Chess and the Limits of Unsourced Data

In September 2026, in Saint Louis, a chess game ended so fast that spectators had not even opened the live board. I was sitting in the commentary booth with three screens lit at once: the live rating list, the game database, and a handwritten notebook file that had followed me for more than twenty years. While the hall argued about a cheating allegation, I did exactly one thing: I checked whether the rating figure the anchor had just read out actually existed on the official list. It did not.

Within eighteen hours, that error passed through four television stations. None of the four meant to lie. They simply repeated a number someone had read before them.

Everything on a chessboard is data waiting to be read — if you are willing to sit down. The trouble is that very few people are willing to sit down, and even fewer are willing to ask where the number came from.

Context: the most data-dense sport

Chess is the most data-dense discipline I have ever covered. A football match gives you twenty-two players and ninety blurred minutes that must be reconstructed through a model. A chess game gives you every move recorded exactly, together with thinking time, engine evaluation, and the history of thousands of comparable games already sitting in a database. In theory, this is a paradise for the analyst.

In practice it is the opposite. Precisely because there is so much data, people forget that data must have provenance. An Elo figure has value only when it comes from the official rating list of the international chess federation. A live rating is trustworthy only when it comes from a real-time tracking platform with a time stamp. A move-quality index means something only when the tool that computed it can be named. Remove those three lines and what remains is a feeling dressed up in numbers.

In 2026, when the entire over-the-board calendar froze, I spent six months digitising handwritten notebooks covering 2026 to 2026. Two thousand four hundred games, one row each. Once the data was aligned, a pattern surfaced: young players produced their best results not in games where they held the white pieces, but in games where they were forced to defend during the first twenty moves. The cause was not talent. It was that they had prepared a specific defensive system and repeated it until it became reflex.

I published that conclusion. I kept the method. It is a bad habit I know well, and I will return to it later.

Analysis: four decisive indices

These four are what I rely on most when reading an elite game.

First, average loss per move measured in hundredths of a pawn — the measure of move quality, lower being better. A top player usually keeps this below twenty units in a classical game. What matters is that it spikes not in the opening but around moves thirty to forty, when the clock starts running low.

Second, the match rate with the engine's choice — the percentage of moves identical to the one the strongest analysis tool selected. This index is easily misread. A young player may hit eighty per cent while a former world champion hits only sixty, and that does not mean the young player is stronger. It means the young player picks lines the engine finds familiar, while the former champion picks lines the engine rates lower but which are harder for a human opponent to handle.

Third, performance rating in a tournament — the rating level corresponding to actual results rather than reputation. It is the only index that lets you compare two players at two different moments without being distorted by the overall strength of the field.

Fourth, the draw rate — a gauge of how watchable an event is. A tournament with a seventy per cent draw rate is usually dismissed as dull. But a draw rate means something only alongside the average number of moves per game. A draw after twenty moves and a draw after seventy moves are different products, even though the statistics table records them identically.

Two factors are left out of almost every report: opening preparation and time distribution. Opening preparation is not measured by how many moves a player has memorised, but by how many alternative lines have been prepared for each main variation. A player with three options for the same line loses less time on move fifteen, and the time saved on move fifteen gets spent on move thirty-five. The distribution of thinking time across phases is the least published index and the one that says most about the final result.

As for formats, chess splits into four tiers: classical, rapid, blitz and bullet. Error grows exponentially as time shrinks. A player who keeps average loss below fifteen units in classical can balloon past forty units in blitz. This means any conclusion drawn from a blitz game cannot be transferred directly to a classical game. People transfer it directly anyway, every week.

Finally, the gap between over-the-board and online results. Technically these are not the same environment. A player can win online titles three years running and never reach the quarter-finals of an over-the-board event. The cause is not cheating; it is the interaction of tournament-hall pressure, travel, and the fact that you are not allowed to leave your chair.

The contrarian angle: silence is not cleanliness

This is the point I most want to make clear.

When an analysis system fails to extract data — because the source is blocked, the file is empty, the connection dropped — it usually returns an empty result. And an empty result, after enough layers of processing, becomes a report reading "no issues detected". No red flag is raised, because there is nothing to raise a red flag about. This is the most dangerous class of failure in any analytical pipeline. Chess is the sport most vulnerable to it, because every number here is checkable, so readers assume the writer checked.

Missing data does not produce neutral analysis; it produces a gap presented as a conclusion. In a discipline where ratings, head-to-head records and prize funds are all searchable, that gap will be filled with guesswork, and guesswork in chess gets caught as precisely as a wrong move.

The same thing has happened on a larger scale. An accusation that was never proven before any competent authority was still enough to reshape how the public saw a player for years. What was damaged was not a rating figure but the entire frame of reference for reading rating figures.

I have been guilty of the opposite error myself. For years I published conclusions while withholding my method. I said I did not want others to depend on me, but the real reason was simpler: keeping the formula kept me always right in the reader's eyes. That is not how a data person should operate. A conclusion without a method is a conclusion that cannot be challenged, and anything that cannot be challenged cannot be corrected.

Takeaway

I no longer believe in miraculous comebacks on the board; I believe only in the conversion rate of an advantage. A player who holds a small edge and converts seven of ten games will win the title. A player with a prettier edge who converts only three of ten will finish second, and will be praised far more.

The work ahead is not adding more data. The work is labelling the provenance of every number and keeping the status "unverified" exactly as it is, instead of letting it drift into "no problem". If you read a piece of chess analysis today and find no source name, no publication date, no definition of the index, you are reading a feeling, not an analysis. The only question left is simple: will you ask for the source from the first number, or from the eighteenth?