The rating and matchmaking system is confusing, and a good overview does not appear to be available. The official overview1 is incomplete and does not answer a number of common player questions and concerns. This information below is gathered from sporadic developer posts and videos, salted with my own experience and experiments, various forum threads, and watching streams. Note that since Blizzard does not give exact algorithms, I do have to fill in some gaps, or leave some items unknown. Finally, Blizzard has developed a habit of making changes without notice, so just because something was true in the past does not guarantee that it is still true.
I will edit this post if information becomes available before the end of this season. Next season I will make a new post and include changes.
Whenever you see a superscript number, look to the references section and follow the link to see the support for the statement in question.
Changes in Season 19
There were no reported changes to the matchmaker.
Competitive now has a map pool of 12 maps, and starts on a Thursday rather than the 1st 1.
See the later sections for a more in-depth exposition, as well as references to supporting material.
How do I maximize my Skill Rating?
1) Win games
2) If platinum or below, perform well, statistics-wise, with the chosen hero.
In my last game, my Skill Rating went up/down by X. Why did this happen?
The details of Skill Rating (SR) movement in a match depends on whether a player wins or loses, the Match-making-rating (MMR) mismatch between the two teams, the mismatch between the player’s SR and MMR, whether a player is new, whether a player is active, the player’s statistical performance, the player’s position on the ladder, and possible bugs. Add it all together and it is nearly impossible to determine why a player’s SR has a specific movement in a particular match. Sometimes over the course of twenty to one hundred matches it can be seen that one particular cause is dominant, but that often depends on careful data collection and some guesswork.
Summarize matchmaking, rating, and progression for me
Each player has a competitive MMR for each role (tank, dps, and support)5. Role MMR is a hidden number6 that goes up when a player wins, and down when that player loses. How much it goes up and down is a complicated formula, based on many factors, but the essential truth of wins are good and losses are bad is certain7.
That MMR is then used to match people with and against people of similar MMR, in an attempt to create a 50% match8.
Here is the key part:
If a player’s MMR is wrong and too low, he will be more skilled than the matchmaker expects, so the odds to win will be greater than 50%, he will win more games than he loses, which will cause his MMR to rise over many games played. He will then be placed with stronger and stronger opponents (and stronger and stronger allies) until his MMR is correct, and his win percentage approaches 50% (with some random oscillation around 50%)9.
Once a player finds himself trapped in a rating range, the only way to break out is to improve as a player and play enough games to overcome any random factors10,11.
What is SR then? SR is a visible and friendlier approximation of MMR. It has no more meaning with respect to matchmaking than tier icons6. However, MMR and SR are closely linked12, so inspecting SR typically gives a reasonable estimate of a player’s MMR.
Matchmaking is based on hidden match-making-rating (MMR), not skill-rating (SR)6. Competitive MMR is separate from the MMR of other modes13. MMR is separate per role (tank, dps, and support). MMR has a value and an uncertainty2, and is stored in units of standard-deviations. That is, 99.74% of the population is between -3.0 and +3.06. -3.0 is in low bronze, and +3.0 is in high GM. Do not freak out about MMR being stored in standard-deviations. It could be stored in any other scaling without affecting matchmaking.
The game uses MMR to determine whether a match is fair, and attempts to match teams such that the game is fair (each team has a 50% chance of winning). Upon victory, the winners receive a bump up in SR and MMR, and the losers get bumped down. At times (off peak-hours or very high/low rated players), finding a fair match will be impossible in a reasonable amount of time, and the matchmaker will match teams based on a best fit. However, note that if no matches can be made with an expected win percentage between 40% and 60% (the infamous Brazilian six stack playing at 3 AM), then no match will be made and queue times will become extreme to infinite. In the middle ground, top (or bottom) players will have a longer, but not infinite queue, to attempt to find a better match11.
