PanzerXO, on Nov 29 2012 - 03:50, said:
Doesn't WOT also arbitrarily differentiate XP per tank within tiers and give greater XP rewards (again, inevenly distributed in tier) between Premium tanks? I think it would be very difficult to look at an average XP per tier for a purely FTP player compared to one who ran a mix of grind tanks and gold tanks. Extrapolate that to trying to equalize performance between players (global stats) some of whom never play a gold tank, while others are addicted to their Types and SPs (etc.).
I think parsing the data that feeds XP (damage done, kills, caps, etc.) probably is the way to go - rather than adding in another data point that is itself already an amalgamated number derived from the other data points.
- another question; is your data pulling damage dealt directly along with damage dealt on spots (i.e. the main XP generating benefit of scouts) or just damage dealt directly?
Damage is just damage. :D Nowhere to get spotting damage... sadly.
hammer91, on Nov 29 2012 - 05:23, said:
I agree with the 60% cut off.
I do think xp/game is a good measure of performance, in my opinion MORE important than WR which I consider more a team stat. I agree with Garbad on weighting XP/game more.
Platooning is the biggest stat padding problem in the game. I would keep in mind that even though you put the 60% cutoff in, those that platoon still get the additional 50% xp bonus on the extra wins which gets blended in to their total XP. But at least your cutoff mitigates some of the platooning exploit.
Your work is very interesting. I appreciate that you are willing to lay out your methodology so that it can be judged on it's merit unlike the "secret formula" nonsense that noobmeter came up with. Is there a way we can check our rating with it?? I play virtually all my battles solo and am curious what my numbers would be.
Your analysis of the G and Havok players was quite interesting.
I have a high Global Ranking (which Garbad points out is completely without value since it only reflects volume) and I'm curious to see how your new system compares to the developers Global Ranking sytem.
Hammer, you can simply load your data using the efficiency calculator and use a calculator or excel to put the numbers into the equation I gave above... another way, like I said would be to download the xls, paste your clans data into page 2, and all the data for page 1 will be auto-calculated, including WNEfficiency.
Garbad, on Nov 29 2012 - 05:36, said:
Its a safe assumption that the vast majority of stat whores run prem all the time. I'd suggest including an exp/game metric (weighted slightly more than DPG) and see which has better correlation. I bet the exp/game improves it.
Will try it and see what happens. Don´t liek XP since low tiers get very little, plus the prem account problems, plus the nasty bonus XP some tanks (like 13 90) get. Will try it to see what happens though.
Banana_Hammock, on Nov 29 2012 - 05:51, said:
Really, at the end of the day, isn't exp per game by itself an efficiency rating? It takes into account everything we've been trying so hard to recreate with these efficiency calculators. Damage, kills, spots, damage from spots, cap/defense points, etc. It even takes into account win rate (because of the win bonus), and things impossible to track with efficiency like the bonus from damaging a higher tier target. The only downfall is the inability to differentiate premium from non-premium (or like many players, myself included, a mix of the two), as well as being unable to ask the developers of this efficiency calculator the exact formula.
Maybe a little mysterious, but really it's already given to us.
Yes, but like I said it has a million problems, becuase the XP on the profile page is AVERAGE TOTAL XP. This is modified like I said above by prem accounts, tier played, and each tank has a different multiplier which is also unfair. If we had NORMALIZED, PRE-TANK BONUS, PRE-PREM ACCOUNT BONUS, XP, then I wouldn´t be doing this. :D
TxSteppeWolf, on Nov 29 2012 - 06:57, said:
I did this analyze way back in spring, here is my feedback to you
1- Please do us a favor, next time you put x-y; x shows independent variable and y shows dependent variable as a custom; you run your model based on winrates as independent factor; is that really what you did ?
2- You want to measure each variable's contribution, you will see quickly many of that very highly coorelated. Damage vs Frags etc. If you want to measure individual contributions, you need reduce the coorelation between, transform them, redefine them combine them in a meaningful way
3- You need to normalize the damage, a tier 2 and tier X damage isnt same; simplest is damage done over damage taken
4- Garbad's data is limited to certain tanks, you cannot generalize to all tanks. You need to add a dummy variable or a fraction variable that show how much each player plays spg etc, otherwise it is apples oranges
5- You collect clan data, these are skewed towards better players, it is not random. They tend to platoon a lot, so their win rate would be higher than what it should really be.
6- In your analysis you use Rsquare, better also use t-stats. You propose a regression model, a causation not correlation. R is for coorelation
But keep up the work,
1-There is no truly independent variable.
2-That is what I did, notice WNEfficiency has a better correlation to winrate than any of the individual factors that compose it.
3-I´ve said it several times, both damage and frags ARE normalized for average tier played. Frags are normalized in a linear y=ax+b fashion, while damage is normalized y=ax^b fashion.
4-I didn´t want to complicate things, and try a simple formula similar to efficiency and see how well it did. Obviously adding data like percentage of games played in each kind of tank would improve the efficiency, but I would need a profile parser, and I don´t have the time to do that. You are most welcome to do it yourself! :D
For each tank type, each stat could be weighed differently, and everything would become more precise, I agree.
5-I don´t mean to be rude, but we have some pretty sub-average players in our clan :D I think the formula measures them quite fairly.
6-You are welcome to use the data and perform whatever tests you like! Download the xls and play around with it all you like!
DracoArgentum, on Nov 29 2012 - 11:21, said:
Why not just use win rate?
Because it is worthless for less than a couple hundred battles. I'm not running the numbers again but IIRC at 100 battles the 85% confidence interval for win rate is +-7% Thats the difference between a horribad and a great player. Since we often want to compare two players tank to tank we need a stat that correlates to win rate but needs no more than 100 games to become genuinly predictive.
I like what you've done OP. What wold be nice is to see 100 game samples and how they correlate to someone's overall score. If the deviation is low enough we might finally have something that works tank to tank. I suspect you'll need to use tank specific damage per game vales though.
Well that is kind of tough, could be easily done with WOT Statistics, collect stats for 100 games, calculate WNEff. Nevertheless, comparing that to overall account is a tad tricky. I for myself was a complete BAD@TANKS for my first 5000 games, so my stats have been rising steadily from then, and I am always above my average. I guess it is the same way for others.