Machine learning reveals when changing shot direction pays off in elite tennis

Changing the direction of a shot is one of the most instinctive weapons in elite tennis, yet its true value has been surprisingly hard to measure. A new study published in BMC Sports Science, Medicine and Rehabilitation has built a stroke-level machine learning framework to quantify exactly when redirecting the ball helps a player and when it quietly costs them the point. Analysing thirty Grand Slam men’s singles matches from 2023 to 2025, the researchers found that redirection is not uniformly beneficial: its tactical efficiency depends on the interaction between directional change, stroke execution, movement demand and the depth at which the ball lands.

The dataset covered 33,975 strokes from baseline rallies played on hard, clay and grass courts. Every stroke was labelled as Attack, Neutral or Defense, giving the model a three-state picture of each rally situation. The team then trained an XGBoost model to estimate what they call expected rally status, abbreviated xRS, which outputs the predicted probabilities of those three states for any given stroke. These probabilities were converted into an expected stroke performance value, or xSPV, a single number that expresses the immediate tactical worth of a shot in its specific context.

Redirection efficiency was then defined as the difference in mean xSPV between redirected strokes and direction-maintained strokes. By comparing these two groups across surface type, stroke technique, shot location depth and movement demand, the researchers could see precisely where changing direction adds value and where it subtracts it. The ΔxSPV analyses showed that positive effects were mainly observed for forehand and backhand strokes, under low-to-moderate movement demands, and when sufficient shot depth was maintained. In other words, a redirected shot hit off balance or left short in the court can be worse than simply maintaining direction.

The underlying rally-status model performed respectably on a held-out test set, achieving an accuracy of 0.7811, a weighted F1-score of 0.7802 and a log-loss of 0.4046. SHAP analysis, a technique that attributes model predictions to individual input features, identified position–technique interaction, movement demand, shot direction, stroke technique, stroke position, shot location depth and opponent-related context as the main contributors to rally status. This means the model did not simply learn that aggressive shots are good; it learned that the value of any stroke emerges from how its technique interacts with where the player is standing, how far they must move and what the opponent is doing.

One of the study’s central findings is that redirected shots shifted the predicted probability distribution of rally status in a specific way: they increased the mean predicted probabilities of Attack and Neutral and reduced the probability of Defense. That pattern sounds like an unambiguous endorsement of redirection, but the ΔxSPV results complicate the story. Because the value of a stroke is context-sensitive, redirection under high movement demand or with poor landing depth can push a rally toward less favourable states even while redirection in comfortable conditions pushes it toward more favourable ones.

To check that the framework measures something real, the researchers assessed criterion-related validity against point outcome. Higher xSPV was associated with greater odds of winning the point, with an odds ratio of 1.079 per 0.1-unit increase in xSPV, and the association was strongest near point termination. That last detail matters: the metric tracks the immediate tactical state of a rally, so it should predict outcomes best in the moments just before a point is decided, which is exactly what the data showed. Point-winner versus point-loser comparisons were corrected using the Benjamini–Hochberg false discovery rate procedure to guard against spurious findings across many simultaneous statistical tests.

Robustness was examined across 21 alternative utility-weight specifications, and the main ΔxSPV patterns remained stable. This matters because any value metric built from predicted probabilities depends on how much weight is assigned to each rally state. If the conclusions had flipped whenever the weighting scheme changed, the framework would be little more than a modelling artefact. Instead, the finding that redirection helps most on forehands and backhands, at low-to-moderate movement demands and with sufficient depth, held across all 21 specifications.

The practical implications reach beyond the scoreboard. Coaches and analysts often treat directional change as inherently aggressive, encouraging players to redirect the ball whenever an opening appears. This study suggests a more conditional view: the tactical efficiency of shot redirection depends on the interaction between directional change, stroke execution, movement demand and landing-depth quality. A player stretched wide who redirects with a stretched, low-quality contact may hand the initiative back to the opponent, while the same redirection from a balanced position with depth can genuinely increase the probability of an attacking rally state.

The authors position the xRS–xSPV framework as a context-sensitive measure of immediate tactical value, suited to retrospective tactical evaluation in elite tennis rather than live prediction. Because it operates at the level of individual strokes, it can isolate the marginal effect of a single tactical choice, something aggregate statistics such as winner counts or unforced errors cannot do. As stroke-level tracking data becomes standard across professional tours, frameworks of this kind could shift tactical analysis from intuition about aggression toward calibrated estimates of when aggression actually pays.

Subject of Research: Machine learning analysis of the tactical efficiency of shot redirection in elite men's tennis baseline rallies

Article Title: Quantifying the tactical efficiency of shot redirection in elite men’s tennis rallies: a stroke-level machine learning framework

Article References: Gao, T., Li, Y., Yu, S., Zhan, Z., Cao, J., & Liu, W. (2026). Quantifying the tactical efficiency of shot redirection in elite men’s tennis rallies: a stroke-level machine learning framework. BMC Sports Science, Medicine and Rehabilitation. https://doi.org/10.1186/s13102-026-02197-4

Image Credits: AI Generated

DOI: 10.1186/s13102-026-02197-4

Keywords: tennis, machine learning, XGBoost, shot redirection, tactical analysis, Grand Slam, sports science, rally status, stroke-level analysis, performance metrics, Quantifying, tactical