Data / entry 13 / September 9, 2026 / 10 min read

What three games can and can't tell you about a title race

I run this thing every few weeks mostly to keep myself honest in arguments. Three rounds in, here is what 50,000 simulated seasons think, and more usefully, how little they are willing to change their mind so far.

Arsenal and Manchester City have played three games each and won all six between them. Nine points apiece, plus five goal difference apiece. City are top because they have scored seven and Arsenal six. That is the entire gap the table is willing to give you.

My model also has them close, but it does not have them equal, and the way it gets there is worth explaining. After round three Arsenal go from 51.7% to 52.8% title probability. City go from 37.5% to 37.8%. Both went up, partly because the rest of the league sorted itself out a little. But Arsenal picked up a full point and City picked up less than half of one, off identical records. That difference is not the model being cute. It is the model doing the only job I ask of it.

Why I bother with priors

You cannot fit anything useful on three games. Nine matches of goals is noise wearing a jersey. Coventry have shipped 4 in 3, which does not make them a bottom-five defense for the next eight months, and Hull have kept three clean sheets, which does not make them Europe-bound.

So I do not start from zero. Every team walks in with a prior on attack and defense, built from last season's goals for and against, and then the three games we have actually seen pull that prior around. That is all Bayes is in practice. You write down what you already believe, you write down how sure you are, and then you let the new stuff move you by exactly as much as it has earned.

Under the hood it is a hierarchical Poisson goals model, Dixon-Coles family. Attack rating and defense rating per team, log scale, centered on the league. Expected goals in a fixture come from home attack against away defense plus a home term, which fit at +0.22 in log space, call it a 25% bump, which is about where Premier League home advantage has always sat. I fit it in PyMC and then simulate the remaining fixtures 50,000 times, because I want a distribution, not a single number I can be wrong about with confidence. Returning teams get a moderately wide band around last year's numbers. Squads turn over roughly a third every summer, so last season is a hint, not testimony.

The promoted teams deserved better than a shrug

My first pass gave Coventry, Ipswich and Hull the same generic weak-team prior. I did not like it once I looked at it properly. Those three played 46 Championship games last season and told us very different things about themselves, and I was throwing all of it away.

Coventry won the division by 11 points, 97 scored, 45 conceded, best defense and second-best attack down there. Ipswich came second and were never really in trouble. Hull came sixth and got through the play-offs with a defense that was below average for the Championship, never mind this league. Three different teams. Now they get three different priors, built off that record and shrunk halfway toward a generic promoted baseline, because Championship dominance does not transfer cleanly upward but it is not nothing either.

Hull arrived as the promoted side I trusted least and leave round three with the lowest relegation risk of the three, below Ipswich and below Coventry. It is the same pattern you always get, belief moving fastest where it was thinnest. The difference from my first version is that this time the thinness was earned rather than assumed.

Where I disagree with the market

Once the promoted priors were honest I wanted an outside check, so I lined the model up against live market win probabilities for gameweek 4. Everything below is my home-win probability minus theirs, in points.

Nine of the ten land within about 8 to 13 points, which is roughly what I want to see. If I disagreed with a liquid market everywhere, the correct conclusion would be that I have a bug, not an edge. The exception is Chelsea against Hull. The market has Chelsea at 79.3% at home to a promoted side, which is the normal read. I have them at 55.0%, basically a coin flip that leans Chelsea.

I am not trying to be clever there. It falls straight out of the earlier decision: Hull have the best start of any promoted team and that landed on a prior with actual signal in it instead of a blanket discount. I genuinely do not know who is right, me or the market, and that is fine. Writing the disagreement down now is the only way to find out later. Tottenham against Everton is the other gap, 13 points toward Everton, which is not shocking given Spurs have not won yet.

The fan, and what it is actually saying

This is the 50,000 seasons unrolled week by week from where Arsenal and City sit right now, both on 9 points. Dark line is the median season, the bands are the middle half and the 10th to 90th range.

The separation is not one hot simulation, it is the shape of tens of thousands of them. I like the fan because it refuses to let me pretend I know a final points total. It says here is the range, here is where it leans, do with that what you like.

The title race

City's three: 2-1 over Bournemouth, 4-1 at Palace, 1-0 over Coventry. Perfectly good week to week. But the model does not grade results, it grades rates, and rate-wise those scorelines sit slightly under what last year's City numbers say they should be doing against that opposition. Arsenal's three (3-0 Coventry, 1-0 at Villa, 2-1 Chelsea) read as less flashy and grade out cleanly in line with the best defense in the league last season. Hence the point of title probability going one way and not the other.

Belief moves fastest where I knew least

Arsenal's defensive rating barely budges after three games because 38 games of history is a heavy anchor. Hull's moves a long way. Identical evidence, wildly different revision, purely because of how sure I was going in. That is the whole mechanism, visible in one chart.

The model learns fastest about what it was least sure of, which is also the exact situation every title-race model is in during September.

The full picture

What this thing does not know

Three games is still three games. There is nothing in here about injuries, fixture pile-ups, new signings settling, or a manager getting sacked in November. Newcastle and Chelsea both changed manager over the summer and the model only finds out about that if it shows up in the goals column. Home advantage is one league-wide number, which is almost certainly wrong ground by ground. And I am using raw goals, not xG, which would separate real chance creation from finishing luck. That is the upgrade I want most and have not made, mostly because the free xG sources I trust were not reliably reachable and I was not paying for an API to publish a blog post.

None of that changes what I actually take from this. The point of a Bayesian setup in September is not that it is right. It is that it knows, team by team, how unsure it ought to be, and it does that better when the priors underneath it came from real games instead of a guess.