Unpredictable Bounces, Predictable Results

These days, the grass court season is the awkward stepchild of the tennis calendar. It takes place almost entirely within a single country’s borders, lasts barely a month, and often suffers from the absence of top players who prefer to rest after the French Open.

The small number of grass court events makes the surface problematic for analysts, as well. The surface behaves differently than hard or clay courts and rewards certain playing styles, so it’s reasonable to assume that many players will be particularly good or bad on grass. But with 90% of tour-level matches contested on other surfaces, many players don’t have much of a track record with which we can assess their grass-court prowess.

I was surprised, then, to find that grass court results are rather predictable. Elo-based forecasts of ATP grass court matches are almost as accurate as hard court predictions and considerably more effective than clay court forecasts. Even when we use “pure” surface forecasts–that is, predicting matches using ratings which draw only on results from that surface–grass court forecasts are a bit better than clay court predictions.

I started with a dataset of the roughly 50,000 ATP matches from 2000 through last week, excluding retirements and withdrawals. As a benchmark, I used official ATP rankings to make predictions for each of those matches. 66.6% of them were right, and the Brier score for ATP rankings over that span is .210. (Brier score measures the accuracy of a set of forecasts by averaging the squared error of each individual forecast, so a lower number is better. To put tennis-specific Brier scores in context, in 2016, ATP rankings had a .208 Brier score, and aggregate betting odds had a .189 Brier score.)

Let’s break that down by surface and compare the performance of ATP rankings, Elo, and surface-specific Elo. “F%” is the percentage of matches won by the favorite–as determined by that system, and “Br” is Brier score:

Surface  ATP F%  ATP Br  Elo F%  Elo Br  sElo F%  sElo Br  
Hard      67.3%   0.207   68.0%   0.205    68.5%    0.202  
Clay      66.1%   0.211   67.1%   0.211    67.0%    0.213  
Grass     66.0%   0.215   67.6%   0.207    68.5%    0.207

All three rating systems do best on hard courts, and for good reason: official rankings and overall Elo are more heavily weighted toward hard court results than they are clay or grass. Surface-specific Elo does best on hard courts for a similar reason: more data.

Already, though, we can see the unexpected divergence of clay and grass courts, especially with surface-specific Elo. It’s possible to explain overall Elo’s better performance on grass courts due to the presumed similarly between hard and grass–if a player excels on one, he’s probably good on the other, even if he’s horrible on clay.  But that doesn’t explain sElo doing better on grass than on clay. There are 3.3 times as many tour-level matches on clay than on grass, so even allowing for the fact that players choose schedules to suit their surface preferences, almost everyone is going to have more results on dirt than on turf. More data should give us better results, but not here.

We can improve our forecasts even more by blending surface-specific ratings with overall ratings. After testing a wide range of possible mixes, it turns out that equally weighting Elo and sElo provides close to the best results. (The differences between, say, 60/40 and 50/50 are extremely small on all surfaces, so even where 60/40 is a bit better, I prefer to keep it simple with a half-and-half mix.) Here are the results for weighted surface Elos for all three surfaces:

Surface  ATP F%  ATP Br  
Hard      68.6%   0.202  
Clay      68.0%   0.207  
Grass     69.8%   0.196

Now grass courts are the most predictable of the major surfaces! Even when we use a weighted average of Elo and sElo, grass court forecasts rely on less data than those of the other surfaces–the surface-specific half of the grass court forecasts uses less than one-third the match results of clay court predictions and less than one-fifth the results of hard court forecasts. In fact, we can do at least as well–and perhaps a tiny bit better–with even less data: A 50/50 weighting of grass-specific Elo and hard-specific Elo is just as accurate as the half-and-half mix of grass-specific and overall Elo.

Regardless of the exact formula, it remains striking that we can predict ATP grass court results so accurately from such limited data. Even if one-third of ATP events were played on grass, I still wouldn’t have been surprised if grass court results turned out to be the least predictable. The more a surface favors the server–and it’s hardest to break on grass–the tighter the scoreline will tend to be, introducing more randomness into the end result. Despite that structural tendency, we’re able to pick winners as successfully on grass as on the more common surfaces.

Here’s my theory: Even though there aren’t many grass court events, the conditions at those few tournaments are quite consistent. Altitude is roughly sea level, groundskeepers follow the lead of the staff at Wimbledon, and rain clouds are almost always in sight. Compare that homogeneity to the variety of hard courts or clay courts. The high-altitude hard courts in Bogota are nothing like the slow ones in Indian Wells. The “clay” in Houston is only nominally equal to the crushed brick of Roland Garros. While grass courts are almost identical to each other, clay courts are nearly as different from each other as they are from other surfaces.

It makes sense that ratings based on a uniform surface would be more accurate than ratings based on a wide range of surfaces, and it’s reassuring to find that the limited available data doesn’t cancel out the advantage. This research also suggests a further path to better forecasts: grouping hard and clay matches by a more precise measure of surface speed. If 10% of tour matches are sufficient to make accurate grass court predictions, the same may be true of the slowest one-third of clay courts. More data is almost always better, but sometimes, precisely targeted data is best of all.