"Oh you won few games in a row, lets see you carry this!!!"

Oh no, as you can see from the length of my posts, I LOVE to talk about this - but what hypothesis would you like explained? The post(s) I deleted were about the details, and all the work that goes into explaining them is a bit lost if all you get is insulted :) But happy to chat about the big picture, so to speak...

Now, to answer your question, I THINK they base the 50-50 chance on MMR, understood as "mean skill" as measured by the system.

And I also THINK that they do not take into account how the performance of one player affects others', nor how "consistent or inconsistent" a player is, i.e. you can think about the system how I envision it as something like a "mean skill" (that's your MMR in an ideal world, i.e. in cases where the system is able to accurately measure your actual skill... which is a topic for itself, as this is difficult to approximate and, likely, impossible to be 100% accurate) and then a "standard deviation" around that mean (that's how consistent you play), and then some sort of "influence" of one players' performance on others, e.g. if one deviates too far from the mean, other players will also play worse (that is what you observe, when the game gets sometimes lost because somebody starts feeding and then the rest gives up). Then you get the actual performance of all players in a given game, and the sum of that would be the performance of the teams... the one with the higher performance wins.

I believe Riot only uses the MMR (i.e. the mean skill), and tries to create two teams with an equal (or, depending on queue times, more or less equal) total MMR. I also believe they try to make the individual MMR differences as small as possible (i.e., not have 1-1-1-1-6 but rather 2-2-2-2-2...).

As implied, there is no proof for any of what I think and believe. There is not even a HINT that Riot provides. The above comes from a LOT of research in probability theory and a personal passion for chess rankings... and as you know, those have been the basis for "ELO systems" in online games, but most of them are very flawed (since the ELO system is simply, in it's basic form, not suited for anything BUT a 1-vs-1 game of skill... and keep in mind, it is not only intended to create "fair competition" in chess games - it is intended to adjust the actual outcome in point changes to how fair the game WAS in the first place). It is basically how you would have to start designing a system that could work (and we have designed one with many parameters that is, potentially, much better, or it at least tries to be... although let's not start talking about what "better" even means... is it fair games? is it the best players can rise very quickly? is it all players play as much as possible and become addicted to the game? is it low queue times? There are so many variables that you can try to optimize, so you have to decide what you want to do first and then see whether your system is actually "good" at it... as I have said many times before, if I am Riot, I would want to design a matchmaking system - and a MYTH around it - that makes people THINK they have a chance to climb if they are good and play A LOT, and makes games on average just not TOO unfair, so that people don't quit altogether out of frustration. That's it! And under that assumption, the current system is probably working extremely well... and by the way, that also explains why they don't disclose how they do it... other games do that, they even try to do more than I assumed above, for example incorporate standard deviation around skill mean, or certainty about estimates, etc.).

In Riot's defense, though: as it is very difficult or even impossible, as said above, to actually correctly estimate the skill of players, and also the underlying "skill spread", if you want (that means: are the actual player skill values 1-100, with a normal distribution around 50? Or is it 0 - 10,000, with a logistic distribution around 200? or is it 1 - 3, with most players being 2? IS it even some sort of value, or is it more complicated?) - it's more possible in a 1-v-1 situation, but VERY difficult in a 5-v-5 situation, unless you have a ridiculously high amount of games, and even then, I don't know, it still depends on the degree of influence of one player's performance on those of the others... (which is why I said before, if I wanted to solve this, I would use a fixed group of 100 players, all Riot employees, and have them play exclusively within that group... and record ALL their games, for a LONG time... and have some sort of "judge group" give a "non-numerical" estimate of how well they play... and then take it from there). Without that, designing a system that has people play a lot is probably the best you can do. And IF that system even allows for progress over time, i.e. the more you play, the better you get, the higher you LIKELY / OFTEN climb, that means it is a good system!

And, even more: you CAN argue that, since it is so difficult to actually ESTIMATE player skill correctly, I will ASSUME my system to be wrong about it... which in turn CAN mean, under certain circumstances, that the majority of players (since for all players, the system assumes the WRONG actual skill, as MMR) can "climb" by beating the system, so to speak, i.e. by being "better then the system thinks"... and if you use THOSE assumptions (i.e. your probabilistic model assumes the system is always wrong), the system I assume above does not work THAT badly... for many players, it will get to a point where the influence of the actual skill on the ranking is high... BUT "many" does not mean 99.5%, it means more like 60%... AND those 60% will change over time, i.e. you can drop in and out of that group.

And LASTLY, certain things are NEVERTHELESS clear and they can be proven:

(a) there is NO WAY in hell that the ranking is ANYTHING close to "linear", as the 100-point divisions and tiers seem to imply... it is WAY more likely that the skill difference is something like "bottom tail" (those who play so badly that they loose games by themselves), huuuuuuge middle, where really RNG has a higher influence on winrate than skill (i.e. one dude gets lucky and has a 53% winrate over 350 games, the other gets unlucky and has 47% over 350 games... is that super likely? No! Would I bet money on either happening? No! But it WILL still happen and in less extreme boundaries, it is very likely, especially with the ridiculously huge player population in LOL - again, we can prove it with a probabilistic model), and then some tiny fraction of elites.

(b) there ARE statistical models that EXIST that can model 5-v-5 situations, influence of unknown variables on each other, etc. etc. - are they being used by Riot? No idea! Since they don't tell us what they are doing...

So, my biggest point in all these posts is: this is actually a super-interesting thing to fully understand, and it would be excellent to develop something that works "best" according to what you want it to do... but since Riot does not TELL us anything, we have no way to judge THEIR system :( But still, back to my original point, if we ourselves spread the myth that "somewhere it is proven to be awesome", that just does not make any sense.

Oh, one last point: the model also proves, beyond any doubt, that what people consider to be "ELO hell" HAS to exist... where and when depends again on many factors, but it is "extremely unlikely" that it does not exist at all. It is very likely that there are actually multiple "elo hells" between all the tiers of the system :)

it uses the same elo system it has been using since season 1, that is also used in various other competitive games or sports, you win you climb, you lose you fall, not sure what's hard to understand about it. players with a high elo are good, players with a low elo are bad, if you win a lot, you'll get matched against other good players with a high elo and games will be more difficult and you'll start losing more.

Oh you are of course right, my apologies. It IS technically evidence, and yes, it is it's own kind of evidence: also known as "the kind of evidence you throw into the garbage can, especially when it comes from anonymous people on the internet".

And what does that change exactly?

If you know what "requirements" are needed to get punished and you also know how to not trigger those requirements, then it doesn't matter if it's a machine or a human.

ahh so i just misread mb