|
On August 22 2014 22:00 Gahlo wrote:Show nested quote +On August 22 2014 21:43 Usagi wrote:
The fact is you cant correlate "I am playing better" to "I have a good KDA" because you can win games by other factors and that affects your kda, thus, looking at "improving your kda" to win more games is a terrible simplistic point of view, and not, kda doesnt predict wins, because you can not have a set kda before the win or loss is determined.
You cant be like "Oh ill be 12-3-4 this game and then I will win" I'm pretty sure you have a much better chance to win the game 12-3-4 than 3-12-4.
It depends. I remember Slayer or somebody purposefully dropping ranks by playing super shitty but also getting ridiculous KDA/CS scores. You can get an amazing KDA... and still be not where you need to be the whole game. Then your team loses the game.
On August 22 2014 22:31 Sufficiency wrote:
No one said anything about a causal relationship. It's simply correlation between winning and kda.
Edit: Sufficiency, you were saying earlier that if your Death/Assist per minute ratio is higher than X, you can blame your team. We're saying no, you can't.
|
I don't see why not. If winning is strongly correlated with kda, and you have a great kda yet still lost, something else went wrong. Maybe your teammates are too heavy, maybe you threw... but something else happened that made the gsme result surprising.
|
Or maybe you played like shit and were in the wrong place at the wrong time the whole game. Or maybe you were toxic despite being good. Or maybe you, as an initiate support, didn't do your job and played like a pussy the whole game. Or maybe you, as an ADC, didn't actually auto-attack and just ran leaving your teammates to dry. Or a million other things.
The fact is that endogeneity is a problem since you're trying to make this conclusion which isn't purely predictive.
|
GrandInquisitor
New York City13113 Posts
Here's what I plan to do with Sufficiency's data, on a per champion basis:
1. Overall winrate
2. K/D/A in wins
3. K/D/A in losses
Now what I'd like to do is come up with this formula:
Q = a * K + b * D + c * A
Where the data derives values for a, b, and c such that Q is most correlated with W/L.
Does this make sense? My goal is to show the importance of each statistic on X champion, i.e., a high kill Nasus is very correlated with winning but a high kill Sona is not. On the other hand, it is possible that the data is distorted, i.e., a high kill Sona has a very high win rate because if Sona is getting tons of kills then you're probably roflstomping.
|
Problem: Kda is one of the easiest stats to calculate and is used ad-nasium by the community because of this.
G.Inquisitor: how can I modify this to make a stat that more accurately reflects playing well, because this statistic IS so accessable
Sufficiency: I'll test your number against millions of games to see if it IS in fact more indicitive of a well played game (higher win%)
that's it, that's whats happening pls stop trying to make it into something it's not, KDA will always be at fault of not being a win condition. It however can tell you something about the game and will be continued to be used because it's one of the most accessible stats to the average player trying to gauge their performance.
lets stop trying to overcomplicate things.
I guess I can see where this came off as negative, I was trying to support the effort, whist trying to clarify for the people trying to clamp it down by means of unrelated variables.
|
GrandInquisitor
New York City13113 Posts
I like how you proclaim these self-evident truths as if you are Moses from God or something. Please, someone else feel the need to post about how KDA doesn't predict team performance. Add in some anecdotes about how that one time someone chased someone else and wouldn't take a tower or something and that's why you lost your promos.
I expect KDA to be rather poorly correlated with win/loss. But I think it is more helpful if we discuss what insights can be gleaned from masses of KDA data (how poorly correlated? does the degree of correlation vary by role? by champion? by time? can we calculate some kind of "DF" quotient that is better correlated?), instead of just declaring that "KDA has nothing to do with whether you win or lose!" ad nauseum.
|
On August 22 2014 20:49 739 wrote:
I just copy pasted that from some thread of Diamond Shaco player on LoL forums.
I think that Elder > Statik > Hydra/Ghostblade > IE > LW would be good? Or you just want BoTRK?
Or do I want Feral Flare on Shaco?
Runes ad marks, armor/flat hp seals?, as glyphs, ms quints? Or just standard same runes as for adc? full Ad + armor + mres?
Quick runthrough for Shaco...
Masteries: 21/0/9 for buff duration or 27/0/3 taking MS in Utility tree and CDR% etc. in Offense tree
Runes: AD or MS quints, AD reds, AS yellows, AS blues... Personally I use MS quints, Hybrid Pen reds, AS yellows, CDR blues (Shaco's W, E, R, Shiv and Feral Flare etc. sum up to a lot of AP dmg)
You barely take any damage in jungle because you can cancel the attack animation of the big creeps by constantly switching the aggro between Shaco and the boxes.
Items:
- Feral Flare vs Elder Lizard... Lantern gives you more offensive stats for ganks, lets you solo drake at level 6 / lets you clear jungle overall faster. Elder Lizard if you wanna go for full CDR% (and Hydra AoE applies EL passive). Matter of preference imo, most go EL though.
