FRC 2016 TrueSkill Ratings

[jtrv]

Hello,

I've been working on a project that focuses heavily on statistics within FRC. One thing I wanted to do was include an Elo or TrueSkill rating system for FRC.

Since TrueSkill can handle team games and Elo is not inherently designed for teams, I decided on TrueSkill for the first iteration. I created this model using TheBlueAlliance API for the data and a Python implementation of TrueSkill, with default values for all TrueSkill variables. These can be adjusted if I need them to be, and I may fiddle with them in the future to see if the model can be improved in some way. Currently, the only matches analyzed are from official, non-offseason events that are not the Week 0 event.

I modeled the ratings twice - once unscaled and once scaled. Unscaled means that the ratings you see are ratings that TrueSkill uses for calculations. The starting value is 25.0. Scaled means I multiplied all ratings by 60, since most Elo systems have a starting value of 1500.0 (and 25 * 60 = 1500). The scaled version may be easier to relate to if you have seen other Elo systems, such as the classic chess, or online games like LoL or WoW.

The links to the ratings are quite large in size- opening on a mobile device may take a little while (3130 rows of a tsv file).

----

I know someone will say "These rankings don't mean anything!" - it always happens with Elo based systems. These aren't meant to be a solid ranking system - just a ballpark figure. I'm not saying Team A is strictly better than Team B because they have a higher rating. Just enjoy the numbers.

[Abhishek R]

This is pretty interesting.

Why is the starting value 15? It seems to me that the scaled ratings may be a little inflated, since in chess and games like League it seems like the very best players are in the 2400 range, but here there are not insignificant number of teams above that threshold (which may be accurate, I was just curious).

[jtrv]

Originally Posted by Abhishek R

This is pretty interesting.

Why is the starting value 15? It seems to me that the scaled ratings may be a little inflated, since in chess and games like League it seems like the very best players are in the 2400 range, but here there are not insignificant number of teams above that threshold (which may be accurate, I was just curious).

Ah, shoot. That was a typo. The initial value is 25, not 15.

I believe Microsoft mentioned why in their research docs on TrueSkill. I can't quite remember off the top of my head, but it might be in there. As for the count of how many, from my experience, the data seems to have be just under what is typical for those rating cutoffs. For example, we see 116 teams above 2200 rating, or 3.7% of teams. From my experience on WoW years ago, 2200 was typically around top ~5%, but would fluctuate between rating resets each season.

[Lil' Lavery]

I'd be interested to see this numbers plotted against other metrics of team performance (OPR, district points, win%, etc).

[jtrv]

Originally Posted by Lil' Lavery

I'd be interested to see this numbers plotted against other metrics of team performance (OPR, district points, win%, etc).

I have some other very general and unorganized stats in this Gist, if you want to look. It doesn't compare between different stats but it has some cool stuff. Not all of them are meant to represent anything (e.g. most auton in a single event - not a great metric due to different # of games per event, later event in the season generally has better performance, etc., most of these were just curiosity on my part). I don't have district points modeled yet (one of the only things I don't besides award recipient numbers). I'm planning to. But I don't know anything about how it works so I'd need to spend a few minutes reading before I went at it (NY no districts )