The Empty Board in Singapore: Gukesh, Ding Liren and the Data Trap of Chess

Singapore, the evening of 12 December 2026. Game fourteen of the World Chess Championship. Ding Liren, the defending champion, held the white pieces in an endgame that every analytical engine returned to the same verdict: a draw. Around move fifty-five, he played a move that sent the evaluation bar straight out of the balance zone. Gukesh Dommaraju, eighteen years old, playing black, did not need long to convert that error into a win. The final score was 7.5-6.5. The world title changed hands in a single move that the player himself knew was wrong.

I watched that game from Chengdu, the live feed on my left screen, the engine evaluation bar on my right. What made me stop was not the move itself. What made me stop was the speed at which the chess content industry rushed to explain it. Within six hours of the game, hundreds of analyses appeared. Most of them were about psychology. Ding was tired. Ding lost focus. Ding had no motivation left. Every one of those conclusions could be right. And not one of them had a single line of data behind it.

Three years earlier, in April 2026 in Astana, it was Ding Liren himself who beat Ian Nepomniachtchi in the rapid tiebreak, 2.5-1.5, after fourteen classical games ended 7-7. At that time, nobody said Ding lacked nerve. At that time, people called Ding the embodiment of endurance. The same man, the same skill set, two opposite conclusions. The data on the board barely changed. The story around it changed completely.

The gap between what the board actually says and what the chess content industry retells: that is the subject of this article. To measure that gap, I have to rebuild the data map of the chess world over the past three years, and I have to be explicit about which parts are verifiable and which are carefully labelled inference.

A Title Detached From the Number One Spot

Professional chess today runs on two separate axes. The first axis is the title. The second axis is the rating.

Magnus Carlsen has held the number one position on the FIDE rating list almost continuously since July 2026. In July 2026, he announced he would not defend the world title in Astana 2026, citing a lack of motivation. Since then, the holder of the world championship and the holder of the highest rating have been two different people. Ding Liren, then Gukesh Dommaraju, hold the title; Carlsen holds the rating. This is the most important structural feature of the chess world today, and it is the feature most analyses ignore when they call either man the number one.

The Elo system was developed by Arpad Elo and adopted by FIDE in 2026. An Elo rating does not measure absolute playing strength. It measures the probability of winning against an opponent with a defined rating. The standard error of an Elo rating after a short tournament is very large, typically in the range of plus or minus a few dozen points. In other words, a twenty-point Elo gap between two elite players sits inside the noise band. Any commentary that asserts A is stronger than B on a gap of under thirty Elo points is selling readers a conclusion the data cannot support.

One more thing about ratings. Elo is not a measurement; it is an iteratively updated estimate. Every game adjusts the rating toward the actual result, at a rate that depends on the K-factor. FIDE uses different K-factors for different age groups and rating bands. For elite players, the K-factor is typically low, around 10. That means a fourteen-game match only moves an elite player's rating by a few dozen points, even when results diverge completely from expectation. The system is designed for stability, and that stability means the rating responds slowly to real changes in form.

FIDE was founded on 20 July 2026 in Paris. In 2026 the organisation turned one hundred, and the world championship match in Singapore sat inside a series of centenary events. The match ran from 25 November to 12 December 2026, comprising fourteen classical games. The fourteen-game format was fixed from the Carlsen-Caruana match of 2026, after a chaotic period in earlier cycles. Before 2026, the world championship had been split between two organisations for more than a decade. The unification of the title in 2026 is the precondition that makes all modern statistics comparable with one another.

The final point of the format needs stating clearly: the tiebreak. When fourteen classical games end level, the title is decided by rapid games, then blitz if needed. This is a rule with major data consequences. It means a match is defined by two different sets of metrics, with two different error distributions, and the final result may belong to the second set. Anyone analysing a world championship match using only classical data is ignoring the decisive part of the format.

The road to Singapore 2026 ran through the Candidates Tournament in Toronto in April 2026. Eight players, double round-robin, fourteen rounds. Gukesh won with 9/14. This is the tournament where data has the highest value in the entire cycle, because it has a large enough sample, evenly matched opponents, and pressure distributed across fourteen rounds. Every serious prediction model should start here, not from open tournaments with uneven opposition quality.

