By Oliver Fitzpatrick
Sorry, not that kind of model...
It’s that time of year again, a new AFL men’s season. A fresh start. 18 optimistic teams and millions of fans still hopeful that their team could be lifting the cup on that one day in September.
The pantomime drama of the trade and draft periods are over and the time has come for every expert worth their salt to give their take on who are the haves and have nots of the AFL in 2019.
With that in mind, I thought I would do something original and preview the AFL season… {insert laughter here} …although rather than go with what my gut that tells me — that Richmond will be wooden spooners like I predicted in 2017 — I will get a model to do the work for me and then, if it goes horribly wrong I can blame the model rather than my own lack of footy nous.
The Model Predicts the 2019 AFL Season
The model I will be using is based on the Elo method, which is a common rating system used for modelling sports, initially created to rate chess players. The model was trained on AFL matches from 2012–17 and tested last year with quite good success — getting 152 tips correct @ 77%.
The basic premise of any Elo model is that it gives each team a rating and then updates the rating after each match is played based on how the team performed compared to their expectations. My model gives each team three ratings: one for when they play at their home ground, one for when they play away from home and one for when they play at their ‘2nd’ venue (such as Hawthorn in Tasmania). The three ratings interact with each other and change after each match to different degrees depending on where the match was played. The model also considers the number of days break between matches compared to the opposition.
To get initial ratings for each season, each team’s list demographics and changes from the previous year are considered and a random forest modelwas created to predict the number of wins. Information such as average age and average matches are used in the random forest model to predict how a team will perform that year. This prediction is then used in the Elo model to give each team a starting point for each of their ratings.
This means that the predictions before Round 1 are based purely on a team’s list demographics, changes in their list from the previous year and on their fixture. Bearing that in mind, it is likely for early season predictions to be a little tentative and the predicted ladder is likely to change depending on which teams perform well early in the season. These were the results from last year’s predictions just using the random forest model on list demographics:
The model did a surprisingly good job, with Carlton, Brisbane and Richmond proving to be the worst predicted with three wins more or less than predicted. The ladder positions given are also impressive, with the model getting seven of the finals teams correct, only needing to swap GWS in for Adelaide to get eight from eight — and these two teams’ predicted wins were separated by just 0.03 wins.
In fact, in the seven years of data analysed, the model has been found to have a mean absolute error of 1.44, meaning on average the model will get each prediction wrong by that amount in either direction. This shows that a team’s list demographics and changes from the previous year are highly informative as to how they should perform in the upcoming season and can be used with reasonable accuracy. This is why they are so useful to add as a starting point for the Elo model, instead of just using a team’s performance from the previous year.
Using this random forest model, and adding it to the Elo model, the following predictions are made for 2019 after doing 10,000 simulations of the season (remember that the Elo model takes into account the fixture and changes in performance over a season).
The predicted wins are very similar for each model because at this stage the Elo model relies on the list demographics to give each team a rating and have no performances to adjust these ratings from. As the season goes on however, these two predictions will diverge.
Unsurprisingly, the models don’t predict much change in the ladder from last year, with only Essendon and North Melbourne predicted to jump into the Top 8 at the expense of Hawthorn and Geelong. A more detailed look at predicted ladder positions can be seen below.
At this early stage of the season, the model remains open to possibilities with all teams except the Gold Coast, St Kilda and Carlton given more than a 10% chance at making the finals. Geelong, Hawthorn and Adelaide appear to be right on the cusp of finals and all teams above them are given better than a 50–50 chance of making the finals.
As many experts have tipped, Essendon are seen to be the most likely team to make a big improvement due to their off-field recruiting and increase in age and experience of their list.
A more surprising predicted riser are North Melbourne, who often go under the radar and were expected to enter a rebuilding phase after culling so much experience from their list a few years ago. Clearly, their list is in good shape however, and the model expects them to be challenging for a Top 4 position, which would be a surprise to most AFL pundits. They will also likely helped by a softer fixture, with the Elo model giving them more wins than the random forest model.
Collingwood fans are unlikely to be happy — after all, they were within a kick of the flag last year — and while most pundits expect the ‘Pies to be challenging again this year, the model rates them as the eighth most likely to win the flag. There is however, very little between them and third placed Sydney.
At the other end of the table, Gold Coast, Carlton and St Kilda are clearly the weakest teams in the league, and it would take a miracle for them to challenge for a spot in the Top 8. Furthermore, if Port Adelaide and St Kilda perform as the model expects then Ken Hinkley and Alan Richardson will be under enormous pressure to keep their jobs.
How unfair is your fixture really…?
Another thing the Elo model allows us to do is to look at each team’s fixture and rate how easy or difficult it is. The AFL is often criticised for their fixture as some teams play weaker teams twice or have to travel more or less than other teams.
The Elo model can see how distinct the bias is predicted to be this season and which teams are at an advantage or disadvantage. The analysis is done by looking at how an average team would perform with each team’s fixture and gives the average win probability for each game of their season.
The fixture comparison confirms that North Melbourne do have the easiest draw on paper, although clearly the difference between the easiest and hardest fixture is minimal. Melbourne and Collingwood’s tough fixture is another reason why the Elo model doesn’t consider them among the top challengers to win the flag despite most experts expecting both these teams to be challenging this year.
While some teams have harder fixtures than others, the fixture analysis shows that these differences are quite small. The difference of 49.6% and 50.3% for Melbourne and North Melbourne equates to 0.16 expected wins over the course of the entire season. The small advantage afforded to teams like North Melbourne certainly could, in some seasons, be the difference between playing finals and not, but it is very much a marginal difference and certainly not the sole reason a team would make the finals or miss them.
The 22-game home and away season perhaps doesn’t necessarily lead to an uneven fixture, and while there are other reasons to change the length of the season, there isn’t a strong argument that the fixture favours certain teams to a large degree.
Why bother following the footy if we know what’s going to happen?
While the model should do quite a good job at predicting the season, it doesn’t mean we can be certain about what will happen in 2019.
In the 10,000 simulations, Gold Coast finished on top in one of those simulated seasons and so this shows that even the model agrees that anything is possible and that all fans can remain optimistic… for now.
Oliver Fitzpatrick is a cricket tragic and Carlton FC (even more) tragic and can usually be found in the Members Pavilion at the MCG. When he’s not at the ‘G he’s studying a Masters of Statistics and Operations Research at RMIT and tweeting his thoughts @ocfitz1.
