MMR Game Score

[cubone]

Since my original post is pretty busy and there's a lot of updates/comments in there, I thought I'd start a new topic with a new sheet. Aside from the previous sheet of overall player data, winning percentage by teams, and percentage of kills attributed to each player, I got bored and decided to add a new tab with a calculated game score.

This tab is a little slow since it is essentially game data for every player and every game. Aside from the standard info, I've added spawns, strays, number of KOs, plus/minus, etc. I then took a combination of kills (weighted down based on range and with strays completely removed as valid kills), players KO'd, TK penalties, a win bonus for actually winning the round (to hopefully penalize players who just lob kills for their rec at the detriment to their team), and a penalty for dying out quickly based on the player's total time played vs the total round time.

The weights are just kind of an estimate for now. Future things I might add (if I have the energy) would be:

• Use my calculated player scores to weight the kills for each round based on quality of player killed (e.g. killing Commodo is more valuable than killing thesauce)
• Perhaps a penalty for spawnkills?
• Larger penalty for TKs? I essentially just made it the negative equivalent to an enemy kill
• I didn't factor in deaths
• Integrate the overall plus/minus for the team into the score?
• Perhaps some sort of efficiency for high kills/low bombs+bullets fired

https://docs.google.com/spreadsheets...gid=1419459706

Also: Would like to look at kill by kill data to determine 2fer/3fer/4fer/5fer and add that as another data point.

[Rab]

why do I not exist?

[cubone]

why do I not exist?

Looks like a combination of two things:
1. The data starts as of 12/02/2022 and is through mid-May right now
2. I filtered it only for twjd3

[daresound]

Looking forward to the day cubone uses this data to copy us all as bots to keep TW alive, and we get to rewatch exalt's bot losing to tittygripper

[cubone]

Sheet has been updated for 06/16 data provided by BIET

[cubone]

Also integrated the MMR rating from the API

[cubone]

Added a couple of things:

• Scaled MMR rating to have a comparable value to the overall calculated score
• Looked at the difference between rating score and MMR to determine overall over/underperformance relative to MMR rating

[Dwopple]

Holy smokes. You sir are a GOAT.

[Sprackle]

this is great, ty ogron

[cubone]

Decided to add something else I've been thinking about. This is the percentage of total kills by my created player score. Scores are represented as deciles to break the level of players into 10 levels (with 10 being the top 10% of players, 1 being the bottom 10% of players). The idea is to see what percentage of each person's kills come from each level/quality of player.

Tab name in Google Sheets doc is 'Pct Kills by Rating'.

https://docs.google.com/spreadsheets/d/1hmgW5T9blYCuSGSg-8dSPdZHd1LH72Sxc3LoMZE8smI/edit?usp=sharing

[Rab]

Interesting, there'll be patterns to that. Some pros will focus on opponent pros. Some ppl will focus on eliminating the weakest players first. Maybe you can work out which tactic is best (sort them by win% ?).

There's a timezone component too, games in my timezone contain the full range of ability players, but games later on when I go to bed tend to be all high level players, dunno what happens later on.

[cubone]

Next step I think will be integrating these into the individual game scores I calculated. Obviously a game of 15 kills of top tier players is much more valuable than killing 15 bottom tier players. I'll weight kills something like

10-8	100%
7-5	80%
4-1	50%

[MHz]

Next step I think will be integrating these into the individual game scores I calculated. Obviously a game of 15 kills of top tier players is much more valuable than killing 15 bottom tier players. I'll weight kills something like

10-8	100%
7-5	80%
4-1	50%

Ostensibly this makes sense, but I think persecuting weaker players - and moving to 5v4 situations - is a winning strategy, so you'd have to be careful about punishing that in your stats. Havok is a good example of a player who constantly tries to harass the "better" players, but gets boned 70% of the time.

Sometimes buzzing/shadowing the top opposing talent can be good, it's a strategy I use too, but the ultimate goal is to get the opposing team to 50 combined deaths, not fight honourably.

[Dwopple]

May have been answered directly or indirectly already, but is this live data or static?

Edit: big fan of your work, please keep it up as long as you're motivated!

[cubone]

Ostensibly this makes sense, but I think persecuting weaker players - and moving to 5v4 situations - is a winning strategy, so you'd have to be careful about punishing that in your stats. Havok is a good example of a player who constantly tries to harass the "better" players, but gets boned 70% of the time.

Sometimes buzzing/shadowing the top opposing talent can be good, it's a strategy I use too, but the ultimate goal is to get the opposing team to 50 combined deaths, not fight honourably.

I think that's a fair point, but in terms of creating a game-by-game score to try to rank the best games, I think getting 20 kills against Turban/Ease/Zidane would be a lot more meaningful than 20 kills against Stinging Metal/coupe/avantgarde. I agree that it's not necessarily a bad thing (it's actually a solid strategy) to hunt lower level players, but if we're weighing 2 games against each other, kills vs high level players should have more weighting.

In terms of ranking players overall, I think it's a good point and you really can't penalize someone for killing the enemies they were given.


Also, to add here, these ideas are kind of what are floating around in my head that I would like to implement. If people have other thoughts/ideas, I'm open to them.

• Implementing player ratings into game-by-game score (as outlined above)
• Go through kill-by-kill to determine 2fer/3fer/4fer/5fers
• Look at player coords for each player and each kill/death to see how close they are to other teammates (it's not perfect but provides 80-100 datapoints of the coords of every player in game to determine people that stick with the team and those that don't)
• Perhaps come up with some kind of metric by MMR/my rating to determine who is over/underplaying their rating (is a 3000-level player significantly outplaying other similarly rated players)
• I haven't added my time-wasting metrics to the Google Sheets doc so I might add that
• Discussed with autopilot maybe creating a few metrics for each player based on certain stats
  • Could be something like:
    • Offense (kills/kills per minute/range/level of player killed)
    • Teaming (avg distance to teammates/TKs/effect on winning percentage of other players)
    • Clutch (wins as last player alive/KOs/2fers,3fers,etc, win %)
    • Defense (average range of deaths/% of games alive more than XX% of round)