Intransitive player dominance, where player A beats B, B beats C, but C beats A, is common in competitive tennis. Yet, there are few known attempts to incorporate it within forecasting methods. We address this problem with a graph neural network approach that explicitly models these intransitive relationships through temporal directed graphs, with players as nodes and their historical match outcomes as directed edges. Our model (65.7% accuracy, 0.214 Brier score) forecasts competitively with established rating systems such as Weighted Elo. Although it does not improve on the baseline in unconditional accuracy, a forecast-encompassing test shows that it carries complementary information. A combined forecast significantly outperforms Weighted Elo, and there is some indication that the gain is larger on the intransitive matchups our model targets. A graph-based representation of player interactions thus captures a forecasting signal that transitive rating systems discard, even between players who share no common opponents.
Figures & tables
Symbol
Description
Indices
s
Target court surface for which the graph is built (clay, grass, or hard)
sk
Court surface of historical match k
i,n
Snapshot indices: arbitrary ( i ), prediction target ( n )
k
Historical match index
u,v
Player indices
Table 1: Mathematical notation, grouped by role.
Figure 1: Overview of the graph construction process.
Figure 2: The walk-forward validation process. After each pass, predictions are generated and the model is retrained; the historical graph grows as matches are added. For illustration, the schematic is not to scale: in practice, the historical graph and sliding training window span several hundred snapshots.
Parameter
Search Space
Optimal Value
Model Architecture
K , Chebyshev filter order
1, 2, 3
1
Hidden units
16, 32, 64, 128
16
Number of layers
1, 2
2
Use activation function
True, False
False
Directional weighting q
0.10 – 0.25
0.24
Table 2: Search spaces and selected values for the graph-construction and architecture hyperparameters.
Figure 3: Brier scores from 250 parameter optimisation trials, with Pareto-optimal trials highlighted. Colour indicates receptive field size ( K×L hops).
Variant
Δ Brier
Remove a component
Dominance graph (entire)
+0.0350
Edge-age decay
+0.0118
Multi-hop receptive field ( L=2 )
+0.0095
In/out-degree node features
+0.0058
Cross-surface edges
+0.0043
Table 3: Component ablation of the tuned configuration on the validation set: pooled Brier-score change for each variant.
Model
Elo
WElo
BT
PS
Gender
Surface
Count
Acc
Brier
Acc
Brier
Acc
Brier
Acc
Brier
Acc
Brier
Men
Clay
1710
0.667
0.209
0.643
0.219
0.653
0.215
0.637
0.217
0.708
0.190
Men
Grass
741
0.676
0.205
0.680
0.205
0.687
0.202
0.649
0.212
0.714
0.185
Men
Hard
3604
0.656
0.214
0.659
0.215
0.661
0.214
0.659
0.217
0.683
0.199
Women
Clay
1539
0.663
0.210
0.668
0.211
0.669
0.209
0.650
0.215
0.700
0.192
Women
Grass
738
0.654
0.218
0.659
0.221
0.664
0.218
0.649
0.223
0.663
0.207
Table 4: Model Performance vs Baselines: Men’s and Women’s Results by Surface.
Figure 4: Calibration curves for the men’s and women’s predictions, showing predicted probabilities versus observed frequencies. Perfect calibration would align with the diagonal y=x line.
Figure 5: An intransitive triangle among Federer, Nadal, and Davydenko. Edge weights are career head-to-head match-win proportions. Each directed edge indicates a consistent head-to-head advantage.
Figure 6: Two intransitive cycles involving Maria Sakkari and Aryna Sabalenka, before their match in the 2021 Abu Dhabi WTA Women’s Tennis Open. Each directed edge indicates a head-to-head dominance advantage.
Figure 7: Distribution of evidence-weighted intransitivity I∗(Auv) across the test set, by surface and gender, restricted to matches with I∗>0 . Mean values: hard (men 2.94, women 2.99), clay (men 2.58, women 2.52), grass (men 1.95, women 2.09). Sample sizes: hard (men n=1,007 , women n=941 ), clay (men n=433 , women n=387 ), grass (men n=171 , women n=165 ).
WElo right
WElo wrong
Total
Model right
7,036
877
7,913
Model wrong
899
3,240
4,139
Total
7,935
4,117
12,052
Table 5: Correctness 2 × 2 between our model and WElo on the test set.
