The Model Called Betis Right Against Real Madrid

Published 5 September 2026 · it called Ipswich wrong on the same afternoon · figures read from the log the same day

On Friday, before either match kicked off, our model published two three-way splits. Real Betis at home to Real Madrid: 36% Betis, 29% draw, 35% Real Madrid. Ipswich at home to Liverpool: 40% Ipswich, 28% draw, 32% Liverpool.

Betis won 1-0. Liverpool won 2-0.

We are writing about both in the same article, on purpose. Publishing the first one on its own is how a track record becomes a highlight reel.

Thirty-six against thirty-five is not a prediction

Start with what the Betis number actually was. The model put the home side one point in front of the away side, with twenty-nine points of draw sitting between them. That is not a call. That is the model saying, in the only vocabulary it has, that it could not separate these two teams — and then leaning, barely, towards the one at home.

The expected goals underneath say it more plainly still: 1.42 for Betis, 1.38 for Real Madrid. Four hundredths of a goal.

So we are not going to tell you we predicted this result. What we will say, because we think it is the part worth something, is that the model was willing to put a side other than Real Madrid fractionally in front, in public, with a timestamp, before kickoff — and to publish how thin the lean was instead of rounding it into a headline. A model that always defaults to the bigger name is not a model. It is a table of reputations with a percentage sign after it.

The result went the way of the lean. On its own that is worth close to nothing, and the rest of this article is about why.

The match agreed with the number

A model that publishes 36 / 29 / 35 is making one claim above all others: this will be close. It was — far closer than a 1-0 makes it look, and the scoreline hides all of it.

There was no goal at all until the 81st minute, when Parrott, on as a substitute twelve minutes earlier, scored the only one of the match. Real Madrid then had the ball in the Betis net in the 87th, through the substitute Espí, and VAR took it away for an offside in the buildup — not the finish, a phase earlier. Four minutes into stoppage time Kylian Mbappé had a penalty to level it, and missed.

So: ninety minutes, one goal, and two separate moments in the last ten that would each have made this article a post-mortem instead. We did not watch a favourite justify itself. We watched a coin land, having said beforehand that it was a coin.

One correction to our own record

We called the outcome. We did not call the scoreline: our most likely score was 2-1 and it finished 1-0. Right answer, wrong route.

That is the ordinary case rather than a bad day — across the whole log we have the exact scoreline 152 times in 1,720, about nine in a hundred — but we would rather write it down than let "we called it" do work it has not earned.

And the one we got wrong

Ipswich against Liverpool was the more confident of the two calls. The model gave Ipswich 40% at home, four points above what it gave Betis, on expected goals of 1.76 against Liverpool's 1.45. Liverpool won 2-0, and 2-0 was not even the scoreline we thought most likely — we had 2-1.

Two calls, same model, same afternoon, four points apart in confidence. One landed and one did not. Neither of those facts tells you anything at all about whether the model is any good, and anyone who presents you the first without the second is selling you something.

What does tell you something

The question worth asking is not whether we got a match right. It is whether a number means what it says. When we publish 36%, do things like that happen about thirty-six times in a hundred?

That is answerable, and we answer it publicly on the track record page, which groups every settled prediction into confidence bands and prints what the model said next to what actually happened.

The Betis 36% falls in the 30–40% band. That band holds 931 settled predictions. Across them the model's average confidence is 33.9%, and the outcome came in 35.1% of the time.

The Ipswich 40% falls in the band above, 40–50%: 700 settled predictions, average confidence 44.1%, outcomes landing 42.1% of the time.

Both of those are close, and close is the whole ambition. It means that in the range where most of our football lives — the genuinely uncertain matches, which is nearly all of them — the numbers we print are roughly the numbers you get. That is a far more useful property than a good week.

Our headline hit rate is 808 correct results out of 1,720 settled predictions, 47%. Nobody is going to be impressed by that at a glance and we are not going to dress it up. Three-way football outcomes are hard, and most of the matches we cover look exactly like Friday: close enough that the honest answer is a probability, not a name.

Where the same page says we are wrong

The bands are not uniformly kind to us, and this is the one to read if you only read one.

In our most confident band, 65% and above, the model's average confidence across 273 settled predictions is 74.5%. Those outcomes land 59.7% of the time. When our model is sure, it is roughly fifteen points too sure.

We publish that on the same page, in the same table, in the same size type as everything else. It is not a footnote we might forget to add: a track record that only flatters the model is worth less than no track record at all, and the fastest way to find out whether a prediction site believes that is to look for its worst number.

What we cannot tell you yet

Not everything on this site is old enough to be worth reading. Our club ranking carries 347 rows across 327 clubs, and a club has to reach five settled predictions before it appears — because at five the margin on a hit rate is already about twenty-two points, and below that a percentage is decoration.

None of the four clubs in this article has got there. The domestic seasons are four matches old: we have four settled predictions each for Betis and Real Madrid, three each for Ipswich and Liverpool. Whatever we think of how the model reads Real Madrid, we have not earned the right to publish it, so we have not.

A competition gathers evidence about ten times faster, and La Liga has cleared the bar: 32 settled predictions, 50.0% correct, fifteenth of the 32 competitions on our competition ranking. The English top flight has not, yet.

One thing worth knowing about that club table, since Friday's fixtures are about to become European ones: a club's domestic season and its European ties are ranked separately there. They are not the same test. A league is the same twenty sides over and over, on form we have watched all season; a European draw throws two leagues together on very little directly comparable evidence, and averaging the two into one figure hides which is which.

Why any of this

Give the offside flag a different angle in the 87th minute, or move Mbappé's penalty half a metre to the left, and this article is about a model that got it wrong — and every number in it would be exactly the same. That is the uncomfortable part, and it is the honest one. The 36% was never a promise. It was a statement that this fixture would go Betis's way slightly more often than not across many versions of it, and Friday happened to be one of those versions. That is not a caveat we are adding now that we know the result. It is what the figure meant when we published it, at a time when we did not.

Unpredictability is the reason anyone watches football. Measuring it honestly is what this site is for — including on the days it measures us right back.

You can read how the model works, including the places we already know it struggles, and check every claim above on the track record. Tell us where we are wrong at [email protected].

We do not take bookmaker sponsorship or gambling-affiliate advertising, and none of this is betting advice. It is a statistical estimate published for its own sake.

Every probability, expected-goals figure and track-record number here was read from our own prediction log on 5 September 2026; the events of the two matches are the matches' own, watched and checked against the official timeline. The log grows daily, so the figures on /accuracy will have moved by the time you read this; where they disagree with this article, that page is right. Read more about why we publish any of this.