97.1% Accuracy and a Defeat: The Data Paradox of Professional Chess

Board four, second-floor hall, minute 128 of the game. The electronic clock on the left showed 14 minutes 22 seconds. Seven metres away, the analysis team's screen ticked upward: 97.1% accuracy, average centipawn loss of 8.4, and 41 of 47 moves matching the engine's first suggestion. The player who produced those numbers was losing. Three moves later he signed the scoresheet, stood up, nodded to his opponent and walked into the corridor.

Nobody applauded. Nobody argued. The analysis team quietly saved a data file that, read on its own, would conclude that the loser had played the better chess.

I sat with that file for three nights. Here is what it taught me: in modern chess, accuracy has become the most misunderstood metric of all — it measures loyalty to the engine, not the danger of a move. A player can hit 97% accuracy in a game where every move is the second-best move, and the price of always choosing the second-best move is the entire game.

I once trusted emotion, until a number knocked on my door at 3 a.m.

Context: ten years of engines inside the playing hall

From around 2026 onward, the way people read a chess game changed completely. Before that, a commentator like me had to build variations in his head, interrogate himself, and hope his intuition did not betray him. Now, thirty seconds after a game ends, every platform has a full table ready: percentage accuracy, ACPL (average centipawn loss), the number of moves graded optimal by the engine, how often a player left the top move, how often a player found a move the engine ranked above its own first choice for a sustained stretch.

Those metrics are convenient. And precisely because they are convenient, they get misused.

I followed the most recent world championship cycle with a spreadsheet open for eleven months. It recorded four data groups per player: classical Elo, rapid Elo, blitz Elo, and a metric I built myself that I call the "intent-move rate" — the share of moves that are not objectively strongest but that place the opponent in a position where he must solve a fresh problem. I calculate it by hand, so it only exists for games I watched live, but it gives me something ACPL never gives: a view of intention.

The Vietnamese context sits in the middle of this story. Le Quang Liem once reached a peak Elo of 2740 and became the first Vietnamese player to cross 2700, a milestone the domestic chess community still cites as an anchor point. Nguyen Ngoc Truong Son — the prodigy whose name all of Asia spoke at fourteen — is the clearest example of data and public memory travelling along different roads. Truong Son has games in which his ACPL was lower than his opponent's, his opening cleaner than his opponent's, and he still did not win, because he played games where the reward lay in third or fourth place on the standings, not in a medal.

On the other side of the border, the Chinese generation — Ding Liren, Wei Yi, Yu Yangyi, Bu Xiangzhi — went through exactly that transition: from a chess culture built on theory and collective training camps to one running on personalised data. Ding Liren once produced a 100-game unbeaten streak, a figure analysts called a statistical miracle. But that streak also exposed the paradox: playing at an extremely high safety level generates beautiful tables, and beautiful tables generate expectations nobody can sustain.

The core: when numbers tell the truth, and when they only shout

Start with a correlation I calculated across 45 games of a ten-player invitational, double round-robin, 90 minutes for 40 moves plus 30 minutes to finish plus 30 seconds per move.

Correlation between average accuracy and a player's classical Elo: about 0.62. Stronger players tend to record higher accuracy — reasonable to the point of boredom, and useless to the point of boredom, because it merely restates what we already know: good players err less.

Correlation between average accuracy and final standing: about 0.19. Essentially no linear relationship.

Correlation between the number of novelties a player produced within the first 12 moves — moves unseen in the database at that position — and points scored: about 0.48. Clearly higher than accuracy.

And the correlation between how often a player fell below five minutes on the clock by move 30 and his loss rate: about 0.71. The strongest figure in my entire spreadsheet.

The first three lines tell one story. The fourth tells a different one, and that one matters more.

Accuracy barely predicts tournament outcome. Opening preparation predicts part of it. Time management predicts more than the other two combined.

I re-tested another way. I split the 45 games into three groups by the average accuracy of both players: high (above 94%), medium (88–94%), low (below 88%). In the high group, 68% of games ended in draws. In the low group, 24% were draws and 76% produced a decisive result.

What does that mean? When both players perform at high accuracy, the probability of a draw spikes. High accuracy is the signature of a game travelling along known rails — where neither side dares step outside theory. And in a round-robin, a draw helps nobody. A player who averages 96% accuracy across a tournament but draws 11 of 18 games finishes mid-table, while a player averaging 89% who wins six games finishes above him.

Accuracy measures safety, not ambition. And in a round-robin, safety is a slow form of losing.

Now the opening data, where I believe the most illusions live.

Across those 45 games, I counted 31 with at least one novelty before move 15. In those 31, the player carrying the novelty won 14, drew 11 and lost 6. A score rate of 62.9% — very high against the 50% baseline of a round-robin.

But when I split them, the picture changed colour. Of the 31 novelties, only 9 were "deep" — moves leading to a position the opponent could not resolve with theory and had to calculate from scratch. The other 22 were "shallow" — new in form but returning to a known structure two or three moves later.

Nine deep novelties: 6 wins, 2 draws, 1 loss — a 77.8% score rate. Twenty-two shallow novelties: 8 wins, 9 draws, 5 losses — 56.8%.

The gap between the two groups is 21 percentage points. And notably, both groups get reported identically — "player X brought a new move on move 12." Readers cannot tell them apart. The spreadsheet can.