If the matchmaking algorithm determined before the game that it was not fair, the higher ranked team will receive less SR/MMR on a win, and lose more SR/MMR on a loss. The lower ranked team will receive more SR/MMR on a win, and lose less SR/MMR on a loss11. This is standard for Elo (rating system from chess) type systems. The SR/MMR penalty/bonus is designed to be fair, such that if you win 40% of your games with a predicted 40% expected win probability, your SR/MMR does not change11. Draws cause no change in SR and have an unknown effect on MMR.
As a player plays games, MMR uncertainty (for the role played) goes down (with a hard floor for active players). If a player does not play for a while, MMR uncertainty goes up. The higher the uncertainty, the more MMR can move in one game11.
The matchmaker does not match based on recent win/loss15, win percentage16, 17, MMR vs SR differential6, endorsement level, player experience level, high/low statistical performance at a given SR/MMR, expected loot box purchases, or any other wild ideas players manage to come up with and flog mercilessly for years at a time.
Matchmaking and groups
In addition to MMR, the matchmaker attempts to match on ping and group size and to make matches reasonably quickly11.
Predicted win probability (which is used to calculate fair matches and MMR/SR rewards) does not depend on grouping11. This leads to a small handicap for solo players vs groups, and smaller groups versus larger groups, for when the matchmaker cannot find groups of equal sizes to match against each other11.
However, if a player groups with someone with whom they have poor synergy, it acts as a penalty on their practical win likelihood, because they are likely to be matched against groups of equal size, with more synergy and experience. Finally, large groups tend to have longer queue times while the matchmaker tries to find other large groups.
If the match-maker says most games are fair, then why are there so many stomps?
There are many reasons:
Overwatch, as a game, has a tendency to snowball. The winner of the first fight has an ultimate advantage that has a tendency to last the round. Losing teams tend to tilt and start playing poorly, which can continue in the same round and carry into the next round.
Random variables are random. Maybe a cat walked on a keyboard or someone fat fingered an ultimate. Maybe someone who has a 99% up-time for their internet had their 1% failure that day. MMR and predicted win percentage only has validity over many matches, not each individual match.
Not every player tries hard every match. Sometimes this is subtle, like playing with a beer or two too many. Sometimes it is less subtle, like practicing a hero a player is not good at, regardless of team comp or map.
Some players actively break the system, by either hard throwing, that is intentionally losing (as opposed to soft throwing as in item 3), playing on someone else’s account to boost it, or other similar actions. These actions are bannable and should be reported when seen. However, soft throwers and hard throwers can be difficult to distinguish, and Blizzard has to err on the side of caution to avoid banning the wrong people, so it will always be a problem. The in-game report UI gives guidelines on what behavior is or is not bannable. Note that if a player is caught hacking, the game will be cancelled and no win or loss will be recorded for either team34.
New accounts in general (or accounts that have just started playing a new role) have less data and will have a less accurate MMR. There is no required minimum number of games to play, so new accounts can stay “new” for a very long time.
MMR itself is only approximate, and can swing fairly widely.
Okay, stomps are going to happen, but why are there so many win/loss streaks and large SR/MMR movements? Shouldn’t a player’s SR/MMR be fairly stable once he gets to the proper rank?
Win probability changes slowly with rank because there are so many random factors in each individual match. Unfortunately, it follows from this that frequent and long streaks will occur, and a player’s rank will oscillate widely. Essentially, a player will tend to bounce between the range of where he is nearly guaranteed to win and where he is nearly guaranteed to lose. This problem can be analyzed in depth, mathematically17,36.
Skill Rating (SR) should be thought of as a visible and friendlier approximation of MMR, similar to tier icons. MMR and SR are closely linked12. The SR/MMR convergence is implemented by having each match pull SR in the direction of MMR (rescaled to be from 500-5000). That is, if SR is lower than rescaled MMR, a player will win more SR than MMR on a win, and lose less SR than MMR on a defeat.
