a comfortable thing to publish about your own model, but a page that only reported the flattering half would not be worth reading.

What this does not prove

Six weeks of one run-in is a small window, and run-in fixtures are strange ones. Motivation is skewed by then, some clubs have nothing left to play for, and league position is partly a proxy for the same separation the conviction score is trying to measure, so a share of that 2.23 gap is baked in by construction rather than discovered. The xG and injuries rows rest on fewer fixtures than the others. And a factor that separates poorly on average can still be pulling its weight at the margins of individual games.

So this is an observation about 182 fixtures, not a law of football. The honest next step is to run the same comparison again on the 2026/27 record, with a full season and a fresh set of results, and see which of these gaps survives. We will publish that whichever way it goes.

If you want the factor definitions and a worked example of a single fixture scored end to end, that is in how the 8-factor conviction score works. For the other side of the same record, on the result these factors handle worst, see why draws are the hardest result to predict. Every fixture in all four divisions is scored across these 8 factors at Lower League Analyst, with the working shown on each one.

Factor figures from Lower League Analyst settled-fixture record, 14 April to 25 May 2026: 192 settled fixtures, 184 named calls, 182 with a stored factor breakdown. 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

Updated 10 August 2026: the head-to-head separation is corrected to 0.15, and the head-to-head banding is stated over the 175 fixtures that carry a head-to-head score. An earlier version read 0.14, and described that banding as covering all 182.