Tennis Momentum Model (TMM)

• Tennis Momentum Model (TMM) is a quantitative framework that infers dynamic performance trends from point-by-point data to represent in-match momentum.
• It integrates methods such as probability-based analysis, fuzzy evaluation, mixed-effects, and deep learning to predict shifts in competitive advantage.
• The framework informs tactical coaching and real-time decision-making, though challenges remain with heterogeneous definitions and scale mismatches.

Searching arXiv for recent TMM-related tennis momentum papers to ground the article. The Tennis Momentum Model (TMM) denotes a family of quantitative frameworks for representing momentum in tennis as a dynamic, match-state-dependent signal rather than a purely narrative construct. Across recent work, TMM has been instantiated as a momentum evaluation and prediction framework built from point-by-point data, as a probability-based update model, as a hidden-state performance model, as a fuzzy evaluation system, and as a multi-granularity deep architecture spanning points, games, sets, and matches (Li, 2024, Graves et al., 11 Sep 2025, Lei et al., 2024, Li et al., 25 Mar 2025, 2505.21882). Despite substantial methodological variation, these formulations share a common objective: to quantify shifts in competitive advantage, test whether those shifts are statistically associated with outcomes, and use the resulting signal for prediction, interpretation, and tactical analysis (Ma, 2024, Goyal et al., 2020).

1. Terminological scope and conceptual basis

In the tennis analytics literature, momentum is commonly treated as a time-varying advantage, a dynamic performance trend, or the direction of flow of the situation in a match (Graves et al., 11 Sep 2025, 2505.21882, Ma, 2024). Some papers define it operationally through scoring windows and streaks, others through latent-state inference, and others through carryover effects in probabilistic outcome models (Li, 2024, Lei et al., 2024, Goyal et al., 2020). A recurrent theme is that momentum is not directly observed; it is inferred from point sequences, serve context, score differences, technical indicators, or latent state transitions (Lei et al., 2024, Liu et al., 2024).

A central distinction in this literature is between descriptive and inferential uses of the term. In descriptive formulations, TMM produces a scalar or vector score intended to summarize current advantage, often visualized as a match-flow or momentum curve (Ma, 2024, Liu et al., 2024). In inferential formulations, momentum is defined as residual dependence of current outcomes on previous outcomes after controlling for player quality, serve/return context, and match state (Goyal et al., 2020). This shifts the concept from a rhetorical label to a measurable statistical dependency.

The term also requires disambiguation. The phrase “tennis racket effect” appears in rigid-body dynamics and quantum nanorotor theory, where it refers to unstable intermediate-axis rotation and separatrix-induced flipping rather than sports analytics. That literature does not present a sports-engineering or match-analysis Tennis Momentum Model; instead, it studies asymmetric rigid rotors and classically forbidden tunnelling near the separatrix (Ma et al., 2020). In the context of tennis analytics, TMM refers to models of in-match performance dynamics, not to the physics of racket-like rotational motion.

2. Data representations and match-state variables

Most TMM formulations are built on point-by-point match-sequence data, frequently augmented with player metadata or technical statistics (Li, 2024, Peng et al., 2024, Li et al., 25 Mar 2025). Common temporal resolutions are explicitly points, games, sets (Li, 2024), while multi-granularity frameworks extend this to the match level (2505.21882). Several studies use Wimbledon 2023 as a principal empirical setting, especially the Alcaraz–Djokovic final, while also validating on additional men’s matches, women’s matches, US Open data, or cross-tournament corpora (Li, 2024, Lei et al., 2024, Liu et al., 2024, 2505.21882).

Feature sets differ by model class but recur around a common core. Scoring-state variables include set difference, game difference, point difference, score difference, consecutive scores, and serve order (Li, 2024, Li et al., 25 Mar 2025). Technical and event variables include aces, double faults, unforced errors, winners, break points, net points, serve speed, serve width, serve depth, return depth, rally count, and distance run (Ma, 2024, Lei et al., 2024, Liu et al., 2024). Probability-based models additionally use historical serve success and rolling in-match serve success (Graves et al., 11 Sep 2025).

Some papers construct momentum from explicitly engineered indicators. One fuzzy framework begins with 22 indicators $x_1,\dots,x_{22}$, including Number of wins, Average winning time, Serve score, ACE number, Unforced errors, Net success rate, and Average run distance (Li et al., 25 Mar 2025). Another study defines ten evaluation indicators $X_1$ to $X_{10}$: Serve Advantage, ACE Incidence, Unforced Errors, Scoring Advantage Winning Points, Running Distance, Winning Dishes and Sets, Return Depth, Serve Depth, Receiving Speed, and Forehand Incidence (Liu et al., 2024). HydraNet instead uses 32 shared features grouped into Serve, Return, Psychology, and Fatigue, with player-specific variables such as $p1\_ace$, $p2\_return\_depth$, $p1\_set\_diff$, and $p2\_distance\_run$ (2505.21882).

