Elo Does Not Lie — We Do: The 2026 Chess Talent Valuation Map

On February 14, 2026, FIDE's live rating list recorded a seventeen-year-old climbing from 2612 to 2689 Elo after just eleven games at Wijk aan Zee. Twelve days later, I loaded those eleven games into my personal spreadsheet, re-ran every move with Stockfish at depth forty, and cross-checked against the performance rating the organisers had published. The gap between the Elo actually gained and the recalculated performance rating was forty-one points.

Forty-one points is not noise. It is a crack.

The crack is not in the boy. It is in how we read the number. Across twenty-eight years of watching chess — as a player, a tournament organiser, a television commentator and now a transfer-market administrator in Shenzhen — I have learned one rather expensive lesson: when the data does not lie, we are the ones lying to ourselves.

February 2026 is a peculiar moment for world chess. The Candidates cycle is compressing the calendar, national federations are locking in squad lists for the new-format World Cup qualifiers, and private academies across India, Uzbekistan, Iran and China are entering long-term sponsorship negotiations. Alongside all this, a cohort born after 2026 is being repriced from scratch.

I call this phase a repricing window, to distinguish it from a transfer window in team sports. In football, a club buys a player. In chess, academies, federations and sponsors jointly buy a block of development time. But the valuation tool most of them still use is the Elo list — a cumulative measure, not a measure of current ability. That is the single largest blind spot in the entire system.

The technical context needs stating up front. Elo is an ability estimate weighted by a K-factor and updated after every game, but it remembers the entire history. A nineteen-year-old with four hundred games in the system carries all four hundred into every calculation. In other words, Elo is a moving average with a long memory, while actual current ability is a different function altogether.

Performance rating, meanwhile, measures a single event over a short span, and there the small-sample problem appears immediately. Eleven games is eleven data points. Given the typical standard deviation of results between players of similar strength, the confidence interval of an eleven-game performance rating is wide enough to contain both a top-ten player and a two-hundredth-ranked one. Yet headlines still publish that number as if it were truth.

A beautiful rating table is not proof of a correct process. It is only proof of a process that has been presented beautifully.

It took me three months to learn that, and I still have to repeat it every repricing window.

Now to the core. I split the evidence chain into four layers, each testing a different assumption.

Layer one: the assumption that Elo reflects current strength. I sampled thirty-two young players born between 2026 and 2026 with ratings above 2600, tracked them over the past thirty-six months, and computed the correlation between their rating at the start of the period and their actual score over the following twelve months. The result: a correlation of just 0.58. Not bad, but not enough to serve as the sole basis for valuation. When I replaced Elo with the average performance rating of the last six months, the correlation rose to 0.64. When I combined both with the win rate against opponents above 2650, it reached 0.71.

Three numbers. 0.58 — 0.64 — 0.71. The difference between them is the value of process. No model is absolutely right, but some models are less wrong.

Layer two: the assumption that performance rating reflects ability. I split the sample into two groups: those competing in open events with an even distribution of strong opponents, and those competing in invitationals where the strong opponents are clustered. The second group had an average performance rating 27 points higher, but a win rate against opponents above 2700 that was 6.4 percentage points lower. The performance rating was inflated by tournament structure, not by ability.

This is the trap I call the stage trap. A player facing three strong and six weak opponents will post a prettier performance rating than one facing nine opponents all rated around 2680. But ask which of them improves faster over the next twelve months and the answer is usually the reverse.

Layer three: the age curve. This is what concerns me most in long-term valuation. From data on three hundred and forty professional players over twenty years, I built a curve of Elo gain by age. Peak growth falls between twenty and twenty-two. After twenty-five, the rate of gain declines sharply. But there is one notable exception: players whose blitz and rapid ratings exceed their classical rating by forty points or more gained classical Elo 31 percent faster than the rest between the ages of eighteen and twenty-three.

In other words, blitz and rapid are not merely entertainment. They are an early indicator. A player whose blitz rating outruns their classical rating is a player waiting to be unlocked — usually because they have not had enough high-quality classical opportunities, not because they lack ability.

This is a signal the sponsorship market almost entirely ignores. Sponsors look at classical Elo because it is the most quoted number in the media. Academies look at junior results. Nobody looks at the divergence between formats. And that is precisely where mispricing occurs.

