Too Long, Didn’t Read
Initial competitive skill rating (SR), for accounts that have never played competitive, is approximately 2350. This leads to new accounts being placed closer to high gold that is correct.
There is substantial luck based variation in initial placement.
Quick play has no influence on starting SR.
Initial placement is not dominated by statistical performance.
For the first eighteen games, including placements, SR change, for average players, is approximately described by SR change = -3.85 * game + 99.34.
Top players can gain more per game, and have exaggerated gains for up to twenty-five games or so. Bottom players likely gain less.
Blizzard has not revealed many details about how initial placements work. For this reason, there have been a number of contradictory theories that are accepted by large segments of the community. Therefore, I decided to create a new test account and take careful and thorough data, available at https://docs.google.com/spreadsheets/d/1Fk0_ZYtkevjX647PV9mZZH6Qt6jY8jWHDNQ8T1m4F7Q/edit?usp=sharing, in order to better understand the initial placement process.
Does Quick Play Matter?
One question is whether a player’s starting 25 levels has any influence on where a player is initially seeded in competitive.
Several forum posts indicate that new players start in high gold (or maybe low plat) at least for top players.
Jeff Kaplan has also stated that competitive MMR is separate from quick play MMR, but he did not really address the question of initial competitive placement (https://us.battle.net/forums/en/overwatch/topic/20760837867#post-3).
I wished to verify that quick play does not affect placements, and verify it for low ability accounts. In this account, I played quick play (and only quick play) atrociously for 25 levels. Think simulated 6 year old child. I insta-locked Mercy when possible, didn’t heal, didn’t dps that well either, etc. I never spoke in voice or responded in chat. I’d still get medals fairly often through Mercy’s self healing (yes, the healing in the low tiers of quick play is that bad). My overall win rate was 24%. It is difficult to judge the SR of people down there because so few of them play competitive, but I did inspect all profiles and (towards the end) opponents were usually in bronze, and occasionally in silver. Most accounts had less than 100 levels played. I generally played during prime times (evenings and weekends) because I quickly noticed that if I played in the early morning on a weekday, I would be placed into games that were not appropriate.
I then played my first ever competitive placement matches at the beginning of season 9 (to assure more active competitive players), and played to win. In each game, I inspected the career profiles off all opponents, and recorded their most recent competitive SR. I also recorded player level and grouping, but didn’t see anything interesting there (player level doesn’t matter and was much higher than I was seeing in quick play; groups tend to be matched against groups of equal size). The data is at https://docs.google.com/spreadsheets/d/1Fk0_ZYtkevjX647PV9mZZH6Qt6jY8jWHDNQ8T1m4F7Q/edit?usp=sharing. In the first match, 9 of the twelve players had a rating in previous seasons, with an average rating of 2315. It is apparent that I started in high gold. So at this point, we can definitively say that quick play does not affect competitive, even in initial placements.
Are Placements Results Dominated by Statistical Performance?
I’ve seen people say this on the forums from time to time, but never with any support. My data certainly contradicts this hypothesis. Looking at the team averages in my placement games (which are admittedly a rough measure), the team averages went up on a win, and down on a loss, for every game except number 10. My understanding (and data) is that placements are like the rest of the season, but with wider swings. That is, win or loss is most important, with performance being a smaller modifier on top of that.
After taking the data, I looked at the change in my SR per match (visible on the second sheet after the data tab, the analysis tab), for games 11-20, and noted that the relationship was linear until game 18 or so, when the games have typical established account movement (20-30 SR). This is exciting. It means that we can make a fit (with performance related outliers removed) and extrapolate backwards to determine how much SR is gained or loss each match, even during placements. From this, I find SR change = -3.85 * game + 99.34. I fill in the values in the table, under the column “Estimated or Actual SR” and find that my SR for the first game was 2366, which is in good agreement (considering the roughness of the calculations) with my estimate of the SR for the first game based on averaging all known SRs (2315).
Other player results
Lately there has been a rash of streamers leveling up smurfs. See https://docs.google.com/spreadsheets/d/1abyFItI1HLNKOmyUKadbcMPnzwFV3TGiOIIYB7HvowE/edit?usp=sharing, PlatChat 11 (Kabaji), Rainbow 10 (Kabaji), and Dmum 11 (Custa). Generally, these confirm that initial SR in competitive is around 2350. However, the equation I derived does not produce very good agreement, likely because these streamers have much higher performance than me. However, winning or losing is still more important than performance alone (see how the ranks of the opponents change with each match). They also can have exaggerated SR gains and losses for up to 25 games. Note that none of these players had great placements records. This is because they made special challenges (such as having chat choose which hero is played) to keep things interesting, and there is a lot of randomness in late season placement matches.
M0rchPonkey below shows what happens when someone sweeps their placements, and has good statistical performance. He placed 3590, which is higher than I would have predicted from my data alone. He also repeated the test of throwing quick play, with the same result, that quick play does not affect competitive.
DongQuixote has done a whole rash of accounts (Best approach to place high on new account?), and has shown that initial placements can be all over the place for a given player: 3550 (10-0), 3267 (7-1-2), 3003 (6-4), 2780 (7-3), 2600 (6-4), 2300 (6-4). Which drives home that a player cannot be accurately placed in only ten games. It also repeats that sweeping placements can cause players to place in low masters.
Since new accounts start in high gold for their first match, this leads to accounts being placed closer to high gold than they should be. Poor players tend to be placed silver / low gold and fall down. Better players tend to place plat / diamond and grind up. Gold players (such as myself) are likely placed okay. However, good or bad luck can throw initial placement off, as 10 games isn’t really enough to place people accurately. The extreme range is roughly high bronze to low master, but most players will place closer to high gold than this, regardless of where they belong. Note that even top 500 players can have trouble winning all placements and hitting masters immediately.
I suspect that new accounts actually start at the median SR for the community, which according to Jeff Kaplan is somewhere in gold (Competitive Mode Tier Distribution). This makes sense in a strictly mathematical sense. An unknown player is equally likely to be above as below median. However, I generally expect a new player would be below median, because they have no experience. But a smurf would generally be above median. So what really matters is where the median new player eventually ends up which depends on how many smurfs vs new players we have. What I suspect is that it is much more likely for new players to sink down into silver/bronze than to shoot up to diamond+. However, without access to Blizzard’s data, it is hard to be sure.
Also realize that it is a very poor new player experience to place low gold / high silver and then sink down into bronze where he belongs. For this reason, I think it is best to start new accounts lower, perhaps 1500, and work them up quickly if performance metrics imply that they belong much higher.
There is also the question, why is the 25 levels of quick play data thrown away? On one hand, it is more pure to have zero influence of quick play on competitive. On the other hand, it makes the competitive experience worse for everybody. When you see a player with less than 20 competitive games (lifetime) on your team, you have no idea what ability player he actually is, but there is a very good chance that he is incorrectly placed. For the person on the new account, he will often have a poor new player experience as he grinds up or falls down to where he belongs.
There is the problem that new accounts can level up entirely in arcade or custom games, so I propose that new accounts be required to grind 100 (or however many is required to get good data) quick play matches before entering competitive. This quick play data should be used to seed the competitive data (but not too aggressively, so people don’t try-hard in quick play to a level much higher than they belong).
Caveats and Warnings
These results are not valid before season 6. Before season 6, players’ starting SR was artificially lowered below their MMR to give them an artificial experience of climbing for the first 50 games or so.
Streak bonuses (at least for established accounts) were removed at the beginning of season 5. It is virtually certain that streak bonuses existed during placements prior to season 5.