How the numbers are made
The games
Every game of every world title match from 1886 to 2024, as ChessFlex ships them (from PGN Mentor's event files; each match's score checked against the record). The FIDE knockout championships of 1999–2004 are not matches and aren't included. Forfeited games have no moves and count only in the score.
The coach
Every position went through the ChessFlex coach exactly as the app runs it when you follow a game in Live: its own engine, its three best moves, its time budget. A side's chance is the coach's winning chances, as shown by each player's name in the app: an expected score from the evaluation, so 50% is level and a draw counts as half.
A move's cost, accuracy and blunders
A move's cost is how much of its own side's chances it gave up. A blunder costs 20% or more, a mistake 12%, a dubious move 6%: the thresholds of the app's own game review. Accuracy per move is 103.17·e^(−0.0435·cost) − 3.17 (the formula Lichess made popular), averaged over a player's moves.
The check
Every move the records name was looked at again by Stockfish 19 at depth 24, on the same scale. A move stays only if Stockfish agrees it cost 15% or more (and a missed win only if Stockfish also saw a win). Stockfish's numbers are never shown; the coach's are.
The match model
Each game's result from the two ratings: White's expected score 1/(1+10^(−(k·gap + edge)/400)), with a fixed draw rate; k = 0.35, edge = 40 rating points and draw rate = 0.68, fitted by maximum likelihood on 456 rated title-match games since 1972. During a game: what title-match positions at the same chance and stage actually ended as. The match chance counts every way the remaining games can go. In the back-test each match is left out of its own fit. A tied match goes to the tiebreak at the rating expectation.
The markets
Polymarket's public price history and trade feed (2024), and Kalshi's public prices (2026, bid–ask midpoint, recorded when this page is built). Quoted for information only. ChessFlex has no tie to either.