# Soccer / Football Analytics — arXiv Literature Review

**Corpus:** 555 papers (311 core soccer/football analytics + 244 tangential)
**Span:** 2007 – 2026  |  **Source:** arXiv, retrieved 2026-10-06
**Companion files:** `keywords.json`, `abstracts.json`, `index.json`, `INDEX.txt`

---

## 1. Scope & method

This review covers the arXiv corpus of soccer/football analytics papers. Each paper was
classified as **core** (genuinely about soccer/football) or **tangential** (mentions soccer
in passing but is really about prediction theory, physics, other sports, etc.) using a
keyword-density classifier over title + abstract. The 311 core papers drive the analysis
below; the tangential ones are retained in the folder but excluded from the review.

Methodology was traced by scanning each abstract for technique signatures (linear/Bayesian
stats, CNNs, GNNs, transformers/LLMs, RL, computer vision, XAI) and bucketing by era.
Papers were then clustered into research streams by their extracted keywords.

---

## 2. Publication trend — a near-exponential rise

| Year | All | Core |
|------|----:|-----:|
| 2007–2016 | ~52 | ~33 |
| 2017 | 18 | 11 |
| 2018 | 31 | 18 |
| 2019 | 17 | 7 |
| 2020 | 31 | 20 |
| 2021 | 56 | 31 |
| 2022 | 52 | 27 |
| 2023 | 56 | 31 |
| 2024 | 82 | 52 |
| 2025 | 88 | 44 |
| 2026 | 78 | 42 |

The field was essentially dormant before ~2017 and has grown roughly **15× since**. The
inflection point (2018–2021) coincides with the arrival of deep learning, the proliferation
of event-tracking data (Opta, StatsBomb, Second Spectrum), and the SQuAD-era transformer
wave. 2024–2026 is the current peak, dominated by vision-based tracking, LLM applications,
and RL agents.

---

## 3. Methodology evolution

Technique presence by era (core papers whose abstract mentions the technique):

| Technique | 07–15 | 16–19 | 20–22 | 23–26 |
|-----------|------:|------:|------:|------:|
| Traditional stats / linear / Bayesian | 2 | 5 | 9 | 15 |
| Deep learning (CNN/DNN) | 0 | 3 | 12 | 23 |
| Graph neural networks | 0 | 0 | 0 | 4 |
| Transformers / LLMs | 0 | 0 | 2 | **32** |
| Reinforcement learning | 1 | 0 | 11 | 15 |
| Computer vision / tracking | 1 | 4 | 15 | **41** |
| Explainable AI | 0 | 0 | 4 | 7 |

**Reading the curve:**
- **2007–2015** — classical statistics: Poisson goal models, logistic regression, copulas,
  early random forests. The "analytics" era of the FiveThirtyEight style.
- **2016–2019** — deep learning arrives; first CNNs for tracking and prediction; RL begins.
- **2020–2022** — computer vision becomes mainstream (tracking, detection); RL matures.
- **2023–2026** — **transformers/LLMs explode (2 → 32)** and computer vision peaks (41).
  GNNs appear for the first time (spatio-temporal action detection, passing graphs). XAI
  emerges as a distinct concern as models get less interpretable.

The single biggest methodological shift is the **2023→2026 transformer surge**: LLMs are now
applied to match forecasting, tactical text analysis, and even evaluating other LLMs on
soccer reasoning (e.g. WorldCupArena).

---

## 4. Research streams

Papers cluster into these streams (240/311 core papers fall into ≥1; many span two).

### 4.1 Tactics / formations / play (124) — the largest stream
The dominant area. Covers formation analysis, off-the-ball movement, defensive role
classification, possession structure, passing networks, and network-science of play. Recent
work emphasizes **structured data formats** (e.g. VERSA verified event data) and
**double machine learning** for causal effects of formations. This stream absorbs a lot of
the general "tactics" keyword, so it's broad.

### 4.2 Reinforcement learning / agents (52)
Multi-agent DRL for tactical decision-making, humanoid-robot soccer skills, and hierarchical
RL for vision-guided play. A large sub-cluster is **simulated/robotic soccer** (RoboCup,
humanoids) — real but distinct from broadcast-data analytics.