This effect is clearly visible if SR is different from rescaled MMR by 50 or more. At this differential, a player will win about 29 SR on a win, and lose about 23 SR on a loss 18. For active players who have not recently dc’d or decayed (though decay has now been removed from the game), this effect has never been shown. I therefore conclude, that for active players, rescaled MMR is always within 50 of SR.
Typical SR movement has range of 20-30. Neutral SR movement (where the teams are balanced, there is no performance SR, etc.) is 2418. Typical MMR movement is difficult to quantify as it is invisible, but is likely similar, but rescaled, so approximately 6*24/5000 = 0.029. If an account is inactive for a while, after the player returns SR/MMR gains will be exaggerated again for a time 11. For this reason, players that only do placements each season will tend to see high volatility in their rating each season.
In addition to Elo type effects for high rated players (in which they gain less SR/MMR because there are no fair matches available), there is an extra SR debuff for high ranked players to prevent them from getting to the SR limit of 5000. The concern here is that players would start to pile up at the limit of 5000 and break the leaderboard. This debuff starts at 4200 and gets progressively stronger as the player ranks higher 11,19.
If a player has never played competitive, his initial competitive SR is around 23502,20 (and initial MMR is 0). For new players, MMR and SR have high uncertainty2. Because initial placements depend on only 5 games, initial placements have substantial random causes and are frequently inaccurate. When a player has data in one role, but not another, the data from the first role is used to inform the second (and third), and give it an initial value2. These other roles start with high uncertainty, though, to help them move more quickly to where they belong.
If MMR are SR are so similar, why have both?
Let’s summarize the differences between MMR and SR:
- Not human readable
- Blizzard’s best estimate of your skill
- Only affected by what happens in a match
- Used in matchmaking
- Human readable
- Chases rescaled MMR (SR gains more on a win / loses less on a defeat for MMR > SR)
- Affected by what happens in a match and DCs
- Not used in matchmaking
The real purpose of MMR is to provide Blizzard’s absolute best estimate of each player’s ability, while SR is to give a friendly indication of rank to players, as well as penalizing for DCs (without affecting matchmaking).
However, it is probably worth coming up with a simpler system to prevent much of the confusion that circulates around these terms. It would also help immensely if Scott would be more consistent in his use of the terms.
< 500 Bronze, SR not listed 500-1499 Bronze 1500-1999 Silver 2000-2499 Gold 2500-2999 Platinum 3000-3499 Diamond 3500-3999 Master 4000-5000 Grandmaster Top 500 players in region.
For season 3, the rank distribution was Bronze: 6%, Silver: 22%, Gold: 34%, Platinum: 23%, Diamond: 10%, Master: 3%, Grandmaster: <1%21. For season 8, the rank distribution was Bronze: 8%, Silver: 21%, Gold: 32%, Platinum: 25%, Diamond: 10%, Master: 3%, Grandmaster: 1% 22.
As a player gains SR, he will be promoted to the next tier when appropriate. He will drop out of each tier as he loses rank. However, there is some loss protection for Diamond and below. After a loss (but not a win), the game will check to see what a player’s current tier and skill rating are. If his skill rating has not exceeded the minimum for his current tier for the past five games, he will be demoted23. For Master and Grandmaster, a player will be demoted immediately if his skill rating is not high enough. Competitive point rewards are based on the highest rank achieved during the season for each role23. Ranks and SR are wiped each season (but not MMR).
There are four top 500 leaderboards, one for each role, and a combined one. It becomes available two weeks after the season start. To be eligible, a player must have played at least 25 games that season, all from one region, for a given role (or 75 games for the combined leaderboard) 2. The player must have had SMS protect enabled for all of these games24, 25. The top 500 spray and icon are not awarded unless a player is in top 500 at the end of the season26.
But I just read this post from a developer, and it said matchmaking is based on SR, not MMR
Jeff has been decidedly clear that matchmaking is done on MMR alone. Scott has been decidedly less clear. The contradiction must be resolved somehow, and I decide in favor of SR being used in no way during matchmaking for the following reasons:
Whenever MMR and SR are discussed in the same post (including by Scott), it is to say that SR is not used in matchmaking.