- Full CDR% Shaco is basically like a stealthed AD Kassadin lategame
- Build order is okay, my personal preference for the standard splitpush/assassin Shaco is: jungle item + t2 boots > Hydra > Ghost Blade + Shiv (both involve Avarice Blade for GP10) > Trinity / IE (sell the jungle item at full build for the other item)... Basically anything that boosts your burst: Hydra for auto reset, Shiv active with guaranteed crit from Shaco's Q, Trinity Sheen proc, IE bonus crit dmg etc.
- Best defense items are GA (you can Q right after revive so Shaco is unkillable), Randuins, BV or one of the QS items depending on team comps
Can't think of anything else usefull to say right now.
*Edit - Summoners: Smite/Ignite to potentially snowball out of an early game advantage, Smite/Exhaust for later team fights, Smite/Flash or Ghost etc. possible; I prefer to go Smite/Teleport with 21/0/9 masteries.
|
On August 22 2014 23:03 GrandInquisitor wrote:
Here's what I plan to do with Sufficiency's data, on a per champion basis:
1. Overall winrate
2. K/D/A in wins
3. K/D/A in losses
Now what I'd like to do is come up with this formula:
Q = a * K + b * D + c * A
Where the data derives values for a, b, and c such that Q is most correlated with W/L.
Does this make sense? My goal is to show the importance of each statistic on X champion, i.e., a high kill Nasus is very correlated with winning but a high kill Sona is not. On the other hand, it is possible that the data is distorted, i.e., a high kill Sona has a very high win rate because if Sona is getting tons of kills then you're probably roflstomping.
Do you know about principle component clustering (not principle ccomponent analysis, CLUSTERING)? It will suit your needs I think.
|
Having a good KDA is conductive to winning but there are to many factors around it to use it as a basis for pretty much anything.
Honestly I dont see how this entire line of reasoning leads to anything other then,"I often win if I have a good KDA".
Which is a pointless statement considering you have limited control over how good your KDA is. It even opens up the trap of playing games to maximize KDA which may well be counter productive to actually winning the game.
|
On August 22 2014 23:03 GrandInquisitor wrote:
Here's what I plan to do with Sufficiency's data, on a per champion basis:
1. Overall winrate
2. K/D/A in wins
3. K/D/A in losses
Now what I'd like to do is come up with this formula:
Q = a * K + b * D + c * A
Where the data derives values for a, b, and c such that Q is most correlated with W/L.
Does this make sense? My goal is to show the importance of each statistic on X champion, i.e., a high kill Nasus is very correlated with winning but a high kill Sona is not. On the other hand, it is possible that the data is distorted, i.e., a high kill Sona has a very high win rate because if Sona is getting tons of kills then you're probably roflstomping.
I feel you if you want to look at K/D/A you also have to look after overall Kills/Deaths in the team for that game. Don't know if that's possible with that data collection method.
|
With my data, no it is not possible. Usually you can only find two or three players' s kda in a game.
It's a bug with my crawling algorithm. I really need to rewrite the algorithm.
|
|
|
On August 22 2014 23:19 GrandInquisitor wrote:
I like how you proclaim these self-evident truths as if you are Moses from God or something. Please, someone else feel the need to post about how KDA doesn't predict team performance. Add in some anecdotes about how that one time someone chased someone else and wouldn't take a tower or something and that's why you lost your promos.
I expect KDA to be rather poorly correlated with win/loss. But I think it is more helpful if we discuss what insights can be gleaned from masses of KDA data (how poorly correlated? does the degree of correlation vary by role? by champion? by time? can we calculate some kind of "DF" quotient that is better correlated?), instead of just declaring that "KDA has nothing to do with whether you win or lose!" ad nauseum.
Are you replying to me? Because I was never claiming that KDA doesn't predict team performance. I'm also not proclaming self-evident truths. I said Slayer has intentionally lost while having a high KDA before. Here's the thread:
http://www.teamliquid.net/forum/league-of-legends/441311-there-and-back-again-diamond-i
Also, I think you're mis-interpreting my argument. I'm not claiming KDA has poor predictive power on win/loss. I'm claiming KDA is a poor reflection of individual play. That is, just because you have a high KDA does not mean you played well. Just because you played well does not mean you had to have a high KDA.
|
As I said before I never expected Riot to allow open sponsoring by a porn site on an LCS team.
