Forecasting the Winner of a Live Tennis Match
cs.LG
Quick verdict
Trace hybrid (Markov+HGBM) hits 77.84% all-point accuracy on 2023-24 Grand Slams; solid chronological design, but no significance tests for the narrow gaps.
Summary
The paper builds live win-probability models for tennis that update as a match unfolds, using 8,222 Grand Slam matches (ATP and WTA, about 1.5 million point-level states) from Jeff Sackmann's public point-by-point data, split chronologically (2011-2021 training, 2022 validation, 2023-2024 test). Five models are compared: a symmetric Markov score-recursion baseline, an Elo-asymmetric Markov variant, a serve-shrink Markov model applying Bayesian shrinkage to in-match serve rates, a histogram gradient-boosting (HGBM) baseline, and Trace, which feeds the two Markov probabilities, their logits and difference, plus live features into an HGBM. Trace achieves the best accuracy at all checkpoints (76.06%/82.15%/88.34% at 25/50/75% progress; 77.84% across all points) and the lowest log loss at every checkpoint (e.g., 0.2002 at 75%). The authors honestly benchmark against DeepTennis under a leakage-prone non-chronological split, where Trace reaches 81.30% across all points. Main limitations: no significance tests or confidence intervals for the narrow margins (Trace leads serve-shrink by only 0.16 percentage points across all points), no betting-odds baseline despite Kovalchik's finding that markets are strongest, ambiguous checkpoint definitions, only 96% Elo coverage, and calibration assessed only visually for Trace.
Review complete
Assessment at a glance
5
risk flags identified
● high 0
● medium 2
● low 3
4 strengths · 5 weaknesses · 5 author questions
AI text signal = Binoculars zero-shot detector score over 20 sampled passages (Fast-DetectGPT co-reported) — absence does not prove human authorship, and a flag does not imply misconduct.
Levels reflect verification complexity — not a quality score. Evolve never rates your paper.
Key claims under review
- Trace, a hybrid stacking two Markov recursions and live features into an HGBM, is the most accurate live win-probability model among five evaluated, with 76.06%/82.15%/88.34% accuracy at 25%/50%/75% match progress and 77.84% across all points on the 2023-2024 test set.
- Live serve-rate updating (serve-shrink) mainly improves the quality of probability estimates (log loss) rather than flipping predicted winners (accuracy), relative to a purely Elo-based Markov model.
- A strictly chronological split is essential: prior DeepTennis results (79.5%) came from a leakage-prone non-chronological split, and under that same setup Trace reaches 81.30% across all points.
- Pure machine learning on live features (HGBM) is weak early in matches but strongest among individual models late in matches, motivating the hybrid design.
Strengths
Strict chronological evaluation
The 2011-2021/2022/2023-2024 split (§2.3) prevents temporal leakage, and the authors go further: they diagnose DeepTennis's leakage-prone 2014 test year and re-run all five models under that same non-chronological split (Table 6, Trace 81.30% all-points) to make the comparison meaningful.
Clean ablation ladder of information sources
The five-model design (§3) isolates score structure (symmetric Markov), player strength (Elo-asymmetric), in-match serve updating (serve-shrink, κ grid 40-640), raw ML (HGBM), and their combination (Trace, Eq. 17), so each gain in Tables 4-5 is attributable to a specific information source.
Dual metrics plus calibration
Reproducibility posture
Weaknesses
No uncertainty quantification
Missing betting-odds baseline
Ambiguous checkpoint construction
Underspecified data handling
Visual-only calibration
Full Review Report
$1.99 · one-timeUnlock the complete preliminary analysis Evolve produced for this paper.
Report preview
Scope & Claims
The paper targets live (in-play) win-probability forecasting for Grand Slam tennis, motivated by live betting (Introduction cites ~80% of tennis wagering as in-play). It claims that a hybrid stack - the two Markov recursion outputs plus live features fed into a histogram gradient-boosting model (Trace) - outperforms both pure Markov and pure ML approaches, and that hybrid modeling is 'a practical approach to live tennis forecasting.' Headline numbers (abstract, Table 4): 76.06%/82.15%/88.34% accuracy at 25%/50%/75% match progress and 77.84% across all points, with the lowest log loss at every checkpoint (Table 5: 0.4753/0.3530/0.2002).
The remaining sections cover Methodology Assessment, Statistical Evidence, Results & Reproducibility, Limitations & Risks, and Verdict & Recommendations.
- All 4 strengths & 5 weaknesses, analyzed in detail
- 5 risk flags with severity, section & reviewer questions
- Methods & statistical evidence assessment
- 5 questions a peer reviewer would ask
- Full six-dimension evaluation report & recommendations
One-time purchase · this paper only
One-time purchase, this paper only. AI-generated preliminary analysis — not peer review.
Frequently asked questions
What accuracy does the Trace model achieve in this paper?
On the 2023-2024 Grand Slam test set, Trace reached 76.06%, 82.15%, and 88.34% accuracy at 25%, 50%, and 75% match progress respectively, and 77.84% across all point-level predictions, with the lowest log loss at every checkpoint (e.g., 0.2002 at 75% progress).
What data does the paper use?
The study uses Jeff Sackmann's public Grand Slam point-by-point data plus ATP/WTA match results: 8,222 matches and about 1.5 million point-level states, split chronologically (2011-2021 training, 2022 validation, 2023-2024 testing).
Is Trace better than DeepTennis?
Not directly comparable: DeepTennis reported 79.5% overall accuracy on a leakage-prone non-chronological split (test year 2014 within 2011-2019 data). Under that same split, this paper's Trace reached 81.30% across all points, suggesting competitiveness, but the evaluation setups differ.
Is the code for this tennis forecasting study available?
Yes. The paper links to a public GitHub repository (github.com/cx-57/live-tennis-research), and all underlying data come from Jeff Sackmann's public datasets, so the pipeline can be re-run without proprietary inputs.
About this review
This is an AI-generated preliminary review by Evolve.Pub, produced from the openly available text of the paper. It is not peer review and does not represent the authors' views. The canonical record of the paper is at the publisher link above.
- Figures 1-4 are judged from captions and surrounding text only; the plotted curve values themselves are images not recoverable from the text extraction. The body text and all tables/numbers appear complete through reference [23].
Go deeper on this paper
Wrote a paper? Stress-test it before reviewers do.
Upload your manuscript for the same structured scrutiny — reasoning, statistics, evidence, figures, and reproducibility.