Advanced Soccer Analytics, Open-Sourced

you don't have to reinvent the wheel!

Joris Bekkers

Contents

• Introduction
• Career Path


• Open-Source
• PySport
• Kloppy
• unravelsports
• mplsoccer
• etc.


• Research & Innovation
• Pressing Intensity
• Graph Neural Networks

Introduction

• Joris Bekkers
• 35 y/o
• Breda, Netherlands


• Football Analytics Consultant
• Research
• Development
• Deployment


• PySport Boardmember
• Unite Sports Analytics Community
• Grow and Maintain Open-Source

Open-Source Soccer Analytics!

Why?

• Learn out in the open
• Learn from your mistakes
• Collaborate and connect with industry pro's
• Share knowledge
• Don't reinvent the wheel!

Open Source Resources

• kloppy
• floodlight
• socceraction / soccerdata / ETSY
• databallpy
• OpenSTARLab
• UFMG Sports Analytics Lab
• KAIST
• unravelsports
• mplsoccer
• Soccer Analytics Handbook
• penaltyblog
• Friends of Tracking

What is kloppy?

Load, Filter, Transform, Export Standardized Soccer Data.

Supported Providers

Provider	Event	Tracking	Public Data
Hawkeye (2D)
Metrica
PFF
SecondSpectrum	PR #437
Signality
SkillCorner
Sportec
StatsPerform / Opta
Tracab
DataFactory
StatsBomb
WyScout

Implemented Not yet implemented

How does it work?

Kloppy basics using Open DFL Dataset

Open DFL Dataset (Bassek et al. 2025)

Documentation (under construction)

Load Event Data

from kloppy import sportec

dataset = sportec.load_open_event_data(
    match_id="J03WN1", 
    coordinates="secondspectrum"
)

home_team, away_team = dataset.metadata.teams
print(f"{home_team.name} v {away_team.name}")

>>> 'VfL Bochum 1848 - Bayer 04 Leverkusen'

Meta data

Within the metadata we can find a lot of information, for example the player_id of each player.

home_team, away_team = dataset.metadata.teams
for p in away_team.players:
    print(f"#{p.jersey_no}.", p, " - ", p.player_id)

>>> '30. Jeremie Frimpong  -  DFL-OBJ-002GOI'
>>> '27. Florian Wirtz  -  DFL-OBJ-002GBK'
>>> '23. A. Hložek  -  DFL-OBJ-J01G0J'

Florian Wirtz' Events

player_id = "DFL-OBJ-002GBK"

player_events = dataset.filter(
    lambda event: (
        event.player.player_id == player_id if event.player is not None else False
    )
)

>>> <EventDataset record_count=72>

mplsoccer!

from mplsoccer.pitch import Pitch
pitch = Pitch(pitch_color='grass', line_color='white', stripe=True)
fig, ax = pitch.draw()

Documentation

from mplsoccer import Pitch


def heatmap(xs, ys):
    pitch = Pitch(
        line_zorder=2,
        pitch_type="secondspectrum",
        pitch_length=dataset.metadata.pitch_dimensions.pitch_length,
        pitch_width=dataset.metadata.pitch_dimensions.pitch_width,
    )
    fig, ax = pitch.draw()
    ax.set_title(f"#{player.jersey_no} - {player.last_name} - {player.team.name}")
    pitch.kdeplot(xs, ys, ax=ax, cmap="YlOrRd", fill=True, levels=100)


transformed_events = player_events.transform(
    to_orientation="ACTION_EXECUTING_TEAM"
)

xs = [event.coordinates.x for event in transformed_events if event.coordinates is not None]
ys = [event.coordinates.y for event in transformed_events if event.coordinates is not None]

heatmap(xs, ys)

Advanced Soccer Analytics

Pressing Intensity

Advanced Soccer Analytics

Pressing Intensity

Pressing Intensity

Requirements

Use positional tracking data Multipurpose / Extendable No ML / prediction model Not too complex Not too simple Measurable at every moment Summable Measure pressing in the whole "system”

Pressing Intensity

Requirements

Use positional tracking data Multipurpose / Extendable No ML / prediction model Not too complex Not too simple Measurable at every moment Summable Measure pressing in the whole "system”

Total Pressure = 1 - [(1 – 86%) * (1 – 19%)] = 89%

What is the chance 🏃‍♂️^defense will reach 🏃‍♀️^attack or ⚽️ within
some amount of time given their current direction and speed?"

