From xG to META: A Data Journalist's Notes on Reading Esports

Russia Night was the first time I saw a number that could hurt.

I was nineteen, a second-year student in Busan, typing every German shot into an xG model I had written in Python. Twenty-three shots. The result appeared on screen: 1.32 xG, zero goals, a 0-2 defeat. Comparing it with the highlight reel I had just watched, I realised the naked eye had been fooled by what I call "the feel of the ball". Eighteen of those twenty-three shots came from outside the box. The reigning champions were not eliminated by a miracle. They were eliminated by a sequence of tactical decisions that only data could see.

That was the day I understood something that later became the foundation of my entire career: the emotion of the crowd is a data source with a very large error term. People remember the shot that hit the post, not the eighteen aimless shots before it. People remember the conceded goal, not the thirty minutes of meaningless possession. And people, including the younger me, mistake what is easy to see for what matters.

Years later, when I moved from football into reporting esports for the Korean market, I carried that same model with me. I carried the scepticism too. And I discovered something interesting: esports has more data than football, yet reads its data worse.

Context: when data is abundant and the foundation is thin

Esports is a paradox. It is arguably the most thoroughly digitised sport in existence. Every click, every position, every item purchase is logged by the server. Publishers hold a data trove that any football federation would envy. While I had to struggle with cameras and algorithms to estimate xG, in esports everything is already sitting in the log.

But abundant data does not mean correct analysis. On the contrary, I would argue that the sheer excess of esports data is producing a new disease: the disease of extracting a single figure from a small sample and building a large conclusion on top of it. A player posts a high rating across three games and is instantly a "rising star". A team wins its group and is instantly a "title contender". Three games. Three games is a sample so small that I do not permit myself to say anything with it.

Before arguing about wins and losses, I have to question the numbers first. The first question is always: where did this figure come from, and under what conditions was it produced?

In football I ran into the same problem. In 2026, when K League 1 became the first league in the world to resume in front of empty stands, my xG model began to drift. I collected 152 matches and found the home-win rate had fallen from 46.2% in 2026 to 31.6%. I wrote a forty-page report concluding that every ten thousand spectators were worth roughly plus 0.08 expected goals for the home side.

The 0.08 coefficient does not measure the silence; it measures what we lost. What was lost was not the cheering. What was lost was a foundational variable that an entire generation of models had quietly assumed to be a constant. When that variable vanished, every conclusion built on it collapsed.

Esports has an equivalent foundational variable, and it goes by another name: the server version. This is what most esports readers overlook. A team competes on a tournament server whose patch differs from the practice server. A player trains for three months on one META, then walks onto the big stage in another. If I do not record the version the way I record the date of a survey, every comparison I make is a comparison of two different things.

The core: reading a patch the way you read a verdict

Every meta update is a confession by the publisher.

I do not mean this metaphorically. I mean it technically. When a publisher reduces the power of a champion, a weapon, or a map, they are indirectly admitting that in the previous version there was a dominant strategy they considered unhealthy. A patch is a readable document. It does not merely list changes. It tells the story of what had become too strong.

Based on my experience following matches, I process a patch through four fixed steps, and I do not allow myself to skip any of them.

The first step is to determine magnitude. A small patch, adjusting a few percentage points of damage, usually does not change the META. A large patch, altering the mechanics of a teamfight or changing the weighting of a map, can invert the entire priority order. What matters is distinguishing numerical change from mechanical change: the first adjusts, the second restructures.

The second step is to find the beneficiaries. Not the directly buffed, but the indirect winners. When a publisher nerfs a dominant champion, the biggest winner is often another champion never mentioned in the patch. This is where esports analysis most often fails. People read a patch like a scoreboard, when it should be read like a chain of dominoes.

The third step is to cross-check against real data. After predicting, I must verify with pick rate and win rate across at least several dozen games. I learned in 2026 that a prediction without verification is just a belief dressed up in jargon.

The fourth step is to check consistency across regions. A META can be read differently in different regions, because tactical histories and team quality differ. When a champion has a high pick rate in one region and a low one elsewhere, what is being measured may be tactical culture rather than true strength.

I want to pause on the second step, because I consider it both the most important and the most underrated.

Imagine a patch that reduces the damage of a top-lane bruiser. The surface reaction: that champion got weaker. But a deeper consequence: the team no longer needs to spend resources to deal with it. Those resources move elsewhere. Perhaps pressure on the bottom lane rises. Perhaps the importance of early teamfights falls, pushing strategy toward the late game. The entire rhythm of a match can shift, and people will assume it was all about "form".

This is why I say esports reads its data poorly. It has too many numbers, yet too few people bother to trace why a number changed.

The evidence chain: from patch to result

I build my reading of esports on an ordered chain of evidence, rather than leaping straight to the result.

The first layer is the patch. The second is pick and ban rate. The third is objective control. The fourth is resource differential. And the fifth, the last, is the match result.

This order matters. People usually read it backwards: they start from the result and go looking for numbers to justify it. That is a systematic form of motivated reasoning, and I have refused it since I was twenty.

Pick and ban rate is the most valuable layer of information esports possesses, and it is severely wasted. Here I learned a lesson from myself: a high pick rate does not mean strong. Sometimes it only means easy. A safe, low-risk choice that fits many compositions will have a high pick rate even when it is not the optimal ceiling pick. A high-ceiling but difficult choice will have a low pick rate and appear only among teams confident enough to use it.

