Before we delve into CSR and SBMM, let's take a look at your MMR;
Typically they do this in the shape of a Gaussian (normal) curve.
The system simply aims to sort players relative to each other - so the horizontal scale is completely arbitrary; eg. Novice to Grand Master, 1-50, Bronze to Onyx, 1-1800+, etc.
The simplest ranking system awards a point for a win and deducts one for a loss. But that is slow and inefficient - taking 1000's of games to sort a large population. And there would be a lot of mis-matched games along the way.
More complicated systems can speed things up by allocating more points for unexpected results.
Professor Arpad Elo developed his special ranking system for Chess in the 1960's. It looked at the skill of the two players and made adjustments based on the following;
- the result of the game (win or loss).
- the skill gap between the two players.
- the confidence the system has in each player's skill.
ELO based ranking systems are very fast.
It was first introduced in 2004 (Halo 2).
In 2018 it was then updated to TrueSkill2 - using data from millions of games of Halo 5.
It can ball park your skill level within half a dozen games - and find your accurate rank in about 50 (for 4v4 playlists).
It is not a simple number - it's a probability curve of your most likely rank.
The mean (mu) of the curve is your most likely rank.
The standard deviation (sigma) represents the confidence in the rank;
- there is a 68% chance your rank lies within one standard deviation of the mean.
- there is a 95% chance it lies within two standard deviations.
- and a 99.7% chance it's within plus or minus three standard deviations.
The width of the curve reflects the confidence the system has in your rank;
- the wider the curve (higher sigma) the less confident the system is.
- curves start off very wide (sigma 300) and narrow (low sigma) with consistent play.
- wide curves are volatile and free to move.
- narrow curves become static and tend to resist movement.
- wide curves move faster towards narrow ones.
ELO type rankings boil down to;
- Beating a team ranked above you moves your curve to the right.
- Losing to a team ranked below you moves your curve to the left.
- How much it moves depends on the width of the curves.
It helps calculate how much to change ranks with each win or loss.
You can average player curves to get team skill curves - and then subtract those curves to get the odds of each team for winning or losing the match.
But with Infinite 343 have made the decision to leave some width in your MMR.
This allows it to respond quickly to fluctuations in form (and deal with Smurfs).
The reality is that your skill can fluctuate by a division (300 points) or more over the course of a session. And this sounds a lot - but it's only dropping wins against your rank from 50:50 to 25:75.
These jumps don't make for a good visible rank - hence the use of the CSR instead.
Halo 3 had more of a traditional ELO - and infamously became 'locked' if the curve narrowed too much - leaving players with no choice but to start a new account if they needed to rank up.
They don't make any judgement on how you won or lost.
But Microsoft looked through millions of games of Halo 5 and found that both kill and death rates helped speed up the ranking process.
KPM (kills per minute) and, to a lesser extent, DMP (deaths per minute).
It's not to say other metrics (such as KD or Objective Score) aren't important - but they all contribute to the sum of all the parts of your skill sets that contribute to the win.
Apparently some parameters were actually negative predictors for rank. MS didn't say which ones, but it has been implied that 'assists' were one of them.
Pretty much the only performance metric that adds value to the ranking.
It can't be manipulated like KD.
Eg. Consider an 8 minute game where;
- one player gets 4 kills and 2 deaths and another gets 16 kills and 8 deaths.
- both have the same KD (2.0).
- the second player has four times the KPM (0.5 to 2.0).
- not unfair to assume that the first player doesn't have the skill to match KPM.
The gold standard is a KPM between 1.2 - 1.6 against players of your own skill (around 10 to 12 kills in a typical 8 minute match).
And remember that it is only a weighting.
As you can see from the table excerpt (from MS's discussion paper) - doubling your KPM only increases your expectation to win by less than 5%.
KPM is used to identify small fish in a big pond and rank them up faster. Smurfs or new accounts that are genuinely high skill. It's not going to rank you higher.
As you try to rank up it gets harder to increase your KPM against players of your own skill or better.
And finally, keep in mind that there is a lag effect. Your personal performance (high KPM) will nudge your MMR up and this will then pull harder on your CSR in the next game. It has little to no effect on the CSR change in the game just played.
Think of it as a "form" MMR.
Each playlist/mode (eg. Ranked/Oddball) then has it's own "offset" MMR - with their own mu and sigma values.
Playing in one list affects both that list's MMR and the global MMR. And then the global MMR drags on all it's offset MMRs. How far each one is moved depends on the confidence (sigma) in each.
It just means that if you are doing a couple of tiers better in Team Slayer (eg. P3 up to P5) then the system expects a similar jump in Team Snipers (eg. S2 up to S4). You can still be Platinum in one and Silver in the other. It just expects your form to carry over.
The classic scenario is someone sweating to their eyeballs in Ranked - and pushing this and their global MMR up. So when they drop into Socials to relax and "chill" - they find all those modes on the upper end of expectation.
But Objective games are just Slayer with a side quest. The best time to run the objective is while the other team is respawning.
Josh Menke summed it up when he said something along the lines that "anyone can run the flag".
You only need a KPM of around 1.5 to justify your rank. Chasing a higher KPM or (even worse) KD at the expense of the Objective will just end up with more losses - and that will drive your rank more than a few kills.
My points from this are;
1. It's funny.
2. 343 honestly wouldn't have expected someone to do this.
3. Part of the problem is that the bots were initially given way too high MMR - and after this they were demoted by 343.
I'd love to have access to plots of MMR vs CSR. With overlays of KPM.
It would not only be insightful into your own performance - but would also demystify the whole ranking process.
I think the main problem is that people just don't accept that;
- a) their skill fluctuates
- b) that skill curves plateau (and are VERY hard to elevate from there).