The problem about the ratings, it does not say what you really should do and what you are doing wrong to get better/worse.
First day i iwas 700 and i lost a few games with my lp banned friends then it was 650. I started solo playing and had 3 great games, but it did not affect my score. Today, i dont really know how, my score is 85 = Copper ... From silver to, bronze, and now copper. How bad am i?
I was solo matchmaking months ago, was in "high bracket" in every 1 games of 2.
Then i recieved some invites and played with noobs, from that point, rarely reached to high bracket i guess. (it doesnt show anymore)
Every f**** game, i try to do my best, i see players like me, but 2 of em always are braindead. Feeding or quitting. In my team, or enemy team...
When i stack my wins like 5-6 games, my team mates start being idiots and enemy team with try hards. And i start losing like 10 games untill im %49 - %50 ... I was %51 once...
How will your rating system decide in those cases? http://dotabuff.com/matches/408064295 In this game i played safe and focused on my farm, yet my team refused to have BKB and farmed sblades instead. And we lost it.
http://dotabuff.com/matches/408339059 And in this game Naga suddenly told us "we will not win this one" and quit. I tried to hold the game but you rarely win 4 against 5...
It seems to me, if you farm, get kills, you will go higher. No supports can go higher compared to carry pickers.
Also, you need to be lucky about team mates if you want to seem "good" player whether if you are actually good or not.
@Sampson
The data is known and clearly correlates with players skill level. I will add additional signals soon but to say the data is meaningless is ignorant. Spend 30 minutes analyzing this problem and you may start seeing the picture.
The problem with using unknown data is the fact that you make huge sweeping assumptions.
For example I could be fighting Dendi on alt account with his account set to private and I will not get credit for stomping one of the top players in the world because his in game stats for that ONE game might seem low and therefore seem like a weaker player. Im sure your system would work IF there was no unknown data but the fact that almost 50% of the data is missing and you are just filling the gaps with assumptions leads the website to lose all credibility
Hmm. Considering i had high bracket games and won, Dota 2 high bracket players are not that really good, they cant even handle a "copper" player.
I will be waiting for additional signals from now on. If you create your own MMR system and ladder, you should be real clear about if you care and how much your algorithm cares about winning / kdas and also topics like hero performance pool, how many games you played etc.
Edit: Did not see your post, sorry. Thank you for your kind interest.
I think the problem is that people are focusing on being in "high" bracket or "very high" bracket previously and thinking their rating should be top 1%.
The high bracket was huge and is mainly average level players, even lower end very high players are not very good (relatively). You have to remember there are hundreds of thousands of dota players, being top 10% still means there are probably 75k to 100k people better than you.
The system seems spot on for me. I solo queue top 1-10 pages, against very high end players. My friend that I consider slightly better than me is about 1620 to my 1560 which seems pretty damn accurate.
I'm also 49% wr overall, I had to teach myself dota and lost a ton early on. Built myself up to queueing vs top tier, so overall winrate is not the end all be all here. At least for me this system very accurately puts me where I am on Dota's internal system.
You are assuming I can't approximate Dendi's alt account skill level. This assumption is wrong. Each players performance in any given match is counted individually. If you played against Dendi, it will show in the replay and algorithm will grant you more points as you've played against a Dendi level player.
no you really can not guess any persons skill level based off ONE game the fact that you believe that you can just shows how your flawed system came about
PS Sampson: I just checked your dotabuff and who you play against. Haven't seen any pros or first pagers - only 49-51% winrate opponents and people who get 3500 in TMM rating - doesn't look like a first pager to me.
^
Sampson is right, the fact that you ignore my post indicates that you are well aware thatn your algorithm is flawed, thats the reason you hide it. Also Just looking at your toplist shows, that it is completely flawed in every way.
