Stattraction in 2020

Hello everyone. I thought I’d write about my changes for AFL 2020, what I didn’t end up changing, and why I’m not posting predictions here for AFLW and Counter-Strike yet.

Stattraction Dual Ratings: 2020 Edition is ready to give the AFL Men’s season a crack, and I thought I’d give a little summary on what has changed from 2019. There’s actually not a great deal of changes in this changelog, since I invested a lot of time in features that didn’t end up coming to fruition (for this year, anyway).

• Changes to Home Ground Advantage: My turbulent relationship with HGA has continued into another year. Last year, I had a changing general HGA value that was applied to each team that was playing at their home ground against a team that wasn’t playing at their home ground (or ground local to them). This year, HGA is calculated differently for each team, which I hope will make things more accurate. I do plan on moving to specific venues in the future but man I love procrastinating and it’s not going to be ready this year.
• End-of-Year Rating Decay: At the start of each year, matches from previous years decay in weighting by a set amount (this year it will be -37%). This means that ratings may change more rapidly than usual at the start of the year.
• More testing: My 2019 model was not tested very thoroughly. I adjusted the model based on tip accuracy and MAE over the previous 5 seasons. This year, the model was also tested on bits, and the testing period was increased to 20 seasons (with more weighting given to performance in more recent seasons). This showed that my ratings system changed too rapidly, so expect it to be a bit calmer this season.
• Simulation uncertainty: The further in the future a match is, the less certain the model will be about its result. I really should have added this years ago. As we don’t have any results in 2020 yet, this won’t really be a factor in simulations yet, but as we get further into the season it will be turned up.

“Surely you could have come up with more than this?”

• I planned to add rain impact on ratings. Sadly, collecting the data for this will be a long exercise and based on the minimal data I have, rain doesn’t have a big impact on scoring. Sure, in heavy rain, scoring is obviously affected, but how many matches like that do we usually get in a season? It’s uncommon enough that there isn’t enough data to make any changes.
• Changing to player-based is a lot harder than it sounds and, honestly, I don’t think it makes a huge difference to model accuracy. It’s more accurate in theory but any implementation of it from me will probably make my model less accurate, if anything. I’ll continue to look into it, but I think most teams are more than capable of covering for star players when they are injured.

“What about those other projects you’ve mentioned in the past?”

• AFLW: This league still has a large number of factors that make predicting the season extraordinarily difficult. The teams play very few matches each season, they commonly change large amounts of players as new blood pushes into the system, completely new teams are being added regularly and there aren’t very many matches to test with (the 100th AFLW match was only just played this season). I am going to be posting my tips weekly on Twitter but these should be taken with several thousand grains of salt until these issues are eventually resolved by the AFL.
• Counter-Strike: I think I’ve been working on this for a good 16 months now, but it is no closer to being finished. I have a competent rating and ranking system, but converting this to be able to predict matches and tournaments is proving to be quite a challenge. I still have a few ideas to make this workable, but I might have to start the whole project from scratch if they don’t pan out. I will get this done eventually.
• Other AFL things: I say this every year, but I do hope to do more than just predictions for AFL this year. The sport is so rich with potential data and there are so many projects I could do, but time is a resource and I don’t have as much of it as I would like.

I think that covers everything.