When Data Is Not Enough: The Verification Discipline of an Esports Analyst

2:47 a.m. in Shanghai. On the third screen, a spreadsheet is still empty. No team names, no metrics, no patch number, no timestamps. In the top-left corner there is only one label: "esports". Every other cell is blank.

Ten years ago, that scene would have kept me awake. Today it does not. I close the spreadsheet, reopen the raw source, and search for the answer to a single question: is the fault in the writer, in the data extraction system, or in the source article itself? In the esports analytics trade, an empty sheet is not a failure. It is a result — and that result must be reported honestly, never padded with guesswork.

That is the most expensive lesson I have learned from five years standing between two esports industries: Germany, where I was born, and China, where I work. Both share the same disease. When there is no data, people tend to invent data. And in esports, where everything moves three times faster than in football, that disease spreads faster than any patch.

Data does not lie, but it learns to hide the thing that matters most.

A Young Industry with an Old Dataset

Esports entered the open-data era almost two decades later than football. By the time European football data providers were covering national leagues in the early 2000s, large-scale esports tournaments only began publishing event-level data systematically in the mid-2010s. League of Legends has Oracle's Elixir for minute-by-minute timeline data; Counter-Strike has HLTV with round-by-round statistics; Dota 2 has OpenDota and Stratz opening APIs to the community. Valve publishes The International match data, Riot Games opened Data Dragon to third-party developers.

But that opening carried a trap. Data grew, standardisation did not keep pace. Each game title defines a different metric. "First-contact fight participation" in League of Legends cannot be compared directly with "opening kill rate" in Counter-Strike. Even within the same title, providers compute differently. One counts kills after a tower has fallen, the other counts from the moment a fight breaks out. That small discrepancy, multiplied across thousands of matches, produces two datasets that cannot be placed side by side.

I began my career as a competitor and tournament organiser, then moved into esports media. That phase taught me something no classroom teaches: esports does not lack data, but it lacks trustworthy data. And the gap between those two is where an analyst's career is shaped.

In football, I once spent an entire World Cup manually recording possession share, passes into the final third, and touches inside the box for every match. I found that possession — the metric most celebrated by the media — often concealed the truth rather than revealing it. Since then I have never used a raw metric as my main argument. That principle, carried into esports, becomes many times more important.

The Architecture of a Data Pipeline

A complete esports data pipeline has four layers. The extraction layer pulls raw data from publisher APIs or community tracking platforms. The cleaning layer removes duplicates, standardises team names, and handles cancelled or rescheduled matches. The verification layer cross-checks at least two independent sources before numbers enter the model. The modelling layer is where analysis actually happens.

The most common mistake among newcomers is jumping straight from extraction to modelling. They take a statistics table, run a few calculations, and publish a conclusion. In most cases what they publish is a conclusion that is mathematically correct but athletically meaningless, because the verification layer was skipped.

I once witnessed a textbook case. A match dataset was fed into analysis. Every important field was empty: no patch name, no team, no player, no specific event. There was only a single topic label. Technically, the pipeline completed. Analytically, it never began.

The correct handling in that situation is not to guess which game, which team, which patch. The correct handling is to state clearly: insufficient information to assess. That is an answer, and it is the most professional answer available at that moment.

Variance is not the enemy — it is the mirror that reflects the arrogance of prediction.

The Discipline of an Empty Answer

In data analysis there is an underrated rule: the null-value handling rule. When input is missing, the analyst must state "insufficient information" rather than filling the gap with speculative content. It sounds simple, but it is extremely hard to enforce, because market pressure always pushes for content, for conclusions, for predictions.

In esports, that pressure is even greater. A season lasts a few months, but the calendar is packed almost year-round. The community demands content every day. When there are no major matches, people still want analysis. That is when conclusions born from nothing become most dangerous.

I have a professional habit: whenever I encounter a claim made without sources, I log it, tag it, and wait to see whether it survives the following weeks. Across many seasons, the survival rate of unsourced claims past one month is so low that I no longer use them as reference points.

What is worth noting is that those claims are usually very appealing. They are written fluently, with emotion, with narrative. They lack exactly one thing: verifiability.

Event-Level Data and the Death of Flashy Metrics

In League of Legends, KDA was once the most quoted metric. But KDA says nothing about where a player applies map pressure, when he leaves lane, or how he traded a death for a major objective. A player with a 4.0 KDA on a control-oriented roster may have lower tactical value than a 2.5 KDA player on an aggressive roster.

Event-level data is what answers real questions. It records every kill with timestamp, location, and gold difference at that moment. It lets the analyst reconstruct causal chains, rather than only looking at the final score.

In Counter-Strike, the analogous measure is rounds won off the opening kill. A team can win 16-13 while winning only four of ten rounds that began with a direct opening fight — meaning they won through round control and economic pressure, not through duel skill. Recognising that difference is the foundation for judging whether a win is repeatable.

In Dota 2, early-game data — gold difference at minute 10, rune control, lane win rate — shows which team truly understands how to build an advantage. At The International 2026, OG became the first team to win the event twice in a row, after their 2026 title. The media called it a miracle of creativity. But when I reviewed the early-game data, they led opponents in gold difference at minute 10 across most of their playoff matches. Their creativity did not happen in a vacuum — it was built on a very consistent numerical advantage.

