Separate Elo ratings for the same player by playing surface — hard, clay, or grass in tennis; turf or grass in soccer.
A single Elo rating averages a player's results across all conditions. For chess this is fine — the board doesn't change. For tennis it isn't. Rafael Nadal on clay is a different player from Rafael Nadal on grass, and a single overall Elo will systematically misprice his matches at the French Open and Wimbledon. Surface Elo solves this by maintaining separate ratings per surface.
The mechanics are identical to standard Elo, except every match updates only the rating for the surface it was played on. A player's hard-court Elo updates after hard-court matches; the clay rating sits frozen until the next clay event. The result is three or four parallel ratings per player that capture surface-specific true talent.
Predictions use the surface of the upcoming match. The expected score function is unchanged — the logistic on the rating gap — but the gap is computed using the surface-specific ratings. For a clay-court major, the Nadal vs. Djokovic projection uses clay Elos; for Wimbledon, it uses grass.
Surface Elos have been the public standard in tennis modeling for over a decade. They are well-correlated with bookmaker prices and produce calibrated forecasts on tennis-data.co.uk and Jeff Sackmann's tennisabstract. The same idea generalizes to other surface-sensitive sports — golf has variants that condition on course type, and soccer has versions that separate turf from grass. The principle is always the same: when conditions reliably change outcomes, ratings should be conditional on those conditions.
Our tennis model uses surface Elos seeded from Jeff Sackmann's published files and updated nightly from tour results. Surface-specific ratings produce visibly better calibration on clay events than blended ratings.