Chess and the Data Problem: When an Empty Analysis Table Still Produces a Legend

At round nine of an open tournament in Chiang Mai, a twenty-one-year-old player sat motionless in front of the board for eleven minutes. He did not touch a piece. He did not write on his scoresheet. The hall was quiet enough to hear the digital clock tick each second. On the organisers' screen, the evaluation slid from +0.4 to -1.8 and then crawled back to +0.2. Nobody in the hall understood what had just happened, including the man sitting opposite him.

Those eleven minutes appear in no official scoresheet. A scoresheet records moves. It does not record silence. That is the foundational problem of modern chess analysis: we are building enormous data structures on a foundation that only preserves the visible tip.

Chess has lived through a measurement revolution over two decades. From a sport of paper scoresheets and notation, it became a discipline with near-perfect digital records. Every game at every major event is logged, cross-referenced against opening databases, and graded by engines.

Three metrics dominate every report today: Elo, the coefficient of relative strength; live rating, updated in real time from results at an ongoing event; and performance rating, the rating level corresponding to a player's results at a specific tournament. Deeper down sit ACPL, average centipawn loss per move, and engine match rate, the share of a player's moves matching the engine's first choice.

This is a toolkit more powerful than anything the game has had before. It is also a toolkit more easily abused than anything the game has had before.

As a chess commentator for VTC between 2026 and 2026, I once believed the quality of a game could be reduced to a single number. Three years behind the microphone at the classic finals of that era taught me otherwise. Only when I left the commentary booth for a coaching staff did I understand the mechanics of the distortion.

ACPL is an average. It sums every centipawn loss in a game and divides by the number of moves. A player who makes one fatal error at move forty and plays perfectly for the other thirty-nine moves will post a cleaner ACPL than a player who fights through sixty moves of a complex position without a single serious mistake.

The metric rewards simplicity and punishes complexity. It cannot distinguish an error in a balanced position from an error in a lost one. It does not know that in some positions the engine's best move is precisely the move that forces the opponent to think longest.

Engine match rate is even more counter-intuitive. A creative player who finds a plan the engine never considered across the first thirty moves is penalised by the metric. A player reciting memorised theory is rewarded. The instrument we use to measure originality is quietly subsidising imitation.

Back to those eleven minutes in Chiang Mai. What mattered was not the final move but the silence before it. Across those eleven minutes the young player's clock drained from twenty-three minutes to twelve. He was not calculating a forced line. He was rebuilding his spatial map after his opponent had just altered the pawn structure.

Fourteen seconds — enough to redraw the opponent's entire defensive map. In chess the equivalent window is rarely fourteen seconds; it is usually three to fifteen minutes. That is the interval between an opponent's unexpected move and the moment you must accept that the old position is dead.

In the 120-page report I completed in 2026, after 250 days spent coding 380 games, I found what the season never records: repetition. The players who lost most were not the weakest calculators. They were the ones who took longest to recover after a position changed, and who routinely made their final decision with under five minutes on the clock.

I called it the structure-repetition ratio: the number of times a player must rebuild a piece configuration within a single game, divided by total moves. Among players with a high ratio, the probability of a serious late-game error was markedly higher than among the rest, regardless of rating gap.

Not the move, but the gap before the move appears. The spatial map never lies — it only exposes what we want to believe. And what we most want to believe, in every analysis room, is that chess quality can be reduced to a tidy series of numbers.

Live rating is one of the most interesting media inventions the chess world has produced. It turns a game between two unheralded players at an open into a minute-by-minute event, because the number updates continuously on online platforms.

But live rating creates an effect I have not seen in any other sport: round-number thresholds become a tactical variable. As a player approaches 2700, 2750 or 2800, the structure of the events they enter changes. They avoid tournaments with dense strong fields. They favour opens with weaker opposition. They withdraw from final rounds when the risk of losing points outweighs the gain.

This is mathematically rational behaviour. It is athletically absurd.

The World Championship qualification structure compounds the problem. Places at the Candidates Tournament — the event that decides the title challenger — are allocated through several routes: the World Cup, the Grand Swiss, the FIDE Circuit, and one place reserved for the highest-rated player not otherwise qualified. Each route carries a different competitive density, and each route generates different incentives.

A player entering small opens to farm rating points is playing an entirely different optimisation game from a player who throws himself into the World Cup, where the knockout format means any game can be the last. Both are chasing the same place. Only one of them is playing chess in the purely competitive sense.

The largest structural difference between chess and most team sports sits here. In football, the world champion and the top-ranked team can differ, but both are determined inside one continuous competitive system. In chess, the two positions are determined by two entirely separate mechanisms.

The world number one by rating is determined by accumulated results across hundreds of games. The world champion is determined by a match lasting several weeks, once every two years. For nearly two decades the two coincided, and the media grew used to treating them as one.

Since Magnus Carlsen declined to defend the title, the separation has become permanent. Carlsen remains world number one by rating, while the crown passed to Ding Liren and then to Gukesh Dommaraju — who took the title in December 2026 in Singapore and became the youngest world champion in history at eighteen.

As information, this is a large gap. Media need a linear story, and chess is supplying two parallel tracks. The result is that reports are forced to pick one, usually calling Carlsen the uncrowned champion or Gukesh the champion who is not the strongest player. Both labels are accurate as data and wrong in substance.

The Elo system and the championship format measure two different capacities. Elo measures the ability to sustain performance across a large sample of opponents. A title match measures the ability to prepare for one specific opponent in a limited window. A player can excel at the first and hold no advantage at the second. Many sports accept this as ordinary. Chess calls it a paradox.

