Impact Rating v1.1 [UPDATE]
[v1.1] Changelog: CTL and SND points per kill are calculated using expected rounds won/lost (xRound) instead of actual points scored. xRounds are based on Team Kill Percentages (e.g., ATL had 49% of the kills, so we expect them to score 47% of the points, historically). This change removes some of the noise from low-scoring games’ Impact Ratings.
Some issues were pointed out (rightly) about how Impact Rating is calculated in CTL and SND. This example (that I tweeted myself, smh), shows a huge gap between the “eye test” and Impact Rating:
My thoughts
Simp went 33–26 and lost 3–1, with a far lower rating than Preistahh, who went 16–28. In Control, kills play an important role in determining the winner of each round. This example shows a big blindspot when only considering kills and rounds.
Noisy Scoring
In a game of Hardpoint, teams can score pretty much any percentage of the available points. E.g., a team that wins 250–50 scored 250 / 300 = 83% of the points. This is not the case in CTL. There are only 6 possible scores in CTL: 3–0 (you score 100% of the points), 3–1 (you score 75%), 3–2 (60%), 2–3 (40%), 1–3 (25%), and 0–3 (0%). These are huge leaps compared to HP, and we see a similar phenomenon in SND.
A single gunfight could be the difference between winning 3–2 (scoring 60% of ALL the points) or losing 2–3 (scoring 40% of the points). In Hardpoint, to swing from winning with 60% of the points to losing with 40% would require the score swinging from 250–166 to 166–250 or an 84-pt swing! That will never happen with one kill, but it can in CTL!
In the ATL vs BOS example, ATL scored 25% of the points (1 of 4 potential round wins), despite getting 49% of the kills (88 of 178 engagements). Scoring 25% of the points is the equivalent to losing a HP game 83–250 (83 / (250 + 83) = 0.25). IF Simp had gone 33–26 in an HP and ATL had lost 83–250, he deserves a very low rating, in my opinion. But this is not HP!
Teams won 47% of rounds in CTL when they got 49% of the kills last season, or 1.89–2.11 in “expected Rounds Won” (xRounds) in this game. I would argue that, based on this fact, we are not only over-punishing ATL, but over-rewarding BOS for a close game.
How can we fix this?
Rounds won in CTL and SND are noisy, so we need to try toextract more signal from this noise by using Team Kill Percentage directly.
While rounds won is a noisy metric, there is still a strong correlation with team kill % (TKP, the percentage of kills each team had). Instead of simply considering rounds won in SND and CTL, let’s consider xRounds too, using TKP as the predictor. In the ATL example, ATL got 49% of the kills, and we predict ~47% of the rounds if the game were played 1MM times.
Predicting xRounds from TKP:
Two simple formulas here, both derived from MW2 data:
CTL xRounds % = -1.79 + TKP x 4.57
SND xRounds % = -0.49 + TKP x 1.97
Using the ATL v BOS example above, ATL TKP = 88 / 178 = 49%, so
ATL xRounds % = -1.79 + [0.49 x 4.57] = 46.9% of 4 Rounds Played
46.9% of the 4 rounds played would be 1.88 xRounds. We replace the 1–3 score with a “noise-corrected” 1.88–2.12. NOW we do what is listed in Impact Rating v1.0
- Points Per Kill (PPK) = Points Scored / Team Kills
- Points Against Per Death (PAPD) = Opponent Points Scored / Team Deaths
- Impact = Player Kills x PPK — Player Deaths x PAPD
1.88 xRds / 88 Team Kills = 0.0213 PPK
2.12 xRds / 90 Team Deaths = 0.0236 PAPD
Player Kills x 0.0213 PPK — Player Deaths x 0.0236 = Raw Impact
Wrapping Up
See adjusted values below. Let me know if they pass the “eye test” better!
Note for the nerds:
I had previously tweeted these coefficients [here] to convert Raw Impact into Impact as shown above (centered at 1.0 with an s.d. = 0.33). Because of v1.1 changes, the coefs for CTL and SND are now 1.626 → 1.43 and 2.88 → 2.74, respectively.