On the other side of the map, 2026 brought a data-governance dispute. The Freestyle Chess Grand Slam Tour, initiated by Jan Henric Buettner with Carlsen's participation, organises Fischer Random events, also known as Chess960. In early 2026, FIDE and the Freestyle organisers publicly disagreed over the use of the phrase world championship for the tour final. FIDE argued that only it has the right to award world championship titles in chess. Freestyle argued that Chess960 is a different discipline and is not subject to that authority.

The dispute sounds administrative. It is fundamentally a data problem. When two organisations award a title with the same name, every statistic behind it becomes non-comparable. An analyst tracking a player's form must separate classical, rapid, blitz, online and Chess960 data into distinct sets. Mixing them produces a beautiful and meaningless index.

What the Data Actually Shows

Back to Singapore. I reconstructed all fourteen games and cross-checked them against both players' rating lists before the match.

Before the match began, Gukesh had a classical Elo around 2780, Ding Liren around 2720. A gap of roughly sixty points. Under the Elo formula, sixty points corresponds to a win expectation of about 58 to 59 percent in a head-to-head pairing. In a fourteen-game format that includes rapid tiebreaks, that expectation is not enough to state anything with certainty. This is the point commercial prediction models tend to hide: they sell clients a probability, while the real uncertainty of a fourteen-game match is far larger than the number they publish.

The more interesting thing lies in the structure of the games.

In the first seven games in Singapore, Ding Liren held the white pieces four times. His rate of choosing safe opening plans was unusually high compared with his own data from Astana 2026. He prioritised stable, low-branching systems that minimised the calculation load required in the middlegame. This is the signature of a player managing energy, not a player attacking.

In the last seven games, with the score still level, that strategy stopped working. Gukesh refused draws. He repeatedly pushed games into long endgames, where errors accumulate and where age and stamina become real variables. Gukesh is eighteen. Ding Liren is fourteen years older. In a classical game lasting six hours, each hour of physical disparity is multiplied by the number of moves that must be calculated.

Game fourteen was the logical output of that strategy. In the endgame, the position was balanced. The engine evaluation hovered near zero. Pressure accumulated across the previous thirteen games turned a simple technical decision into an impossible one. Ding erred. Gukesh won.

If you have read this far and find everything fits the Ding was tired story, read it again. What I described is not a psychological conclusion. It is a structural observation: the fourteen-game classical format favours younger players and penalises older ones, independent of form. This can be measured by tracking average game length, average move count, and average thinking time in the second half of a tournament.

In the data I recorded across the games, average game length in the second half was markedly higher than in the first half. The number of games exceeding forty moves rose. Thinking time for both players after move thirty rose. Those are countable signals, and they all point the same way.

Now to the part where the data is insufficient for conclusions.

A popular metric in the analytical community is ACPL, average centipawn loss per move. It measures the average deviation of actual moves from the best move the engine suggests. Lower ACPL is better. A single large error in one game can push a whole tournament's average ACPL up disproportionately.

This is where many chess analyses fool themselves. They take a player's average ACPL over a tournament, compare it with another player's, and conclude that one is more accurate. The problem lies in the distribution. ACPL is not normally distributed. It has a long right tail driven by individual blunders. Comparing two means from two heavily skewed distributions, on a sample of just fourteen games, yields no statistically meaningful conclusion.

The same applies to engine match rate, the percentage of a player's moves matching the engine's first choice. This metric depends heavily on game phase. In the opening, elite players all have very high match rates, because they have memorised known variations. In the endgame, the rate falls for everyone. Comparing two players on this metric without splitting by game phase compares two different things.

One more metric to mention: tournament performance rating. It is calculated by averaging opponents' ratings and adding the score overperformance. With fourteen games and eight to fourteen different opponents, the error of this metric is large enough that two players twenty points apart in performance still fall inside the same confidence interval.

There is a historical reason these metrics became popular late. In 2026, IBM's Deep Blue beat Garry Kasparov in an official match. In 2026, DeepMind's AlphaZero learned chess from scratch by self-play and defeated the strongest engines of the time. Since then, engines have become the standard yardstick for every move. But a standard yardstick for a move is not a standard yardstick for the quality of a game, let alone the quality of a player over years. Evaluating moves with an engine is a technical measurement. Inferring a conclusion about a human being from it is an unlicensed logical leap.

