Which stat is important to win a map in OWL?

 Overwatch League (OWL) provides player hero stats (phs): https://overwatchleague.com/statslab

The phs are aggregated at the map level, so here I compare the stats of map winner and loser to analyze which stats are importantly correlated to win the map.

It is tricky to directly compare asymmetric stats, varied by maps and teams.

For example, a team used Sombra at Lijiang Tower, but didn't use her at King's Row. Sombra has unique stats such as 'Enemy hacked', that other heroes don't have. Thus, it is appropriate to analyze the importance value of 'Enemy hacked' at Lijiang Tower, but not at King's Row. Even here, it is worth noting that it may be hard to generalize the importance of the 'Enemy hacked' stat in cases when the opponent doesn't use Sombra at Lijiang Tower.

There are several ways to compare those specific stats, but here I analyze the 'common stats', which all heroes have, such as 'All Damage Done', 'Final Blows', 'Elminations', 'Healing Done', etc.
As the common stats can be applied to almost every hero, I assumed that the players' average of a certain stat at the map level means their team's stat.

OWL provides 36 kinds of common stats for 'All Heroes' (out of total 365 kinds of stats): 
['All Damage Done' 'Assists' 'Average Time Alive' 'Barrier Damage Done' 'Damage - Quick Melee' 'Damage Blocked' 'Damage Done' 'Damage Taken' 'Deaths' 'Eliminations' 'Final Blows' 'Hero Damage Done' 'Melee Final Blows' 'Melee Percentage of Final Blows' 'Objective Kills' 'Objective Time' 'Offensive Assists' 'Shots Fired' 'Time Alive' 'Time Building Ultimate' 'Time Elapsed per Ultimate Earned' 'Time Holding Ultimate' 'Time Played' 'Ultimates Earned - Fractional' 'Ultimates Used' 'Weapon Accuracy' 'Defensive Assists' 'Healing Done' 'Solo Kills' 'Environmental Deaths' 'Recon Assists' 'Multikills' 'Environmental Kills' 'Turrets Destroyed' 'Teleporter Pads Destroyed' 'Knockback Kills']

To normalize the team stats (average stats of players), I put the values in percentiles, of which the value of each team (winner or loser) was divided by the sum of both map winner and loser, so the value of the stats will be mapped into the percentage of the stats (scale of 0 to 1). 

Here are the winner's team stats throughout all the seasons so far (2018, 2019, 2020).

error bar means standard error mean (s.e.m.)

It is plausible that stats like 'Objective Kills', 'Assists', 'Eliminations' and 'Final Blows' are higher than 0.5 (the point where winner = loser), and the stats like 'damage taken' or 'deaths' are lower than 0.5.

Let's check some features in detail.
In Overwatch, you can win a map by taking objectives like the payload and the point. In order to take these objectives, you need to win team fights, and kill enemy heroes.
so I believe that there is no argument that the most critical and direct indicator for winning a map is 'Final Blows' (or simply 'Kills').

This is the distribution of 'Final Blows' (orange: winner, blue: loser; in y-axis, 0: lose, 1: win). As you can find in the figure, 'Final Blows' seems to be clearly separated, which means that higher number of final blows is correlated to win.

Similarly, the opposite concept, the distribution of 'Deaths' is clearly separated as below,

How about damage stats like 'All Damage Done', 'Hero Damage Done' or 'Barrier Damage Done'?

In 'All Damage Done' and 'Hero Damage Done', the distributions seem to be separated but not clear as those in 'Final Blows' or 'Deaths'. Also, the distributions of 'Barrier Damage Done' seem to be not clearly separated, which indicates that the 'Barrier Damage Done' might not be a critical factor for determining a map winner. 

Likewise, the 'Weapon Accuracy' stat seems to be almost inseparable as below,

which indicates that 'Weapon Accuracy' of the winner and loser are almost the same.

Interestingly, 'Healing Done' also seems to be inseparable between the winner and the loser, which implies that both the winner and loser healed nearly the same amount to their teammates.

Unlike other stats, 'Environmental Kills' shows skewed distributions,

This means that the most of environment kills were 0 for both winner and loser, which is understandable if we think that environmental kills rarely occur in OWL (the distributions are not symmetrical because I replaced N/A values to 0).

So far, I compared the percentages and distributions of common stats between the map winner and loser. 

Then would it be possible to predict a map winner using only these stats? If so, which stats among these would be important for winning a map?

To classify a map winner by the stats, I applied a logistic regression model, which shows around 92% accuracy. It means that the model can predict the map winner with 92% accuracy.

Let's take a look at which stats have high correlation to winning in order to predict the winner.

Here are feature coefficiencies, simply 'Importance' for the winner prediction,

As we found in previous distributions figures, 'Deaths' and 'Final Blows' were top 2 critical features for the winner prediction. Also, 'Healing Done' or 'Environmental Kills' seems to be less important for the prediction in this model. 

Let's take a look the latter half of the 2020 season data,

It is very similar to the all time importance data which implies that the matches of the latter half of the 2020 seasons followed the trend of the entire OWL season.

Would the importances of certain stats look different in the GOATS meta? I believe that GOATS was dramatically different from any other previous OWL meta and it will not make a return after the 2-2-2 role lock implementation on Stage 4 of 2019. 

Let's take a look at stage 2 of 2019 season when GOATS dominated OWL.

The model predicted the map winner with 96% accuracy. Unlike previous data, 'Assists' seems to be the most important feature to predict a map winner, probably because of the synergy effects of 3 supports (Lucio, Zenyatta and Brigitte) and CC ultimates from Rein and Zarya (i.e. Earth shatter & Grav). 

Indeed, both 'Offensive Assists' and 'Defensive Assists' were clearly separated between the winner and the loser.

So far, I listed the feature importance of the common stats between map winners and losers. As everyone predicted, the 'Final Blows' and 'Deaths' stats were critical and distinguishable stats from the winners and losers, which means that a team should get more 'Final Blows' and less 'Deaths' to win the map (too obvious). 

With these pipeline, next time I will post the important stats of specific Hero such as Sombra's 'Enemies hacked', Mercy's 'Damage Amplified' or Widowmaker's 'Critical shots' rather than common stats.

For further analyzing, please find the source information: https://drive.google.com/drive/folders/1Ex-t0jrhXwL0v58JC35uZFOdGgJh8ZaH?usp=sharing