### 4.3 Player tracking / computer vision (50)
The fastest-growing technical stream. Broadcast-to-minimap game-state reconstruction,
multi-object tracking, ball-free event detection via possession-path inference (PathCRF),
off-ball defensive role detection, and long-term player tracking with graph hierarchies.
This is where the CV peak (41 papers, 2023–26) lives.

### 4.4 Player valuation / scouting / transfers (40)
Market-value prediction with XAI, LightGBM + FIFA data, transfer-decision forecasting, and
future-quality/value development models. Increasingly framed as explainable so scouts can
trust the numbers.

### 4.5 xG / shot & scoring models (32)
Expected-goals modelling from first principles and data; position-adjusted xG; counterfactual
xG ("what if they took the shot?"); and contribution metrics (restricted Shapley value for
player contribution). A mature, well-developed subfield.

### 4.6 Explainability & interpretability (32)
A cross-cutting concern rather than a standalone topic: SHAP-based xG, XAI market-value
models, interpretable prediction. Grows as deep models displace transparent statistics.

### 4.7 Action & event detection (24)
Spatio-temporal action spotting via GNN + 3D CNN, unsupervised player classification,
game-state + action detection, ball-free event inference.

### 4.8 Match outcome prediction (18)
A benchmarked, well-studied problem. From NLP/ML baselines to distributional forecasting
("Forecasting Soccer Matches through Distributions") and generalizable cross-league
frameworks.

### 4.9 Betting / odds / markets (13)
Premier League outcome forecasting, market-calibrated accelerated-failure-time models for
in-play betting, systematic ML review of odds-setting, and LLM evaluation arenas for
betting reasoning.

### 4.10 Injury / load / health (6) — the thinnest stream
Women's-soccer injury risk (SoccerGuard), fatigue assessment from video, head-kinematics and
brain-tissue deformation from heading. Relatively under-explored on arXiv compared to the
other streams — a potential gap.

---

## 5. Cross-cutting themes

1. **Data is the bottleneck, not the model.** Many papers contribute datasets/benchmarks
   (176 papers tagged "dataset/benchmark") — SoccerNet-Tracking, FOOTPASS, VERSA,
   OpenSTARLab. The field is as much about standardising event data as about modelling.

2. **From prediction to explanation.** The arc runs from "who wins?" (2007–2015) to
   "why, and what should we do about it?" (2023–2026). XAI, counterfactuals, and causal
   (double ML) methods reflect clubs wanting actionable insight, not just accuracy.

3. **Vision is closing the loop on tracking.** The frontier is reconstructing full game
   state from broadcast video without dedicated sensors — minimaps, possession paths,
   off-ball roles.

4. **LLMs are the new wildcard.** 32 transformer/LLM papers in 2023–26, applied to
   forecasting, tactical reasoning, and meta-evaluation — an area that barely existed in
   2020.

5. **Robotics is a parallel track.** A meaningful share of RL work is simulated/humanoid
   soccer (RoboCup), which shares methods with broadcast analytics but targets a different
   end goal.

---

## 6. Notable gaps / opportunities

- **Injury & health** is thin (6 core papers) — big commercial interest, little arXiv output.
- **Causal inference** is emerging (double ML) but still rare — most work is predictive, not
  causal.
- **Women's football** is under-represented relative to men's.
- **Cross-league generalisation** is a recurring acknowledged weakness; few robust
  transferable models.
- **Real-time / in-play** modelling is younger than pre-match prediction.

---

## 7. How to use this corpus

- **Browse by topic:** open `INDEX.txt` (human) or `index.json` (machine) — 33 topics.
- **Filter by keywords:** `keywords.json` maps each PDF to a keyword array.
- **Read abstracts:** `abstracts.json` has full abstracts + authors + PDF URLs for all 555.
- **Stream membership:** `_streams.json` lists every core paper under its research stream.

*Generated 2026-10-06 from local metadata; no papers were re-downloaded for this review.*