There're smurf accounts with a 55% winrate which have a higher rating than professional players, which is completely laughable.
you keep making the assumption that if you are a good player you HAVE to have your profile taken off private. Thats all you keep doing making assumptions and trying to pass them off as facts
Yes, people need to calm down. I think this is why Valve doesn't want a MM system - because this turns the community into something that can be seen in this thread. This is a game and should be fun.
PS Sampson - I am really curious - do you really play on 1st page as you said? Because I really don't see any good opponents in your last matches.
@MyDotaSkill.com #650 DSR:
Im studying computer science, i guess that should answer your question.
Also, you seem not to have much experience when it comes to the very high bracket players. There're many people with a winrate ranging from 49-51% on page 1. The winrate doesnt indicate the skill or bracket, i have also seen a 41% winrate guy on page 1 already. When you know how ELO works, then you should also know that winrate doesnt say much.
Also, Sampson does not always play on page 1, it happens rather rarely, but his MMR is rather high.
If Sampson is queuing top pages (not saying he doesn't) he should have previous matches against known stacks - warden, taki, guanzo, etc. I ran into 3 stack of EG (fear mss jeyo) and it was first game 2nd page two days ago, not even top page.
I entered my profile in the queue yesterday, it was 600 or something, today I checked again and it is 5700. Is something wrong? Have I just got to wait?
I am not mad at all, every single post you have been trying to turn this into 'Sampson do you not like your rating?'
I have told you countless times where exactly your system is flawed yet you keep trying to change the topic. If you want to say that you take your opponents skill level into account when you are ranked I will continue to advocate against your misleading system. As long as 40%+ of the data from games is missing and you are just doing guesswork you should make that clear
@Blunt - strange because I've just went through his profile and all of his opponents seem to have around 3.5k in TMM rating ( that is calculated on Valve's internal Elo rating ).
3.5k doesn't seem to be the top level.
Besides, he scored above 1000 - and anything above 1000 can be considered to be pro level.
^Have to gone through all of the unknown players data or are you just making up numbers again? Are you going to answer the question or are you going to continue to try and dodge the issue?
One thing im noticing on Sampsons profile is that 70% of the players on average are anon.
Just clicking through my last 7 games 8/10 players on average are reported in my games, and those are mainly top 1-7 pages. I've never seen someone queuing high end with so many players anon.
I've told you two times and I will say this last time: I don't care about profile privacy. How you play is all in the replay. It doesn't matter if Dendi plays on his own account or on account of your grandmother, he will display the same characteristics.
Did you understood the third time? Or it's still too hard?
@MyDotaSkill.com #650 DSR:
I have seen him on page 1 once, so he is not completely wrong. Since this is a team game you cannot really take their TMM into account, since if people don't synergize well, then the result wont be as good as it should be.
Also i don't think you understand Sampson's point. He doesnt question your system at this point, but more the data your gathering. His main issue here is, that he's concerned about wroing data which is used to predict a score.
So you are personally watching each replay? No you are taking the data from the match history and that can only show so much. You will only get ONE match from the unknown player so trying to find a trend in ONE match? as I said before please tell me which player was in the data I linked (keep in mind I am leading you into a trap but you know exactly which trap I am leading you into which is why you want to dodge this one)
The entire world is a bunch of unobservable data. If you actually believe you cannot extract meaningful inferences from incomplete data sets, then I feel very bad for you (if you are in a technical field). If you're not in a technical field, then I would kindly ask you to stop pontificating about something you never studied.
@the creator of the system
I understand why you are choosing a variable rating, it is an easy way to deal with the extremely challenging problem of people partying up. But I think there has to be a better way to determine it. Off the top of my head, you could disregard stacked games outright by trying to detect people who queue together a whole bunch (seems unfeasible computationally) or you could try to detect whether each player "held his own" in the games before rewarding him the full amount of points associated with it. The next task you have (now that you have avoided the partying problem in theory) is to stabilize the rating. Here is how I would do it:
Have the incremental function g(last rating, last new game) that finds a new rating with a single additional game, and initialize everyone to a starting rank. Each player has a term "e" associated with them. Then, when someone plays a game,
target rank = g(last rank, last new game)
deviation = |target rank - rank|
e = adjust_e(e, deviation)
rank = (target rank - rank)*e + rank // simple update rule
e should be between 0 and 1. As your confidence increases, you adjust their ranks more slowly. adjust_e( - ) should most likely have memory in that it should take several odd matches all tending in one way (overapproximating/underapproximating) to increase it a lot.