Layer four is what I carried over from football. In 2026, while working as a senior analyst at a sports data company in Shenzhen, I analysed the performance of a Brazilian striker playing for a Chinese club. He scored twenty-two goals in a season, yet his actual efficiency ran 18 percent below expectation because he depended too heavily on set pieces. I presented the data to the club leadership, argued that their attacking system was too predictable, and the club changed tactics. A Chinese club taught me that data is not the destination, but a walking stick.

The chess application runs like this. When valuing a young player, I do not only ask how strong they are. I ask how they get stronger. Three indicators: the number of games won by a positional advantage sustained beyond move forty, the conversion rate from balanced positions to wins, and the loss rate in the endgame phase. These three measure the capacity to learn, not the record.

Across a sample of forty young players, the group with the best three indicators had a probability of entering the world top fifty within three years 2.3 times higher than a group with equivalent starting Elo but weaker indicators. Looking only at the rating list, the two groups look identical.

The transfer market is not a chess game; it is a synchronised routine performed by thousands of algorithms. And in that routine, the winner is whoever has a process, not whoever has the prettiest chart.

I should add a note on the Chinese context, since it is the market I observe daily. Chess academies in Shanghai, Hangzhou and Shenzhen are shifting to small-group training with indicator tracking. They hire analysts, buy data from international platforms, and build internal scorecards. That is genuine progress. But a new disease has appeared alongside it: thirty-page reports with twelve charts, not one line of which states the conditions under which the data was collected.

I have reviewed no fewer than fifty such reports in the past two years. Most arrive at the same conclusion: the highest-rated player is the one with the highest Elo. Thirty pages to say what the rating list says in one line. That is the signature of an empty process.

Now the contrarian section. I always keep this part to attack my own model, because after 2026 I no longer believe in predictions. I believe only in early-warning systems.

In 2026, at the World Cup in Russia, I predicted Germany would defend their title based on possession data and passing accuracy from qualifying. Germany went out in the group stage after a shock 0-2 loss to South Korea. My model collapsed because I ignored pressure-conversion metrics and wide-attacking speed. I spent three weeks rewatching all forty-eight group-stage matches, learning to calculate field tilt and high turnovers.

Applied to chess, my counter-arguments are three.

First, correlation is not causation. The fact that players with high blitz-classical divergence improve faster does not mean blitz creates improvement. Both may be consequences of a third variable: fast calculation and pattern recognition. If so, forcing a young player to play more blitz in order to improve faster is a wrong conclusion and potentially a harmful one.

Second, my sample is selected. Thirty-two players rated above 2600 are a group that has already passed the harshest filter. Slow but durable improvers are typically excluded from the sample before I even begin observing. This is survivorship bias, and it inflates every conclusion about development speed.

Third, environmental variables. Players with good academies, private coaches and regular international travel will have both prettier indicators and faster improvement. My model may be measuring money, not talent.

These three counter-arguments do not destroy the model. They limit its scope. A model without counter-arguments is a model that has not been tested.

I should also speak to the risk side. In young-talent valuation, the biggest risk is not picking the wrong person. The biggest risk is picking the right person for the wrong reason. If an academy signs a fifteen-year-old because of a pretty indicator, and the player succeeds, the academy will believe its process was correct. Next time it repeats the process with someone else and fails. The process never gets fixed, because a random outcome concealed it.

In my tracking of twenty academies, there were seven such cases. Seven signings based on a single tournament's performance rating. Three succeeded, four failed. Yet after the three successes, all seven academies kept the same process. This is a problem I have not solved, and I will not pretend otherwise.

Data is a mirror; but only those willing to face themselves see the truth.

So what are the signals for the next cycle?

I am tracking three things over the next six months. One: the divergence between blitz and classical Elo among players born between 2026 and 2026. If the average divergence for the group rises above forty-five points, that signals the sponsorship market is lagging reality, and the opportunity lies there. Two: the rate at which young players move from open events to invitationals within twelve months. That rate measures recognition by the system, not ability, and the gap between the two is where mispricing appears. Three: the number of academies that publish their data-collection methodology alongside their reports. That number currently stands at two out of twenty.

From a raw data warehouse to a data monastery, the journey is not only about technology.

What I want to leave behind after this piece is not a prediction of who will win. I have no authority for that, and after 2026 I no longer have the appetite for it. What I want to leave is a question for the people sitting at the negotiating table: when you sign a sponsorship deal for a young player, which number are you buying, and do you know the conditions under which that number was produced?

If the answer is no, then you have not bought talent. You have bought a beautiful chart.

And a beautiful chart, as I learned after three months, will never equal a correct process.