Victor Morson, on 09 November 2013 - 11:44 PM, said:
Which both you and the people who are programming ELO have outright butchered through a fundamental misunderstanding of WHICH stats your are using to figure out your averages.
Calculating individual PLAYER worth based on TEAM outcomes is a flaw in the very basis of everything you are saying, which renders all the rest of it entirely moot. You cannot gather clean data about individual player performance from a completely random team, with every single other random factor (such as tonnages) further muddying the water.
You aren't getting good input. All the math and links about how statistical analysis in the entire world can't change that fact.
Ultimately win/loss comes down to more how lucky you were in pulling an average solid team, or if you brought your own teammates into the match. Period. Even if things even out over time to a 1:1 via all the random elements are you are suggesting, this just means that players in every game will have an artificially high win and loss rate based on their random factors. End result?
Newbie McNewbieton will likely win 50% of his games due to random factors. If he is really bad and causes some losses by himself somehow, that will only influence things a TINY amount. He'll still be lumped in with other players that had a similar team experience.
Again, bad input data, bad output data. You can't statistically analyze things clearly when there is this much out of ANYONE'S control at play.
PS: They often do grab statistics from skewed sources like this in real life..... when they want to push a specific viewpoint. They will purposely pull "muddy data" that they can skew to show a different picture. Statistical analysis is one of the very, very easiest things to get wrong... or manipulate. And the way ELO is done here is very muddy indeed.
The vast majority of the stat he listed (without any firm evidence of the 60-70 figure really) was done in 4-mans, a whole other ballgame.
A critical point in here - gaining Elo is not a reward for skill. It's just a metric reflecting how often you drive a win rather than a loss. Cbill and XP are rewards for kills, damage, assists, spotting and the like. Win/loss is the superior metric specifically because you can't game it - right now for example I'm leveling my tbolt. I do better in terms of winning when I've got 2PPCs and 2xAC5s. I get more kills and help my teammates survive longer. For XP gain though I'm actually loaded with 3xMLs, 2xSRM6 and an LB10X. While I don't win quite as often or get kills quite as often I do more total damage and when I do catch someone with their armor down I pile in the component destruction. My match score average is higher. This is more a byproduct of exploiting the successes of my teammates though, not some sign of my being a better player. You'll likely find that the highest Elo scores are held by light pilots. Lights excel at disrupting enemy tactics and drawing disproportionate attention. One spider can shut down 3 LRM boats and pull several heavy mechs off the line in a match, letting the spiders teammates exploit that opportunity for a victory. The spider may have done minimal damage but his ability to help his team win is what matters to Elo.
So your main concern is the belief that your impact in 12 v 12 isn't statistically significant enough to represent you in Elo? I'd say that statistically a bigger impact than your teammates on your win/loss score is your mech build, how rested or distracted you are and other personal factors but for the sake of simplicity of this debate we'll go with the core of your concern - teammates and your impact on the odds of a win/loss.
First, the difference between small and large sample sized statistics and why Las Vegas casinos make money.
Casinos make money because 'luck', as a statistical representation of probability modeling, doesn't play the long game. Ever. So while some people may win the odds are inevitable that the house wins out. So inevitable that the house inevitably wins out BIG. That's because in gambling statistical 'average' pays the house. Win/loss might swing wildly in increments of 10 matches but over 100, 300, 500 matches probability theory wins with relentless mathematical precision. If it didn't no casino could make money in the long run.
For another thing your team is not as equally probable to be someone who downloaded MW:O by complete accident and is in fact earnestly trying to play Barbie Pet Rescue while being confused about all the weapons fire they're taking. Pretty much everyone on both sides in both matches is going to be outside their first 25 matches. There's a small probability that you might get a newbie but as a given rule if you're a couple hundred drops into any weight class you're not likely to see newbies.
For yet another your own perception of how well you play vs how often you win is absolutely untrustworthy. That's human nature and how your brain works. It'll remember when you do well but your team didn't and you lost even though you got 4 kills but forget every time you got killed early and your team carried the match.
Your impact on your 12 man team in pugs is ~8.333%. 100/12 = 8.333. That's your impact on your teams overall success or failure as a statistical average if you were in fact absolutely average. What's that equate to?
Let's use an analogy. Suppose you were on a ship out in the ocean and you set a course for a port 5,000 miles away. The ship is a huge cruise ship with 12,000 people on it. Every 30 minutes you and 11 other randomly selected people are required to adjust the course either port or starboard 8.333%. If everyone was random the ship would eventually hit the port - or at least pretty close to it. While in the short run there might be stretches where the ship would end up pointed way off course probability theory dictates that it'll average back out to stay on course.
You, however, like to go right. Out of every single 'shift' in adjusting the course you are present for every one of them - you are the only absolute constant in measuring the course correction. Some other people might show up repeatedly but you are accounted for in every match. Suppose you however ALWAYS adjust the course to starboard. That means that since every single other entry will, mathematically, average out after enough 'turns' you will cause a persistent 8.333% shift to starboard. Result? After the first 20 or 30 sets of adjustment you might notice a bit of a starboard shift. After 100 it'll be pronounced. You will in fact have driven the ship ~833.333% off course at this point. So every 25 hours, statistically you would have whipped that big cruise ship into 8 and 1/3rd circles - even though there are 11 other people making adjustments every 15 minutes.
Matchmaking in MW:O is certainly more complex and it would be arrogant to assume that anyone is 100% beneficial to their team every game. That's why it takes a good 100 matches just to get you in the right zone and not unreasonable to say ~500 matches to get you well seated in your appropriate Elo weight class. If you were dropping with the same 11 other people against the same 12 other people it would go far more quickly as there is less variance to account for.
This works however in pugs for the same reason that Google and Facebook work - probability theory helps scrub out random static by dint of volume of data. The more player telemetry there is the more accurate everyones matchmaking is.
How about another example?
www.random.org
On the right is a random number generator. Remember how your performance on your team equates to 8.333%? Certainly not all players are equal. So generate 12 numbers between 1 and 16.666%. That's the 'value' of your team.
Do that 2 or three times. You'll probably see a swing of about 10 points - you'll end up with totals between 90 and 110. You might even get an outlier in the 80 to 120 range.
Do that about 100 times. Add up all 100 of them and divide by 100. You'll end up with a number ranging from 99 to 101.
Now, do that again but only randomly generate 11 numbers each time and for that 12th man always make it 10.
You'll end up with a number ranging from 101 to 102. Why? that extra 1.666% difference between you and statistical average plays out clearly. In fact the more times you roll it the more pronounced it gets. At 500 matches you'll be at a 508.315 instead of 500.
Does that make more sense? There are functional limits on how much any one person, for good or ill, can influence a win/loss in a match so the swing created by your teammates is certainly less than -100 to +100%. It is however perfectly direct to plot and identify, unlike KDR, damage, score or other personal metrics.
It certainly does have a margin of error but that's another place Elo shines - the equation is self-correcting. Since win/loss is a 1/0 factor and not a sliding scale if your Elo gets pushed too quickly in the wrong direction probability theory will ease it back.
Win/loss is the only viable metric for determining an Elo score. Everything else is too subjective. Win/loss in a pugging environment is the only reliable indicator that can be used to place and rank players. Everything else is less reliable and more prone to manipulation.
1