I still believe the system's broke. My rating went down from 1641 to 1547 and I still think that's about 800 points too high. I'll see how it adjusts over the course of the next few days and see if it levels out.
There seems to be a pretty big bug with gathering data. I think non-ap matches get wrongly fetched. That's why some good players get bad scores. I will fix it as soon as I find the source
@Vandal
Interesting, finally someone who writes on topic! Hit me up on steam when you have time.
"this site will die out within a month or so like all the previous ranking sites"
Yep, i guess this is just for fun,i dont think anyone is taking this seriously.
I am pretty sure this is really far away of representing people real skill, a lot of <1000 are a lot better than the ones rated higher.
If you want to know if you are good or bad just go and ask someone who you know is really good. Simple.
Anyway i give the creator credit for doing all of this, just wait for valve to release more info i guess if you want make this right. Now, i think you should try to make the queue go faster because right now is really slow.
This shows how desesperate the comunity is for being rated, valve should look how much importance people is giving for a ramdon system and act in consequence.
@ MyDotaSkill I guess I can wait, I have 4/20 games that are with friends who are 2-3 wins only. That might skew it. But I doubt I'm gonna get a huge increase in rating? FYI I went 2/4 out of those games. Just wondering if you ever thought about say a rating like on heroes. Say one of each role. Get the average stats on carry, support, ganker roles. rank people accordingly to a set % above average etc..
example: Antimage
W L K D A gold LH Deny GPM
90 13 5 9 10 9.6k 95 19 454
So 1000 rating should do 100% of that; 1200 should do 130%, 1400 should do 150%, 1600 do 175%, 1800 do 200%
Do most played Suport & ganker also then average them. Also add wards bought OFC with a HIGH value!
I'm not in cs but I have a few thoughts about this topic:
I used to play in www.dotacash.com, I think it is the biggest platform in dota 1 before www.throneit.com died. Dotacash first used an ELO rating where it first simply calculates by win rate, + 15 if you win and - 15 if you lose, under the condition that total elo from both teams are equal.
From the public rants about the ELO system not being viable, dotacash spent a long time inventing a new method, called "true skill rating", where it pretty much prevented stacking problems because if they stacked, they would win very little TSR or even lose some TSR.
Dotacash had t1, t2, t3 respective tiers which was alright, as you move up in tiers the skill difference becomes drastic, especially between t2 and t3, t1 and t2 is not much but still a noticeable difference at times.
They have their algorithm on their website I believe, maybe you can check that out.
As for your own rating system, I think it's pretty good. For that I think it is probably near impossible to create a system so accurate (within 5% error) in calculating a player's skill. I believe your system relatively good at estimating a player's general performance within the last 20 games, but I think perhaps increasing the number of games would be abit more accurate. ~100 would be ideal.
I personally think the percentage of total kills is a huge influence on the individual's skill level, this does not matter if you are a carry or a support, if you have been involved in 75% of the team's kills then you are playing the game right, wouldn't matter if you have died the most on your team if you are a support such as cm. This perhaps would not apply for a heros such as spider.
I really wish dota 2 had a tier system, or even a vouch league. I wouldn't say the match making system is "broken" right now, but there are too many 1 sided games which I believe it should not happen if both teams are balanced.
Sampson is just being picky to the needle about the precision of the stats, it will never be as accurate as what he asked. Sure dendi can play on alt and perform poorly in one game, that's fine. The reality is, dendi played poorly in that game, hence it will affect his rating, but those games will probably occur rather rarely, so it wouldn't affect dendi's skill level.
