This article is not about the players themselves, but what stats make a good player. I truly believe that more and more viewers are believing stats have some value. But so what? Stats have value; how much value? Based on what? Which stats are better? How can we compare players that play completely different roles? There are a few questions I want to address in this article. Let’s go one-by-one.
We have spent the better part of...forever, regarding KD ratio as the king of stats. This is for a couple of good reasons:
I would call this elegant in its simplicity. But what does the best KD mean? It means, literally, that the player is the best at getting kills without dying. He can achieve that only two ways: getting more kills or dying less than his enemies. Do teams win games for highest KD? When a player reaches a certain level of dominance, do they call the game off? No. Granted, KD is strongly correlated with winning, but kills are not essential to winning. In fact, Red Reserve became popular last year for their high win rate in games in which their whole team died more often than the opponent. If you said RR wasn’t good last year because their KD wasn’t good, you’d have been disappointed by most of their placings. So what does make a player the best? I would argue it is his ability to give his team a chance to win. Winning is the ultimate goal and we should judge our players on that. But we can’t just go and assign every player on the team with the most wins as the best player, that doesn’t make sense. All players on a team are not equal, producing varying stats in each game. A player can have a great game and still get the loss. Luckily, we can use math to help us determine which stats lead to high probabilities of wins.
As a math-less example, let’s say these are two players on Team Codstats.gg after a big loss:
Who gave us a better chance of winning? JP did, of course (just like he does when we actually play together). But how much better? And how accurately could we even say by knowing just our KD? We can determine this pretty easily using data from the last event.
Let’s first build basic models that just use one stat: KD, Kills per Minute (KPM), Score per Minute (SPM), or Damage per Minute (DPM) to predict whether a player won or not. The model will be trained by being fed just a singular player's stats and the game result. That is, teammate scores don’t matter to the model. We can go even further and say that the model has to assume league average teammates because it must assume every performance it cannot see is the average of everything it has seen. Since predicting Win-Loss is binary, I’ll use Log loss as my measure of accuracy. Log loss just means that the penalty for being wrong is the log of how far my prediction was from the truth (100% chance being a win and 0 being a loss). The higher the score the better the predictor. Here are the results for Hardpoint ONLY:
So wait a second...KD provides the best single Log Loss values in our models for ALL three modes (provided that you assume KD is by far the best indicator of SND success)? I was at first sort of shocked by this. This was NOT the case in previous COD titles. So why in BO4?
Kills used to be awarded to the player who put the finishing touches on the dying player. This led to times where the killing player was responsible for a tiny proportion of the total damage done. In BO4, the player responsible for the largest proportion of the damage gets the “kill” no matter what. This is HUGE. This makes kills mean so much more. If you are tagging guys a lot from range, you are still getting kills if you do the most damage.
While SPM, Dmg/Min, and KPM all use one stat divided by the duration of the game to normalize them, KD takes into account a players kills and deaths. This is a huge advantage when we think about it logically. If I asked you to guess the area of a rectangle but only told you the width, you’d be in a lot of trouble… Let’s add another variable to our other factors: SPM + Deaths per Minute (DPM), Dmg/Min + DPM, KPM + DPM and see how they compare to KD at predicting wins
There we have it. Once we even the playing field and let every stat have as much info as KD, we see that normalizing stats by duration is a better predictor of winning that KD alone
We can actually use a players Score, Kills, Deaths, and the Duration of the game to come up with an estimate of how likely his team was to win based exclusively on a single scoreline. Pretty sweet. Let’s look back at our example of the Codstats.gg team in our epic loss from above. Using my logit model built above, using score per minute and deaths per minute as the two input parameters I get:
There you have it. Even in a devastating loss, JP manages to provide 29x the benefit to the team. Would you have been able to tell by looking at our two scorelines that JP was providing the team 29x the value? I doubt it.