|
|
On December 16 2014 12:42 Sufficiency wrote:
Not really. It's not a dick measuring contest.
I think Goumindong obviously has some expertise on econometrics that the rest of us here lack. I think it would be nice to see how an economist would conduct a study on "how important is gold lead at 10 minutes". I think there is a lot to be learned here.
The problem with the simple statistics is that its hard to disentangle the gold from the composition or the skill. So it really only tells us that we are winning and not whether or not we have a particular advantage. I am not sure there is a better way to answer the question though because actually getting a value on the gold is tricky. And because sometimes we don't mind a full estimate
Also the only solution i can think of is not particularly feasible. Basically you take each game and regress gold difference as a time series. Then you take the values from that regression and use them for any other regressions you want(summary statistics will give you a broad idea of how powerful any gold lead snowballs, you should be able to see if the rate of gold change has any effect on win%, and i think once you had the rate of gold change and the effect of gold itself you could find some sort of independent effect on win% but not sure, differences in champions in the game would be interesting and so forth). Ideally you would want to use a vector auto regression (on both sides gold and xp totals) but a simple time series on the lead would probably be just fine.
I haven't actually looked at time series in a long time so am not sure precisely how you would want to set it up (i know the base model you would want to run of G_+1=B*G_0 +e won't work because B will too high and breaks the math. The variable has to be stationary and I am not sure precisely what I would want to run without some more thought) and i know there will heteroskedasticity problems that I am not quite sure how to solve in a time series case though i think its equivalent to the normal case. Sample size of the time series doesn't matter much since we have hundreds of thousands of games.
Fortunately the Heteroskedasticity problems won't be an issue for the larger set since the main thing it does is widen your error bars when corrected for and we are either working in a situation where we don't care about error bars or because we ran 200k regressions they're so small it doesn't matter whether you used HC corrected standard deviations or not.
The most challenging issue would probably be structural breaks. Games in which structural breaks occur could potentially break the math (again), as could any game with particularly wide gold swings. I don't know of a way to systematically examine for structural breaks. Though if such a method did exist it could be an interesting way to look at champion power (basically if you have a champion which has a strong lategame or early game there may be a point in which the structure of the game breaks, by examining the distribution of these breaks you could figure out when champions had power spikes more accurately than just looking at their win rates by time.
Anywho, here is something you could try that wouldn't be difficult computationally. Take a champion and look at the density of their game lengths as well as the win % as the lengths. I suspect that you will find troughs in density at or right before a champions powerspike (for champions with powerspikes post 25 minutes or so) and peaks when they fall off with the win % basically following those numbers
|
On December 16 2014 14:02 Goumindong wrote:Show nested quote +On December 16 2014 12:42 Sufficiency wrote:
Not really. It's not a dick measuring contest.
I think Goumindong obviously has some expertise on econometrics that the rest of us here lack. I think it would be nice to see how an economist would conduct a study on "how important is gold lead at 10 minutes". I think there is a lot to be learned here. The problem with the simple statistics is that its hard to disentangle the gold from the composition or the skill. So it really only tells us that we are winning and not whether or not we have a particular advantage. I am not sure there is a better way to answer the question though because actually getting a value on the gold is tricky. And because sometimes we don't mind a full estimate Also the only solution i can think of is not particularly feasible. Basically you take each game and regress gold difference as a time series. Then you take the values from that regression and use them for any other regressions you want(summary statistics will give you a broad idea of how powerful any gold lead snowballs, you should be able to see if the rate of gold change has any effect on win%, and i think once you had the rate of gold change and the effect of gold itself you could find some sort of independent effect on win% but not sure, differences in champions in the game would be interesting and so forth). Ideally you would want to use a vector auto regression (on both sides gold and xp totals) but a simple time series on the lead would probably be just fine. I haven't actually looked at time series in a long time so am not sure precisely how you would want to set it up (i know the base model you would want to run of G_+1=B*G_0 +e won't work because B will too high and breaks the math. The variable has to be stationary and I am not sure precisely what I would want to run without some more thought) and i know there will heteroskedasticity problems that I am not quite sure how to solve in a time series case though i think its equivalent to the normal case. Sample size of the time series doesn't matter much since we have hundreds of thousands of games. Fortunately the Heteroskedasticity problems won't be an issue for the larger set since the main thing it does is widen your error bars when corrected for and we are either working in a situation where we don't care about error bars or because we ran 200k regressions they're so small it doesn't matter whether you used HC corrected standard deviations or not. The most challenging issue would probably be structural breaks. Games in which structural breaks occur could potentially break the math (again), as could any game with particularly wide gold swings. I don't know of a way to systematically examine for structural breaks. Though if such a method did exist it could be an interesting way to look at champion power (basically if you have a champion which has a strong lategame or early game there may be a point in which the structure of the game breaks, by examining the distribution of these breaks you could figure out when champions had power spikes more accurately than just looking at their win rates by time. Anywho, here is something you could try that wouldn't be difficult computationally. Take a champion and look at the density of their game lengths as well as the win % as the lengths. I suspect that you will find troughs in density at or right before a champions powerspike (for champions with powerspikes post 25 minutes or so) and peaks when they fall off with the win % basically following those numbers
You really don't want to include win/loss, do you?