I strongly believe stats is a long-term calculation, short term stacks wouldn't mean as much since there can be losing/winning streaks. I personally think 20 games are way too little, or perhaps provide a "life time rating" and a "past 20 games" or "past 100 games" etc rating, would be much more ideal.
i actually think the system works... from what ive read (your DSR), i can see it being accurate because you all suck anyway
first page my ass try to play well first remember its fucking volvo MM 1st page we are talking about, there is bound to be stupid fucks there (i'd say 30% which in this case includes you guys).
i also think they may be another bug like the one fixed for which low ratings (as my 774 before bug fix) can get their accurate rating (Now 1796)
The problem about the ratings, it does not say what you really should do and what you are doing wrong to get better/worse.
First day i iwas 700 and i lost a few games with my lp banned friends then it was 650. I started solo playing and had 3 great games, but it did not affect my score. Today, i dont really know how, my score is 85 = Copper ... From silver to, bronze, and now copper. How bad am i?
I was solo matchmaking months ago, was in "high bracket" in every 1 games of 2.
Then i recieved some invites and played with noobs, from that point, rarely reached to high bracket i guess. (it doesnt show anymore)
Every f**** game, i try to do my best, i see players like me, but 2 of em always are braindead. Feeding or quitting. In my team, or enemy team...
When i stack my wins like 5-6 games, my team mates start being idiots and enemy team with try hards. And i start losing like 10 games untill im %49 - %50 ... I was %51 once...
How will your rating system decide in those cases? http://dotabuff.com/matches/408064295 In this game i played safe and focused on my farm, yet my team refused to have BKB and farmed sblades instead. And we lost it.
http://dotabuff.com/matches/408339059 And in this game Naga suddenly told us "we will not win this one" and quit. I tried to hold the game but you rarely win 4 against 5...
It seems to me, if you farm, get kills, you will go higher. No supports can go higher compared to carry pickers.
Also, you need to be lucky about team mates if you want to seem "good" player whether if you are actually good or not.
@Sampson
The data is known and clearly correlates with players skill level. I will add additional signals soon but to say the data is meaningless is ignorant. Spend 30 minutes analyzing this problem and you may start seeing the picture.
the " How bad i am ? " made me laugh, didnt read rest tho
@Loverman
Your rating is lowered because you stack with bad players. Go play alone or with people from your skill level and you will get accurate reading.
The problem with using unknown data is the fact that you make huge sweeping assumptions.
For example I could be fighting Dendi on alt account with his account set to private and I will not get credit for stomping one of the top players in the world because his in game stats for that ONE game might seem low and therefore seem like a weaker player. Im sure your system would work IF there was no unknown data but the fact that almost 50% of the data is missing and you are just filling the gaps with assumptions leads the website to lose all credibility
Hmm. Considering i had high bracket games and won, Dota 2 high bracket players are not that really good, they cant even handle a "copper" player.
I will be waiting for additional signals from now on. If you create your own MMR system and ladder, you should be real clear about if you care and how much your algorithm cares about winning / kdas and also topics like hero performance pool, how many games you played etc.
Edit: Did not see your post, sorry. Thank you for your kind interest.
I think the problem is that people are focusing on being in "high" bracket or "very high" bracket previously and thinking their rating should be top 1%.
The high bracket was huge and is mainly average level players, even lower end very high players are not very good (relatively). You have to remember there are hundreds of thousands of dota players, being top 10% still means there are probably 75k to 100k people better than you.
The system seems spot on for me. I solo queue top 1-10 pages, against very high end players. My friend that I consider slightly better than me is about 1620 to my 1560 which seems pretty damn accurate.
I'm also 49% wr overall, I had to teach myself dota and lost a ton early on. Built myself up to queueing vs top tier, so overall winrate is not the end all be all here. At least for me this system very accurately puts me where I am on Dota's internal system.
