How Topspin Elo treats table tennis results

A useful table tennis rating has to solve three problems: format, field strength, and recency. The model starts with a standard Elo update, then adjusts confidence by match format and tournament tier. This is the first pass at a sport-specific rating, and it is the number we cite when the official ranking points and the underlying level disagree.

The goal is not to replace the ITTF ranking. The ranking is the official record and it is what determines seeds and entries. The goal is to have a second number that reacts to opponent strength faster than the ranking does, so that form gaps show up before they show up in points.

The base Elo update

The foundation is the standard Elo update: each player has a rating, a match produces an expected score from the two ratings, and the ratings move toward the actual result by a fraction of the surprise. Win when you were expected to lose and your rating jumps; win when you were expected to win and it barely moves.

That core is unchanged. What changes is how much each match moves the rating, and that is where table tennis is not chess. A best-of-seven at a Smash and a best-of-five at a lower-tier event are not the same amount of evidence about a player's level, and the update size has to reflect that.

Format and field strength

The first adjustment is for match format. Longer matches are more informative because the better player wins more reliably over more games, so a best-of-seven result moves the rating more than a best-of-five. The K-factor, which controls how much each match moves the rating, scales with format length within a bounded range so that one format cannot dominate the rating on its own.

The second adjustment is for tournament tier. A win at a Smash counts for more than a win at a feeder event because the field is stronger and the path to the win is harder. Tier acts as a multiplier on the K-factor, which means a deep run at a major moves the rating more than the same number of wins at a smaller event. This is what stops a busy schedule at weak events from inflating a rating.

Recency and confidence

The third problem is recency. A player's level is a moving target, and a rating that weights a result from eleven months ago the same as one from last week is too slow to pick up form changes. The model decays older results so that the rating reflects recent level more than historical level, without throwing away the longer-term signal entirely.

Confidence is the other half of recency. A rating built on thirty recent matches is more trustworthy than one built on three, and the model tracks how much evidence sits behind each rating. When we cite an Elo estimate alongside a ranking, the confidence is part of what we are reading: a steep slope on low confidence is a hint, while the same slope on high confidence is a finding.

What the rating is and is not

The rating is a form-sensitive complement to the official ranking, not a replacement for it. It is most useful exactly when the ranking is least useful, which is when a player's current level has moved faster than the ranking window can reflect. That is the gap this model is built to fill, and it is why the analysis on this site leans on it whenever seed position and current form point in different directions.

It is still a first pass. The confidence model is conservative, the tier multipliers are tuned by hand, and live ingestion is not finished, so the numbers currently run on sample ranking snapshots. The shape of the model is right; the calibration will tighten as real match data lands.