Tracking Data and Serve Strategy:
Part 2

Peter Tea

Players generally have lower risk tolerance when they are backed into tight situations. High pressure points in a match can rattle a player's nerves and effect their serve decisions.

And it is within these high-pressure situations where a player turns to their most comfortable and trustworthy tennis weapons. In other words, they become more predictable. To investigate, let's see how Roger Federer approached his pressure points at Roland Garros.

We formally derived match pressure using match score and applying some fancy math equations. Key points like break points and points approaching break (eg: 15-30) were assigned high match pressure while other less important points like 0-0 or 15-0 were assigned low match pressure.

In our study, Federer was more predictable on high pressure points. This meant more T serves in the Deuce court and more Wide/Body serves in the Ad. He was less predictable on low pressure points. Interestingly, all opponents in the study's Federer matches were right-handed. Federer's serve strategy on these high pressure points was to pound the backhand side.

Clearly a lot of factors influence a player's serve direction decisions. But from a scouting perspective, it is valuable to describe a player's serve tendencies under specific match scenarios.

For example, if Federer were trailing in a match (say down 0-1 in sets, 5-6 in games and facing break point) against a right-handed returner at Roland Garros, which direction would he serve?

To accomplish this herculean task, we fit a complicated Bayesian mathematical model to the data. The model gets fed information of all factors described thus far (player handedness, serve number and current match score) along with a host of other factors we haven't mentioned (like previous serve direction and ball-toss related characteristics). We also tell the model which player is serving so that it can search for any style differences.

After considering all these factors, the model then makes its best guess at serve direction. That is, the model learns the patterns we've specified above and then predicts which direction a player will serve. We kept it simple and classified serve direction into three categories based on ball bounce location: Wide, Body and T.

For the curious reader, the Bayesian model is set up as a Multinomial logic where the log odds of each serve direction is modelled against the set of described factors. For example, the log-odds of a player aiming down T is modelled as:

Log ((Probability(Serving T))/(Probability(Serving Body))) = server_name + server_handedness + serve_number + current_score +etc...

We fit a similar equation for the probability of serving Wide, and then algebraically solve for the individual probabilities of serving Wide, Body and T.

Now, returning to our Federer scenario, here's what our model predicted:

On first serve, the prediction is close between Wide or T. But on second serve, the model clearly favors a Wide serve – a direction that tactically would pull the right-handed returner off court and target their weaker backhand side.

Prediction is a difficult task to perform accurately. For instance, the Bayesian model might work well for some players, but less so for others. And with tennis filled with temperamental swings, tactics can easily change on a whim, which is tough to predict. But one clear example of a player who makes tactical adjustments is 20 time grand slam champion, Novak Djokovic. In the French Open final of 2020, Djokovic mixed in corner-to-corner serves against the king of clay, Rafael Nadal.

This strategy was not effective with Novak quickly losing in straight sets. But in their next encounter at the French Open in the 2021 semifinal, Djokovic completely changed tactics. In the 2021 semis match, Djokovic hounded Nadal's backhand with his serve: a strategy that would later hand Novak the victory in 4 sets.

Tracking data can help highlight any player's tendencies, whether it's your opponent's tendencies or even our own. And being aware of our own patterns can help us diagnose our exploitable weaknesses before the opponent even notices. As Boris Becker exclaimed after facing an infuriatingly prepared Agassi: "It's as if he's reading my mind!"