Screenshot of the interactive per-team graph (Morocco). The interactive graph (
output/grafo_tactico.html) is deployable via GitHub Pages.
EN: Network graph analysis on real football passing data to reveal tactical patterns that traditional stats miss. Uses StatsBomb Open Data (event-level passes with actual player names and locations) to construct directed graphs and calculate centrality metrics.
ES: Análisis de grafos de red sobre datos reales de pases de fútbol para revelar patrones tácticos que las estadísticas tradicionales no muestran. Usa datos abiertos de StatsBomb (pases a nivel de evento con nombres reales de jugadores y ubicaciones) para construir grafos dirigidos y calcular métricas de centralidad.
StatsBomb Open Data (mandatory attribution), Free, event-level football data. The data is made available by StatsBomb under the StatsBomb Open Data License, which requires that any published analysis or derivative work credits StatsBomb as the data source. Please cite StatsBomb when reusing or referencing this dataset.
- Real pass events with player names, locations, and outcomes
- World Cup 2022 match: Canada vs Morocco (match_id 3857276, 2022-12-01)
- Only completed passes are kept as edges (failed passes are excluded), one directed graph per team
P1 - Single Point of Failure: Is there a player whose removal would collapse the team's possession network? Betweenness centrality identifies these structural vulnerabilities.
P2 - Traffic monopolies: Does one player concentrate a disproportionate share of transitional traffic? Opponents can target this player to disrupt buildup.
P3 - Failed connections: Are there missing links between midfielders and forwards that traditional stats don't show? The graph reveals where tactical communication breaks down.
src/graph_builder.py builds a nx.DiGraph from StatsBomb passing events:
- Nodes = players (real names from match data)
- Directed edges = actual pass direction between players
- Edge weight = accumulated pass frequency
| Metric | What it measures |
|---|---|
| Betweenness Centrality | Who acts as indispensable bridge (the SPOF) |
| PageRank | Who is most connected in the passing network |
| In/Out Degree | Ball reception vs distribution volume |
Centrality variants, which is which:
- Betweenness unweighted (
top_betweenness): every pass hop counts as 1. Answers "who sits on the most minimum-hop paths", this is the one used for node size in the interactive graph and for the headline rankings. - Betweenness weighted (
top_betweenness_weighted, distance = 1/weight): frequent pass lanes become short edges, so shortest paths prefer high-frequency connections. Answers "who bridges the frequent-pass cores". NetworkX treatsweightas distance, hence the inversion. - PageRank / degree: use pass frequency as weight (frequency = importance).
The pipeline has no synthetic fallback: without data/raw/statsbomb_passes.csv it stops with a warning and builds nothing, so no centrality can come from fabricated edges. The old passing.csv FBref/United fallback was removed.
PyVis renders the graph as an interactive HTML file (output/grafo_tactico.html) with force-directed layout. Node size reflects betweenness centrality.
Each figure below is recomputed from the regenerated per-team outputs (sources in parentheses).
Morocco (flourish_arc_diagram_morocco_network_metrics.json, flourish_arc_diagram_morocco.csv)
- 16 players, 100 aggregated passing edges from 310 completed passes, 3 communities
- Sofyan Amrabat has the highest betweenness (0.1100, unweighted), key connector; Achraf Hakimi Mouh leads PageRank (0.1244)
Canada (flourish_arc_diagram_canada_network_metrics.json, flourish_arc_diagram_canada.csv)
- 16 players, 140 aggregated passing edges from 480 completed passes, 3 communities
- Tajon Buchanan has the highest betweenness (0.0740, unweighted); Kamal Miller leads PageRank (0.1380)
Match totals (data/raw/statsbomb_passes.csv)
- 869 pass events → 790 completed (79 failed excluded) → 240 aggregated edges across both teams
- Completed passes by length: 140 short (<10 m), 449 medium (10–25 m), 201 long (≥25 m), medium/short dominate possession phases, long balls appear in transitions
Mide (un partido: Canada vs Morocco, StatsBomb 3857276):
- Estructura de la red de pases por equipo: nodos = jugadores, aristas = pases completados agregados, peso = frecuencia acumulada (790 completados → 240 aristas).
- Centralidad de betweenness (sin peso y con distancia 1/peso), PageRank y grado: quién es puente indispensable y quién monopoliza la distribución.
- Comunidades de jugadores detectadas por Louvain (3 por equipo).
No mide:
- No estima efectos tácticos causales: es la descripción de la red de un solo partido (n=1 por equipo; sin comparación entre partidos ni controles).
- No incluye xG/xT, pases esperados ni eventos que no sean pases: solo eventos de pase de StatsBomb Open Data.
- No predice resultados ni proyecta rendimiento futuro: las centralidades son descriptivas de ese partido.
git clone https://github.com/alvarosalinaso/tactical-narrative-graph-analysis
cd tactical-narrative-graph-analysis
python -m venv .venv
source .venv/bin/activate # Linux/Mac
pip install -r requirements.txt
python src/graph_builder.pyOpen output/grafo_tactico.html in your browser.
networkx>=3.0
pyvis>=0.3
pandas>=2.0
statsbombpy>=1.0
- Manchester United Analysis, Descriptive before/after analysis of managerial changes
- Chilean Video Games, Market analysis with scraping
- World Cup 2026, Interactive dashboard