Premier League 2026-27 Predictions, Built with Maia

Arsenal are 63% favorites to win the Premier League. Man City sit at 31%. Between them, that's a 95% share of the title, and the gap comes down to one number: Arsenal concede 29% fewer goals than City do.

That's the headline from a model I built by talking to Maia, Matillion's AI data automation platform. Ten thousand simulated seasons, opponent-adjusted team ratings, and enough historical relegation data to explain why three newly promoted clubs are about to have a very bad year.

I've written about this kind of project before, when Maia predicted the 2026 World Cup. Same approach here, pointed at a full 38-game domestic season instead of a knockout tournament, plus the Championship promotion race underneath it.

Arsenal Win It on Defense, Not Attack

City actually have the better attack. Their expected output is 1.67 goals a game against Arsenal's 1.55, roughly 8% more firepower. If you only looked at that number, you'd back City.

But Arsenal's defense is in a different league entirely. Their rating of 0.39 against City's 0.55 means they're conceding 29% fewer goals, and that gap is what decides the title in the model's simulations. Arsenal finish first or second in 90% of the 10,000 runs. City manage that in 79%. Neither side misses the top four in more than 3% of scenarios.

City's squad losses turned out to matter less than I expected going in. Rodri left for Barcelona, Bernardo Silva moved on, and City spent big to cover the gap: Semenyo, Cherki, Marmoush, and Reijnders arrive with a combined FPL price north of £29m. That reinvestment keeps City the most expensive squad in the league by a wide margin. It just hasn't fixed the defense.

Liverpool sit a clear third, but 15 points off the pace, with a defensive rating (0.75) nearly double Arsenal's. Man United and Chelsea are fighting over the last Champions League spot, separated by a single predicted point, which the model treats as close to a coin toss.

One week in, though, there are already reasons not to take the table too literally. Manchester United were the model's fourth-placed team, but opened with a 2-0 defeat at newly promoted Hull. That's exactly the sort of result a 10,000-season simulation can accommodate but a neat pre-season table cannot explain away.

City have also already had a small reality check of their own. They scraped past Bournemouth 2-1 after falling behind, needing two late goals to take the points. It does not change the model's view of the title race, but it is a useful reminder that an 83-point projection is a probability-weighted forecast, not a script for every Saturday afternoon.

The Pack Fighting for Europe

Six points cover six clubs from 6th to 12th: Brighton, Aston Villa, Brentford, Newcastle, Nottingham Forest, and Tottenham. The model gives Brighton the edge at 58 points, but don't read too much into the exact order here. Any of the six could land anywhere in that range.

Newcastle are the more interesting story than their 9th-place projection suggests. They lost Bruno Guimarães to Arsenal, Tonali to AC Milan, and Gordon to Liverpool, roughly £215m of talent walking out the door in one window. Their squad value swung from a strong positive rating to firmly negative, and the model has them at just a 5% chance of top four as a result. That's a squad that got visibly worse on paper, and the numbers picked it up cleanly.

Everton, by contrast, are the model's definition of a nothing team this year. Almost no European chance, almost no relegation risk, most likely finishing somewhere between 12th and 17th and staying there.

The opening round is already a good illustration of why the model is more comfortable with the group than the exact ordering. Brighton beat Aston Villa 4-0, while Brentford beat Tottenham 3-0. Both results fit within the model's broad view of a volatile European pack, but neither was the sort of result you'd infer from simply reading down the predicted table.

Why the Promoted Clubs Are Almost Certainly Going Down

This is where the model gets blunt. Ipswich, Coventry, and Hull are all predicted to go straight back down, and the gap between survival and relegation isn't close. Sunderland finish 17th on 46 points; Ipswich, in 18th, sit on 26. That's a 20-point cliff between the last safe spot and the first relegation spot, wider than any promoted cohort the model has backtested.

The first weekend has already made this section more interesting. Hull beat Manchester United 2-0, while Ipswich beat Sunderland 2-1. Coventry lost 3-0 to Arsenal, so the promoted trio certainly haven't followed the script uniformly. Two of the three started with wins. One round is nowhere near enough to overturn the historical case, but it is a useful reminder of just how much uncertainty sits underneath that 99% relegation figure.