I think time series is interesting, but remember when you win the game you almost always have a big gold spike (because you probably aced the other team then took some towers) so I am not sure how the end point will be like. I think the last term will correlate with win/loss so strongly that one can substitute for the other. At the end of the day you might be looking at the same thing.
|
On December 16 2014 12:55 Gahlo wrote:Show nested quote +On December 16 2014 12:51 JazzVortical wrote:
Can we go back to harshly criticising every move Riot makes, instead of arguing about statistical garbage?
Although, I like the new Void Portal item in theory. It got a buff on the PBE that gives bonus AD and AP based on your armour and magic resist when you are near the portal. In addition, if it goes though as is on PBE, it will be only the 2nd item currently in the game that gives both Armour and Magic Resist. Neat. It got changed today to buff voidspawn after the 3rd instead.
Ah so they have. Added a bunch of champ balancing as well, but again to champs that have recently fallen out of favour. Elise is getting buffed people, granted it is only her damage vs monsters, but still.
Also another Sona nerf. Is she actually even strong? I never see one.
|
skarner buffs. i dont think that will be enough for me to play him again which makes me sad. i used to love playing skarner.
those renekton changes are interesting...
|
On December 16 2014 14:20 Sufficiency wrote:Show nested quote +On December 16 2014 14:02 Goumindong wrote:On December 16 2014 12:42 Sufficiency wrote:
Not really. It's not a dick measuring contest.
I think Goumindong obviously has some expertise on econometrics that the rest of us here lack. I think it would be nice to see how an economist would conduct a study on "how important is gold lead at 10 minutes". I think there is a lot to be learned here. The problem with the simple statistics is that its hard to disentangle the gold from the composition or the skill. So it really only tells us that we are winning and not whether or not we have a particular advantage. I am not sure there is a better way to answer the question though because actually getting a value on the gold is tricky. And because sometimes we don't mind a full estimate Also the only solution i can think of is not particularly feasible. Basically you take each game and regress gold difference as a time series. Then you take the values from that regression and use them for any other regressions you want(summary statistics will give you a broad idea of how powerful any gold lead snowballs, you should be able to see if the rate of gold change has any effect on win%, and i think once you had the rate of gold change and the effect of gold itself you could find some sort of independent effect on win% but not sure, differences in champions in the game would be interesting and so forth). Ideally you would want to use a vector auto regression (on both sides gold and xp totals) but a simple time series on the lead would probably be just fine. I haven't actually looked at time series in a long time so am not sure precisely how you would want to set it up (i know the base model you would want to run of G_+1=B*G_0 +e won't work because B will too high and breaks the math. The variable has to be stationary and I am not sure precisely what I would want to run without some more thought) and i know there will heteroskedasticity problems that I am not quite sure how to solve in a time series case though i think its equivalent to the normal case. Sample size of the time series doesn't matter much since we have hundreds of thousands of games. Fortunately the Heteroskedasticity problems won't be an issue for the larger set since the main thing it does is widen your error bars when corrected for and we are either working in a situation where we don't care about error bars or because we ran 200k regressions they're so small it doesn't matter whether you used HC corrected standard deviations or not. The most challenging issue would probably be structural breaks. Games in which structural breaks occur could potentially break the math (again), as could any game with particularly wide gold swings. I don't know of a way to systematically examine for structural breaks. Though if such a method did exist it could be an interesting way to look at champion power (basically if you have a champion which has a strong lategame or early game there may be a point in which the structure of the game breaks, by examining the distribution of these breaks you could figure out when champions had power spikes more accurately than just looking at their win rates by time. Anywho, here is something you could try that wouldn't be difficult computationally. Take a champion and look at the density of their game lengths as well as the win % as the lengths. I suspect that you will find troughs in density at or right before a champions powerspike (for champions with powerspikes post 25 minutes or so) and peaks when they fall off with the win % basically following those numbers You really don't want to include win/loss, do you? I think time series is interesting, but remember when you win the game you almost always have a big gold spike (because you probably aced the other team then took some towers) so I am not sure how the end point will be like. I think the last term will correlate with win/loss so strongly that one can substitute for the other. At the end of the day you might be looking at the same thing.