UnravelSports Blog

More Open-Source!

pip install unravelsports

make sure you have Python 3.11 and unravelsports 0.3.0+

The unravelsports package aims to aid researchers, analysts and enthusiasts by providing intermediary steps in the complex process of converting raw sports data into meaningful information and actionable insights.

• Polars DataFrame Conversion
• Graph Neural Network Training, Conversion and Prediction
  [Bekkers & Sahasrabudhe (2023)]
• Pressing Intensity
  [Bekkers (2024)]

1. Load Kloppy & Convert

from kloppy import sportec

from unravel.soccer import KloppyPolarsDataset

kloppy_dataset = sportec.load_open_tracking_data()

kloppy_polars_dataset = KloppyPolarsDataset(
    kloppy_dataset=kloppy_dataset
)

KloppyPolarsDataset

2. Pressing Intensity!

from unravel.soccer import PressingIntensity

import polars as pl

model = PressingIntensity(
    dataset=kloppy_polars_dataset
)
model.fit(
    start_time=pl.duration(minutes=1, seconds=53),
    end_time=pl.duration(minutes=2, seconds=32),
    period_id=1,
    method="teams",
    ball_method="max",
    orient="home_away",
    speed_threshold=2.0,
) 

>>> PressingIntensity(n_frames=143)

Model Output in Polars

model.output

Polars Documentation

Probability to Intercept

>>> [x for x in model.output.to_pandas().iloc[0]["probability_to_intercept"]].round(2)

[[0.07 0.   0.   0.   0.   0.03 0.01 0.   0.   0.   0.23]
 [0.28 0.   0.   0.   0.   0.02 0.   0.   0.   0.   0.  ]
 [0.   0.   0.   0.   0.   0.   0.   0.   0.   0.   0.  ]
 [0.09 0.   0.   0.   0.   0.1  0.06 0.   0.   0.   0.05]
 [0.   0.   0.   0.   0.   0.   0.   0.   0.   0.   0.  ]
 [0.   0.   0.   0.   0.   0.09 0.6  0.   0.   0.   0.  ]
 [0.   0.   0.   0.   0.   0.83 0.08 0.   0.   0.   0.  ]
 [0.02 0.   0.   0.   0.   0.02 0.   0.   0.   0.01 0.  ]
 [0.   0.   0.   0.   0.   0.   0.   0.   0.   0.   0.  ]
 [0.   0.   0.   0.   0.   0.   0.   0.   0.   0.   0.71]
 [0.3  0.   0.   0.   0.   0.06 0.   0.   0.   0.   0.03]]

3. Video

Full Example

Advanced Soccer Analytics

Graph Neural Networks

Graph Neural Network

Counter Attack Predictions

GNN Architecture

• Quick Start Guide
• From Kloppy to GNN with Polars
• BigDataBowl Example
• What are GNNs (FAQ)

PFF Open World Cup Tracking + Event Data

• 64 Full Matches of World Cup 2022 Tracking + Event Data

0. Install "beta" version of kloppy and unravelsports

pip install git+https://github.com/PySport/kloppy.git

pip install git+https://github.com/UnravelSports/unravelsports.git

1. Load Kloppy ...

from kloppy import pff

match_id = "10510"

dataset = pff.load_tracking(
    raw_data=f"{match_id}.jsonl.bz2",
    meta_data=f"{match_id}.json",
    roster_meta_data=f"{match_id}.json",
)
home_team, away_team = dataset.metadata.teams
print(f"{home_team.name} v {away_team.name}")

>>> Croatia v Brazil

1. ... and convert to Polars DataFrame

from unravel.soccer import KloppyPolarsDataset

kloppy_polars_dataset = KloppyPolarsDataset(
    kloppy_dataset=dataset
)

2. Add Labels

kloppy_polars_dataset.data = (
    kloppy_polars_dataset.data
    .join(
        some_label_dataframe.select(
          ["game_id", "period_id", "frame_id", "label"]
        ), 
        on=["game_id", "period_id", "frame_id"],
        how="left"
    )

Where to get Labels?

Currently unreleased

from kloppy import pff

match_id = "10510"

event_dataset = pff.load_event()

>>> Croatia v Brazil

Kloppy's Pattern Matching Functionality

Pattern Matching

3. Convert to Graphs

from unravel.soccer import SoccerGraphConverterPolars

kloppy_polars_dataset.add_dummy_labels()
kloppy_polars_dataset.add_graph_ids(by=["frame_id"])

converter = SoccerGraphConverterPolars(dataset=kloppy_polars_dataset)

For each frame...

For each frame...

