The ATP ranking adds up a player's results over the last 52 weeks. That is fair for deciding seedings, but it is slow to show form: a player who won most of his points ten months ago and has lost since can sit next to one who just beat two top-10 players. Our ELO ratings are built to tell them apart.
What is an ELO rating?
ELO comes from chess. When two players meet, the winner takes rating points from the loser. How many depends on how surprising the result was: beating a higher-rated opponent earns more than beating a lower-rated one.
One season at a time
Every January the ratings restart from the ATP ranking points (1500 plus a tenth of a player's points, so a player on 5,000 points starts at 2000). From there they move after every match of the season. Because of this reset, compare ratings within one season, not across seasons.
Why the ranking is not enough for predictions
Take two players both ranked around 15th in the world:
Same ranking, different form
Player A (ranked 15th):
- Lost in the first round of his last 4 tournaments
- Won most of his points 10 months ago
- Has lost 6 of his last 8 matches
Player B (ranked 15th):
- Just reached back-to-back semifinals
- Beat two top-10 players in the last month
- Won 8 of his last 10 matches
The ranking says they are equal. Their ELO ratings will not be. Player B has been taking points from strong opponents, Player A has been giving them away.
How much one match can move the rating
In classic ELO every match counts the same. In ours, the size of the change depends on where and when the match was played:
Tournament level
A Grand Slam match moves the rating two and a half times as much as an ATP 250 match. Masters 1000 matches count 1.8 times, ATP 500 matches 1.2 times.
Round
Late rounds count more: a final 1.5 times as much as an early round, a semifinal 1.3 times, a quarterfinal 1.2 times and the round of 16 1.1 times.
Rating gap and upsets
When two players are more than 400 points apart, the change grows with the gap, up to twice the usual size, and by another half again if the lower-rated player wins. Every change stays within fixed limits, so no single match can swing a rating wildly.
What else we track about each player
Alongside the rating we keep a few simple records for every player. They do not change the ELO itself, but they go into our prediction models:
- Upset wins: how often a player beats higher-rated opponents
- Dominant wins: the share of wins by two sets or more (2-0, 3-0 or 3-1)
- Close matches: how often a player's matches go to a deciding set
- Average opponent rating: how strong the competition has been
ELO vs. the ATP ranking
What each one tells you
ATP ranking
- Points from a player's best 19 tournaments
- 52-week rolling window
- Fixed points per round
- The same points whoever you beat
Our ELO ratings
- Changes after every match
- Depends on who you beat
- Weighs the tournament and the round
- Bigger swings for upsets
- Restarts from the ranking every season
The ratings our models use
We keep more than one rating per player:
| Rating | How it works |
|---|---|
| Points-started ELO | Starts the season from ATP ranking points, then moves with results |
| Equal-start ELO | Everyone starts the season at 1500, so the rating reflects this season's results only |
| Race ranking | Points won this calendar year, like the ATP Race |
| Performance score | A score for each win that adds up opponent strength, how clearly the match was won and any upset, weighted by tournament and round |
All of them are inputs to our prediction models, together with the surface and form indices and head-to-head records.
Note: ELO is one input among many. Tennis stays unpredictable, and no rating removes that.
See the ratings at work
Our tennis predictions combine ELO with surface indices, head-to-head records and many other inputs to give win probabilities for every ATP match.
In short
The ATP ranking tells you what a player has earned over the past year. His ELO rating tells you how strong he is this season: who he has beaten, at which tournaments and in which rounds.