Interpretable ensemble machine learning identifies player-specific performance indicators in elite tennis

In the world of elite sports, performance analysis is the art of turning the chaos of a match into clear evidence. Coaches and scientists watch games not just to see who won, but to understand how the victory was built. They break a game down into thousands of tiny moments: a serve, a return, a footstep, a mistake. The goal is to find which of these moments actually matters most. However, a single number or a simple average can be misleading. Just because a player hits a lot of winners does not mean that specific action caused the win; sometimes, the most important factor is something subtle that only reveals itself when you look at the whole picture. This is where modern computer science steps in. Instead of relying on one way of thinking, researchers can now use many different computer programs at once. These programs act like a team of experts, each with a slightly different way of seeing the data. By listening to all of them, rather than just the loudest voice, analysts can find patterns that a single method might miss. This approach is becoming vital in sports like tennis, where the difference between winning and losing is often hidden in the details.

A team of researchers at Wuhan Sports University and Wuhan University applied this team-based thinking to the career of Novak Djokovic, one of the greatest tennis players in history. They wanted to know exactly which technical skills and tactical choices were the strongest drivers of his success in winning a set. To do this, they gathered a massive amount of data from 239 sets played on hard courts at major Grand Slam tournaments between 2019 and 2025. They tracked thirty-four different indicators for every set, ranging from obvious things like the number of double faults to more specific details like the speed of a backhand shot or the number of times a player approached the net. The researchers split this data into two groups. They used the older data from 2019 to 2024 to teach twelve different computer algorithms how to predict the outcome of a set. Then, they tested these programs on a fresh, unseen group of forty sets from 2025 to see if the lessons held up in the real world.

The researchers did not just pick the single best computer program. Instead, they tested every possible combination of the twelve algorithms, creating thousands of different "teams" or ensembles. They found that the most accurate team was not a giant group of all twelve, but a specific four-model ensemble. This selected group of four programs worked together to predict whether Djokovic would win or lose a set with remarkable precision, achieving an accuracy of 90 percent on the new 2025 data. When the team looked at what these programs had learned, they found that the most important factors were exactly what experts had long suspected. The ability to break an opponent's serve, the total number of points won, and the control of unforced errors were the top drivers of success. These results confirmed that the computer models were not making things up; they were recovering the established truths of the game.

However, the true value of this study appeared when the researchers looked for the things that a single computer program might have ignored. When they asked each of the twelve individual programs to rank the importance of the thirty-four indicators, they found a surprising disagreement. One specific indicator, the number of backhand returns, was ranked as a top-five factor by the combined team, but it appeared in the top five for only three of the twelve individual programs. If the researchers had simply picked the single best-performing program, they might have missed this signal entirely. The combined approach revealed that the frequency of backhand returns was a subtle but consistent marker of success for Djokovic, a pattern that was too quiet for any single algorithm to hear on its own.

The study also clarified how these patterns work. The researchers showed that when Djokovic made more backhand returns, it was associated with a higher chance of winning the set. They visualized this relationship and found that the positive effect became more noticeable when the number of backhand returns in a set reached around twenty-two. It is important to note that this number is not a magic target or a rule that every player must follow. It is simply a description of what happened in these specific matches. The researchers were careful to explain that they could not say why this happened. It could be that opponents served more to his backhand, or that he had more opportunities to return, or that the length of the rallies changed his strategy. The data showed a link, but it did not prove the cause.

This work demonstrates a new way to study sports performance. By using a team of different computer models and comparing their answers, analysts can separate the noise from the signal. They can confirm what is already known, like the importance of breaking serve, while also shining a light on quieter, player-specific habits that might otherwise be overlooked. The study does not claim that these findings apply to every tennis player or every surface. The results are specific to Novak Djokovic and the hard-court matches he played during these years. Yet, the method itself is a powerful tool. It offers a reproducible way to look at repeated performance data, ensuring that the conclusions drawn from a game are based on a broad consensus of evidence rather than the narrow view of a single analytical tool. For the curious observer, it suggests that in the complex dance of elite sport, the most important clues are often found not by listening to one expert, but by asking many.