The model measures this using 27 teams promoted over the last 10 seasons, and the pattern holds almost without exception. On average, a promoted side's attacking output drops to about 54% of its Championship level once it's adjusted for league quality, while it concedes at 181% of its old rate. Only one team in that entire sample, Leeds in 2020-21, came close to keeping its attacking numbers intact, and even that side still conceded at 183% of its previous rate.

Each of this year's three has its own version of the same problem. Hull came up through the play-offs with the weakest points total and goal difference of the three, and their Championship attack rating (0.76 against a league average of 1.0) was already below par before facing Premier League defenses. Coventry had a brilliant Championship campaign, 95 points and a +52 goal difference, but recent history says that doesn't travel: Leicester won the Championship with 97 points in 2023-24 and were relegated on 25; Ipswich won it with 96 and went down on 22. Ipswich this time round get the best survival odds of the three, at 10%, mostly because their season was more balanced rather than spectacular.

The one thing worth flagging as a live risk to the model itself: history says there's roughly a 40% chance an established club goes down alongside the promoted trio. Leicester dropped from 8th to relegated in a single season not long ago. The model gives Sunderland the highest exposure among the established 17, at 3.2%, largely because their attack (0.80) is the weakest of any non-promoted side and they're juggling Europa League football on top of survival.

The Championship Looks Like a Straight Bounce-Back

One level down, the three clubs relegated from the Premier League last season, West Ham, Wolves, and Burnley, are strong favorites to go straight back up.

West Ham win the division outright in 43% of simulations on the back of the biggest squad in the Championship. There's still roughly a one-in-four chance at least one of the relegated trio misses the top six, and if that happens, Burnley are the most likely to be the one that slips, with Southampton and Middlesbrough best placed to take advantage.

How the Model Actually Works

The build follows the same shape as the World Cup project: team ratings from an opponent-adjusted Poisson model, fitted on 10 seasons of match history, then run forward 10,000 times per league to generate a full table of outcomes rather than a single guess.

A few things feed into the ratings beyond raw results. Squad market value, weighted through Fantasy Premier League pricing, captures the summer's transfer activity, which is how Newcastle's sell-off and City's reinvestment both show up correctly. For the promoted and relegated clubs, the model applies a quality-gap adjustment learned from those 27 historical promotions, converting a team's Championship form into a realistic Premier League equivalent instead of assuming last season's numbers carry over untouched.

I asked Maia to backtest the whole thing before trusting any of it. Across 8 Premier League seasons, the model's average position error came out to 3.01 places, and it called the champion correctly in 5 of 8. It's strongest at the top and bottom of the table and weaker in the mid-table scramble, which tracks with how tightly bunched that group of six clubs is this year. The Championship backtest, over 2 seasons, ran to a wider error of 5.6 to 6.1 places, which makes sense given a 24-team division with more variance and less data per club.

Where I'd Trust This, and Where I Wouldn't

High confidence, but not certainty: Arsenal and City occupying the top two. The model has never missed that pairing across 8 backtested seasons. Liverpool third. The promoted clubs remain the model's strongest relegation call, but Hull and Ipswich winning on opening weekend are an early reminder that 99% in a simulation isn't the same thing as 99% certainty in football.

Arsenal, at least, have given the model an encouraging opening weekend. They beat newly promoted Coventry 3-0, while Coventry became the only one of the three promoted clubs to lose. That does not prove the forecast right, but it is a useful reminder that the opening round has cut both ways.

City's opening weekend was less convincing. They scraped past Bournemouth 2-1 after falling behind, needing two late goals to take the points. Again, that is not evidence against the 31% title probability in any meaningful statistical sense. It is simply the sort of result that makes a pre-season forecast worth revisiting rather than worshipping.

Moderate confidence: Arsenal over City for the title itself. A 63/31 split is a real edge, not a coin flip, but if Erling Haaland stays healthy and Arsenal's defense drifts even slightly, that number moves. Man United over Chelsea for fourth is close enough to call it a genuine toss-up.

Where I'd be least surprised to be wrong: the 20-point gap between 17th and 18th place is more extreme than anything in the backtested history, and every previous promoted trio has had at least one team clear 34 points. If the opening weekend is any indication, that part of the forecast deserves watching closely.

That's perhaps the biggest lesson from the first weekend. The model can tell us what happens most often across 10,000 simulated seasons. It cannot tell us which of those seasons we're actually living through.

None of this took long to build. That's the actual point, same as it was with the World Cup model. The interesting part of a project like this stops being whether I can get the data and code working, and starts being how many questions I have time to ask.