EP Win Rate Statistics

[Kaiserklein]

Mitoe is right, nowadays it feels like everyone wants to play passive and go fortress no matter what, when in fact there's a lot of more aggressive colonial options. Brits being an obvious example, but also civs like france, iro, china, even germany sometimes.

[Kawapasaka]

Show hidden quotes

Wtf Brit players?

1) Most active top players dont bother to play it because many counter picks (erik inactice, dia grinding unr)
2) Std map set gave klondike, HP, mendo 24/7
3) Manor nerf undeserved

4) GUA plays British into every match up

That doesn't explain the 39% win rate against Dutch, historically one of the easiest Brit MUs
But a small sample will do that

That's all me vs Sheckler.

[Dsy]

I think British should be buffed. Its played a lot and performing very bad.

[RefluxSemantic]

It might be wise to apply some statistics to this data, as without knowing the uncertainty of a certain point of data that point of data is essentially meaningless. I'm not great at statistics but as far as I could tell the optimal way to approach this problem is to consider this a binomial distribution. Then, as far as I can tell, it seems reasonable to assume that 2 * the standard deviation gives a 95% confidence interval, the general norm for significance (slight sidenote: it seems reasonable to me to approximate it this way, as the binomial distribution approaches a bell curve for large numbers, but if anyone knows how to do this more cleanly please let me know). So according to my 'research' on the internet, the standard deviation of a binomial distribution like the coinflip is determined by sd = sqrt (n * p * (1-p) ) with n the number of tries and p the probability of succes (a win). Using this and finding out the 95% confidence interval in winrate gives:

Civilization	winrate +- 2 sigma
Iro	55.8 +- 3.89%
Spain	54.7 +- 4.75%
Russia	53.6 +- 2.99%
Portuguese	53.3 +- 4.90%
Aztec	52.6 +- 6.35%
India	52.1 +- 3.48%
Dutch	51.2 +- 3.42%
China	51.1 +- 3.85%
Germany	50.8 +- 2.89%
Japan	49.4 +- 3.05%
Ottoman	48.4 +- 4.04%
France	48.3 +- 2.71%
Sioux	47.0 +- 5.37%
British	42.2 +- 2.59%

Funnily enough only 3 civs differ significantly from 50%: brits with at best 45.79% winrate, iro with at worst a 51.91% and russia with at worst a 50.61% winrate. But it also kind of shows that it's hard to say much about balance using this. Aztec could be anywhere from the top civ with a large margin to the second worst civ for example. I only consider the British result to be odd and actually require some sort of explanation (which could be that british is too weak atm or that weaker players tend to play british).

[[Armag] diarouga]

You have to take the players into account.
For instance the top 8 has a 70% winrate regardless of the civ.

[bwinner]

It might be wise to apply some statistics to this data, as without knowing the uncertainty of a certain point of data that point of data is essentially meaningless. I'm not great at statistics but as far as I could tell the optimal way to approach this problem is to consider this a binomial distribution. Then, as far as I can tell, it seems reasonable to assume that 2 * the standard deviation gives a 95% confidence interval, the general norm for significance (slight sidenote: it seems reasonable to me to approximate it this way, as the binomial distribution approaches a bell curve for large numbers, but if anyone knows how to do this more cleanly please let me know). So according to my 'research' on the internet, the standard deviation of a binomial distribution like the coinflip is determined by sd = sqrt (n * p * (1-p) ) with n the number of tries and p the probability of succes (a win). Using this and finding out the 95% confidence interval in winrate gives:

Civilization	winrate +- 2 sigma
Iro	55.8 +- 3.89%
Spain	54.7 +- 4.75%
Russia	53.6 +- 2.99%
Portuguese	53.3 +- 4.90%
Aztec	52.6 +- 6.35%
India	52.1 +- 3.48%
Dutch	51.2 +- 3.42%
China	51.1 +- 3.85%
Germany	50.8 +- 2.89%
Japan	49.4 +- 3.05%
Ottoman	48.4 +- 4.04%
France	48.3 +- 2.71%
Sioux	47.0 +- 5.37%
British	42.2 +- 2.59%

Funnily enough only 3 civs differ significantly from 50%: brits with at best 45.79% winrate, iro with at worst a 51.91% and russia with at worst a 50.61% winrate. But it also kind of shows that it's hard to say much about balance using this. Aztec could be anywhere from the top civ with a large margin to the second worst civ for example. I only consider the British result to be odd and actually require some sort of explanation (which could be that british is too weak atm or that weaker players tend to play british).

The best aproach would be to use bayesian interferences. Although what you did is fine and the confidence interval indeed works the same as for a gaussian for big numbers binomial.

[Goodspeed]

It might be wise to apply some statistics to this data, as without knowing the uncertainty of a certain point of data that point of data is essentially meaningless. I'm not great at statistics but as far as I could tell the optimal way to approach this problem is to consider this a binomial distribution. Then, as far as I can tell, it seems reasonable to assume that 2 * the standard deviation gives a 95% confidence interval, the general norm for significance (slight sidenote: it seems reasonable to me to approximate it this way, as the binomial distribution approaches a bell curve for large numbers, but if anyone knows how to do this more cleanly please let me know). So according to my 'research' on the internet, the standard deviation of a binomial distribution like the coinflip is determined by sd = sqrt (n * p * (1-p) ) with n the number of tries and p the probability of succes (a win). Using this and finding out the 95% confidence interval in winrate gives:

