Glicko-2

Elo gives every competitor a single number. Glicko-2 gives three: a rating, a rating deviation (how uncertain we are about it), and a volatility (how erratic their results have been). That extra structure matters.

• A new player with a high RD moves a lot per game — we're still learning who they are.
• A veteran with a low RD barely budges from a single upset.
• RD grows during inactivity: certainty decays, so a player returning after a long break is treated as less known again.

I experimented with football ratings in an AI-assisted dataset project, replaying match history chronologically. A rating is a candidate feature, not evidence of predictive expected value until it has been evaluated out of sample against a specified comparison.

The general lesson: a point estimate without its uncertainty is half a measurement. Same reason I won't act on a model until I validate before automating.