You could just exclude the last term.
I would actually love to include win/loss but its tying the effect that is hard, unless you don't want a measure of advantage from gold and just want a measure of advantage when you have that much gold. I just don't think the second question is that interesting, you can't really extrapolate for a fight due to gold advantage more than you can from just looking at items/levels.
There are similar problems measuring the effects of items (basically you capture the effects of winning and losing if winning/losing effects your item choice) but looking at items might be a better path
|
Bearded Elder29904 Posts
Well, yeah.
Morellonomicon
Recipe cost increased to 880 from 680
Total Cost increased to 2300g from 2100g
Escalated quickly :D And they increased hp amount of jungle monsters ? Goddamn.
|
Just did an ARAM as a Rek'Sai with an ally Orianna.
The ball delivery system is real, holy crap. Burrow W - E tunnel - W Unburrow - Ori R, just wrecks people.
|
On December 16 2014 15:08 Goumindong wrote:Show nested quote +On December 16 2014 14:20 Sufficiency wrote:On December 16 2014 14:02 Goumindong wrote:On December 16 2014 12:42 Sufficiency wrote:
Not really. It's not a dick measuring contest.
I think Goumindong obviously has some expertise on econometrics that the rest of us here lack. I think it would be nice to see how an economist would conduct a study on "how important is gold lead at 10 minutes". I think there is a lot to be learned here. The problem with the simple statistics is that its hard to disentangle the gold from the composition or the skill. So it really only tells us that we are winning and not whether or not we have a particular advantage. I am not sure there is a better way to answer the question though because actually getting a value on the gold is tricky. And because sometimes we don't mind a full estimate Also the only solution i can think of is not particularly feasible. Basically you take each game and regress gold difference as a time series. Then you take the values from that regression and use them for any other regressions you want(summary statistics will give you a broad idea of how powerful any gold lead snowballs, you should be able to see if the rate of gold change has any effect on win%, and i think once you had the rate of gold change and the effect of gold itself you could find some sort of independent effect on win% but not sure, differences in champions in the game would be interesting and so forth). Ideally you would want to use a vector auto regression (on both sides gold and xp totals) but a simple time series on the lead would probably be just fine. I haven't actually looked at time series in a long time so am not sure precisely how you would want to set it up (i know the base model you would want to run of G_+1=B*G_0 +e won't work because B will too high and breaks the math. The variable has to be stationary and I am not sure precisely what I would want to run without some more thought) and i know there will heteroskedasticity problems that I am not quite sure how to solve in a time series case though i think its equivalent to the normal case. Sample size of the time series doesn't matter much since we have hundreds of thousands of games. Fortunately the Heteroskedasticity problems won't be an issue for the larger set since the main thing it does is widen your error bars when corrected for and we are either working in a situation where we don't care about error bars or because we ran 200k regressions they're so small it doesn't matter whether you used HC corrected standard deviations or not. The most challenging issue would probably be structural breaks. Games in which structural breaks occur could potentially break the math (again), as could any game with particularly wide gold swings. I don't know of a way to systematically examine for structural breaks. Though if such a method did exist it could be an interesting way to look at champion power (basically if you have a champion which has a strong lategame or early game there may be a point in which the structure of the game breaks, by examining the distribution of these breaks you could figure out when champions had power spikes more accurately than just looking at their win rates by time. Anywho, here is something you could try that wouldn't be difficult computationally. Take a champion and look at the density of their game lengths as well as the win % as the lengths. I suspect that you will find troughs in density at or right before a champions powerspike (for champions with powerspikes post 25 minutes or so) and peaks when they fall off with the win % basically following those numbers You really don't want to include win/loss, do you? I think time series is interesting, but remember when you win the game you almost always have a big gold spike (because you probably aced the other team then took some towers) so I am not sure how the end point will be like. I think the last term will correlate with win/loss so strongly that one can substitute for the other. At the end of the day you might be looking at the same thing. You could just exclude the last term. I would actually love to include win/loss but its tying the effect that is hard, unless you don't want a measure of advantage from gold and just want a measure of advantage when you have that much gold. I just don't think the second question is that interesting, you can't really extrapolate for a fight due to gold advantage more than you can from just looking at items/levels. There are similar problems measuring the effects of items (basically you capture the effects of winning and losing if winning/losing effects your item choice) but looking at items might be a better path
OK I think at the end I want to know what you would do to find this instrumental variable. Because, frankly, ditching winning/losing is not acceptable for me since:
1. At the end of the day, winning/losing is all that matters. Gold growth/etc is just an intermediate step.
2. If you are OK removing the last term, you probably want to remove the second last term too, because winning/losing correlates with the last term, and the last term correlates with the second last term - the argument becomes recursive. It's not clear to me where the cut-off is for a "proper" model by your standard.
|
On December 16 2014 15:14 739 wrote:
Well, yeah.