Civilization	winrate +- 2 sigma
Iro	55.8 +- 3.89%
Spain	54.7 +- 4.75%
Russia	53.6 +- 2.99%
Portuguese	53.3 +- 4.90%
Aztec	52.6 +- 6.35%
India	52.1 +- 3.48%
Dutch	51.2 +- 3.42%
China	51.1 +- 3.85%
Germany	50.8 +- 2.89%
Japan	49.4 +- 3.05%
Ottoman	48.4 +- 4.04%
France	48.3 +- 2.71%
Sioux	47.0 +- 5.37%
British	42.2 +- 2.59%

Funnily enough only 3 civs differ significantly from 50%: brits with at best 45.79% winrate, iro with at worst a 51.91% and russia with at worst a 50.61% winrate. But it also kind of shows that it's hard to say much about balance using this. Aztec could be anywhere from the top civ with a large margin to the second worst civ for example. I only consider the British result to be odd and actually require some sort of explanation (which could be that british is too weak atm or that weaker players tend to play british).

There is much more uncertaintly due to players of certain levels playing certain civs more than others. People mentioned this earlier but a civ like Brit doesn't really get played at the top level, so there will be a disproportional amount of games in there where the Brit player was significantly worse than their opponent. With data as meaningless as this, statistics don't make it any more meaningful.

What you would need is a data set where games are only counted if the difference between the players' elo ratings is low enough, say 100 JP elo points or less.

[RefluxSemantic]

Funnily enough only 3 civs differ significantly from 50%: brits with at best 45.79% winrate, iro with at worst a 51.91% and russia with at worst a 50.61% winrate. But it also kind of shows that it's hard to say much about balance using this. Aztec could be anywhere from the top civ with a large margin to the second worst civ for example. I only consider the British result to be odd and actually require some sort of explanation (which could be that british is too weak atm or that weaker players tend to play british).

There is much more uncertaintly due to players of certain levels playing certain civs more than others. People mentioned this earlier but a civ like Brit doesn't really get played at the top level, so there will be a disproportional amount of games in there where the Brit player was significantly worse than their opponent. With data as meaningless as this, statistics don't make it any more meaningful.

What you would need is a data set where games are only counted if the difference between the players' elo ratings is low enough, say 100 JP elo points or less.

At which point the data set would probably become very small and unusable I suspect. I think an easy way to test the hypothesis that weaker british players are overrepresented in the data is to actually gather the average ELO for each civ. That could at least give a vague indication as to whether something like that can explain the low british winrate.

[Goodspeed]

Ah yeah you can actually do that because due to the way elo works you know that a 100 elo difference means 66% win rate. This would actually make the data much more meaningful. You'd still have some issues (like some civs being easy to exploit but hard to win with at the top level), but I daresay you would be able to draw some conclusions based on it.

Average elo per civ also would be a pretty interesting stat on its own. Someone do this pls

(Would have to be JP elo since EP elo ladders aren't close to settled)

[pecelot]

Spain still needs to be buffed, though

[zoom]

Great job on this, guys! It's interesting.

I think British should be buffed. Its played a lot and performing very bad.

From what I've seen it's played little. It remains more popular with less good players, though.

[[Armag] diarouga]

These stats are meaningless because it depends so much on the player.
Some guys have 80% regardless of the civ while some have 40%

[Guigs]

Great job on this, guys! It's interesting.

I think British should be buffed. Its played a lot and performing very bad.

From what I've seen it's played little. It remains more popular with less good players, though.

Problem with brits is counterpick anyway, any good Brit map is good for Russia or india

[Hazza54321]

The map pool isnt ideal for brits and tbh civs like jap iro (and german on low res) beat brit aswell.

[deleted_user]

Because most players play primarily the standard euro civs and because most players are bad, and because most good players are not limited by familiarity, it makes perfect sense.

[zoom]

Great job on this, guys! It's interesting.

I think British should be buffed. Its played a lot and performing very bad.

From what I've seen it's played little. It remains more popular with less good players, though.

Problem with brits is counterpick anyway, any good Brit map is good for Russia or india

It's certainly one problem. It isn't very relevant outside of events, though, which is where this data-set comes from, I gather.

[BrookG]

I have always signaled that referencing winrates is tricky. Total picks are the most important in that aspect, they can throw off the rate. And then there is popularity of a civ. The more popular a civ is the more likely it is their winrates are low to mid. There are more opportunities one can mess their game up.

[Goodspeed]

There are also more opportunities for their opponent to mess up?
Popularity only has an effect on win rate if it's different per skill level.

[BrookG]

There are also more opportunities for their opponent to mess up?
Popularity only has an effect on win rate if it's different per skill level.

That's why I mentioned mid too.

[Le Hussard sur le toit]

As was mentioned already by someone else, the only way to give some sense to those data is to compare for each civ the expected score based on elo to the actual results, and see which civs performs generally worse than expected. The way elo works, only the relative strengths of the players count.

[chronique]

Where are the entire data set? (not just the win% of eache civ).

[Goodspeed]

There are also more opportunities for their opponent to mess up?
Popularity only has an effect on win rate if it's different per skill level.

That's why I mentioned mid too.

Why not high?

[BrookG]

There are also more opportunities for their opponent to mess up?
Popularity only has an effect on win rate if it's different per skill level.

That's why I mentioned mid too.

Why not high?

Because bad players are more than pros

[BrookG]

Where are the entire data set? (not just the win% of eache civ).

I have it here for you. The data is from EP6 games since mid August. The data is compiled from lemmings121 and Buckethead, and I simply included some visualisation. The csv is here.
For more data look here: viewtopic.php?f=19&t=17351

[chronique]

8 game aztek vs spain :O