AFL 2023 Season Preview

2023 Initial Ratings and Model Explaination

The 2023 footy season is upon us and it's time for wannabe modellers to break out their coding skills and put their brain and football knowledge to the test.

I have refined and improved my previous models for sports like rugby, baseball and tennis to come up with what I believe to be an accurate and realistic model for the AFL season.

My model works mainly from elo as its base rating system but includes variables like a fluid home ground advantage multiplier, expected score (or xScore) and previous head to head results to skew the probabilities one way or the other.

The Elo Rating System

Elo works on a couple main principles. Firstly is that it is a zero sum rating system, which means whatever points you add to one team you must take the same number of points away from the other team. 

Secondly there is no "perfect" rating for each team as it is fluid and changes after every game. 

For this reason two teams may play 10 matches against each other one after the other and if one team wins the first 3 in a row by chance, elo thinks they are more likely to win again, even if the two teams are perfectly matched. This is the "fog of war" of rating a seemingly random sporting event.

To begin the 2023 season the base ratings for each team are as follows (1500 is average):

Adelaide: 1394.598696

Brisbane: 1622.062897

Carlton: 1503.239146

Collingwood: 1607.779784

Essendon: 1397.882242

Fremantle: 1605.159967

Geelong: 1761.015706

Gold Coast: 1491.946616

GWS: 1380.363269

Hawthorn: 1395.800915

Melbourne: 1613.031611

North Melbourne: 1197.275315

Port Adelaide: 1521.628644

Richmond: 1597.931034

St Kilda: 1482.133823

Sydney: 1680.221747

West Coast: 1198.581527

Western Bulldogs: 1549.347059

These ratings are not too out of the ordinary as we think back to last year where North Melbourne and West Coast were by far the worst clubs. Geelong won the Grand Final by 200 goals, have only got better in the offseason and rightly are the only club in the 1700s for rating.

As a base to these ratings and to get a simple probablilty for each game between two teams, you must use the formula:


It is not easy however to say who should win each game based just on these ratings which is why my model takes this as a start point but does many more calculations to the numbers to come up with a (hopefully) more representitive outcome for each match.

Similarly you cannot take the top 8 elo ratings and say they will be the finalists for 2023 as the ratings will swing and change through the year and some schedules will be harder earlier or later depending on what our overlords at AFL House feel like doing on any given year.

2023 AFL Predicted Results

After running 10,000 simulations on the AFL season we can say one thing for certain, nothing is certain.

Every team finished at least one simulation in every position on the ladder. Of course some were less likely to happen than others, for example North Melbourne finished first 64 times but finished 18th 2,606 times.


The raw numbers are above and I will highlight some key ones of note.

Geelong are the standout, winning the minor premiership 24.2% of all seasons, finishing top four 57.4% of the time.

Other hopes for the minor premiership are Sydney (12%), Brisbane (8.3%), Melbourne, (8.2%), Collingwood (7.7%), Richmond (7.1%) and Fremantle (6.8%)

North Melbourne are slightly more likely to win the wooden spoon with a 26.1% chance compared to West Coast's 21.2% chance. GWS, Hawthorn and Adelaide are thereabouts on 7% and Essendon is the only other club above 5% with a 5.9% chance.

Looking at the top four now and it looks to be Geelong (57.4%) and Sydney (41.3%) with the best hopes, the only ones above 40% with Brisbane, Melbourne, Collingwood, Richmond and Fremantle all within 1.5% of a 30% chance at the double chance spots.

The same culprits as above are in the hunt for finals albeit with an obviously much higher percentage chance at claiming one of the eight spots. 

Those around the 50/50 mark for the top eight include Carlton (43.5%), Fremantle (55.6%), Port Adelaide (48%) and the Western Bulldogs (49.5%).

2023 Final Ladder Prediction

1 - Geelong 14.3 wins
2 - Sydney 13.3 wins
3 - Brisbane 12.6 wins
4 - Melbourne 12.6 wins
5 - Collingwood 12.5 wins
6 - Richmond 12.5 wins
7 - Fremantle 12.4 wins
8 - Western Bulldogs 11.8 wins
9 - Port Adelaide 11.7 wins
10 - Carlton 11.5 wins
11 - Gold Coast 11.2 wins
12 - St Kilda 10.8 wins
13 - Essendon 10.5 wins
14 - Adelaide 10.5 wins
15 - Hawthorn 10.4 wins
16 - GWS 10.2 wins
17 - West Coast 8.6 wins
18 - North Melbourne 8.5 wins

How Good is the Model?

Throughout the year I will post tips and reviews of them to compare my model to those elsewhere such as squiggle.com.au and on betting sites with the aim of being good enough to join the Squiggle ranks in 2024.

Model rating will happen in a few different ways; one is by bits, a score based on how confident you are in a prediction, punishing overconfidence but potentially rewarding risk. 

MAE or Mean Absolute Error which is the mean average of all your tipped margins compared to actual results. For example if I tip Geelong by 15 and they win by 5 my MAE for that game is 10, if their opponent won by 15 my MAE would be 30. A MAE of under 25 is excellent, 30 is good and anything above is not idea.

And lastly good old fashioned tips where I will pick one team to win and record a point or no points depending if they won or lost.

Follow along here or on my twitter @zoonmattau where I will be tweeting footy nonsense all season long. Good luck if you are a fellow modeller or tipster!

Comments

Post a Comment

Popular posts from this blog

Model Performance - Round 3

How Did the Model Perform? Round three threw up quite a few unexpected results, with somewhere between 4 and 5 underdogs getting the victory, depending on which source you use. For the PPP model  it was the best round so far to date in terms of tips and MAE with a -0.11 bit round being second only to round 1's -0.07 bits.  It is too easy to say this year has been unpredictable and not easy to model, the evidence for this however is in the Squiggle  model competition, with only four models in the net positive on bits. Having three straight rounds of negative bits does notr bode well for your model but so far in 2023 we can all be given a pass while the ratings and teams find their running order and settle down going into the first turn and back straight. If the betting markets are your preferred yardstick then we can safely say we have the bookies covered on MAE and bits, only being one behind on tips. For round three alone, the model tipped 5/9, lost 0.11 bits and had a M...
Read more

Model Performance - Round 1

How Did the Model Perform? Throughout 2023 I will be comparing my model to those in the Squiggle model competition with hopes of competing in it next year.  The Squiggle competition is a very well respected group of modellers who all predict the outcome of every match in the AFL season and then analyse their performance using various methods. In this post  I explained how these were but for those in need of a refresher we have bits (confidence based picks), MAE (the average of how many points you were wrong by) and just general tips (how many games did you pick the winner). On Squiggle the top models by bits are The Footycast  who leads the way with 0.79, Elo Predicts  who is second on 0.59 and AFLalytics  a close third on 0.58. My model finished round one in a would be 17th place with -0.07 bits. Not to worry, 23 rounds of AFL football to go. Looking at MAE now and we have  The Footycast   again in first with 27.13, The Cruncher in second with ...
Read more