What is a teamfight? There’s a good chance you already have a picture of it in your mind. Some might say a teamfight is when players use ultimate abilities, others might say a teamfight is where multiple players are eliminated. But before we try to clearly define teamfights in Overwatch, let’s take a step back to help educate new viewers about the concept.

Overwatch is primarily a game of objectives, rather than strictly eliminations and deaths. To win a map, you must capture more objectives than your opponents, and to do this teams must battle. These battles are what we call teamfights: two teams engage and one team emerges victorious, freeing them to progress towards an objective unhindered. In a way, teamfights are like possessions in basketball, crossed with rounds of boxing. They have dynamic length, one team attacks while the other defends, and both sides are trying to beat the crap out of each other to score a point. Eliminations and deaths—or not dying, rather—are the things that make you more likely to win a teamfight; hero damage, healing, ultimates, and crowd-control abilities are the currency a team spends to achieve those eliminations.

To understand what goes on before, during, and after teamfights, we need to define two important moments: the start and the end of a teamfight. As mentioned before, not every teamfight has the same duration. Some may be over quickly, and some may last for several minutes, especially in overtime situations.

Creating a method for determining the start and endpoint of teamfights is essential but also difficult, and this is where Blizzard Data Science comes in. Thanks to the work of one of our resident data scientists, the Burke Model—named after its creator, William Burke—was developed to find these critical teamfight endpoints, using statistics from the first season of the Overwatch League. The model itself makes use of scary things like calculus and will be refined over time, but the key takeaway is that it works by driving down into the most basic stats in Overwatch that contribute to eliminations in teamfights.

Objectives are won with teamfights, teamfights are won with eliminations, and eliminations cannot happen without a combination of damage, healing, ultimates, and more. Not every teamfight begins with an elimination; therefore it makes sense to attempt to capture those pre-death moments where a purely elimination-based model would fail.

In the Burke Model, this has all been considered. The model considers the beginning of a teamfight to be the snapshot in time when both teams’ combined output rate of carefully chosen statistics exceeds a certain threshold. The teamfight continues as long as this rate of output is maintained or exceeded. For the teamfight to end, the output must drop below the threshold for a prescribed amount of time. It looks a little like this:

It is important to note that this is a living model that will be continuously improved as we develop new techniques to evaluate the endpoints of teamfights. However, the output will always remain the same: its purpose is to automate the process for determining teamfight endpoints. These endpoints have already led to the creation of a new suite of statistics that will help us tell the stories of the league’s best teams and players.

One of those statistics is simplistic in nature, but vital in importance: determining the winner of a teamfight. In order to build better statistics that take teamfight wins into account, I decided that the definition of a teamfight win must be simple and understandable. I found that assigning a win to the team who scored more final blows than deaths accomplished this goal. This does carry the caveat that it is possible under this definition to draw a teamfight, however.

I deal with draws by utilizing filters. Draws in this model tend to occur when it detects short teamfights that are likely not actually teamfights, and these should be eliminated after further improvements to the model. By simply setting a minimum teamfight duration of 15 seconds, I can drop the rate of teamfight draws to under 7% of the total. If I’m considering win rates for general purposes involving a large data sample size, I filter out draws entirely.

Now to the meat: with teamfights and teamfight win rates, we can do some pretty cool stuff. Not only can we determine how often teams take teamfights and how often they win them, we can also do this for players as they’re subbed in and out of team lineups.

The following graphic shows how often every team has been engaging in teamfights, as well as a specific example of how London’s team win rate looks while individual players are in the lineup:

We can see that teamfight pace doesn’t necessarily correlate to win rate: the Washington Justice have the lowest teamfight pace and win rate, but the Dallas Fuel’s league-high pace hasn’t done them many favors. However, we’re only two weeks into a single stage. Last season, the higher a team’s teamfight pace was, the lower their teamfight win rate tended to be. I reasoned that although weaker teams were less aggressive than their opponents, stronger teams were forcing them to take teamfights faster than they would have preferred. This will be something to keep an eye on as the 2019 season progresses.

Moving on to London’s specific player breakout, we can see that London’s decision to substitute Ji-Hyeok “Birdring” Kim out in Week 2 paid off, as their teamfight win rate with Hee-Dong “Guard” Lee has been much higher. This is the type of analysis that teamfight data unlocks, and these observations simply scratch the surface of the fun things we can do with teamfight data.

Ben "CaptainPlanet" Trautman is the statistics producer for the Overwatch League global broadcast. Follow him on Twitter!