@Sampson
You are assuming I can't approximate Dendi's alt account skill level. This assumption is wrong. Each players performance in any given match is counted individually. If you played against Dendi, it will show in the replay and algorithm will grant you more points as you've played against a Dendi level player.
I solo que into page 1-2 and I have seen others as well and they end up in the 800-1200 range
Right now all the 'system' is doing is guessing
@MyDotaSkill.com #650 DSR
no you really can not guess any persons skill level based off ONE game the fact that you believe that you can just shows how your flawed system came about
Do you stack?
Your skill level is approximated from 20 games, not one.
No the unknown players in a game you will only get data from ONE game
L K D A gold LH Deny GPM
13 5 9 8 9.6k 95 19 254
these are the stats from one game of a player who is one of the best in the world
PS Sampson: I just checked your dotabuff and who you play against. Haven't seen any pros or first pagers - only 49-51% winrate opponents and people who get 3500 in TMM rating - doesn't look like a first pager to me.
^
Sampson is right, the fact that you ignore my post indicates that you are well aware thatn your algorithm is flawed, thats the reason you hide it. Also Just looking at your toplist shows, that it is completely flawed in every way.
There're smurf accounts with a 55% winrate which have a higher rating than professional players, which is completely laughable.
@Mydotaskill
you keep making the assumption that if you are a good player you HAVE to have your profile taken off private. Thats all you keep doing making assumptions and trying to pass them off as facts
And you are assuming I care about profiles being private.
L K D A gold LH Deny GPM
13 5 9 8 9.6k 95 19 254
tell me which player that data came from
come on I can wait
Blunt @ Fnatic do you have experience in machine learning?
Because this isn't A+B=C/2*0.3434
If you have some expertise in this, then why not.
Calm down, bitches! I think this rating is just for fun.
Yes, people need to calm down. I think this is why Valve doesn't want a MM system - because this turns the community into something that can be seen in this thread. This is a game and should be fun.
PS Sampson - I am really curious - do you really play on 1st page as you said? Because I really don't see any good opponents in your last matches.
@MyDotaSkill.com #650 DSR:
Im studying computer science, i guess that should answer your question.
Also, you seem not to have much experience when it comes to the very high bracket players. There're many people with a winrate ranging from 49-51% on page 1. The winrate doesnt indicate the skill or bracket, i have also seen a 41% winrate guy on page 1 already. When you know how ELO works, then you should also know that winrate doesnt say much.
Also, Sampson does not always play on page 1, it happens rather rarely, but his MMR is rather high.
@mydotaskill.com
can i possibly get mine?
If Sampson is queuing top pages (not saying he doesn't) he should have previous matches against known stacks - warden, taki, guanzo, etc. I ran into 3 stack of EG (fear mss jeyo) and it was first game 2nd page two days ago, not even top page.
Jeyo had his accoutn anon for this one too
http://dotabuff.com/matches/405835651
I entered my profile in the queue yesterday, it was 600 or something, today I checked again and it is 5700. Is something wrong? Have I just got to wait?
@MyDotaSkill.com #650 DSR
I am not mad at all, every single post you have been trying to turn this into 'Sampson do you not like your rating?'
I have told you countless times where exactly your system is flawed yet you keep trying to change the topic. If you want to say that you take your opponents skill level into account when you are ranked I will continue to advocate against your misleading system. As long as 40%+ of the data from games is missing and you are just doing guesswork you should make that clear
@Blunt - strange because I've just went through his profile and all of his opponents seem to have around 3.5k in TMM rating ( that is calculated on Valve's internal Elo rating ).
3.5k doesn't seem to be the top level.
Besides, he scored above 1000 - and anything above 1000 can be considered to be pro level.
^Have to gone through all of the unknown players data or are you just making up numbers again? Are you going to answer the question or are you going to continue to try and dodge the issue?
One thing im noticing on Sampsons profile is that 70% of the players on average are anon.
Just clicking through my last 7 games 8/10 players on average are reported in my games, and those are mainly top 1-7 pages. I've never seen someone queuing high end with so many players anon.
