Discover more from Cannon Stats
First Projections for the 2023-24 Premier League
Looking ahead to next season and looking back at the performance from last season.
the 2023-24 season is streaming towards us after what feels like an incredibly short summer break. With that comes a new round of doing weekly simulation updates for the Premier League starting today.
For this season I haven't made any changes to the model, last January I added in a small adjustment for net transfer spend that will decay as time progresses from the end of the window but that is really the only thing I have done for a while in the model. I remain generally happy with how it performs given the minimal inputs and lack of real granular data and I didn't see any big return on investment tweaks that I could have implemented in my limited time this summer. Later in this article, I will go through and do a little bit of a retrospective on how things performed last year.
Let's take a look at what we are dealing with to start the season:
Directionally I think that everything here is in the ballpark of reasonable and that makes me happy. It tracks fairly well with the betting odds which is another good check and I think matches my general starting points for most teams without any massive surprises.
Manchester City starts the season as they do seemingly every year now as the clear favorites to win the League. Arsenal start the season as the main chasing team, followed by Liverpool. One of the big departures from last season is that the teams in 3rd and 5th are being given more than a token chance of being in the title hunt this season.
This is down to the gap at the top shrinking.
Last season both Manchester United and Liverpool were head and shoulders clear of the teams chasing them, this year Manchester City are clearly the favorite but they also start the season with a lower rating than they started last season with (159 to 154). The gap between the top teams and bottom teams is continuing to grow with another year where the midtable is pretty barren, replaced by a knife fight in the relegation zone where a large number of teams could be dragged down.
I think the biggest ones that I am not sure about here are Chelsea (promising youth comes with big error bars), Brighton (replacing 2/3 of midfield is hard), Spurs (do they get to keep Harry Kane?), and Brentford (the model doesn’t know about Toney’s suspension).
Here is how the overall spread looks for each team by simulation (I do 10,000 of them for each run).
My model is putting Arsenal at an average of 81 points, the 75th percentile outcome is 87 points and the 90th percentile at 93 points. On the downside the 25th percentile is 76 points and the 10th percentile outcome is 69 points.
I am always curious how this differs from others so drop your prediction in the poll below (I think I am in the 83-87 range and feeling optimistic).
How the model performed last season
I always like to look back at how my model did this time last year too and like I said at the start of the post I was very happy.
A good place to start is with the points predicted for each team compared to the final tally.
My biggest miss was Chelsea, I was not expecting that team to have a total collapse from third place to almost getting dragged into the fringes of the relegation fight. This also makes them one of the more interesting teams for next season.
Newcastle was another big miss and I think that one is a little more understanding but with a model that potentially knew more could have had a smaller miss. My model didn’t know they had new owners and that the investment would be so good as to take a team that was one of the worst into third best.
Arsenal was another miss and I am not sure it should have been any wildly more optimistic here. The simulation was already baking in some improvement for the team and they just ended up in my mind having something close to a 90th percentile outcome. Maybe with hindsight for teams that have a young overall average age the season before that could be a bonus factor the next but I would need more research in helping to think through what the factor should be.
On a per-match simulation basis, my model seemed to do really well and it makes me very happy to see that.
Things tracked pretty well staying pretty close to having the actual outcomes match the predicted outcomes. The biggest weirdness came in the 75-80% bucket where I predicted this outcome 21 times and it happened 10 times. These were some of the biggest shock results of the season and I think my model has tended to really favor strong teams against weak teams (this was an area I had vastly underrated in the past).
I am not too worried about that and I think it will be more of a fluke than a long-term problem.
I also was in a bit of a contest last year with other people who publish these types of simulations and ended up with very similar results. My weekly updates had results not too far off the betting odds and my prediction from pre-season rated pretty well coming in 5th of 9 people (still would have liked to do better).
Anyway, I think my takeaway is that this model does a decent job and that something pretty simple can still be pretty powerful, it isn't going to go win me money gambling based on it (it is not designed to either) but I think it does a fine job giving an idea for where the season might go and the general magnitude of the odds.