Morellonomicon
Recipe cost increased to 880 from 680
Total Cost increased to 2300g from 2100g
Escalated quickly :D And they increased hp amount of jungle monsters ? Goddamn.
This needed to be done though, it never needed the price reduction in the first place.
|
Banner on Soraka is pretty great, especially if you have baron. Just sits back and wrecks the other team's tower while you heal it constantly.
|
On December 16 2014 15:40 JazzVortical wrote:Show nested quote +On December 16 2014 15:14 739 wrote:
Well, yeah.
Morellonomicon
Recipe cost increased to 880 from 680
Total Cost increased to 2300g from 2100g
Escalated quickly :D And they increased hp amount of jungle monsters ? Goddamn. This needed to be done though, it never needed the price reduction in the first place.
I kind of want to see the passive removed or changed to something more..... inconsequential, if you will.
|
On December 16 2014 15:40 JazzVortical wrote:Show nested quote +On December 16 2014 15:14 739 wrote:
Well, yeah.
Morellonomicon
Recipe cost increased to 880 from 680
Total Cost increased to 2300g from 2100g
Escalated quickly :D And they increased hp amount of jungle monsters ? Goddamn. This needed to be done though, it never needed the price reduction in the first place.
I didn't mind morello's cost tbh, the problem was there were no alternative at that pricepoint that is worth getting for most mages
|
wtf that raist guy that was a "diamond smurf" on these forums a couple months ago has 1500 viewers??? how did that happen
|
Bearded Elder29904 Posts
On December 16 2014 16:07 Slusher wrote:
wtf that raist guy that was a "diamond smurf" on these forums a couple months ago has 1500 viewers??? how did that happen
151-156 Silver II, lel.
|
I solved it, I remembered how to internet
+ Show Spoiler +https://twitter.com/LiveBotDetector/status/544751320415354880
|
On December 16 2014 16:00 wei2coolman wrote:Show nested quote +On December 16 2014 15:40 JazzVortical wrote:On December 16 2014 15:14 739 wrote:
Well, yeah.
Morellonomicon
Recipe cost increased to 880 from 680
Total Cost increased to 2300g from 2100g
Escalated quickly :D And they increased hp amount of jungle monsters ? Goddamn. This needed to be done though, it never needed the price reduction in the first place. I didn't mind morello's cost tbh, the problem was there were no alternative at that pricepoint that is worth getting for most mages
Because alternatives are costed more appropriately (Grail). 2100 is stupidly cheap.
|
On December 16 2014 16:35 JazzVortical wrote:Show nested quote +On December 16 2014 16:00 wei2coolman wrote:On December 16 2014 15:40 JazzVortical wrote:On December 16 2014 15:14 739 wrote:
Well, yeah.
Morellonomicon
Recipe cost increased to 880 from 680
Total Cost increased to 2300g from 2100g
Escalated quickly :D And they increased hp amount of jungle monsters ? Goddamn. This needed to be done though, it never needed the price reduction in the first place. I didn't mind morello's cost tbh, the problem was there were no alternative at that pricepoint that is worth getting for most mages Because alternatives are costed more appropriately (Grail). 2100 is stupidly cheap.
I would say the closest equivalent is mpen boots+haunting guise.
|
On December 16 2014 15:39 Sufficiency wrote:Show nested quote +On December 16 2014 15:08 Goumindong wrote:On December 16 2014 14:20 Sufficiency wrote:On December 16 2014 14:02 Goumindong wrote:On December 16 2014 12:42 Sufficiency wrote:
Not really. It's not a dick measuring contest.
