repeekgg / browser-extension

Browser extension that enhances the FACEIT experience and adds useful features
https://repeek.gg
GNU General Public License v3.0
137 stars 58 forks source link

Estimate ELO gain/lose isn't working properly #40

Open Onigeshi opened 5 years ago

Onigeshi commented 5 years ago

https://www.faceit.com/en/csgo/room/1-67f68fb9-6d97-4156-b31b-26904c697f0e

I forgot to screenshot the lobby before, but after the last update the margin of error has increased exponentially.. in the lobby linked above it was showing that we will gain +15, yet we gain only +12.. remember that before this last update I've never had a margin of error bigger than +/-1 like it was mentioned anyway..

I'll try to screenshot a lobby before and after it ended and edit this post.. at the moment I forgot to make a screenshot before so you could have the ELO of the players and the averages..

timche commented 5 years ago

We didn't change the calculation. Sometimes it can happen that the margin is higher, but rarely.

Onigeshi commented 5 years ago

https://imgur.com/a/RhT9AN5

@timche like I told you, it may not be your fault, it may be a change made by faceit that broke things.. but after the last update this is a thing.. and its not that rare.. its been for like 4-5 games in a row..

timche commented 5 years ago

Okay, that's weird. Will have a look at this, thanks.

MaBauMeBad commented 4 years ago

+1 still a thing

airen29 commented 3 years ago

Allright. So it does work good if a player plays solo game.

Once there is a stack of 3+ people it calculates elo wrong.

timche commented 3 years ago

Since FACEIT has introduced Elo gain/loss penalties for teams with too much Elo difference, there's currently no way for us to make precise calculations.

h0lmiew commented 2 years ago

Since FACEIT has introduced Elo gain/loss penalties for teams with too much Elo difference, there's currently no way for us to make precise calculations.

From playing and experiencing ELO penalty from many many games, it looks to me like there is a pattern in which the K-Factor is scaled depending on the spread of your teammates ELO in a 5-man stack, into three buckets.

The normal K-factor is 24, which is then scaled as usual if the average ELO of both teams is not even. So for an even match with 50%/50% for both teams, you win or lose 24 ELO points, if there is no ELO penalty applied.

But, the K-Factor is scaled to 16, or 8, depending on how wide the spread of ELO the team members have, in terms of highest to lowest, if too wide.

The following buckets are the ELO span threshold (from lowest to highest rated player) for a 5-man stack team:

E.g. if you have a 3200 ELO and an 1100 ELO player, you have a span of 2100, and will win or lose roughly 8 ELO.

If you take these ranges into account when predicting ELO to win or lose, I believe it should fit.