4. Custom Graph Dataset

from unravel.utils import CustomSpektralDataset

graphs = CustomSpektralDataset(graphs=converter.to_spektral_graphs())

>>> CustomSpektralDataset(n_graphs=100_000)

Under the hood (Graph)

graphs[0]

>>> Graph(n_nodes=23, n_node_features=15, n_edge_features=6, n_labels=1)

• 23 nodes
  • (22 players + ball)
• 15 node features
  • (e.g. location, speed, direction)
• 6 edge features
  • (e.g. distance between 2 players, speed difference between 2 players etc.)
• 1 label
  • (0 or 1)

Under the hood (Node Feature Matrix)

graphs[0].e

>>> 
array([[0.62, 0.37, 0.05, 0.55, 1.  , 0.33, 0.12, 1.  , 0.  , 0.  , 0.61, 0.99, 0.79, 0.1 , 0.  ],
       [0.5 , 0.2 , 0.07, 0.92, 0.78, 0.45, 0.16, 1.  , 0.  , 0.  , 0.68, 0.97, 1.  , 0.52, 0.  ],
       [0.5 , 0.49, 0.01, 0.81, 0.9 , 0.42, 0.01, 1.  , 0.  , 0.  , 0.5 , 1.  , 0.5 , 1.  , 1.  ],
       [0.57, 0.38, 0.03, 0.91, 0.78, 0.36, 0.09, 1.  , 0.  , 0.  , 0.59, 0.99, 0.87, 0.17, 0.  ],
       [0.75, 0.42, 0.03, 0.67, 0.97, 0.21, 0.21, 1.  , 1.  , 0.  , 0.6 , 0.99, 0.59, 0.01, 0.  ],
       [0.32, 0.25, 0.02, 0.54, 1.  , 0.59, 0.2 , 0.1 , 0.  , 0.  , 0.61, 0.99, 0.82, 0.88, 0.  ],
       [0.37, 0.45, 0.  , 0.81, 0.9 , 0.53, 0.12, 0.1 , 0.  , 0.  , 0.53, 1.  , 0.61, 0.99, 0.  ],
       ...
       [0.03, 0.5 , 0.13, 0.73, 0.94, 0.82, 0.4 , 0.1 , 1.  , 0.  , 0.5 , 1.  , 0.5 , 1.  , 0.  ],
       [0.59, 0.53, 0.01, 0.89, 0.81, 0.34, 0.07, 1.  , 0.  , 0.  , 0.48, 1.  , 0.38, 0.01, 0.  ],
       [0.34, 0.6 , 0.04, 0.53, 1.  , 0.56, 0.15, 0.1 , 0.  , 0.  , 0.45, 1.  , 0.32, 0.97, 0.  ],
       [0.39, 0.22, 0.02, 0.71, 0.95, 0.53, 0.18, 0.1 , 0.  , 0.  , 0.64, 0.98, 0.92, 0.77, 0.  ],
       [0.51, 0.49, 0.15, 0.73, 0.95, 0.41, 0.  , 0.1 , 0.  , 1.  , 0.5 , 1.  , 0.5 , 1.  , 0.  ]])

• A row for each "object" (player or ball)

Under the hood (Edge Feature Matrix)

graphs[0].e

>>> 
array([[ 0.  ,  0.  ,  1.  ,  0.5 ,  0.5 ,  1.  ],
       [ 0.14,  0.01,  0.87,  0.84,  0.63,  0.98],
       [ 0.12, -0.  ,  0.91,  0.22,  0.66,  0.97],
       ...,
       [ 0.15, -0.  ,  0.97,  0.32,  0.02,  0.62],
       [ 0.18, -0.  ,  0.77,  0.92,  0.3 ,  0.96],
       [ 0.  ,  0.  ,  1.  ,  0.5 ,  0.5 ,  1.  ]])

• A row for each "connection"

Under the hood (Adjacency Matrix)

graphs[0].a

>>> <Compressed Sparse Row sparse matrix of dtype 'int64'
	with 287 stored elements and shape (23, 23)>

• Describes connections

5. Draw the rest of the Owl!

split_test_train_validation model = CrystalGraphClassifier() model.fit() model.evaluate() model.predict()

Full Example

6. Combine Open-Source!

Contributing?

• Learn how to use git / GitHub
• Open Issues
• Reach out to package owners / maintainers / contributors
• Create a Pull Request with a good idea
• Build your own open-source tool(s)

• This stuff isn't easy
• At least you don't need to reinvent the wheel
• Don't be afraid to make mistakes
• Try to break things
• Try to build things that others have not built
• Embrace the continuous learning process

• Licenses

• This presentation is made with Marp

Thanks!