A statistical branch of the literature treats outcomes themselves as the primary data object. In generalized linear mixed effect models, the response is binary at the set, game, or point level, and momentum is encoded by lagged outcomes and their interactions with serve/return context (Goyal et al., 2020). This line of work is especially important because it frames momentum as a carryover structure rather than an engineered scalar score.

3. Core model families

The TMM literature is methodologically heterogeneous. It does not converge on a single canonical algorithm; rather, it defines a modeling family.

Probability and score-based formulations

A score-window model defines momentum at round $n$ as

$P(n)=a_1 M(n)+a_2 N(n),$

where $M(n)$ uses the three rounds before and after and $X_1$0 uses the seven rounds before and after (Li, 2024). This formulation adds exponential streak adjustments, with $X_1$1 for the three-round component and $X_1$2 for the seven-round component, where $X_1$3 (Li, 2024). The paper assigns

$X_1$4

while noting that the weighting is chosen reasonably / subjectively rather than fully optimized (Li, 2024).

A probability-based TMM defines momentum as a time-varying scoring advantage built from historical scoring probability, instant scoring probability, and efficiency (Graves et al., 11 Sep 2025). Historical serve probability is

$X_1$5

instant serve probability is computed within the match using empirical Bayes estimation, and long-term momentum interpolates from $X_1$6 to $X_1$7 according to match progress $X_1$8 (Graves et al., 11 Sep 2025). Efficiency is defined as

$X_1$9

so that shorter rallies and aces correspond to higher efficiency (Graves et al., 11 Sep 2025). The final TMM multiplies the long-term probability term by $X_{10}$0 (Graves et al., 11 Sep 2025).

Multi-criteria evaluation models

Several frameworks treat momentum as an evaluation score derived from multi-criteria weighting. An AHP-based unsupervised model constructs a judgment matrix $X_{10}$1, derives weights $X_{10}$2, and checks consistency via

$X_{10}$3

with acceptable ranking when $X_{10}$4 (Li, 2024). That study reports CR = 0.085 (Li, 2024).

A TOPSIS-based framework computes a closeness coefficient

$X_{10}$5

using indicator weights $X_{10}$6 over number of sets / discs, number of games / innings, number of points scored, and serving side (Ma, 2024). The resulting situation flow curve is interpreted as a time series of relative performance advantage.

An EWM-GRA system first derives entropy weights and then combines them with Gray Relation Analysis to construct a performance score

$X_{10}$7

using a sliding window of size 10 (Liu et al., 2024). Momentum is then defined as the trend of the EMA-smoothed performance score, with smoothing factor $X_{10}$8 and sliding period $X_{10}$9 (Liu et al., 2024).

A fuzzy TMM builds a two-tier evaluation system with first-level groups for physical fitness, serving proficiency, winning capability, and overall score, and maps normalized variables into the evaluation set

$p1\_ace$0

with numerical score

$p1\_ace$1

(Li et al., 25 Mar 2025).

Statistical carryover and mixed-effects models

A distinct TMM interpretation models momentum as carryover effects from previous sets, games, or points (Goyal et al., 2020). The generic GLMM is

$p1\_ace$2

with a match-specific random intercept (Goyal et al., 2020). At the game level, lagged game outcomes, service indicators, and their interactions encode context-dependent momentum. At the point level, the model includes lagged point outcomes up to three lags, lagged service indicators, pairwise interactions among lagged points, and a three-way interaction among the three lagged point outcomes (Goyal et al., 2020). In this formulation, TMM is not a handcrafted score but a structured dependency model.

Machine-learning and deep-learning systems

Data-driven TMMs use ensemble learners and sequence models. One framework combines SVM, Random Forest, and XGBoost with Bayesian optimization, PCA-based compression, SHAP, CUSUM, and Monte Carlo simulation (Peng et al., 2024). Another uses HMM to infer hidden momentum states, applies EMA to smooth the signal, validates the significance of the latent variable with XGBoost, and predicts momentum swings using LightGBM with SHAP interpretation (Lei et al., 2024). A further system uses a sliding-window performance score, defines strategic and psychological momentum, predicts outcomes with Lasso-Ridge-based XGBoost, and models game fluctuation through Deep_LSTM and the derivative of winning rate (Zhai et al., 2024). A 2025 framework uses a BP neural network optimized by PSO, with momentum metric $p1\_ace$3, change-point labels $p1\_ace$4, and shift intensity $p1\_ace$5 as predictive inputs (Du et al., 1 Sep 2025).

