Long think, wrong think: Olympiad edition
Every chess player has heard the adage: think long, think wrong. And we’ve all probably had the corresponding experience: thinking a long time in a tough position, and finally playing a bad move, often after considering a better one earlier.
Is this a real effect, or just our imagination? Let’s take a look at how this played out at the recently completed 2026 Olympiad in Samarkand. There were over 8,000 games and 700,000 moves across 11 rounds, all of which were analyzed with machine learning at chessds.com. And from the live coverage, clock data is available too. That means we can use Expectation Points Added (XPA, a move-by-move play quality metric that compares play to what we’d expect of a modern GM) to estimate play quality as a function of thinking time. And what do you know, the adage is true?
Shown here on the x-axis is thinking time, on a log scale. On the y-axis is XPA for each move. These are grouped into bins of similar thinking time, within which the average XPA is plotted, with a 95% confidence interval produced by bootstrapping (2,000 player-level resamples). What we see is that lower XPA is correlated with longer thinking time. Players did tend to give up more game expectation after thinking longer, and chose the strongest moves less often.
Across ratings
One thing to note is that my models are built to predict GM-level play. But the Olympiad has a lot of lower-rated players, and so they could be the source of this observation. Indeed, cutting this result by rating level, the strongest players show an attenuated effect. But it’s still there: the worst move-level performance occurs on the moves where players thought the longest:
Complex positions
Correlation doesn’t equal causation – it’s of course not likely that thinking longer actually causes weaker play. Instead, difficult positions cause both. Positions may be difficult for various reasons; what’s difficult for one player might be easy for another. The XPA model produces a value that is something like position complexity: expected loss.
More details are in the original XPA blog, but positions with high expected loss are those where a game-altering mistake is deemed more likely by move prediction modeling. The signal is there, as longer thinks do tend to occur in positions with higher expected loss. But the actual average loss of players in these positions is higher than what the model expects:
Contested positions
Long thinking time also tends to happen in contested positions. Here’s thinking time now on the y-axis, with the x-axis showing game expectation for the player making the move. Note the dip around 0.5 – these buckets contain many clearly drawn positions. The peaks correspond to positions where the player is at an advantage or disadvantage that is not yet decisive:
Other events
This is a generalizable finding. So far I have limited coverage of events with clock data, but there are a number of events that have been analyzed live by now. It’s true for all the classical events covered live so far:
Back to the lab
While models are predicting higher expected loss in long-think positions, they are not doing a great job predicting the actual played move:
Long-think positions are a significant source of potential error in XPA, and thus I’m spending time now using them as teaching examples, with many promising findings for move prediction. Some observations:
- One strong predictor of long thinking is how surprising the previous move was. This is a feature going into the next generation of move prediction.
- It’s not hard to train a model that identifies long thinks. But interestingly, the long thinking the model could predict was not associated with higher-than-expected mistakes. There’s something extra to learn about the real situations in which players err.
- The move model used on the site today was trained to optimize a move-ranking objective – the model tried hardest to learn how to pick the most likely moves, sacrificing some accuracy on others. But for XPA, it’s more important to get expected loss right, and that means prioritizing accurate probabilities across all candidate moves.
This has been a great source of data for the next round of move prediction, which is training as I write this! Lots of new features going in; one of many updates I’m looking forward to shipping soon.