When decay was still in the game, Master+ players would regularly see decayed “diamond” players in their matches. And when someone like Seagull decayed down to diamond on stream, he would still be placed in grandmaster / top 500 matches. If SR was either used for matching, or the limits on matching were restricted by SR, this would not have happened.
SR and MMR are closely linked, except for decayed players (who only exist in diamond+) so saying matching is done on one or the other is a distinction without much of a difference. This makes Scott being sloppy with the terms not as surprising and makes the answer to this question less important.
Generally people who are insistent that SR and MMR are both used in matchmaking are using it to infer some broken or rigged system with horrible and implausible outcomes that are contradicted by careful analysis, common sense, and unrelated Blizzard posts.
In platinum and below, SR/MMR gains are adjusted up or down based on the performance of the player. This is a small modification11. This is done based on a numerical comparison of measurable quantities such as elims, deaths, assists, damage blocks, ults cast, etc. between a given player and other players of that hero at that MMR27. Generally it is assumed that the measured stats are those visible to the player, but that has never been confirmed by Blizzard. Most of the details of this implementation are fuzzy and not published (probably to reduce exploits). This performance measure is correlated but not identical to “on fire” calculations. “On fire” compares a player to his teammates, while SR/MMR bonuses compare a player to other players that are not in the current match, but in a similar skill bracket, and playing the same hero27.
There are two camps in the community, those who don’t want their SR/MMR to be so heavily influenced by those in their group, and those who worry that having a performance based system will cause people to not play the objectives / win conditions and instead go stat hunting. In addition to people who go stat hunting, people can unintentionally be at the wrong rank because their good / bad play is not reflected in their stats.
Starting in season 8 (January 2018), the performance modifier was removed for diamond and above33. Blizzard has started to realize that having a motivation other than winning causes all sorts of non-ideal behavior and effects. The immediate forum response has been to request this change for lower tiers as well (at least whatever tier the poster happens to be in). However performance modifiers do help move new players where they belong much faster than an Elo type system, so I propose that the performance modifier be removed for Bronze to Platinum players after a sufficient number of competitive games have been played. Blizzard can use their data to determine how long it takes for people for their rank to stabilize, which would be a good place for a cutoff.
Players’ 5 placement matches will start with Blizzard’s best estimate of each player’s MMR for matchmaking purposes. If the player has been active in a role, MMR at season start will be very close to where it was at the end of the previous season.
An active player’s season starting SR will generally be close to his previous season ending SR, with some adjustment based on 5 placement matches, which confirms that most players’ season ending MMR and season starting MMR are close to each other.
If a player is inactive, (or inactive in a particular role) his initial SR may show a large jump from the last time that he played, because MMR uncertainty increases when a player does not play11.
If the player has not played since the implementation of role queue, his initial MMR takes several months worth of data into account2. If a player has not played a role, it will estimate that player’s MMR from how other “filling” players have typically done, but start the player off with a high MMR uncertainty2.
Win Streak Bonus / Loss Streak Penalty
Prior to season 5 (June 1st, 2017), there was a substantial bonus to SR for winning many games in a row (about 4 or more), scaling all the way up to 150 SR for one win. To keep things balanced, penalties existed if a player lost many games in a row. This was mostly removed because it lead to people who got lucky/unlucky being thrown far from their true rank28. Currently, streak bonuses exist, but in a more limited form, and are quite difficult to trigger. The current design is much less likely to throw people out of their rank incorrectly, while still allowing accelerated motion for people that really don’t belong there.