At the current technical frontier, HydraNet defines a learned Momentum Score (MS) through

$p1\_ace$6

combining point momentum, explicit momentum, implicit momentum, adversarial representation learning, and multi-granularity classification at the point, game, set, and match levels (2505.21882).

4. Temporal structure, streaks, and latent dynamics

A broad consensus in the literature is that momentum is temporally structured and nonstationary. One probability-based TMM explicitly frames tennis as non-iid, arguing that point outcomes are not independent and identically distributed because earlier points influence later points through psychological, tactical, and performance effects (Graves et al., 11 Sep 2025). HMM-based work operationalizes this by treating momentum as a hidden state evolving under the Markov assumption, with

$p1\_ace$7

and observation independence

$p1\_ace$8

(Lei et al., 2024).

Streak structure is a recurrent modeling device. A contingency-table framework defines $p1\_ace$9 and $p2\_return\_depth$0 as winning and losing streaks of length $p2\_return\_depth$1, then tests whether extension probabilities depend on streak length (Du et al., 1 Sep 2025). For 2023 Wimbledon Men’s Singles, the paper reports

$p2\_return\_depth$2

which it interprets as strong evidence against independence (Du et al., 1 Sep 2025). In the same study, the estimated probabilities $p2\_return\_depth$3 are not constant across streak lengths, supporting nonlinear streak dynamics (Du et al., 1 Sep 2025).

Other models encode streaks directly in the score function. Psychological momentum has been quantified through aces, unforced errors, double faults, and winning or losing streaks, with the Fibonacci sequence used to magnify the effect of longer consecutive point runs (Zhai et al., 2024). Dynamic scoring models use exponential amplifiers $p2\_return\_depth$4 and $p2\_return\_depth$5 to increase the effect of consecutive scoring within short and long windows (Li, 2024).

The multi-granularity perspective emphasizes that momentum is not a single-scale phenomenon. HydraNet distinguishes point momentum, historical momentum, explicit momentum, and implicit momentum, and reports different predictive regimes at the point, game, set, and match levels (2505.21882). A plausible implication is that TMM should be viewed less as one scalar process than as a hierarchy of coupled temporal processes.

5. Validation strategies and empirical findings

The literature devotes substantial attention to the question of whether momentum is statistically meaningful rather than random. Validation strategies include correlation analysis, mixed-effects inference, nonparametric hypothesis testing, predictive uplift, and change-point analysis.

A summary of reported findings is useful because the empirical evidence is distributed across distinct methodological traditions.

Paper	Validation mode	Reported result
(Li, 2024)	Correlation / surface fitting	$p2\_return\_depth$6 for a poly22 model relating momentum-related variables and winning rate
(Ma, 2024)	Logistic regression accuracy	81.6% overall accuracy for Player 1; 82.3% for Player 2
(Lei et al., 2024)	Predictive comparison	Proposed momentum model: 81.30% accuracy, 81.27% AUC; random momentum baseline: 54.29% accuracy, 54.28% AUC
(Liu et al., 2024)	Hypothesis testing	Mann–Whitney $p2\_return\_depth$7; K-S $p2\_return\_depth$8
(Du et al., 1 Sep 2025)	Point prediction with momentum features	Best AUC 0.7443 for $p2\_return\_depth$9
(2505.21882)	Multi-granularity classification	WID point-level AUC 0.9919, match-level AUC 0.9511

One strand of evidence comes from predictive uplift. In the HMM-based framework, treating momentum as a Gaussian stochastic process yields 54.29% accuracy and 54.28% AUC, whereas using the HMM+EMA momentum variable yields 81.30% accuracy and 81.27% AUC (Lei et al., 2024). The paper interprets this as evidence that momentum is not merely random noise. Likewise, a BP+PSO system reports that adding $p1\_set\_diff$0, $p1\_set\_diff$1, and $p1\_set\_diff$2 sequentially improves AUC from 0.7125 to 0.7443 (Du et al., 1 Sep 2025).

Another strand comes from formal hypothesis testing. The EWM-GRA study generates 1,000 random datasets as a no-momentum null sample and compares them with the observed momentum sample. It reports a Mann–Whitney U-test $p1\_set\_diff$3-value of 0.0043 and a Kolmogorov–Smirnov $p1\_set\_diff$4-value of 0.00128, rejecting the null of no significant difference (Liu et al., 2024). A separate machine-learning fusion model applies the run test and reports, for match 2023-wimbledon-1312, $p1\_set\_diff$5 for p1_momentum and $p1\_set\_diff$6 for p2_momentum, while turning-point series are not significantly non-random (Peng et al., 2024).

