It's possible to check.

Unfortunately I can't find a recent example on 1v1 ladder. Although I found a couple of Malakkan's games where his rating changed by less than 1 point after he won, like against HalfMoon just today. So that's not great already.

Is all working good if you don't get even 1 full point after you win?

But let's take a look at last season of seasonal ladder.



The upper one is how the season ended (these are raw 'bayeselo' ratings, add 1300 to every player to get what Warzone displays)

The bottom one is how the rating would look like if we cut the 89thlap vs T54321 game (89thlap's second to last game, 89thlap won). All the other games and results remain unchanged.

You can see that if that game never happened, 89thlap's "bayeselo" rating would be 4 points higher.

Now, I'm not complaining about how seasonal works. Bayeselo actually works pretty good there (compared to alternatives). But I'm showing you that it's possible to lose points after winning a game under Bayeselo system.

Now, I'm not complaining about how seasonal works. Bayeselo actually works pretty good there (compared to alternatives). But I'm showing you that it's possible to lose points after winning a game under Bayeselo system.

Due to proper matchmaking with low rating differences between players it is super rare to lose points after winning a game on the 1v1 ladder. Actually it is close to impossible. I did the calculations once when I was rated 2300+ and found that even if I would have been matched up with the lowest rated player (~900) I would "only" lose 1 point. On the Seasonal however you will be paired up with pretty much anyone when you are behind on games. Also rating differences can be much bigger due to the 65 extra points per game which makes the scenario of losing points after winning a game much more likely. I am fine with BayesELO on the seasonal though, I think you could improve matchmaking to decrease the issue of very unfavorable matchups.

Would you mind explaining why BayesELO is better on Seasonal than regular ELO? My understanding was that Seasonals would be benefecial for regular ELO since all players will have the same game count at the end of the season. Or is the issue with people dropping out early / joining late and hence not finishing 20 games?

Would you mind explaining why BayesELO is better on Seasonal than regular ELO? My understanding was that Seasonals would be benefecial for regular ELO since all players will have the same game count at the end of the season. Or is the issue with people dropping out early / joining late and hence not finishing 20 games?

If you haven't read it already the site for BayesElo is a quick and good read listing both pro's and con's. One thing that is not mentioned here already, is that both ELO and BayesELO assume draws are possible. (However, in WZ that is not the case ofc). Also, Bayes let's you give an advantage to first pick (this is set to 10 elo for WZ, iirc).

https://www.remi-coulom.fr/Bayesian-Elo

The 1v1 ladder ratings on the whole are actually highly accurate at predicting match outcomes. The approximate mean prediction error over a sample of more than 100,000 games is <2%.

That surprised me. Is this <2% true across all the rating 'groups'? For comperision, this is better then FIDE's chess rating's ability to predict which can be off by ~5%. https://en.chessbase.com/post/sonas-overall-review-of-the-fide-rating-system-220813
(The article by Jeff Sonas is much more in depth, i recommend a read if you have time. It has some clever way of analyzing a ratings performance when ranking player and ti would be interesting to see how WZ's rating would hold up to a similar stuff.)


Edited 2/28/2019 23:32:33

If you haven't read it already the site for BayesElo is a quick and good read listing both pro's and con's. One thing that is not mentioned here already, is that both ELO and BayesELO assume draws are possible. (However, in WZ that is not the case ofc). Also, Bayes let's you give an advantage to first pick (this is set to 10 elo for WZ, iirc).

Thanks for sharing, but I have read this before. It doesn't really answer my question regarding the Seasonal unfortunately. However it is doing a good job promoting Bayesian ELO which kind of fortifies my opinion of trying to rather adjust the current calculation method instead of replacing it.

The main problem of Bayesian elo is, that it weights all games equally. But in case of your current skill games which are played in the recent past should be weighted in more.

The idea of expiring games after 5 month is a bad fix of this massiv drawback. Very often the rating of a player is ruined by games which are pretty old. Pure nonsense!

The Bayesian elo works for the seasonal ladder, but for all other ladders it is pure garbage. The disappointing thing is that several Forum threads already addressed this issue and Fizzer hasn't done anything yet. And as long as the rating system isn't changed the MDL is the way better 1v1-ladder.

Edited 3/1/2019 10:57:48

The MDL will always be superior to the 1v1 ladder no matter what rating system it uses.

"Another problem with BayesElo and current matchmaking is that it sometimes matches players with such a big rating difference that the better player will lose rating even after he wins."

Agreed with you guys that this is ridiculous and should not happen under any situation. :P

I also have to wonder more of the arguments against changing the algorithm because I'm sure Fizzer prefers this system for a valid reason, and I definitely agree that the algorithm encourages runs, but at the same time there are already "ladder rules" that have been put in place to discourage players from starting alt runs, but even otherwise I don't think it's easy to get 1st place on a run unless you get ridiculously skewed matchups and maybe a couple of lucky wins during the final matchups.

I read the Bayes promotion article but I still don't get what there is to weight in: An ELO system where you can lose points for a victory in even a corner case is just garbage by definition, no matter how well it elsewise approximates the strengths of the player field.

Also predictivity is only half the issue. There's reason Overwatch, League, etc., have weird rating systems that probably aren't great on prediction. Game/incentive design is also important. The fact that MDL not only happened but succeeded is basically a major upset against the 1v1 Ladder itself.

MDL in general and the MW rating system is brilliantly designed to keep you motivated, whether you want to play competitively (since the rating system and the '10-games condition' remove all those boring rants about stalling, run and rating manipulation) or just play casually and get good matches on good templates (what QM most of the time fails to do).

Unless they really love MME and don't care about seeing people with overinflated rating getting temporarily above them, the 1v1 ladder certainly fails to keep good players engaged for a long time.

There is however a design flaw which is annoying with the current implementation and would probably discourage even the most monomaniac MME fan : the Boston run pattern.


There are currently 2 players that I don't want to be matched with if I expect to have a fair rating. They both are good players, who rely pretty heavily on gambles/smart predictions and the outcome of a game against them is often unpredictible. That's fine, I like such opponents.

The issue is that after a while, they get bored of Warzone, stop logging in and start a nice boot streak. Which will only stop after 50 (!) boots to get them to a rating below 1000 in less than 2 months (both players generally have 5 ongoing games of course). If you are unlucky enough to have lost against them while they were active, your rating will incredibly suffer from it due to BayesElo.

That has been discussed dozens of time already I guess, but just removing a player from the ladder after a few boots would solve the issue.

(One of those players is obviously Boston, who has this interesting rating shape displayed above. The other one is that Dutch player -hello / later / KakkieG / Ricky87 who keeps creating new accounts with the same pattern. And you guessed it right, I lost twice to him last month :-(

I can't agree enough with Malakkan, if we want a noncontroversial way to improve the rating system. Lets automatically remove people after a couple boots. Last time I lost on the ladder. I lost to a player who was just under a 2000 ELO. 5 days later he began a streak in which he got booted from approximately 46 out of his next 47 ladder games. Ouch...


Edit: And the one win was a boot, where neither of them made a pick.

Edited 3/8/2019 15:41:03