Turning PPDA into an Esports Measure

In football, PPDA — passes allowed per defensive action — is the most reliable metric for pressing intensity. During the pandemic, when global football halted, I taught myself to code and built a database of thousands of matches across top European leagues and multiple World Cups. I combined PPDA with the location of the first ball contest to create a composite defensive-compression index.

The backtest result stayed with me. Leicester City won the English league in the season the media called an emotional miracle. But when ranked by the compression index, they sat in the leading group. The emotional story is not wrong, but it obscured a real tactical structure.

I carried that principle into esports. Here I built a similar index: how often a team actively initiates the first contact per unit of map resource. That index measures tactical intent, not just outcome. A team that presses constantly and still loses is fundamentally different from a team that defends passively and loses.

Measuring intent matters more than measuring outcome, because outcomes are far more exposed to variance. A lucky fight can decide an entire game. But intent repeated across dozens of games reveals a team's true nature.

One season is a statistical sample. A decade is evidence.

Backtesting: A Single Goal Does Not Lie

Every conclusion of mine passes a backtest against historical data before publication. The rule is simple: if a hypothesis does not hold on older data, it is not allowed to appear in a forecast about the future.

But backtesting is not only for confirmation. It is for destruction. I have repeatedly seen a theoretically sound hypothesis collapse the moment it touched real data. For example, the hypothesis that teams winning more fights must win matches. Run on a sufficiently large dataset, that correlation revealed something surprising: in the early tournament stage it holds; in the knockout stage it weakens markedly, because by then every team has adjusted and fight quality becomes more even.

That is something correlation never tells you unless you split by stage. The raw correlation between two variables can conceal an important truth: that the relationship changes over time. In an esports season, stages carry completely different psychological pressure, and any metric not split by stage risks misleading.

Fans remember the goal; I remember the probability before the goal happened.

Two Industries, Two Clocks

When I left Germany for China, what surprised me was not the language difference but the difference in behavioural practice data. The European teams I had followed typically structured practice time around clear rest cycles. The Chinese teams I later worked with had higher raw practice hours, but their time allocation for VOD review differed.

This is an observation from behavioural data, not a cultural judgement. Both approaches have strengths. What interested me was that when comparing the ratio of practice time to review time, the two industries showed very different numbers. And in some cases, teams with a low review ratio performed better in the early season, then were caught once other teams hit their stride.

I do not use that observation to say which system is better. I use it to remind that any comparison between two training systems must rest on actual behavioural data, not on feeling. My dual background has value only when it illuminates a measurable difference. Otherwise it is just a personal story.

Esports is not slower than football — it simply runs on a different clock.

The Counter-Intuitive Angle: More Data, More Hidden Truths

A common industry belief is that more data leads to more accurate conclusions. I think that belief holds only within a very narrow limit.

As the number of metrics grows, so does the number of ways to select one that supports a pre-existing argument. An analyst without discipline can cherry-pick metrics to confirm what he wants to say, rather than letting data lead. In esports, where thousands of metrics are available, the capacity for self-deception is higher than in any other sport.

The empty spreadsheet I mentioned at the start is the inverse example of this problem. When data is empty, no one can cherry-pick to prove anything. Honesty is forced to appear. That is why I am always wary of analyses that are too complete, too fluent, too confident. An analysis with no hesitation is usually one that skipped the verification layer.

A correct prediction does not prove a method is right. I once published a model that correctly predicted the champion of a major tournament, so accurately that the model was widely shared. But the same model also predicted another strong team would reach the final, and that team was eliminated in the first knockout round on penalties. If I looked only at the correct result, I would delude myself that the model was more reliable than it actually was.

The lesson is not that the model is weak. The lesson is that every model carries variance, and the analyst's duty is to speak about that variance before speaking about the result. What cannot be measured — psychological pressure before a decisive match, the effect of online criticism, the feeling of a young player at his first international event — lies outside every model. That is why each of my analyses must close with its own warning section.

What Data Cannot See

There is a limit I accepted long ago. Data can tell you what happened and how often. Data cannot tell you what will happen in a moment that has never occurred.

In esports, that is especially true. A patch can invalidate every model built on the previous version. A player transfer can change a roster's chemistry in ways no metric captures. A new coach can bring a philosophy that has never existed in historical data.

When someone tells me a team will certainly win, I know exactly where to look for the highest risk: in the very team considered invincible. Not because they are weak. Because absolute belief in them causes the market to misprice the probability of unfavourable scenarios. In esports, where a patch can land mid-tournament, that probability is far from small.

I call that state by a phrase I use often in my personal notes: "cannot lose". It is the state of a team the public believes cannot be beaten. That state has never existed in data. But it always exists in collective memory.

What Comes Next

Every new season opens a chance to re-test what we believe is true. Patches will change, rosters will shuffle, metrics will be redefined. The only thing that should not change is the discipline of verification.

I will still cross-check two sources before asserting anything. I will still run a backtest before publishing a model. And I will still write a variance warning at the end of every piece, even when it makes the article look less confident.

The empty spreadsheet at 2:47 a.m. will appear many more times. Each time, the question is not how to fill it, but how to be honest that it is empty. In an industry running on speed, that slow honesty is the only asset no one can copy.

The next cycle will begin with a new patch, a new roster, and a new spreadsheet. The only question left: will we use it to understand the match, or to confirm what we already believed?

During the pandemic, I built an empire from numbers no one was watching. It still stands.