I think Goumindong obviously has some expertise on econometrics that the rest of us here lack. I think it would be nice to see how an economist would conduct a study on "how important is gold lead at 10 minutes". I think there is a lot to be learned here. The problem with the simple statistics is that its hard to disentangle the gold from the composition or the skill. So it really only tells us that we are winning and not whether or not we have a particular advantage. I am not sure there is a better way to answer the question though because actually getting a value on the gold is tricky. And because sometimes we don't mind a full estimate Also the only solution i can think of is not particularly feasible. Basically you take each game and regress gold difference as a time series. Then you take the values from that regression and use them for any other regressions you want(summary statistics will give you a broad idea of how powerful any gold lead snowballs, you should be able to see if the rate of gold change has any effect on win%, and i think once you had the rate of gold change and the effect of gold itself you could find some sort of independent effect on win% but not sure, differences in champions in the game would be interesting and so forth). Ideally you would want to use a vector auto regression (on both sides gold and xp totals) but a simple time series on the lead would probably be just fine. I haven't actually looked at time series in a long time so am not sure precisely how you would want to set it up (i know the base model you would want to run of G_+1=B*G_0 +e won't work because B will too high and breaks the math. The variable has to be stationary and I am not sure precisely what I would want to run without some more thought) and i know there will heteroskedasticity problems that I am not quite sure how to solve in a time series case though i think its equivalent to the normal case. Sample size of the time series doesn't matter much since we have hundreds of thousands of games. Fortunately the Heteroskedasticity problems won't be an issue for the larger set since the main thing it does is widen your error bars when corrected for and we are either working in a situation where we don't care about error bars or because we ran 200k regressions they're so small it doesn't matter whether you used HC corrected standard deviations or not. The most challenging issue would probably be structural breaks. Games in which structural breaks occur could potentially break the math (again), as could any game with particularly wide gold swings. I don't know of a way to systematically examine for structural breaks. Though if such a method did exist it could be an interesting way to look at champion power (basically if you have a champion which has a strong lategame or early game there may be a point in which the structure of the game breaks, by examining the distribution of these breaks you could figure out when champions had power spikes more accurately than just looking at their win rates by time. Anywho, here is something you could try that wouldn't be difficult computationally. Take a champion and look at the density of their game lengths as well as the win % as the lengths. I suspect that you will find troughs in density at or right before a champions powerspike (for champions with powerspikes post 25 minutes or so) and peaks when they fall off with the win % basically following those numbers You really don't want to include win/loss, do you? I think time series is interesting, but remember when you win the game you almost always have a big gold spike (because you probably aced the other team then took some towers) so I am not sure how the end point will be like. I think the last term will correlate with win/loss so strongly that one can substitute for the other. At the end of the day you might be looking at the same thing. You could just exclude the last term. I would actually love to include win/loss but its tying the effect that is hard, unless you don't want a measure of advantage from gold and just want a measure of advantage when you have that much gold. I just don't think the second question is that interesting, you can't really extrapolate for a fight due to gold advantage more than you can from just looking at items/levels. There are similar problems measuring the effects of items (basically you capture the effects of winning and losing if winning/losing effects your item choice) but looking at items might be a better path OK I think at the end I want to know what you would do to find this instrumental variable. Because, frankly, ditching winning/losing is not acceptable for me since: 1. At the end of the day, winning/losing is all that matters. Gold growth/etc is just an intermediate step. 2. If you are OK removing the last term, you probably want to remove the second last term too, because winning/losing correlates with the last term, and the last term correlates with the second last term - the argument becomes recursive. It's not clear to me where the cut-off is for a "proper" model by your standard.
I don't have one. If I had one I would have told you it 
1) fair but also not really interesting. Though a tally across patches would be
2) the reason you remove the last term wasn't because of win/loss correlation but because you think the variance on the last term is significantly higher than prior terms (your reasoning by saying an ace was likely) and so we can get an idea of the series better by omitting it
|
|
|
On December 16 2014 16:46 wei2coolman wrote:Show nested quote +On December 16 2014 16:35 JazzVortical wrote:On December 16 2014 16:00 wei2coolman wrote:On December 16 2014 15:40 JazzVortical wrote:On December 16 2014 15:14 739 wrote:
Well, yeah.
Morellonomicon
Recipe cost increased to 880 from 680
Total Cost increased to 2300g from 2100g
Escalated quickly :D And they increased hp amount of jungle monsters ? Goddamn. This needed to be done though, it never needed the price reduction in the first place. I didn't mind morello's cost tbh, the problem was there were no alternative at that pricepoint that is worth getting for most mages Because alternatives are costed more appropriately (Grail). 2100 is stupidly cheap. I would say the closest equivalent is mpen boots+haunting guise.
What?
How on earth are those two items an alternative to Morello? Neither occupy the same intended spot in your inventory (mana regen + CDR item) or fulfil the same role. I can't think of a single champ that would go for that item combination over Morello, except for like Rumble, and he isn't getting Morello anyway. The closest I can think of is Malz, and even then every single Malz I see goes for a mana item first.
They aren't anywhere near alternatives.
|
|
|
|
|
|
EPT