The observations above are not meant to deny the metrics. They are meant to set their proper limits. A responsible analyst states the sample size, the error margin, and which conclusions would be withdrawn if the sample were larger.

Empty Data and the Art of Storytelling

Now I want to discuss what I consider the most serious problem in chess analysis, and the one I encounter daily in my work.

Most chess content online is written in a state of empty input data. No ACPL table. No engine match rate. No phase-by-phase analysis. There is only a final result and a few clipped highlights. From that material, the writer constructs a complete story, with a protagonist, a turning point, and a lesson.

This mechanism is identical to a data pipeline fault I once encountered in my own systems. An extraction process returned a valid but empty structure. No field was broken. The format was correct. There was simply no content. If the system has no mechanism to halt on empty data, it passes that empty structure downstream, and the next stage fills it with whatever sounds plausible.

In chess, that next stage is human. When there is no data, the writer fills the gap with psychology. Ding was tired. Gukesh is young. Carlsen is bored. Nepomniachtchi lacks nerve. Firouzja is inconsistent. Every label sounds plausible, is unverifiable, and has a short shelf life.

I call this analysis from empty data. Its identifying features are clear. It never cites the player's own numbers. It never states the sample size. It never leaves a conclusion that can be refuted. Every assertion is packaged as a remark about human nature, and human nature has no error threshold.

The consequence does not stop at inaccurate content. It creates a false database in the reader's head. Three years later, when a decent analyst writes that Ding Liren's Elo rating declined for specific technical reasons, readers file it in the same drawer as Ding was tired. Two fundamentally different claims are blended into one. By then nobody can tell which had evidence.

For someone who works with data, this is the most dangerous form of contamination. It does not produce wrong data. It produces a gap filled with something that looks like data.

There is a simple test to detect this kind of content. Read an analysis and ask yourself: if the result were reversed, how much of the article would change? A data-driven analysis would change very little, because the structure of the games does not depend on who won. An analysis built on empty data would reverse completely, because its entire argument is constructed backwards from the result.

That is the test I apply to myself every time I write.

What Is Actually Worth Tracking Next Cycle

From here, some judgements, each with a condition that would invalidate it.

The dispute between FIDE and the Freestyle Chess Grand Slam Tour will determine the data structure of professional chess over the coming year. If the two sides reach an agreement on the naming of titles, statistical tables remain comparable. If not, each side will build its own database, and cross-system comparison becomes manual work for every analyst. The invalidation condition for this judgement is a joint statement between FIDE and the Freestyle organisers delineating the scope of titles before the tour ends.

The young generation is shifting the rating frontier. Arjun Erigaisi has crossed 2800. Nodirbek Abdusattorov and Rameshbabu Praggnanandhaa are closing on the top group. As the number of players above 2750 grows, the informational value of any given rating level falls. A tournament with eight players all above 2750 can no longer distinguish them by rating. At that point, people will have to go back to reading the games themselves. The invalidation condition is the emergence of a clearly dominant player with a rating gap of more than fifty points over the chasing group.

Advanced metrics will have to be normalised by format. When classical, rapid, blitz and Chess960 coexist in one ecosystem, shared metrics lose value. The analyst who builds a format-segmented classification system first holds the advantage. The invalidation condition is FIDE and the major platforms agreeing on a common classification standard.

For Gukesh Dommaraju, the next challenge lies in defending the title in a cycle where opponents already have enough data on him. A young champion wins the first time by exploiting an opponent's weakness. The second time, he must win on his own strength, with every one of his games analysed move by move. The invalidation condition is whether he maintains a high win rate against the top group in round-robin events, where opponents cannot be selected.

Ending

Data never lies, but it enjoys testing our patience. Game fourteen in Singapore gives us one wrong move and fourteen games preceding it. Anyone who reads only the wrong move will write about psychology. Anyone who reads all fourteen games will write about the format, about age, about tournament structure. Two articles about the same event, one refutable, one not.

I bet on numbers before the world knows how to read them. And when the data is still empty, I choose silence over inventing a story. On a board with no moves, data is the only audience left, and if it is absent, it is better not to perform.