Supercross Elo Ratings

Feint
Posts
1049
Joined
7/25/2010
Location
Leominster, MA US
1/17/2015 6:57pm Edited Date/Time 5/11/2015 8:53pm
I've been working on adapting the Elo rating system used in chess(among other things) for supercross. I'm doing individual seasons and career ratings, but only this season is ready right now. For the career ratings I am starting from 2006 when the majority of the field switched to 450s. The way the rating works, DNFs hurt a lot. I'm thinking of a way to lessen the impact when it is a mechanical DNF, but that is not implemented yet.

450 Class


250 West Class
|
turbotime
Posts
120
Joined
9/27/2012
Location
Miltowne CA
1/17/2015 8:26pm
What is your algorithm?
Feint
Posts
1049
Joined
7/25/2010
Location
Leominster, MA US
1/17/2015 8:41pm
I'm using the regular Elo algorithm with a K value of 8. Each rider in the main has a "match" against each other rider and their rating is adjusted based on that. I'd like to create something a little better for comparing that many competitors, but that will have to wait until next season. I'm going to let this play out, but do some work in the background for next season.
turbotime
Posts
120
Joined
9/27/2012
Location
Miltowne CA
1/17/2015 10:00pm Edited Date/Time 1/17/2015 10:13pm
Might also be worth considering a scenario where each rider is only being "matched" once against the highest ranked opponent(22 matches total per race instead of 231 matches). For example, from the list above, if Roczen finishes a race in 5th, the 4 riders finishing ahead of him would see an increased rating and Roczen's points would be adjusted down or up depending on where the next highest ranked rider(Dungey) finished. Everyone finishing behind Roczen would see a decrease. One adjustment per race for every racer. This would curb the wild swings you are seeing by adding all matches and IMO give you better results overall. After all, only one rider wins the race.

I would also use a K value of 32.
Feint
Posts
1049
Joined
7/25/2010
Location
Leominster, MA US
1/17/2015 10:17pm
It's not all about winning though. I think your suggestion would cause an even bigger swing, especially at a K value of 32. By your suggestion, A1 would have given Roczen an increase of around 336 if my quick estimate is right. The system I'm using should normalize a little better after a few races when riders settle into ratings that are close to where they belong. If anything, the K value I'm using may be too high.

Here is an update for tonight. Riders that have not made every main are not included. They will be included if they have raced at least 75% of the mains.

450 Class


250 West Class

The Shop

enketchum
Posts
3785
Joined
2/6/2012
Location
CA US
1/18/2015 2:02am
what do all the numbers mean
Feint
Posts
1049
Joined
7/25/2010
Location
Leominster, MA US
1/31/2015 12:43pm
The numbers are the riders Elo rating after each round. Each riders Elo rating is adjusted after each race based on their position relative to other riders. A rider with a low rating who beats a higher rated rider will gain more points than a high rated rider beating a low rated rider. Because of the large number of comparisons being done each round, it's hard to see the effect of a single comparison. I'm showing the ratings for the sake of transparency.

It will take a few more races before riders begin to settle into the ratings they should have. I'm mostly just doing the single season ratings as an experiment. I am working on career ratings that would be a bit more accurate, but the amount of data entry I have to do for that is a bit daunting.

450 Class



Here we can see the effect one bad race can have on a riders rating. Roczen had a decent lead, but his 15th place finish last week actually dropped him down into 4th in the ratings behind Dungey, Canard, and Tomac. Speaking of Tomac, he has made a big recovery from his awful start to the season. I expect to see Roczen's rating recover some this week as I doubt he will have two bad races in a row.



250 West Class



The 250 class has been a bit more consistent with only Justin Hill losing rating. Webb and Bowers are showing themselves to be the class of the field. You can also see how Malcolm Stewart has had a similar recovery in the ratings as Tomac has in the 450 class.
Feint
Posts
1049
Joined
7/25/2010
Location
Leominster, MA US
1/31/2015 3:07pm
I just finished updating my career ratings for the 2008 season. Here is a little preview of the results. I started with the 2006 season(When most switched to 450s), so that is why RC's rating isn't higher. RC and Byrne did not race in 2008 so their bars are greyed out.

disbanded
Posts
6040
Joined
8/26/2007
Location
Denver, CO US
Fantasy
1842nd
1/31/2015 4:44pm
I'm trying to think but nothing happens
Triton
Posts
618
Joined
10/1/2006
Location
Redmond, WA US
1/31/2015 5:59pm
Looks like you are already neck deep in it. Good stuff. Always a challenge to rate against multiples. Have you checked into this? Its variation on ELO and a few others but accurately compares against a group.
http://research.microsoft.com/en-us/projects/trueskill/details.aspx

Also you may want to give Glicko2 a scan. Might be a little better for you if you haven't.
Feint
Posts
1049
Joined
7/25/2010
Location
Leominster, MA US
2/7/2015 7:00pm
I will look into that Triton. I'm going to stick with what I'm using for this season since I already did the work to set everything up.

450 Class



Ryan Dungey continues to be the most consistent top finisher and it shows in his rating. However, one bad race could change things up significantly. Cole Seely has also shown up in the top 5 for the first time this season due to his improvement each week.

250 West



Cooper Webb continues to dominant in the 250 class. He crashed in practice today and it remains to be seen how that will affect his racing tonight. Due to his bad race last week, Tyler Bowers has fallen outside the top 5. I expect to see him recover tonight and regain his spot.
daemon616
Posts
2026
Joined
2/19/2012
Location
Euless, TX US
5/11/2015 8:16pm
Feint wrote:
I will look into that Triton. I'm going to stick with what I'm using for this season since I already did the work to set everything...
I will look into that Triton. I'm going to stick with what I'm using for this season since I already did the work to set everything up.

450 Class



Ryan Dungey continues to be the most consistent top finisher and it shows in his rating. However, one bad race could change things up significantly. Cole Seely has also shown up in the top 5 for the first time this season due to his improvement each week.

250 West



Cooper Webb continues to dominant in the 250 class. He crashed in practice today and it remains to be seen how that will affect his racing tonight. Due to his bad race last week, Tyler Bowers has fallen outside the top 5. I expect to see him recover tonight and regain his spot.
Did you ever finish this for the season!? I love it
Tim507
Posts
3136
Joined
6/8/2010
Location
Oregon City, OR US
5/11/2015 8:53pm
This type of numerical analysis is interesting. When you've completely entered the 2015 SX season the ELO numbers will represent the actual results for the year correct? Yet when you manage the data for many seasons you'll start to see trends?

Knowing that bad results or no results skew the results, what do we learn from that? should you not drop a rider from the problem when or based on a % representative to some baseline?

Post a reply to: Supercross Elo Ratings

The Latest