Will be interesting to see how Valve reacts if they ever qualify for something like TI. (not much reason for them to do anything prior to that. tho the tournament scene will need to act sooner)
|
GrandInquisitor
New York City13113 Posts
Here are some results:
Champions with the biggest gap in "DF" between Wins and Losses:
+ Show Spoiler +Draven 20.14
Akali 20.02
Katarina 19.57
Miss Fortune 19.34
Jinx 19.14
Ezreal 18.71
Twitch 18.66
Vayne 18.65
Quinn 18.49
Swain 18.42
Champions with the smallest gap in "DF" between Wins and Losses:
+ Show Spoiler +Gragas 13.30
Nasus 13.45
Shyvana 13.52
Tryndamere 13.82
Singed 14.04
Olaf 14.09
Shaco 14.10
Malphite 14.11
Poppy 14.19
Garen 14.21
Champions who have the most additional kills in wins (expressed as a percentage, i.e., normalized against the average number of kills for that champion across all games)
+ Show Spoiler +Vayne 129%
Twitch 129%
Sivir 128%
Draven 127%
Yasuo 126%
Orianna 126%
Urgot 126%
Kassadin 126%
Gnar 126%
Ezreal 125%
Champions who have the fewest additional kills in wins (expressed as a percentage, i.e., normalized against the average number of kills for that champion across all games)
+ Show Spoiler +Blitzcrank 112%
Zilean 112%
Amumu 112%
Rammus 112%
Janna 113%
Nautilus 113%
Zyra 113%
Singed 113%
Shaco 113%
Pantheon 113%
Champions who have the most additional deaths in losses (expressed as a percentage, i.e., normalized against the average number of deaths for that champion across all games)
+ Show Spoiler +Anivia 132%
Amumu 129%
Nautilus 129%
Warwick 128%
Janna 128%
Jarvan IV 127%
Nunu 127%
Skarner 127%
Rammus 126%
Malphite 126%
Champions who have the fewest additional deaths in losses (expressed as a percentage, i.e., normalized against the average number of deaths for that champion across all games)
+ Show Spoiler +Gnar 113%
Twitch 115%
Poppy 115%
Karthus 115%
Kennen 115%
Yasuo 115%
Jax 116%
Vayne 116%
Trundle 117%
Darius 117%
|
GrandInquisitor
New York City13113 Posts
It is interesting to me that Gragas tops the list of champions with the smallest difference in DF between wins and losses. I expected him somewhere near the bottom but would not have predicted that he would have led the way ahead of Garen, Singed, Malphite, etc.
The list also seems to confirm the 'hyper-carry' status of Yasuo/Orianna/Urgot [?]/Kassadin/Gnar? As in, of the list of champions that kill the most more when they're winning, these are the only non-ADC champions on that list.
|
|
|
You should adjust KDA per role
In top lane assists should be worth a lot more in general I think, on jungle and bot lane deaths should be a bigger negative because they are usually really bad because they set someone else behind as well
also very high KD on a tank or melee or assassin is a lot better than high KD on a ranged champ because staying alive while doing a lot on a melee probably means you are killing their important carries while cleaning up a fight with a ranged champ mightn't have a big impact.
I mean if you die 3 times in top lane and you are 1/3/2 it seems like a shit score but if their jungle and topkilled you 3 times and you killed him 3 times with your jungle he might be 3/3 and you might be 1/3/2 but it might be good for your team if you dont get any worse than that becaues the top lane soaked the kills while your jungle can carry the momentum to other lanes.
Meanwhile going 1/3/2 on an adc or support is real bad because you might have 6 deaths on bot lane and then thats like a 5/3/1 lucian or something and shits real bad
On August 22 2014 23:43 ghrur wrote:Show nested quote +On August 22 2014 23:19 GrandInquisitor wrote:
I like how you proclaim these self-evident truths as if you are Moses from God or something. Please, someone else feel the need to post about how KDA doesn't predict team performance. Add in some anecdotes about how that one time someone chased someone else and wouldn't take a tower or something and that's why you lost your promos.
I expect KDA to be rather poorly correlated with win/loss. But I think it is more helpful if we discuss what insights can be gleaned from masses of KDA data (how poorly correlated? does the degree of correlation vary by role? by champion? by time? can we calculate some kind of "DF" quotient that is better correlated?), instead of just declaring that "KDA has nothing to do with whether you win or lose!" ad nauseum. Are you replying to me? Because I was never claiming that KDA doesn't predict team performance. I'm also not proclaming self-evident truths. I said Slayer has intentionally lost while having a high KDA before. Here's the thread: http://www.teamliquid.net/forum/league-of-legends/441311-there-and-back-again-diamond-iAlso, I think you're mis-interpreting my argument. I'm not claiming KDA has poor predictive power on win/loss. I'm claiming KDA is a poor reflection of individual play. That is, just because you have a high KDA does not mean you played well. Just because you played well does not mean you had to have a high KDA.
i never updatedthat thread but it only took me like 3 weeks minus days i didnt play to get d1
i was actually more surprised people thought it was going to be an actual challenge for me when i was already a d1
then again i guess im not d1 now but im sort of following the rules for losing uninteionally by duoing with awful players (broken watch and marten) who are extremely good at throwing games because if they lose lane they lose hard and if they win lane they often get caught and throw, if you duo with them bringing higher elo into the game itsmore likely that they lose lane
|
Instead of looking at KDA, why don't we look at gold differentials? (both for team and vs your opposing laner, might even be able to do some sort of crazy equation to weigh how important the gold diff. is for each laner)
|
the goal is to make kda more useful, I tried to explain it earlier but worded it kinda poorly I guess. KDA is one of the easiest stats to find, so he's trying to find a way to extrapolate actual good information from KDA itself. Like their might be better statistical correlation to winning with other stats, but what is the best stat using KDA
|
|
|
|
|
|
EPT