Periodically, there are calls on the forums for all MMR/SR information to be wiped and to start over. The justification here is that because the system has had many flaws, and that MMR/SR takes a very large number of games to move, and that people are misplaced, that it would be good to wipe everything and start over. The chief problem with this is that before the wipe, all players would be thrown together. Top 500 would be playing against bottom 500. This would cause matchmaking and rating quality to fall through the floor until a sufficient number of games are played. And because many people do not play a huge number of games, or play inconsistently, huge inaccuracy would persist for months. And at the end of that, the accuracy would be similar to what it was before the reset, because the underlying system has not changed. Periodically, the developers restate that a reset would be a very bad idea2.
Each time a player leaves competitive matches any time before the Victory/Defeat screen, he will receive a 50 SR penalty for that role, and a 10 SR penalty for the other roles. In addition, with increasing leaves the player will get increasing time bans, up to and including a season ban.
Season bans cannot be reversed. The reason or method of leaving is irrelevant29. If a player plays many games without leaving, the leaving penalty resets downward. The rule of thumb is that if the bans are getting into 8 or 24 hours, a season ban is imminent. If a player receives three season bans, a permanent ban is likely30. The season bans do not need to be consecutive to count toward a permanent ban.
If the player leaves before the game is 30 seconds old, the match is cancelled and no one except the leaver is penalized/rewarded. If a player leaves after the game is 30 seconds old,
According to official statements, players on the leaver’s team gets no consideration due to the leaver. That is, they (and the other team) gain and lose SR/MMR as normal. This is to prevent situations where it is the team’s interest for one person to leave and save everyone else from losing SR/MMR (and the winners from gaining SR/MMR)31,32. However, there are credible reports that this was changed in Season 18, with the introduction of role queue. Players are now gaining less SR when they win, but there is a leaver on the enemy team. In the reverse, when players lose, but there is a leaver on their team, they lose less SR37
Recovery from disconnecting is identical to recovery from decay, before decay was removed from the game 19. That is, after a single disconnect a player will have an SR buff for about 12 games until his SR has caught back up to what it would have been if he’d never disconnected. Because leaver recovery is identical to decay recovery, it is likely that MMR is unaffected by leaving, just as it was unaffected by decay.
If a player leaves and rejoins the match (after rebooting / internet comes back, etc.), he will often be able to save himself from getting the penalty, but it is not a sure thing. He needs to rejoin quickly, and not disconnect and reconnect several times29.
Leaving during or after the Victory/Defeat screen is safe.
Exiting the game while in a queue or a group is not safe. The game does not immediately remove players from queue and group when they exit, so a queue can pop (and a group leader can queue up) before the game figures out the player is gone and the player can be penalized as a leaver. Always leave groups and queues and wait a minute before exiting the game.
(1) Season 19 Competitive Season Updates
(2) Role Queue Update
(4) SR, MMR, and Role Queue
(7) https://us.battle.net/forums/en/overwatch/topic/20745504371#post-3 “In Overwatch, whether your MMR goes up or down is contingent on winning or losing. But there are a number of factors that determine how much that rating goes up or down.” … “All the system does when it comes to matching on skill is attempt to match you with people of a similar number”.
(11) Groups and Matchmaking in Overwatch
(16) Forced lose streak
(18) https://us.battle.net/forums/en/overwatch/topic/20753625906 “When you do come back and actively play matches, you’ll also typically gain more SR from a win until your displayed skill rating and internal matchmaking have again reached ‘equilibrium’ “.
See data at 3000+ Skill Rating Data and Analysis (now including DCs). However, note that decay has been removed from the game, so these effects are now only readily visible after a DC (which is the only known way to drop SR by 50 without affecting MMR).
(19) 3000+ Skill Rating Data and Analysis (now including DCs)
(20) Initial Competitive Skill Rating, Decrypted
(22) Competitive Mode Tier Distribution
(24) GUIDE: Enable SMS Protect (required for PC Top 500) - UPDATED
(34) Good news, cheaters cant ruin your SR anymore
(35) Role Queue enabled for Competitive Season 18!
(37) If you have a leaver in comp do you suffer the same decreaase in SR?
Last edited on 11/8/2019