Elo in tennis: understanding the rating that predicts matches

The ATP ranking tells you who accumulated the most tournament points over the last fifty-two weeks. It does not tell you who is the strongest today, nor who is most likely to win a given match. For that, prediction models almost all use another tool, borrowed from chess: the Elo rating. A single number per player, updated after every match, that rises when you beat a strong opponent and barely moves when you beat a weak one.

This article explains what Elo is, how it is computed, why it predicts better than an official ranking, and above all which adjustments are essential for it to work in tennis, a sport with many new entrants, three surfaces and scores that say more than a plain win. The choices described are those of the Pronocast model; each was measured before being adopted, and some results run against intuition.

Elo in one formula

The principle has two steps. First, the rating gap between two players converts into an expected win probability. With the classic convention, a 400-point gap gives about 91% to the higher-rated player, a 100-point gap about 64%, and a zero gap exactly 50%. Then, once the match is played, each player’s rating is corrected by the difference between the result (1 for a win, 0 for a loss) and the expected probability, multiplied by a K-factor.

Why it works

Beating a player you were expected to beat at 90% earns a tenth of K: almost nothing. Beating them when you had a 10% chance earns nine tenths of K: a lot. The system rewards surprise, and corrects faster the more wrong it was. Over the matches, ratings converge towards a hierarchy where the gap between two players reflects their real probability of beating each other, not their tournament attendance.

What the ATP ranking does not see

The official ranking awards points by round reached, whoever was beaten. A quarter-final reached thanks to two walkovers is worth as much as one wrested from two seeds. It also ignores tournaments outside the main tour beyond a certain threshold, and it forgets everything after a year, all at once. Elo has none of these flaws: it only cares about who beats whom, and it forgets gradually, through the updates themselves.

First adjustment: a K-factor that decreases with experience

In chess, K is often a constant per player category. In tennis, with hundreds of new players every season, a constant does not fit. A beginner should see their rating move quickly, because almost nothing is known about them; a veteran of six hundred matches should move slowly, because one isolated result does not overturn what is known.

The formula used

The model uses K = 250 / (n + 5)^0.4, where n is the number of matches the player has already played. For a first match, K is about 131; after fifty matches, about 50; after five hundred, about 21. The decay is gentle: it never quite freezes a player, which allows a decline or a late breakthrough to be tracked, but it protects established ratings from the noise of a bad day.

The case of injury comebacks

This is the known limit of the approach. A player returning after a year away has kept their previous rating and a low K, that of an experienced player. If they are diminished, the system will take several matches to reflect it. Recent form and the fatigue indicator, read alongside Elo by the model, partly compensate, not fully.

Second adjustment: a surface Elo, initialised on the global one

A player is not the same on clay and on grass. The natural idea is to keep a rating per surface, on top of the global one. It is a good idea, on one condition that makes all the difference: the starting value.

The counter-intuitive result

First version tested: each surface rating starts from a neutral value, like the global one. Measured result: the surface rating predicted worse than the global rating alone. The reason is arithmetic. A player plays three to four times fewer matches on a given surface than overall; a rating that starts from scratch per surface does not have enough matches to converge, and stays noisy for seasons.

The fix

The surface rating is initialised on the player’s global rating at their first match on that surface, then evolves at its own pace. It thus inherits everything the global rating already knows, and only has to learn the surface-specific gap, which is much smaller. In this configuration, surface Elo becomes a useful signal, read by the model next to the global one. On every finished match page, both values are published, and the analysis mentions the surface gap when it exceeds about fifteen points.

Third adjustment: the margin of victory

A 6-0 6-1 win and a third-set tiebreak win count the same in a classic Elo: 1 point for the winner. Yet the score contains information. A player who crushes an opponent of equal rating is probably underrated; a player who scrapes through a tiebreak, probably rated about right.

A simple multiplier

The model applies a multiplier to K, equal to 1 + 2 × max(0, share of games won − 0.5). A player who wins 60% of the games in the match sees their gain multiplied by 1.2; at 75% of the games, by 1.5. The multiplier is applied symmetrically to winner and loser, so that the sum of ratings stays conserved. It changes nothing for a tight win, and speeds up the correction after a rout.

Why games rather than sets

Sets are too coarse: in a best of three, there are only two possible scores. Games won, on the other hand, clearly separate a 6-4 6-4 from a 6-1 6-2, and remain comparable between a two-set match and a five-set one.

What Elo does not capture, and what comes alongside

Elo is the strongest signal in the model, but it is not alone. Three complements are read with it, each to fill a specific blind spot.

• Recent form, the share of wins over the last ten matches: an established Elo moves slowly, form captures a momentum of a few weeks.
• Fatigue, the number of matches played in the preceding days: Elo does not know that a player fought a three-hour final the day before.
• Head-to-heads, weighted by recency with a half-life of one to two years: some playing styles clash, and global Elo does not see it.

One last point deserves a mention, because it was a surprise: serve statistics (aces, first serves), “obvious” as they seem, brought nothing once added. A player who serves well wins more matches, so already has a better Elo. The rating had already absorbed the information.

Reading an Elo on Pronocast

The model’s ratings are published on each player page (global and surface Elo at the last covered match) and on every finished match page, next to the prediction and the result. A few reference points, measured on the matches covered in September 2026: an average Challenger player sits around 1,800 to 1,900, an average ATP or WTA main-draw player around 2,000, and the best players in the world exceed 2,400; a 100-point gap between two players corresponds to roughly a 64% chance of winning for the higher-rated one, before surface, form and head-to-heads.

Finally, keep in mind that an Elo is an estimate, not a measurement. Two players 15 points apart are, for the model, indistinguishable: the match analysis says so explicitly in that case, rather than inventing a favourite.

Today’s Pronocast predictions

The Elo described here is recomputed every morning on all of the previous day’s matches, before the model produces the day’s predictions. Those are listed on the predictions of the day page, and by tour on the ATP, WTA and Challenger hubs. Each prediction is frozen before the match and published with the result once it is over.