I've told you two times and I will say this last time: I don't care about profile privacy. How you play is all in the replay. It doesn't matter if Dendi plays on his own account or on account of your grandmother, he will display the same characteristics.
Did you understood the third time? Or it's still too hard?
End.
@MyDotaSkill.com #650 DSR:
I have seen him on page 1 once, so he is not completely wrong. Since this is a team game you cannot really take their TMM into account, since if people don't synergize well, then the result wont be as good as it should be.
Also i don't think you understand Sampson's point. He doesnt question your system at this point, but more the data your gathering. His main issue here is, that he's concerned about wroing data which is used to predict a score.
@Blunt
Valve's TMM is directly tied to avg team players Elo rating. You can't change it by much even if you win 50 times.
So you are personally watching each replay? No you are taking the data from the match history and that can only show so much. You will only get ONE match from the unknown player so trying to find a trend in ONE match? as I said before please tell me which player was in the data I linked (keep in mind I am leading you into a trap but you know exactly which trap I am leading you into which is why you want to dodge this one)
i dont know what your fucking problem is
nobody remotely decent at this game has or will ever take this seriously
this site will die out within a month or so like all the previous ranking sites
@MyDotaSkill.com #650 DSR:
True, but the rating has nothing todo with it.
Also, i need to know something, might be interesting, also this will prove Sampson's point:
http://dotabuff.com/matches/407899231
pls tell me every DSR, from every player.
winterchillz 1080
SKYNET 262
Mipsu Lii 1299
AznFaiL 1077
Zupport 234
^
i meant from the matchlink...
@Sampson
The entire world is a bunch of unobservable data. If you actually believe you cannot extract meaningful inferences from incomplete data sets, then I feel very bad for you (if you are in a technical field). If you're not in a technical field, then I would kindly ask you to stop pontificating about something you never studied.
@the creator of the system
I understand why you are choosing a variable rating, it is an easy way to deal with the extremely challenging problem of people partying up. But I think there has to be a better way to determine it. Off the top of my head, you could disregard stacked games outright by trying to detect people who queue together a whole bunch (seems unfeasible computationally) or you could try to detect whether each player "held his own" in the games before rewarding him the full amount of points associated with it. The next task you have (now that you have avoided the partying problem in theory) is to stabilize the rating. Here is how I would do it:
Have the incremental function g(last rating, last new game) that finds a new rating with a single additional game, and initialize everyone to a starting rank. Each player has a term "e" associated with them. Then, when someone plays a game,
target rank = g(last rank, last new game)
deviation = |target rank - rank|
e = adjust_e(e, deviation)
rank = (target rank - rank)*e + rank // simple update rule
e should be between 0 and 1. As your confidence increases, you adjust their ranks more slowly. adjust_e( - ) should most likely have memory in that it should take several odd matches all tending in one way (overapproximating/underapproximating) to increase it a lot.
I'm 262. 2400+ games, 50.5% winrate...I blame ELO HELL.
profile: http://dotabuff.com/players/47400225
more accurate profile: http://pubstats.me/47400225/
start of DBR: 1600-1700 end of DBR: 1300s
my estimated rating? 450-900.
Does holding 52% winrate til 1000 games count? How about 51% for 2000 games?
Also keep in mind 0-0 players start off at 1500?
I still believe the system's broke. My rating went down from 1641 to 1547 and I still think that's about 800 points too high. I'll see how it adjusts over the course of the next few days and see if it levels out.
@Skynet:
lol, and i thought my rating was low, hahahahahahahahaha
@SKYNET
There seems to be a pretty big bug with gathering data. I think non-ap matches get wrongly fetched. That's why some good players get bad scores. I will fix it as soon as I find the source
@Vandal
Interesting, finally someone who writes on topic! Hit me up on steam when you have time.
Guys i programmer 5 year experience in html and notepad+ i also graduated from google in 2012
Can you check mine? I've been waiting for a really long time now and the queue seems to be longer.
