Every Friday night, a made-up fan called Dan sends the same message to a WhatsApp group of nine old school friends: “Predictions in by 12, no excuses.” It’s a private league with a fiver each and bragging rights until May. Dan has finished second three years running, which he blames on Ollie, a friend who never watches a match and somehow always wins.
This season Dan tried something new. Before filling in his picks, he started checking AI football predictions, which give each fixture a home-win, draw and away-win percentage plus a likely score. He figured the machine would finally beat Ollie.
By October he’d learned a lot about why Europe’s leagues are so hard to call. He was also still second.
The week his “banker” lost
Dan’s first big lesson came in week five. The model had the home side at about 70% to win. He put them down as his banker, a pick he considered nearly certain.
They lost 2–1. Their first-choice centre-back had picked up a knock in midweek Europa League, and the 20-year-old who came in looked terrified for the whole first half.
That’s squad depth, and it decides more seasons than people expect. Teams playing in Europe can face two games a week for months, so rotation isn’t optional. Clubs with quality on the bench barely notice when a regular is out. Thin squads can collapse when three are out at once.
Money helps clubs buy backups, though sharp scouting and a good academy can go a long way too. The model knew the centre-back was missing. What it couldn’t judge was how nervous his replacement would be.
A team can keep the ball all game and still lose
Ollie, it turns out, has a theory. He backs any team facing a side that “passes it sideways forever.”
He’s not entirely wrong. European football mixes very different styles, and some of them clash badly. A possession team can keep the ball for 70% of a match and still lose to opponents who sit deep and break quickly. A high-pressing side can make life miserable for a team that’s nervous playing out from the back.
Managers keep adjusting, too. A shape that worked last week might leave holes against this week’s opponent, and a half-time switch can turn a game around. Models can learn which styles tend to trouble which teams. They have a much harder time predicting a manager who rips up the plan at 1–0 down.
Thirty-eight games is a long time
Any decent team can win a big match. Keeping it up from August to May is far harder.
Even the best clubs hit bad spells. Two draws and a strange defeat in February can change a title race, particularly when the top three are separated by a couple of points. The clubs that last tend to defend well, have attacking patterns they can repeat, and keep their heads when results go wrong.
Then there’s the rest of the table. A side fighting relegation plays a home game against a title chaser very differently from a mid-table team with nothing left to play for. Dan’s worst weekends usually come in April, when motivation matters more than form.
Long runs of dominance happen as well. Bayern Munich won 11 Bundesliga titles in a row, and then Bayer Leverkusen went the whole 2023–24 league season unbeaten. A model fed on Bayern’s decade would have given Leverkusen very little chance that August.
The number Dan started watching
The stat that changed how Dan reads the table is expected goals, or xG. It estimates how likely each shot is to go in, based on things like distance and angle and whether it came from a header or a clean strike.
Over a couple of months, xG shows whether a team keeps creating good chances or is surviving on a few wonder goals. Dan’s own club sat 12th in October with one of the better xG figures in the league. They were creating plenty and missing sitters. By December they’d climbed to seventh, which xG had more or less hinted at.
It works the other way too. A team near the top with weak underlying numbers may be riding luck that runs out. Shots conceded and set-piece record add to the picture, depending on how a side plays.
So how do the models actually work?
Most prediction systems mix inputs like recent form, home advantage and head-to-head records with goal data, xG and known injuries. Better services refresh their numbers once lineups are confirmed, and that’s often when the most useful information arrives.
What comes out is a set of probabilities, such as 55% home win, 25% draw and 20% away win. That 20% happens one game in five, and Dan has learned that the hard way. It’s the same problem people hit when they try to pick stocks with ChatGPT. A confident estimate sounds like a promise, but it’s only ever a likelihood.
Dan also tried asking a general chatbot for picks one week. It recommended a player who had transferred abroad two months earlier. Unless they’re connected to live data, these tools can be badly out of date, and some can’t even work out today’s date, which is a well-documented reason AI chatbots can’t tell the time. Purpose-built football models that pull fresh team news are a far better starting point.
His group plays for a fiver, but some people bet real money on these numbers. If that’s you, set limits and check your local rules before you start.
Where the money stops mattering
Big budgets clearly help. Rich clubs can afford proven stars, better training grounds and deep squads that cope with injuries. Over a season that adds up.
Expensive squads still fall apart surprisingly often, though. A string of bad signings or a manager who’s sacked in November can undo the budget advantage. Smaller clubs, meanwhile, build around strong academies and sharp scouting, and coach a defensive shape so well-drilled it drives better teams mad. Leicester began 2015–16 as 5,000-1 outsiders and finished as Premier League champions, and every underdog has been quoting them since.
Watching more than one league
Dan used to watch only the Premier League. Now he dips into a few other popular football leagues most weekends, partly to sharpen his picks and partly because a 3–3 in the Eredivisie on a Sunday afternoon is great fun.
Each league has its own personality. Some are tactically cautious, some are wide open, and some have a couple of giants with a brutal fight beneath them. Following a few shows him how differently clubs recruit and how managers adapt, and he spots good young players a season before the big clubs come calling.
It’s also made him more sceptical about predictions. Once you’ve seen a model get surprised in three countries in the same weekend, you stop calling anything a banker.
Ollie, for the record, still hasn’t opened a single prediction page, and he’s top of the group again.
Related: How AI Is Secretly Powering the FIFA World Cup 2026 Fan Experience
