Monke used to bang on about this in the H5 forums a few years ago. It’s almost certainly K/Min.
I don’t get why it isn’t score
As you said sample sizes are small.
The links are mainly on the effect of taking an bot or social MMR into ranking for the first time. We know this is significant. TrueSkill borrows a previous MMR to seed a new playlist.
The third link makes me sad that we don’t have a public API for Infinite. 
And for the first example the only effect I’m interested in is the last phase… playing in ranked after tanking a few games in social. But they only presented two games! - and the team CSR’s were only 16 and 19 apart. Hardly ground breaking differences- especially keeping in mind that we don’t know of any squad weightings… or what the actual MMR was given it was still a fresh / volatile account.
I haven’t seen a trend that really shows a significant influence of social/bots on ranked outside of the new account scenario. It’s certainly not my anecdotal experience and there have been plenty of Reddit posts about people who have tried (eg. the player who generated a 30+ KDA in bots) with no effect.
We know matchmaking can be a bit up and down. Especially out of peak times, when running squads, and post placement.
The bias is that it is only the people who get harder games that post about it.
A while ago I was discussing a case on Waypoint with a player who was ropeable they had lost 20 odd CSR points in a small run of losses. But the fact that they had accumulated pretty much all of those points in the same number of wins just before that had alluded them.
Chance and bias?
And look, I am not naive enough to ignore the chance and bias from my side as well.
We know that there is a shared component to the MMR. I imagine it includes various weightings that contribute to form and time lapsed since the last game. These could well influence your experience, especially if you only play ranked intermittently.
But I would argue that any influence is minimal.
If it really worked we would have clear examples of players exploiting it. Definite cases of lower skilled players showing off their Diamond / Onyx ranks. All they have to do is rinse and repeat bots and ranked over and over… and hey presto they should just keep ranking up without any trouble.
But we don’t.
You may (or may not) get a slight warping of match-making. But in most cases it won’t be noticed. And it certainly isn’t strong enough to affect your rank (for or against).
I stand corrected. I would have thought the fall would be faster than that.
His MMR must have been insane!
I’m sure bots MMR will be removed - or at least no longer used for seeding other playlists.
After placement your MMR curve should still have some width and volatility. Even if your matches are 50:50 (and they shouldn’t all be so if match making is working properly) wins will be rewarded more than losses as you are given a chance to move up and earn your rank.
It only narrows and becomes resistant to change once you start performing consistently at a certain rank (which should be around your skill ceiling).
If, for some reason, you do actually improve in skill after your rank has settled (this is usually a very slow process)- you can still hopefully eke out a rise… but it may be easier to wait for the next placement to free up your rank. This is one of the roles of seasons / placement. It’s the anti-lock for what used to happen in Halo 3.
It’s tricky.
The main time the CSR and MMR are wide apart is post placement. But it would be more disheartening to be given the bigger end of your MMR (or even mean) and lose it when the system loses confidence in your rating.
Would you prefer your ranking is kept hidden until the system has a certain level of confidence in your rank?
Later on it shouldn’t be a problem as your MMR and CSR would only drift out by a handful of points. You know, when you lose 10 points or so on a loss and everyone else only loses 1 or 2.
But the downside is that you will either get uneven matches (your MMR is actually higher than your CSR) and a slower climb to your actual rank (not getting the chance to play the better teams).
And again, apart from placement, the drift shouldn’t be a lot.
I thought it was still considered a curve. One of the GDC presentations talked about how the maths are done so that it never has to actually generate a number that represents your skill.
Of course you can take a number from it. The mean being the obvious one, or the mean - 3xSD if you want a high degree of confidence in your result.
But the important bit is that the shape of the curve influences how the MMR moves up / down the x-axis.
And as you said, it’s essentially a probability distribution. But if your curve is short and wide, the probability of where you choose to pin the tail on the donkey (or slime) starts to fall.
And I still think the biggest mistake 343 did was trying to emulate the CSR on the same x-axis values of the MMR. While they do shadow each other’s values - they do move differently and this causes confusion.
I wish they would change the x-axis scale of the CSR. Not necessarily back to 1 to 50 (this would clash with people’s nostalgia for Halo 2/3). Maybe a 1 to 36 system (Bronze 1 to Onyx 6). Or increase the tiers to 10 per division for a 1 to 60 (Bronze 1 to Onyx 10) - and use the Halo 2 symbols from 41-50 for the Onyx tiers.
Yep. It makes more sense than K or K/D.
The sad bit is the toxic play that has drifted in as players try and bump up their K/D in the hopes of eking out a tiny bit of extra rank.
I’ve had a couple of games where the dominant opposition would drop the oddball in the hope of farming their K/D. But the joke is on them as I hide and try and soak up time to lower their K/min.
Good times :). Not.
When they ran various combinations of kills, deaths, and assists through the simulations they found that raw kills didn’t actually add anything to the rank predictions. If you had a high score or K/D this was usually already reflected in the win.
But they noticed that K/min allowed them to pick the big fish from a small pond. As you increase in opponent level it becomes much harder to have high K/min than it does to have a positive K/D.
It can be subtle;
A score of 15 can be in a three minute game (5 K/min) or 10 minutes (1.5 K/min).
In a 10 minute game a scores of 20:10 and 10:5 have same K/D - but one has twice the K/min.
It’s a weighting that allows them to rank people faster.
If they can explain how going 38-18-16 assets in a tie games makes my rank go down then I’ll listen…
Or how Onyx players go almost negative and never lose csr… it’s broken… but a D5 can lose one game and down 30-40 csr…
Couldn’t see that particular game on Halo Tracker; 38:18:16.
But you do have at least three losses on the first page despite 30 odd kills.
And the key here is losses.
If you want your rank to go up you need to win the game (particularly vs better opponents). Personal performance has some weighting that may mitigate values.
But a loss is a loss.
It’s a team game and the ranking works on the team result.
If you lose to a stronger opponent then rank change can be minimal.
That sounds a bit extreme. I think your CSR should only change by around +/- 15. Are you sure you weren’t looking at the session bar?
Anybody remember the “house of blue leaves” challenge, from OG Reach?
Its kind of related to this.
Except kills don’t tell you how good you are at oddball, that just tells you how good you are at slayer.