"this site will die out within a month or so like all the previous ranking sites"
Yep, i guess this is just for fun,i dont think anyone is taking this seriously.
I am pretty sure this is really far away of representing people real skill, a lot of <1000 are a lot better than the ones rated higher.
If you want to know if you are good or bad just go and ask someone who you know is really good. Simple.
Anyway i give the creator credit for doing all of this, just wait for valve to release more info i guess if you want make this right. Now, i think you should try to make the queue go faster because right now is really slow.
This shows how desesperate the comunity is for being rated, valve should look how much importance people is giving for a ramdon system and act in consequence.
@ MyDotaSkill I guess I can wait, I have 4/20 games that are with friends who are 2-3 wins only. That might skew it. But I doubt I'm gonna get a huge increase in rating? FYI I went 2/4 out of those games. Just wondering if you ever thought about say a rating like on heroes. Say one of each role. Get the average stats on carry, support, ganker roles. rank people accordingly to a set % above average etc..
example: Antimage
W L K D A gold LH Deny GPM
90 13 5 9 10 9.6k 95 19 454
So 1000 rating should do 100% of that; 1200 should do 130%, 1400 should do 150%, 1600 do 175%, 1800 do 200%
Do most played Suport & ganker also then average them. Also add wards bought OFC with a HIGH value!
Also did you see how they did pubstats.me ?
@mydotaskill
I'm not in cs but I have a few thoughts about this topic:
I used to play in www.dotacash.com, I think it is the biggest platform in dota 1 before www.throneit.com died. Dotacash first used an ELO rating where it first simply calculates by win rate, + 15 if you win and - 15 if you lose, under the condition that total elo from both teams are equal.
From the public rants about the ELO system not being viable, dotacash spent a long time inventing a new method, called "true skill rating", where it pretty much prevented stacking problems because if they stacked, they would win very little TSR or even lose some TSR.
Dotacash had t1, t2, t3 respective tiers which was alright, as you move up in tiers the skill difference becomes drastic, especially between t2 and t3, t1 and t2 is not much but still a noticeable difference at times.
They have their algorithm on their website I believe, maybe you can check that out.
As for your own rating system, I think it's pretty good. For that I think it is probably near impossible to create a system so accurate (within 5% error) in calculating a player's skill. I believe your system relatively good at estimating a player's general performance within the last 20 games, but I think perhaps increasing the number of games would be abit more accurate. ~100 would be ideal.
I personally think the percentage of total kills is a huge influence on the individual's skill level, this does not matter if you are a carry or a support, if you have been involved in 75% of the team's kills then you are playing the game right, wouldn't matter if you have died the most on your team if you are a support such as cm. This perhaps would not apply for a heros such as spider.
I really wish dota 2 had a tier system, or even a vouch league. I wouldn't say the match making system is "broken" right now, but there are too many 1 sided games which I believe it should not happen if both teams are balanced.
Sampson is just being picky to the needle about the precision of the stats, it will never be as accurate as what he asked. Sure dendi can play on alt and perform poorly in one game, that's fine. The reality is, dendi played poorly in that game, hence it will affect his rating, but those games will probably occur rather rarely, so it wouldn't affect dendi's skill level.
I strongly believe stats is a long-term calculation, short term stacks wouldn't mean as much since there can be losing/winning streaks. I personally think 20 games are way too little, or perhaps provide a "life time rating" and a "past 20 games" or "past 100 games" etc rating, would be much more ideal.
Just my two cents.
i actually think the system works... from what ive read (your DSR), i can see it being accurate because you all suck anyway
first page my ass try to play well first remember its fucking volvo MM 1st page we are talking about, there is bound to be stupid fucks there (i'd say 30% which in this case includes you guys).
i also think they may be another bug like the one fixed for which low ratings (as my 774 before bug fix) can get their accurate rating (Now 1796)
have 96... i will go kill myself :(