7!
Then I’d like that, yes.
Hi all,
A non-update update. Got majorly addicted to elden ring so have been neglecting the site recently. Am hoping to do a data refresh sometime next week which will probably be the final cut before the new dlc.
Wasn’t sure whether to do the most recent period as its own block or to merge it into the existing block (would prefer to avoid doing both for storage / bandwidth reasons). Any opinions ?
6 Likes
Can you probably set a local data storage? So it only loads once and then from the browser cache?
I actually would like even more to have the raw data. If I have that I can make graphs myself.
Glad to see you again 
Ok added the latest data as a new period/group, will probably merge it into the existing one when I do an update for the new dlc. Please let me know if you spot any problems / mistakes.
3 Likes
Encase anyone is interested here are the win rates for the 28th and 29th April with the new civs for >1250 Elo open maps:
Needless to say this is based on a very very small set of data and is likely to be extremely biased by everyone jumping on the new civs so please try to avoid over interpretation (will update again in a few days time). I wont be updating the main site until I am happy I have enough data for the win rates to be stabilising.
11 Likes
Thank you very much! I like your work a lot.
3 Likes
yeah cool stuff
Interesting to see hindustanis on top already. I wouldn’t have expected to see gujaras in medium range but the others also so far kinda expected on open maps as they are clearly more designed for more closed maps.
Burmese still strong I see. 
4 Likes
So no urgent buffs are needed, but we should keep an eye out for the Hindustanis in the next large tournament. Pretty good launch for a DLC.
3 Likes
As an aside I’m currently trying to implement my own Elo derivations so I can program a separate elo ranking for “open” and “closed” maps. I remember seeing a proposal on the forum for an alternative Elo formula for team games but I can’t find it now, does anyone have a link ?
Though I also can’t remember if the proposed formula was just for match making or if it also applied for Elo redistribution as well. Actually on that note does anyone know what the current team Elo formula is? I’ve just been taking the mean of each team and putting that into the single player formula.
3 Likes
2 Likes
Agreed it’s a great launch, no real massive bugs, the ratha will likely be fixed (using melee arrows)
But this is a tiny sample. Going by this teutons and byz need a nerf 
(up to 70% and 65% winrates)
Thanks for this. Always great
2 Likes
True, true. The Indian civs were picked more often than Teutons though, hence the smaller error bars.
2 Likes
Yer, to build on this, the error bars give an indication of plausible values given the limited sample size. From this it appears Indians are really performing well at the moment. However what’s much harder to deal with is the bias induced from both civ picking and people likely making very random plays atm with the new civs
4 Likes
ah yeah totally, but at least gives us a rough idea of how they are performing
i still think gurjaras have some amazing matchups, they just need to be worked out. likely see tweaks on all the civs either way, even if its due to player experience and not balance
2 Likes
Yeah the thing is like 50 % of the games currently are inner-indian matchups.
So we basically only see how they perform against each other, not against the other civs.
But interesting nevertheless.
2 Likes
Interesting, if anyone here is able to code in R I would really appreciate some help with a code review as I’m getting some results I wasn’t expecting so was hoping to check I’m not making any obvious mistakes.
Essentially I tried to re-implement the Elo calculations both across all matches and split by map class (open vs closed). After implementing this I get model accuracies of:
- Using existing Elos = 61%
- Using my re-derivation across all matches = 55%
- Using my re-derivation split by map class = 55.5%
I was expecting a bit of a drop in my re-derivation due to the fact I lose ~20% of the data due to data integrity issues from aoe2.net but (a) this was a bigger drop than I was expecting and (b) I was expecting a bigger improvement from the map-class split.
If anyone is able and interested to help please pm me!
2 Likes
Could you explain what you actually did? Cause I only understand rail station.
Also you need to know that all functionable rankings have a positive feedback loop, so if you rederivate them you will most of the time have lower accuracy in win predictions than the used ranking method.
Second we know that there are a lot of smurfs currently that compromise the data.
1 Like
Sure,
- So my local database contains data from 2021-Oct-10 to today (missing ~20% games that were removed due to various data issues, I am assuming missing completely at random and thus ignoring it, perhaps naively…).
- I start by initialising everyone’s Elo to be the mean of their Elos in the original data from their first 5 games from 2021-Oct-10.
- I then process each match in order that it occurs calculating the players Elo after the end of the game based on the result. I do this, independently, for “all” matches, then again for only “closed” matches and then a final time for only “open” matches.
To evaluate this I then take the “open”, Elo > 1200, 20Nov2021 - 27Apr2022 cohort and fit 3 logistic regression models to the data; 1 adjusting for the Elo reported from the original data, 1 adjusting for my “all” matches derived Elo and another just using the “open” only derived Elo. I then look at the model accuracy which gives the above figures.
2 Likes
And you use the model of mercy?
Then there is most likely a gain/loss issue involved. Cause if you make this depending on what value you chose for the logarithmic factor players will gain or lose more elo per game on average than the current model, resulting in players who get “better” during that timespan being under- or overrated depending on what you chose.
I’m guessing here cause I still don’t really comprehend fully what you made.