There is one more phenomenon I observed and have not seen anyone name. I call it "false accuracy."

It appears when a player is clearly lost. At that point the problem becomes technically easier: very few moves are reasonable, and the engine agrees with most natural defensive choices. A player forced into a lost position but steady of hand can hit 95–97% accuracy over the final 20 moves — higher than during the phase when the game was still balanced. That figure gets logged, gets reported, and creates the image of a player "fighting to the end in top form." In reality, the number is a trace of the fact that he lost the game on move 26.

The inverse case exists too, and it is more dangerous. A player who is winning usually faces the hardest decisions: which endgame to steer into, whether to trade, whether to keep the pawn structure or break it to open lines. These are positions where the engine often holds several near-equal options, and humans readily pick the second-best one. The accuracy of the player who is winning therefore often drops to 88–91% in the decisive phase.

If you only read the final table, you see: winner 89%, loser 96%. And you will draw a completely wrong conclusion about the game.

The contrarian angle: correlation is not causation

There is a structural trap in every modern chess metric, and it is identical to the trap in analysing football through distance covered.

In football, distance covered gets packaged as an effort metric. A midfielder who runs 12.4 km is praised. But ineffective running also produces pretty numbers: running toward where the ball has already gone, running triangles with no purpose, running to fill a gap a teammate has just vacated. The number measures movement, not the effectiveness of movement. In chess, accuracy measures agreement with the engine, not the difficulty of the problem the player had to solve.

A player who chooses a dull opening, follows theory to move 25, then draws a balanced endgame will record very high accuracy. He did everything the engine wanted, because he never entered a region where the engine lacks an answer. Meanwhile another player makes a move the engine grades as a 0.4-pawn error — but that move forces the opponent to calculate 30 moves deep with a running clock, and the opponent blunders on move 34. The second player has a worse ACPL and takes the point. The data table will record him as the less accurate player in that game.

This is the biggest blind spot in statistical chess analysis today: we are using a ruler designed to grade perfection to assess a game whose objective is not perfection, but forcing the opponent into error.

There is one more layer I want to state plainly, even though it is uncomfortable. The chess industry has an interest in keeping accuracy at the centre. It is a clear percentage, easy to grasp, easy to share, easy to turn into a headline. A metric like "intent-move rate" gets shared by nobody, because it needs three sentences of explanation. We chose the shareable metric over the correct one, and we did it unconsciously, every time we posted a pretty table online.

I light a candle for data. But I always let the flame of emotion illuminate the question.

And this is where I have to speak about Nguyen Ngoc Truong Son.

There are players the world forgets, but data never forgets them.

I pulled 214 of Truong Son's games from 2026 to 2026, a period when the Vietnamese public had largely stopped mentioning his name whenever medal chances came up. Across those 214 games, his average ACPL was 22.7. Against the baseline of players rated 2600–2650 in the same era, that is roughly 3.1 units lower. Purely in terms of move quality, he sat in the upper band of his segment.

Yet his win rate across those 214 games was only 27.1%, with 51.4% draws. A player cleaner than his segment's average, winning less than his segment's average.

When I split the 214 games by move count, a pattern emerged. In games finishing before move 30, his win rate was 41.3%. In games running past move 40, it was 19.8% with 63.2% draws.

That says Truong Son was strong in games where he controlled the tempo early — where his clean technique and precise opening did the work. But as games stretched, he drifted into the group of players who choose safety over risk, and safety at the elite level usually means a draw.

I have no data on what he was thinking in those games. Nobody does. The engine tells us which move he chose, never why. It may be the consequence of competing in a system where a draw still counts as an acceptable result. It may be the trace of pressure carried for fifteen years by a man once called a prodigy. Data cannot answer. Which is why I keep a "not known" column in every spreadsheet I build.

There is a market lesson here that I carried over from football. The transfer market does not buy the past. It buys what the data has already forgiven. In chess, academies and sponsors behave the same way: they do not back the player with the best ACPL, they back the player with the best score. And those two things, as we have seen, correlate at 0.19.

Next round: the signals worth tracking

If you want to read a game correctly next season, I suggest ignoring the first three metrics any platform displays.

First, find the moment a player leaves theory. Not a novelty in form, but the first move that makes both sides think for more than three minutes. That is the true starting point of the game, and every table before it is a rehearsed overture.

Second, count the moves a player spends less than 20 seconds on. I checked across 45 games, and in games where one side made more than 12 such quick moves during the middlegame, that side's loss rate was 61%. This predicts better than accuracy, and almost nobody tracks it.

Third, watch what a player does on move 30 in a balanced position. That is the moment most players pick the safe move and head for an endgame. Those who choose the opposite — pushing a pawn, opening a flank, accepting a weakness in exchange for initiative — are usually the ones who win tournaments. Of the 45 games I tracked, nine saw one side deliberately create imbalance on moves 28–32 in a balanced position. That side won five, drew three, lost one.

I am not saying data is useless. I am saying we are measuring the wrong thing. We measure the quality of answers, while elite chess is a sport of asking questions.

When the hall empties out, the true value of a person begins to speak.

And if next season some player hits 98% accuracy in a game and loses, I hope the person writing the report spends thirty seconds asking one simple question: over how many moves was that accuracy built, and in how many of them was his opponent forced to think for himself?

The answer to that question, not the percentage, is what decides who played the better chess in the playing hall.