Consider the ratings of 2 players after completing 1 seasonal ladder game.

p1 has a rating of 1671 after 1 win over a player with rating 1582
p2 has a rating of 1697 after 1 win over a player with rating 1543

How is it possible for p2 to have a higher rating? p2 finished their game first, is this relevant?

Just a shot in the dark after reading the Bayeselo documentation, but perhaps it comes down to the first-move advantage adjustment?

https://www.warlight.net/wiki/Ladder_Ranks_and_Ratings

Following those instructions you can calculate someones rating.

I can't answer your question cause for me it seems pretty random. In the beginning it's easily possible for a player with 1 loss to have a higher rating than a player with 2 wins too...

Ya, my rating after 2 wins is like 400 points lower than some people with 2 losses

The Rating is a little bit wierd in the beginning.

When I did a test on move order it didn't seem to matter.

It may be because one opponent has a higher standard deviation than the other, meaning that it's more certain of one rating than the other, which affects the rating.

Edited 5/18/2016 17:28:12

https://www.warlight.net/LadderTeam?LadderTeamID=14249

4 losses 1267 Rating

https://www.warlight.net/LadderTeam?LadderTeamID=14231

3 wins 1280 Rating

seems fair cause 3 wins have a bigger Rating than 4 losses.

^+1

lol MoD, you`re just a noob. Admit it ;)

Note how the first guy's opponents all have the same rating (1449).

Edited 5/18/2016 19:23:28

that's quite interesting. 1win and 1 loss are equal to 2 wins here... i wonder if that is really just a super strange coincidence

Per MoD's link, pick order (not move order) does matter. The example above, both p1 and p2 got a win, but p1 had first pick, p2 had second. So it makes sense that p2's win is worth more.

The link, https://www.warlight.net/wiki/Ladder_Ranks_and_Ratings, also describes how you can compute ratings based on actual ladder data. Very useful!

The data for season 23 is at http://data.warlight.net/Data/BayeseloLog4023.txt and presumably is updated regularly.

The variance (not displayed) of the ratings matter as well. Winning against a player with rating 1200 with small variance is different from winning against a player with rating 1200 with large variance. Especially with few games, this makes a large difference.

In Seasonal, you also get points for getting games (up to 20). Unlikely, but it's also possible one of the players joined late and simply had 1 game less assigned up to that point.

player12345: Fizzer has advantage set to 0, so the algorithm doesn't account for pick order.

@Math Wolf: does Bayeselo store variance measurements? I was under the impression that it only stored single values for ratings.

EDIT: nvm just noticed that it does track standard deviation

Edited 5/18/2016 19:55:40

At the end of the season you have a 1300 point boost for having been assigned 20 games (65 points per game).

Edited 5/18/2016 20:05:41

@ vitriol: it does, but note that rating systems don't need to store it while still implicitely using it.

A team with more games will automatically show less variation in its rating as it "pulls" the fitted model towards itself more than a team with fewer games would. (intuitive statistics 101 ;-) but can be proven easily for more straightforward models.) So even if the models stores nor displays them, the reasoning will likely still hold.
Any good rating system will "understand" that an outcome of a game against a player/team of which the rating is well-known gives more information that an outcome of a game against a player/team of which the rating is pretty unknown and adjust for it accordingly.
(BayesElo, classic Elo, TrueSkill etc. all do this.
As a counterexample: the FIFA rankings don't, which is one of the many reasons why it's one of the worst rating systems out there.)

nvm

Edited 5/19/2016 11:04:14