Subset
N
ΔB
η
p -value
No common opponents
745
0.0044
14.6%
0.002
Common opponents, no triad
6,275
0.0017
10.5%
< 0.001
Transitive triads
911
0.0004
8.9%
0.371
Intransitive triads
4,121
0.0012
10.7%
0.028
Table 6: Encompassing gain of the combination model across common-opponent subsets of the test set.
Share (%)
Prior meetings
Pairs
Test matches
0
–
58.2
1
67.4
20.3
2
17.8
9.7
3
7.4
4.9
4
3.4
2.7
Table 7: Distribution of head-to-head meeting counts.
h
Triads
N
ΔB
η
p -value
1
Intransitive
4,121
0.0012
10.7%
0.028
Transitive
911
0.0004
8.9%
0.371
2
Intransitive
1,362
0.0027
20.1%
0.012
Transitive
1,174
0.0009
11.6%
0.233
3
Intransitive
736
0.0031
27.2%
0.034
Transitive
630
0.0026
19.2%
0.055
Table 8: Encompassing gain on intransitive and transitive matches as h (minimum prior meetings per triad edge) increases.
Appendix figures & tables1 asset
Supplementary material from the paper’s appendix.
Appendix
Count
Accuracy
Brier
Dataset
MS
E
MS
E
MS
E
Clay
314
681
0.675
0.649
0.216
0.226
Grass
169
169
0.710
0.710
0.209
0.209
Hard
591
1068
0.638
0.624
0.221
0.228
All
1074
1918
0.660
0.640
0.218
0.226
Appendix
Table 9: Model Performance Comparison By Surface: Extended Dataset.
We present two approaches for predicting tennis match outcomes using topological data analysis and graph theory on ATP singles matches from 2000-2025. The first method applies lower-star filtration to player competitive networks, extracting topological features through persistent homology using four summary methods (VAB, HNAV, HWNAV, OW-HNPV) combined with Modified Band Depth analysis. Algorithmic optimizations including ego graph approximations and triangle elimination enable analysis of about 66k matches. Our Random Forest model achieves 66.2% accuracy (AUC = 0.719) using topological, graph-theoretic, and ranking features. Feature importance analysis reveals that rankings contribute 36.3%, centralities 25.5%, and TDA features 24.0%, with topological features providing complementary signal. When rankings are unavailable, the topology-only model maintains 63.56% accuracy, demonstrating that network-derived features alone capture meaningful competitive structure. The second method uses a modified Katz similarity index with temporal edge weighting, achieving 62.48% accuracy on held-out test data. This work represents the first application of lower-star filtration to tennis prediction, provides systematic comparison of four topological summary methods in sports analytics, and demonstrates that TDA can achieve above-chance prediction using network topology alone while providing additional value when combined with traditional features.
Jake Schwaderer, Alexander Bastien, Omid Khormali +3
Department of Mathematics University of Evansville Evansville, Indiana 47722, USA.
With the rise of live sports betting in recent years, tennis forecasting has expanded from pre-match prediction to models that update win probabilities as a match unfolds. A central challenge in creating such a model is the constant need for models to adapt to score and performance changes. This study examines how pre-match and live information can be most effectively integrated into a model to produce accurate win-probability estimates. The analysis uses 8,222 Grand Slam matches containing a total of 1,505,355 points. Five models were evaluated using a chronological split, with matches from 2011-2021 used for training, 2022 for validation, and 2023-2024 for testing. Trace, a hybrid model, achieved accuracies of 76.06%, 82.15%, and 88.34% at 25%, 50%, and 75% match progress, suggesting that hybrid modeling is a practical approach to live tennis forecasting.
Charles Xie, Aneesh Muppidi
Natick High School 15 West Street, Natick, Massachusetts 01760, USA · University of Oxford Oxford, United Kingdom
Serves, especially first serves, are very important in professional tennis. Servers choose their serve directions strategically to maximize their winning chances while trying to be unpredictable. On the other hand, returners try to predict serve directions to make good returns. The mind game between servers and returners is an important part of decision-making in professional tennis matches. To help understand the players' serve decisions, we have developed a machine learning method for predicting professional tennis players' first serve directions. Through feature engineering, our method achieves an average prediction accuracy of around 49% for male players and 44% for female players. Our analysis provides some evidence that top professional players use a mixed-strategy model in serving decisions and that fatigue might be a factor in choosing serve directions. Our analysis also suggests that contextual information is perhaps more important for returners' anticipatory reactions than previously thought.