http://www.qlranks.com/duel/player/dnt_b_scared = a1r_r41d

Wrong. I share that account with my brother so its not accurate.

Won't use it then. I'd probably add you with default parameters :)

Well :) I'm going to put on about 1 month of practice before DH and then another month before qcon :) so i will have a lot of games played :))

I'm pretty sure that u can already add Cypher and Av3k to that graph.

Will do in an update when the last players are announced. I'd probably also update it when the tournament starts and players have been training for a bit.

Interesting!

Just one thing, this might be due to my monitor/eyes, but the colour-coding could stand to be a bit clearer (wider range of colours). I can't really tell exactly what colour most of them are.

I'll give you a better reply later, but what would you suggest as a gradient? Blue -> Red? Or something else entirely... blue to orange perhaps .. or a three color range: green->yellow->red or something such.

Your choice :)

Thanks for the response :)
As far as readibility goes, the more colours the better! But I guess if you used the whole spectrum it might not be intuitively obvious that, say, green came between blue and red. So something like Green - Red might be best for a combination of contrast and clarity.

Green - Red is bad for about 12% of all males, due to color blindness. And I do believe ESR is mostly males :P I'll go with Blue -> Red then.

Ah yeah, good point, red-green colour blindness is the common one. (I guess I have blue-blue colour blindness :p)

someone's studying statistics, rite? :) - had exactly the same stuff this year and it's actually brilliant to see how it can be applied to Quake :)

Not Stats :) I am kind of like a polymath, spanning Computer Science, Formal Logic, Math and Statistics - but I am strongest in CS.

But yeah, it is cool that you can apply statistical knowledge to Quake. It is also one of the only games where you can have half a million matches as a corpora of data to learn from.

Nice work, but lol @ describing yourself as a "polymath"

Please do update the graph as new players enter though. And if you felt like sharing some of your data & methodology, I'd certainly be interested.

Shouldn't the right side of the mid point be shorter as the longer you don't play Quake Live your max skill limit should decrease a tiny bit. For example you are saying draven could beat anyone there but theres no way that is going to happen.

In principle: yes. There is no reason to believe that a player who did not play becomes better all of a sudden.

But the wide horizontal line is used by the system to estimate how much it believes in the strength of the player. When a player is "new" to the system, we may have obtained an estimate, but we don't really know if it is the real strength. We represent this uncertainty as a separate parameter in the system, namely the confidence of the strength. In this case, it makes sense to say that the player might be better or worse - we don't know.

fair enough lol

i think ur calling me shit but im still gonna plus u

Hey, you may be better than Cooller according to this. You are in the same 95% confidence interval. Same with Party1e - you may be as good as.

However, the system estimates that both of these guys strength is probably above yours. Coollers large range is due to him not playing - but he still keeps his relative strength from where he was.

i wonder how is this estimated, since i've already beaten almost everyone in that list who supposedly i still have no chance even if im playing my best, and they'r playing their worse :|

I think the width of the horizontal line represents uncertainty about their 'true' skill level, rather than the range of levels that they could possibly be at for any single game.

now u've mindfucked me :<

:p
Basically, the wider your horizontal line, the less certain the authors are about their estimate of your skill level.
But to see how likely they think you are to beat better players/lose to worse players, look at the colour, not the width.
If the right-hand side of your bar doesn't make it to the left-hand side of another player's, that means the other guy is almost certainly (99.75%+ confidence, if I interpret it properly) better than you overall, according to the authors of the graph. It doesn't mean that his worst game is better than your best.

u've made it worse now :|

Haha, I'll give up then :p

tl;dr: You suck, but [edit: probably] not as much as Draven.

I'll give it a try:

The horizontal lines ...

... don't tell you how good or bad someone can play when he's at his best or at his worst.
... don't show who you can beat when your line overlaps someone else's line.

... tell you in which interval your strength is with 95% certainty.
... tell you that your strength could possibly be as high as someone else's if your lines overlap.

... don't estimate your chance of beating someone with a higher estimated strength.
That's an entirely different calculation and is represented by the color of your line. Your color is pretty light, which reflects the fact that you've in fact beaten some players with a higher estimated strength than yours.

Let me add to the last point: You've beaten them consistently. That is, you beat them often. How often is often you might ask. The answer is: I don't know yet, but I have a 6 hour calculation in front of me to tune the system so it predicts well.

The opposite is also true: If you lose to worse players consistently you will be lighter blue as well.

Yep, I've left out an adverb like "consistently" or "often" because I couldn't judge how high is volatility actually is. The fact that it can get high due to losses against worse players hinders the interpretation as a dark horse, though.

You're not thinking it through. If you just lost to worse players your strength would be lower and you wouldn't actually lose to worse players.

At the same time if you just beat good players occationally your strength would simply be higher. You could only possibly be a dark horse if you both lose to worse players and beat better ones because else it would just show an accurate strength.

I agree with you.

