luxury-lakehouse/line-breaking-passes

Line-Breaking Passes

Every pass from ~5 million professional soccer events, annotated with whether each pass breaks through opponent defensive lines. Built from StatsBomb Open Data and Wyscout event data, with line-breaking labels derived from StatsBomb 360 freeze-frame and Metrica tracking data. The is_line_breaking flag enables binary classification models that learn to predict line-breaking potential from pass context alone.

Part of the (Right! Luxury!) Lakehouse soccer analytics platform.

Quick Start

from datasets import load_dataset

ds = load_dataset("luxury-lakehouse/line-breaking-passes")
df = ds["train"].to_pandas()

# Filter to confirmed line-breaking passes
lb_passes = df[df["is_line_breaking"] == True]
print(f"{len(lb_passes)} line-breaking passes out of {len(df)} total")

Explore interactively: HF Space demo

What is a Line-Breaking Pass?

A line-breaking pass is a pass that travels through a compact defensive shape, bypassing one or more lines of defenders. They are high-value attacking actions because they disrupt defensive organization and create goal-scoring opportunities in the space behind the defensive block.

Line-breaking detection is computed using Ward hierarchical clustering on opponent positions at the moment of the pass (sourced from StatsBomb 360 freeze frames, Metrica tracking data, and IDSSE Bundesliga tracking data), followed by a cross-product straddle test to determine whether the pass vector intersects the cluster boundary. Three defensive lines (low-block, mid-block, high-press) are identified per frame; a pass may break through one, two, or all three.

The dataset includes all passes — both line-breaking and non-line-breaking — to provide balanced positive and negative examples for machine learning. The is_line_breaking column is the binary target label.

Data Fields

Column	Type	Description
pass_id	string	Unique pass identifier (surrogate key)
match_id	bigint	Match identifier
player_id	int	Player identifier
team_id	int	Team identifier
pass_recipient_id	int	Player who received the pass (NULL for Wyscout and incomplete passes)
competition_id	int	Competition identifier
season_id	int	Season identifier
period	bigint	Match period (1–5)
minute	bigint	Match minute
second	bigint	Match second
start_x	double	Starting X coordinate (StatsBomb 120-yard scale)
start_y	double	Starting Y coordinate (StatsBomb 80-yard scale)
end_x	double	Ending X coordinate
end_y	double	Ending Y coordinate
pass_type	string	Pass type classification
pass_height	string	Pass height (Ground Pass, Low Pass, High Pass)
body_part	string	Body part used
pass_length	double	Pass length in yards
pass_angle_radians	double	Pass angle in radians
pass_outcome	string	Pass result (Complete, Incomplete, Out, etc.)
is_cross	boolean	Whether the pass was a cross
is_switch	boolean	Whether the pass was a switch of play
is_through_ball	boolean	Whether the pass was a through ball
is_complete	boolean	Whether the pass was completed
is_progressive	boolean	Whether the pass moved the ball significantly closer to goal
pass_direction	string	Pass direction (forward, backward, lateral)
is_line_breaking	boolean	Whether the pass breaks through opponent defensive lines
lines_broken	int	Number of defensive lines broken (0–3)
line_breaking_type	string	Type of line break (through, around)
data_source	string	Data provider (statsbomb, wyscout)

Coordinate System

All coordinates use the StatsBomb standard: 120×80 yards. The origin (0, 0) is at the bottom-left corner from the perspective of the attacking team. The x-axis runs along the length of the pitch (0–120 yards), and the y-axis runs along the width (0–80 yards).

Detection Method

Line-breaking detection runs as a distributed compute pipeline via applyInPandas on Databricks, grouped by match_id. For each pass in a match:

1. Cluster opponent positions using Ward hierarchical clustering on the Y-axis to identify defensive lines.
2. Straddle test: compute the cross product of the pass vector against each cluster boundary segment. A sign change indicates the pass crosses the line.
3. Label assignment: is_line_breaking = true if the pass straddles at least one defensive line cluster; lines_broken counts how many.

This method does not require labeled training data — it is a deterministic geometric rule over freeze-frame positions.

Data Sources

Source	Matches	License
StatsBomb Open Data	~3,000	CC-BY 4.0
Wyscout Public Dataset	~1,900	CC-BY-NC 4.0

Coverage includes the Premier League, La Liga, Serie A, Bundesliga, Ligue 1, Champions League, World Cup, and more.

Limitations

• Sparse line-breaking labels: Line-breaking detection requires opponent positional data (StatsBomb 360 freeze frames, Metrica tracking, or IDSSE tracking data). For matches without this data — the majority of Wyscout matches and StatsBomb matches without 360 — is_line_breaking, lines_broken, and line_breaking_type are NULL. These passes are still included in the dataset as unlabeled examples.
• StatsBomb 360 coverage: Approximately 323 StatsBomb matches include 360 freeze-frame data. Line-breaking labels are densest for these matches. IDSSE Bundesliga tracking provides an additional 7 matches with full positional data.
• Geometric approximation: The Ward clustering + straddle test is a rule-based approximation. It does not model defensive intent, pressing triggers, or tactical context beyond raw position.
• Coordinate system: All coordinates are in StatsBomb 120×80-yard scale. Wyscout passes are normalized to this scale; minor rounding differences may exist near pitch boundaries.
• Open data only: Commercial datasets with richer tracking coverage may yield different label distributions and model generalizations.

Citation

There is no formal paper for this dataset. If you use it, please cite the repository:

@misc{luxury_lakehouse_line_breaking,
  title={Line-Breaking Passes: A Soccer Pass Dataset with Defensive Line Labels},
  author={Nielsen, Karsten Skytt},
  year={2026},
  url={https://github.com/karsten-s-nielsen/luxury-lakehouse},
  note={Data sourced from StatsBomb Open Data (CC-BY 4.0) and Wyscout Public Dataset (CC-BY-NC 4.0)}
}

More Information

Explore interactively: HF Space demo

• License: MIT