The mixed-effects literature provides a more cautious but statistically rigorous interpretation. Using Grand Slam singles matches from 2014–2019, one study finds strong evidence of carryover effects at the set, game, and point levels after accounting for ranking, serve percentage, return percentage, service status, and match-specific random effects (Goyal et al., 2020). It reports that winning the previous set increases the odds of winning the next set by about 56% to 109% for men and about 43% to 187% for women (Goyal et al., 2020). At the point level, winning the previous two or three points in a row is associated with the highest estimated odds of winning the next point (Goyal et al., 2020). This suggests that some empirical regularities attributed to “momentum” remain after observable controls, while still stopping short of a definitive causal psychological interpretation.

A final empirical theme is multi-granularity asymmetry. HydraNet reports strong point-level and match-level performance but weaker set-level performance: for Wimbledon (WID), point AUC 0.9919, game AUC 0.8130, set AUC 0.6749, and match AUC 0.9511 (2505.21882). The paper interprets this as evidence that momentum impacts outcomes differently at different granularities, with psychology becoming more important at coarser scales and fatigue emerging as significant at the match level (2505.21882).

6. Interpretation, applications, and limitations

TMM research is consistently motivated by tactical and analytical applications. Several papers argue that momentum trajectories can inform coaching interventions, especially in relation to serve and receive skills, mental toughness, break-point management, net play, and fatigue-aware tactics (Li, 2024, Lei et al., 2024, Liu et al., 2024). A probability-based TMM is explicitly proposed as a real-time analytical tool for adjusting serve strategy, rally strategy, or shot selection as the match progresses (Graves et al., 11 Sep 2025). HMM and LightGBM analyses similarly translate feature rankings into recommendations to improve performance at the net, convert break points, strengthen ace ability, and monitor elapsed-time effects (Lei et al., 2024).

The interpretability of TMM depends strongly on model class. Multi-criteria and fuzzy systems offer directly interpretable weights, indicator hierarchies, and momentum categories (Li, 2024, Li et al., 25 Mar 2025). Statistical carryover models offer coefficient-level interpretability and clearer confounding control, but they do not produce a visually intuitive scalar momentum curve by default (Goyal et al., 2020). Deep systems such as HydraNet offer richer temporal expressivity and explicit adversarial interaction modeling, but implementation details can be more opaque and some formulas are reported as notationally messy (2505.21882).

Several recurring limitations prevent a single definitive TMM formulation from emerging.

Heterogeneous definitions: Momentum is variously treated as scoring-window advantage, psychological force, hidden performance state, carryover dependency, or learned latent representation (Li, 2024, Lei et al., 2024, Goyal et al., 2020, 2505.21882). This complicates direct comparison across papers.

Potential confounding: Some studies explicitly ignore psychological quality, physical condition, venue effects, playing style, or tactical exploitation of opponent weaknesses (Ma, 2024). Others attempt to control for player quality and serve/return ability, but acknowledge that fatigue and matchup effects cannot be fully isolated (Goyal et al., 2020).

Data dependence: Many models are developed on Wimbledon 2023 or a small number of case-study matches, although some recent systems use broader Grand Slam datasets (Li, 2024, Graves et al., 11 Sep 2025, 2505.21882). This suggests that apparent agreement on momentum may partly reflect shared empirical settings.

Model instability and reporting inconsistencies: Some papers report extreme logistic coefficients, inconsistent formulas, or performance numbers that vary across sections (Ma, 2024, Peng et al., 2024, Zhai et al., 2024, Li et al., 25 Mar 2025). This suggests caution in treating all reported equations as implementation-ready without code or supplementary clarification.

Granularity mismatch: Momentum that is predictive at the point level may not transfer cleanly to game or set prediction, and vice versa (Goyal et al., 2020, 2505.21882). A plausible implication is that TMM is better conceived as a collection of scale-specific models with partial cross-scale coupling than as a single universal metric.

In its strongest form, the TMM literature supports three propositions. First, tennis momentum can be operationalized in multiple mathematically explicit ways (Li, 2024, Graves et al., 11 Sep 2025, Li et al., 25 Mar 2025). Second, many of those operationalizations show non-random association with winning probability, point outcomes, or match-flow reversals (Ma, 2024, Liu et al., 2024, Du et al., 1 Sep 2025). Third, the most technically advanced work increasingly treats momentum as a multi-granular, opponent-coupled, dynamically inferred state rather than as a simple count of recent successes (Lei et al., 2024, 2505.21882).