The gap between pick rate and win rate is precisely the gap between what is easy and what is strong — and most esports analysis erases that gap.

Objective control is where I see the most misunderstanding. In many titles, objectives are resources that convert into advantage. The number of objectives secured is often used as proof of dominance. But as in football, what interests me is not the number of objectives, but the cost of securing them. A team that takes an objective for free while the opponent is mispositioned on the map is one story. A team that takes an objective after a five-for-five teamfight is a completely different one.

In other words, I separate "obtained" from "paid for". This is exactly what I once did with my football foundation. In Morocco in 2026, I analysed three knockout matches and found they conceded possession 71.6% of the time yet only conceded one goal, while opponents generated 4.02 xG in total. The most striking metric was a PPDA of 25.1, nearly double the league average of 13.2.

A PPDA of 25.1 — sitting deep is not a concession, it is stretching the pitch. Morocco did not lose in a passive state. They deliberately allowed the opponent to pass in harmless areas, pulled them out of structure, then punished them in transition. Korean media at the time called it "being pinned back". The data said the opposite.

And in esports, the same logic applies perfectly. A team that accepts a small early disadvantage to preserve resources for the decisive phase is making a deliberate tactical choice, not displaying weakness. When I read a data board, I always look at how a team is spending its resources, not merely at how much it owns.

The resource differential layer is where I apply my discipline about sample size. A small gold or experience lead at minute ten, observed in a single game, is nearly meaningless. It only becomes meaningful when I have at least several dozen games with the same matchup and the same version. This is what I learned from the 152-match report in 2026: a number only speaks when I know how many observations it was drawn from.

And finally, the result layer. I treat the match result as the layer with the lowest emotional error but the highest signal variance. A single match is a sample of size one. And a sample of size one, in any science, is never enough to conclude anything.

The contrarian angle: correlation is not causation

This is my favourite part to write, and also the part that loses me the most readers.

In esports, there is a popular belief that the team with better objective control is the stronger team. Statistically this is trivially true. Teams that win more often control more objectives. But the right question is not "do strong teams control more objectives", but "does controlling more objectives make a team strong". These two questions are entirely different in causal terms.

I have seen ranking tables that order teams by average objectives per game. This is like ranking a footballer by goals scored while ignoring chances created. It is correct as correlation, but it does not tell you what would happen if that team changed its strategy.

In esports this is subtler still because of selection. Strong teams tend to pick compositions that let them control objectives early, because they have the skill to do so. Weaker teams are forced into other strategies. So when I see a strong team with high objective counts, I cannot tell whether that is the cause or the consequence of being strong. This is a form of selection bias, and it is everywhere in esports analysis.

Another contrarian angle lies in the transfer market. A transfer fee does not measure talent; it measures the buyer's desire. This holds in football, and it holds even more in esports, where the transfer market is young and opaque.

I remember 2026, when I connected with a sports data company in Lisbon. From that source, I discovered a midfielder who had played only 564 minutes the previous season, far below the 1,200 minutes recorded in his contract. I sent his agent a six-page metrics report. On 8 June 2026, I was the first to report the loan deal with a 2.8 million euro purchase option.

What matters is not the 2.8 million euro figure. What matters is the gap between 564 minutes and 1,200 minutes. That gap is not a lie. It is a different way of counting. And had I not verified the foundation before building the upper floors, I would have written a false story about a player who had not declined at all.

This is why I always replace vague phrases with concrete figures. Not "declining form". But "minutes played down 41% year on year". The first phrase is a judgement. The second is a verifiable event.

In esports I see the same disease. A player is judged "washed up" after one tournament. But if I look at minutes played, games started, the patch they had to play, and their role in the composition, I may see an entirely different picture. Some players look like they are declining when in fact they are playing a role constrained by the META and by their team's strategy.

Let me state plainly something that took me years to learn: most debates about "who is the best" in esports are not debates about truth. They are debates about which metric to choose. And the winner of that debate is usually whoever chose the metric that favours the conclusion they wanted.

Three questions I always ask before writing

First, what is the sample size. Below a threshold I set for myself, I do not conclude.

Second, what is the version. If the data comes from several versions, I must separate them or state the limitation.

Third, what am I measuring. A figure always measures something specific; it does not measure talent, courage, or mentality. If I want to talk about those, I must find a corresponding behavioural metric, never reach for concepts that cannot be measured.

This is the discipline I learned from my first xG model. It is not glamorous. It does not generate compelling headlines. But it is the only thing that keeps me from writing what is false.

I do not write about football. I write about the light that data illuminates. And when I moved into esports, that light still fell on the same place: the moments when the intuition of the crowd says one thing and the number says another.

What comes in the next cycle

I believe the next cycle of esports analysis will revolve around a question few people currently ask: how do we measure the value of a decision that failed.

In football I have xG. A shot that does not go in still carries informational value. Esports does not yet have an equivalent metric for tactical decisions undone by randomness. A failed incursion may be a mistake, or it may be a correct decision ruined by an individual misplay.

When someone builds that metric, the entire way esports is read will change. And I suspect it will come from a small group, not from a major publisher, because publishers have an interest in keeping metrics simple to serve the mass audience.

Until then, I keep my method unchanged: ask the number first, ask the version first, ask the sample size first. And I keep one principle I set on Russia Night: if I cannot verify the foundation, I do not build the upper floors.