The boundary between board and screen is where chess data becomes most fragile. Online and over-the-board results cannot be directly converted, because the conditions differ on fundamentals: control of the device, control of the physical space, and the ability to observe the opponent.

During the pandemic, when the entire competitive system moved online, this became central. Major platforms published periodic fair play reports, and those reports showed a violation rate online far above the over-the-board rate. That is unsurprising. What is surprising is how the community responded to the figures.

A high violation rate does not only mean more people cheat. It also means the detection system is more sensitive, and a more sensitive detection system always carries a false-positive risk. In online chess, a false positive is not merely a statistic. It is a locked account, a damaged reputation, and in many cases a career ended without a proportionate appeal mechanism.

FIDE was founded on 20 July 2026 in Paris and now governs close to two hundred national member federations. Its governance apparatus must simultaneously run two nearly incompatible systems: a slow competitive system with on-site arbiters and paper scoresheets, and a globally distributed online system where games take place on personal computers at thousands of locations.

No anti-cheating system achieves perfect accuracy under those conditions. The problem is not technical capability. The problem is that the chess community has implicitly defined the absence of evidence as evidence of innocence, and has implicitly defined the presence of an anomalous pattern as evidence of guilt. Both assumptions are wrong.

This is where the story returns to the analysis room, and to a lesson I learned the hard way.

In 2026, with football and chess both frozen by the pandemic, I spent most of my time building an automated extraction pipeline for analytical reports. The pipeline worked exactly as designed. It returned a complete structure, correctly formatted, with every field: title, source, article type, domain label, core viewpoints, information points, entities involved, time sensitivity, source quality.

And on some occasions, that structure came back completely empty.

The tables were intact. The title field existed. But the field was blank. The information point list contained no entries. No entities were resolved. No timestamps were assessed. Looking at the output file, it resembled a valid analysis. Read closely, it contained no information at all.

That is the most dangerous failure mode in an entire data chain, and it is not a machine failure. It is a verification failure. A system that returns an explicit error gets fixed. A system that returns a file which looks valid but is hollow gets used, cited, and built upon.

It took me a while to recognise that my error was not in the extraction algorithm. It was that I had never set a minimum content threshold. I checked whether the output was correctly formatted. I never checked whether it contained at least one verifiable piece of information.

In chess analysis, we are committing exactly that error at a far larger scale.

A post-tournament report carries every heading: game list, results table, each player's ACPL, live rating charts, portrait photos, and a concluding paragraph. Every field is filled. But ask that report one question — what changed in this player's reading of positions compared with three months ago — and no field can answer.

That is a file with a complete structure and an empty interior. It looks like analysis. It is not analysis.

The chess industry transmits across three layers. Upstream sits youth development and the talent supply chain. Midstream sit events, players and competitive platforms. Downstream sit content, commerce and derivative markets.

When the downstream layer produces hollow but plausible-looking analysis, the damage does not stop at readers. It travels back upstream. Sponsors read those reports to value a tournament. Federations read them to allocate resources. Youth academies read them to shape training programmes.

An industry making decisions on data files that look full but hold nothing will optimise for the wrong thing. It will sponsor events with attractive broadcast metrics instead of events with high competitive quality. It will produce players who post low ACPL instead of players who read positions well.

In Southeast Asia, the gap shows up clearly in tournament structure. Vietnam's leading players such as Le Quang Liem and Nguyen Ngoc Truong Son have passed 2700 and 2600 respectively, yet the number of domestic events strong enough to sustain a live rating at that level is limited. As a result, most of their peak careers take place abroad, and the data on them is recorded mainly by systems outside the region.

On the other side, the Indian wave — Gukesh Dommaraju, Rameshbabu Praggnanandhaa, Arjun Erigaisi — shows the reverse. A youth development system built methodically over fifteen years produced a generation deep enough to generate its own international-standard events, and from there to generate its own data.

The difference between the two models is not talent. It is who owns the data on their own players.

Even with ownership of the data, the next question remains unanswered: which data is worth collecting. This is where I want to go against the industry consensus.

The prevailing belief is that more data is better, and that anything measurable should be measured. That belief has produced a generation of analysis that knows precisely the ACPL of every move but cannot describe why a player changed style after a heavy defeat.

Our blind spot is not a shortage of data. Our blind spot is confusing data with evidence. A measured value is not automatically a meaningful fact. A high engine match rate is not automatically proof of talent, nor automatically proof of cheating. It is an observation, and every observation needs a hypothesis before it becomes knowledge.

Chess today holds an unprecedented volume of data about itself, while sitting in a state of acute uncertainty about its most important questions: who the strongest player is, whether the online environment can be trusted, and which governance system can protect both the players and the integrity of results.

That uncertainty does not come from a lack of numbers. It comes from our failure to build the verification layer that sits between numbers and conclusions.

In the 120-page report of 2026, the most valuable section was not the tables. It was a short chapter listing what I could not conclude from the data. Every position is a hypothesis until a piece is touched. I wrote that chapter in three days, and it is the only part of the whole report I still reread.

So what happens next season.

When a new cheating case surfaces, and it will, the first question the community raises will be whether the player is guilty. That is the wrong question. The right question is whether we hold enough of a record to answer it in a verifiable way, or whether we are again reading a file with a perfect structure and an empty interior.

I have spent twenty-three years watching this industry, from the commentary booth at VTC to the analysis room in Chiang Mai. That experience taught me one thing, repeatedly: what determines the quality of a sport is not the volume of data it produces, but the honesty of the verification layer it builds on top of that volume.

The spatial map never lies — it only exposes what we want to believe. Coaching does not produce identical players; we produce different paths. And if next season begins again with an analysis table full of filled cells and empty meaning, the party responsible will not be the machine that produced it, but us — the people who read it and believed it.