Duel Swing: a pure firepower metric

Note: if you hate maths and just want to read the Duel Swing explanation, skip this part.

When comparing professional Counter-Strike players, we typically evaluate them using loose categories like overall rating, game sense, firepower, supportiveness, and leadership ability. There are a number of simple-to-measure statistics that we can combine in some way to approximate these categories. For example, you might want to approximate "firepower" by taking kills per round, damage per round, opening kill attempts, etc. and scaling them between 0 and 100 to produce a single number. The value of doing this is obvious—you take different values that all point roughly in the same direction (higher = better) and collapse them into a single data point. But it comes with significant drawbacks.

Whenever you combine metrics to produce a new measure, you're doing what statisticians call a projection. In other words, "summarise this mess of numbers into one direction that I care about." What you lose is how each individual metric contributes and what other directions they may have also been pulling in. For example, a player with high kills per round but low opening kill attempts may end up with a similar "firepower" value to a player with average kills per round and high opening kill attempts (using our trivial version of "firepower" from earlier.) So if both players have a value of 70, you cannot know which of the two types of players it corresponds to, and thus you need to look at the metrics themselves again anyway.

This isn’t a flaw in this particular formula. This will always happen when you use statistics as proxies for concepts instead of trying to measure the concept itself. It is impossible to come up with a "score" or "rating" number that is based on other values that everyone will be happy with, because baked into these formulas are the designer's personal beliefs about how heavily to weight each component. Or in some cases (like PCA clustering), you let the computer do it itself and get some funky results.

Great for Twitter posts, not so great for explainability. See my post below and ask yourself, do you know what this tells you other than that donk is really special? Your pre-existing biases are doing all the work here in conjuring the kinds of stats he must be special in, based on his position on the chart. But if you didn't know CS, you couldn't possibly tell if he was uniquely good or especially bad, nor what kind of stats went into making the two component axis. Ratings like HLTV's 3.0 are functionally the same.

The only way to avoid this problem is by having one metric per concept and knowing exactly what that metric is tracking. That's precisely what we set out to do with Duel Swing.

What is Duel Swing?

Put simply, it is the difference between expectation and reality in terms of gunfights. We talk about gunfights in probabilistic terms all the time: 50/50s, 70/30s, etc. When we watch a match, if we see an AWPer whiff a long range shot onto an opponent with a Deagle, and said Deagle responds with a bullet through the dome, we rightly say "he really should have hit that." It was an obvious whiff followed by an impressive shot. This was only one kill (and one death, respectively), but it's much more impressive than winning an even-matched duel or a duel that you're favoured to win.

Eco frags shouldn't count towards your "firepower" as much as kills vs full buy opponents, and even those shouldn't count as much as donk getting a Deagle ace. So Duel Swing simply assigns each gunfight a prior probability for both players, using their health, armour, equipped weapon, distance from each other, vision status, etc. Then, when the gunfight concludes, the victor gains the probability of their opponent winning the fight and the loser is deducted the rest. So in the AWP vs Deagle case, let's assume the AWPer is assigned an 80% chance of winning the fight (meaning 80% of the time, the player with the AWP wins in that situation). They are awarded +20% if they win, or -80% if they lose. Conversely, the Deagle gets +80% if they win, or -20% if they lose. Big number good, small number bad.

Because it takes into account the weapons of both players, their health, armour, position on the map, and so on, a kind of built-in eco-adjustment happens. When you look at a player's Duel Swing after a map, you can get the cleanest possible answer to the question, "of all the fights they took, did they win more than they should, less than they should, or about the right amount?" In our humble opinion, this is a much better indicator of "firepower" than some convoluted rating with eco-adjustment after the fact.

How should I interpret it?

Because the sum of Duel Swing for a match, map, or series will always equal to zero, it tends to fall into a pretty neat pattern:

• A +500% Duel Swing player utterly dominated their opponents, and usually did so from disdavantageous positions. They were winning clutches on low health, winning fights with pistols and SMGs, hitting really difficult shots, etc.*
• A +250% Duel Swing player won more fights than they were supposed to but didn't completely take over the server.
• A 0% Duel Swing player did exactly as they were expected to, given their resources and situations.
• A -250% Duel Swing player had a few key misses and lost more gunfights than they should have, but didn't completely throw.
• A -500% Duel Swing player is almost always a high-resource player who spectacularly underperformed, often single-handedly losing the game.

*Examples of +500% Duel Swing include Magixx on Dust2 vs FaZe at the 2024 Shanghai Major and Mezii on Inferno vs FaZe at the 2025 Budapest Major. Poor FaZe.

FAQs

Where can I find Duel Swing?

Right now, only here. We're working on integrating it into our site publicly, along with a plethora of other new metrics that we think you'll love. While we do that, you can be sure to find it in our articles and on our podcast.

How does the model calculate the priors?

Using information directly from the demofile (including but not limited to: player positions, health, armour, weapons equipped, vision, and distance from each other), a machine learning model trained on almost a year's worth of matches learned to accurately generate duel outcome probabilities for both players. This was calibrated via the eye-test method on a number of professional matches and benchmarked highly across a litany of measures.

What counts as a "duel"?

For two players to be engaged in a gunfight, at least one player needs to deal damage to the other. The possible outcomes are player one killing player two, player two killing player one, or both players surviving. Grenades and utility damage do not count. (It's in the name, gunfight.)