Elo vs Glicko-2, simulated: Glicko-2 catches a change in strength faster only when it is unsure of the player
Our own simulation, run on 2026-11-06 (sim.py, results.json beside this page). One test player of true strength 1500 plays opponents of known, fixed strength drawn evenly from 1300 to 1700. At game 0 the player's true strength jumps by 200 points, to 1700. Each game is decided by the Elo expected score of the true strengths, the same generator for both systems. We count the
games until the player's rating first comes within 50 points of 1700; the figure is the median of 2,000 runs. Elo uses FIDE's three K values. Glicko-2 follows Glickman's 2022 worked example (one game per rating period, as Lichess rates; system constant tau = 0.5; starting volatility 0.06), with three starting states: a settled player (rating deviation 50), a new player (rating deviation 350), and a player back from a break (rating deviation 50 grown by 24 idle periods).
What we found
- A settled player who improves by 200 points is rated within 50 points of the new strength after a median of 49 games under Elo with K = 20 (FIDE's K for players under 2400) and 51 games under Glicko-2 with a rating deviation of 50: no difference.
- The same jump takes a median of 107 games under Elo with K = 10 (FIDE's K at 2400 and above) and 22 games under Elo with K = 40 (FIDE's K for new players).
- A new player under Glicko-2 (rating deviation 350) is rated within 50 points of their strength after a median of 2 games, against 22 games for a new player under Elo with K = 40.
- A player returning after 24 idle rating periods under Glicko-2 catches up in a median of 40 games, against 49 for Elo with K = 20.
- In 2,000 runs of every case, every rating came within 50 points within 400 games.
What this does not show
One opponent pool (1300 to 1700), one jump size (200 points), one game per rating period, tau 0.5. Longer rating periods and a larger tau let Glicko-2 react faster to a settled player's change; we did not vary them. Catching up is not the same as predicting results better: we measured the lag, not the prediction error.