Maybe this is mainly a problem of the definition of 'dark horse'.
To me, a dork horse is a player with the ability to beat players which are deemed better than him -- and who is NOT equally likely to lose against players which are deemed worse than him.

This "one-sided" definition of dark horse (stable with frequent peaks upwards) isn't reflected in a statistical variable, only the "two-sided" one (frequent peaks upwards and downwards).

http://www.youtube.com/watch?v=RaOb7I4GKNg

Correct. It maps their uncertainty. Ins just need to win a bit more over the better players and his level would increase, just like an ELO in QLranks.

In fact, the system I am using becomes ELO if we set the uncertainty to 0.

Shouldn't the "strength" equal to the qlranks ELO then? Or does the difference originate from the fact that you only track since february?

There are four reasons it does not match up:

I have about 6 months less worth of data so I have less knowledge about the skill of individual players.

I don't start players at 1200 but 1500 points.

The strength scale is not the same, so you can't say that 1200+800 = 2000 ELO is the same as 1500+800 = 2300 ELO. In fact, it gets gradually harder to obtain higher rating the closer you are to the top players.

The system is not ELO. When recalculating the new rating of a player, the current uncertainty of the player is taken into account. So is the uncertainty of all opponents. Finally, the players own volatility plays a role as well. Volatility is a measure that makes the system able to cope with players who are improving faster than what a rigorous system can handle. p0ni is a good example since he is often on the "move" in rating. Hence, you would argue he is kind of a dark horse.

loldraven

I am not sure it is lol :) draven has not played a lot. If he starts playing, then by the time I reupdate this, his rating will have improved a lot. Since he has a high uncertainty, his rating can improve quickly.

So the players with a high volatility (i.e. a light blue color) are the dark horses.

GO WEIRD!

Yep, but the volatility rating here is fairly low for all players. 0.0508 is what most of these players are at, but if you look at the whole set of 31000 players, there are players with 0.0530. Small numbers, but the difference is pretty significant.

I would be deep red since I loose to any scrub I face

There is no red in this ... yet... :)

Its because im not playing!

A bit on how the ranking works:

Let us assume that each player has a real strength. Statisticians like to call that the "parameter" so the parameter of a player is the real strength of that player. Now, we can try to estimate the parameter of a player by assigning him a number, which is the strength of that player. The problem is that we will never know what the parameter is. Even if rapha beat Cooller 3 times in the IEM5 finals, it may still be the case that cooller is better than rapha, but rapha was just fortunate and lucky.

So we opt for an estimation. We look at games the player played in the past and use this to estimate where the parameter is. For simplicity, we only look at a single thing: win/loss. We don't look at item control, nor at accuracy or damage dealt. We don't look at the scores either. Just win/loss. By building a statistical system around this, we can estimate the ranking of a given player.

One such system is ELO, which is what QLranks uses. The system used above is called Glicko2. Glicko2 is hard to calculate by hand, but easy to calculate by machine - when ELO was created hand creation was an important factor. The web page glicko.net links to the relevant descriptions of the model.

The major difference to ELO is two things: First, we track the rating or "strength" of a player. Second, we track an uncertainty of that strength, usually called the "rating deviation" (RD). Players start out with a high RD and as they play games, we lower their RD. If a player is not playing for a period the RD is slowly increased. In ELO, all point transfers are symmetric: If you gain 5 points, the opponent lose 5 points. In Glicko2, it is asymmetric. The reason is that the RD is factored into the result: losing to a worse player with a high RD could just be a sign that the player is not yet shuffled to the right strength, so it won't affect your rating as much. New players start with a very high RD value, so that means they will take large strides when they are updated with a new strength. It makes the system able to quickly shuffle new players to their right strength ballpark.

The third value, the "stochastic volatility" or just V is inspired by typical observations from finance: There may be sudden changes which can not be explained. The volatility is a parameter which is tracked by the system. If a player "fools" the rating system consistently in a way so Glicko2 makes the wrong prediction - then that player is considered to be volatile. Volatility affect the players RD when considering a new rating. It means that p0ni is currently eligible to move more in rating when his rating is updated. Whereas a player like rapha is not since he is highly stable. The volatility makes the system able to cope with the fact that players might improve faster than the system can handle. Which is probably true of new players. Or old players who pick up duel for the first time: they have the aim and as they learn the other aspects they will quickly improve their game.

Finally, among the 31000 players I have, I can currently predict the outcome of a match with 64% correctness. I should probably play for E$R :) But the 64% figure is climbing as weeks go by and I have more information. Also Glicko2 is tunable, so I'll run some measurements to find better parameters. My preliminary checks are that this can improve the 64% a couple percent more. But more on that as we get closer to DHS 2012 :)

You should definitely try and predict the match outcomes, would be interesting to see how your model performs in practice

I would definitely like to see either system applied to the previous events.

Basically a qlranks ladder that only tracks tournament matches from the previous (all?) Quake Live LAN results. That would be very interesting.

You danes sure like to write a lot :D

This is very, very good.