A Unified Server Quality Metric for Tennis

Traditional tennis rating systems, such as Elo, summarize overall player strength but do not isolate the independent value of serving. Using point-by-point data from Wimbledon and the U.S. Open, we develop serve-specific player metrics to isolate serving quality from overall performance. For each tournament and gender, we fit logistic mixed-effects models using serve speed, speed variability, and placement features, with crossed server and returner random intercepts capturing unobserved server and returner-strength effects. We use these models to estimate Server Quality Scores (SQS) that reflect players’ serving ability. In out-of-sample tests, SQS shows stronger alignment with serve efficiency (measured as points won within three shots) than weighted Elo. Associations with overall serve win percentage are smaller and mixed across datasets, and neither SQS nor wElo consistently dominates on that outcome. These findings highlight that serve-specific metrics complement holistic ratings and provide actionable insight for coaching, forecasting, and player evaluation.

💡 Research Summary

The paper addresses a notable gap in tennis analytics: the lack of a dedicated metric that isolates a player’s serving ability from overall performance. While traditional rating systems such as Elo (and its weighted variant wElo) summarize a player’s overall strength, they compress the multifaceted nature of tennis into a single number and do not explicitly model the serve, which is the only shot fully under a player’s control. To fill this void, the authors develop a Server Quality Score (SQS) that quantifies serving quality using point‑by‑point data from Wimbledon and the US Open (men’s and women’s singles, 2018‑2019 and 2021‑2024).

Data and Feature Construction
The authors extract publicly available point‑level data from Jeff Sackmann’s database. For each serve they compute: average serve speed (mph), standard deviation of speed (as a proxy for variability), a categorical “location bin” defined by ServeWidth and ServeDepth, and a location entropy that captures how dispersed a player’s placement is across bins. Only players with at least 20 serves of a given type (first or second) are retained. Continuous features are z‑standardized; location bins are one‑hot encoded.

Modeling Approach
Separate logistic mixed‑effects models are fitted for first serves and second serves. The binary outcome is “serve efficiency”: the server wins the point within the first three shots (serve → return → server’s next shot). The fixed‑effects part includes the standardized speed, speed variability, location entropy, and the one‑hot location vector. Crucially, crossed random intercepts for the server (u_j) and the returner (v_k) are added, each assumed Gaussian with variance components σ² and τ² respectively. This structure adjusts for the quality of opponents faced, ensuring that the server random effect reflects pure serving impact rather than a favorable draw.

The linear predictor for a given player j and serve type s is:

 η_{s,j} = β_{0}^{(s)} + β_{1}^{(s)}·speed_z + β_{2}^{(s)}·sd_speed_z + β_{3}^{(s)}·LocationOneHot + β_{4}^{(s)}·entropy_z + u_{s,j}

The Server Quality Score (SQS) is defined as this linear predictor evaluated at an average returner (v_k = 0). Thus SQS is on the log‑odds scale, directly additive across fixed and random components. Two scores are reported per player: SQS(1) for first serves and SQS(2) for second serves.

Out‑of‑Sample Evaluation
Data are split 80/20 within each tournament‑gender subset, preserving whole matches in a single split to avoid leakage. The training set is used to estimate model parameters and compute SQS for every player. The held‑out test set provides two aggregate outcomes per player: (i) Serve Efficiency (fraction of points won within three shots) and (ii) overall Point Win Percentage (fraction of all points won). For each outcome, a binomial GLM regresses the observed success counts on the corresponding SQS (first‑serve SQS for first‑serve points, second‑serve SQS for second‑serve points). The same procedure is repeated using weighted Elo (wElo) as the predictor, providing a baseline.

Results
Across all four datasets (Wimbledon men, Wimbledon women, US Open men, US Open women) SQS shows a consistently strong positive correlation with Serve Efficiency on first serves (e.g., r = 0.667 for Wimbledon men, r = 0.564 for Wimbledon women). The regression coefficients are also highly significant (p < 10⁻⁴). By contrast, wElo’s correlation with Serve Efficiency is modest (r ≈ 0.15) or even slightly negative in some cases. For second serves, the SQS‑Efficiency correlation drops but remains positive in three of four datasets (r ≈ 0.23 for Wimbledon men) and is insignificant or slightly negative for US Open men. Correlations with overall Point Win Percentage are weaker for both metrics and vary in sign, reflecting that long‑rally outcomes depend on many skills beyond the serve.

Interpretation and Implications
The findings confirm that SQS captures a serve‑centric signal that is largely orthogonal to holistic match ratings. By incorporating measurable serve characteristics and adjusting for opponent return strength via crossed random effects, SQS isolates the intrinsic quality of a player’s serve. Its strong alignment with the three‑shot efficiency metric demonstrates that the model successfully extracts the immediate impact of the serve, whereas overall win percentages dilute this effect with rally dynamics. Practically, SQS can be used by coaches to pinpoint whether a player’s first‑serve or second‑serve game needs improvement, to compare serve profiles across players, or to augment existing Elo‑type forecasts with a serve‑specific component. The two‑dimensional nature of SQS (first vs. second serve) also respects the strategic shift from aggressive first serves to more conservative second serves.

Future Directions
The authors suggest extending the framework to other shot types (return, net play) and to incorporate contextual variables such as surface, match pressure, or fatigue. A unified hierarchical model that simultaneously estimates serve, return, and rally quality could provide a comprehensive picture of player skill composition, further enhancing both interpretability and predictive performance.

In summary, the paper introduces a rigorously validated, statistically sound metric—Server Quality Score—that isolates serving ability from overall performance, outperforms weighted Elo in predicting short‑point success, and offers actionable insights for player evaluation, coaching, and forecasting.