Contextual Football Scouting (Wyscout Integration)

Ranking players by what they do with the space around them, not by how much they touch the ball.

Matteo Vezzoli & Armando Mio — 2026

---

ℹ️ Fork Notice: This repository is a dedicated fork of Contextual Football Scouting.
The complete theoretical framework, mathematical proofs, academic paper, and original tracking-data methodology (StatsBomb 360) are documented in the main repository:
🔗 Corresponding README in the original repository: Contextual-Football-Scouting/README.md

---

Changes in this Fork

This fork adapts the contextual scouting methodology to work with Wyscout seasonal aggregated data (webapp/backend/data/wyscout_sample.csv):

1. Wyscout Statistical Proxies:

   • H1 — Space Control & Value: Replaced 360° convex hull geometry with seasonal proxies for Progression (Progressive passes/90, Progressive runs/90, Passes to final 3rd/90), Reception (Touches in box/90, Received passes/90), Gravity (Fouls suffered/90, Offensive duels/90), and Dangerousness (xG/90, xA/90).
   • H2 — Decision Quality: Replaced event-level freeze-frame xPass/xEPV models with a Decision Quality Index (DQI) evaluated across Possession Accuracy, Duel Robustness, Risk Readings /90 (Key passes/90 + Smart passes/90), and Risk Accuracy (Accurate progressive passes, %).
   • H3 (Off-Ball Movement): Excluded, as off-ball runs and receiver resolution require spatial tracking coordinates not present in seasonal aggregated data.
   • H4 — Player Similarity: Replaced the 11-axis tracking DNA with an 8-axis orthogonal style DNA combining H1 and H2, computing within-role Euclidean distance look-alikes across 1,059 eligible players.

2. 100% In-Memory Full-Stack Webapp:

   • Replaced PostgreSQL / Supabase and offline DB migration scripts with a lightweight in-memory FastAPI backend (webapp/backend/data_module.py) that loads webapp/backend/data/wyscout_sample.csv directly on startup.
   • Sub-5ms response times across all endpoints (/players/, /similar/, /roles/, and /space-control/).
   • Modern React 19 + Vite frontend with responsive radars, multi-criteria filtering, and head-to-head similarity overlays.

3. Expanded Player Pool:

   • Expanded the evaluation pool from 272 players (StatsBomb 360 minutes threshold) to 1,697 total players (1,059 outfield players with $\ge 300$ minutes) from the Wyscout seasonal sample dataset.

---

Original Project: ArMat-Analytics/Contextual-Football-Scouting