Stathletes - NESSIS 2025

​

1 Data Generation Process

1 – Dataset Overview

2 – Tracking Demo

3 – AI Models

4 – From Detections to Reliable Tracking Data

5 – Puck & Player Coordinate Data Sample

2 Feature engineering & metric derivation

1 – Signal Smoothing & Motion Calculations

2 – Metrics Overview: Speed & Motion

3 – Tools for Measuring Defensive Structure

4 – Gap Closure Metrics

5 – Roaming Metrics

6 – Beyond Voronoi: Region Control

7 – Visualizing Region Control

8 – Pass Completion % Model

9 – Visualizing Pass Completion % Model

10 – Action Value Models

11 – Optimal Action Value (OAV_t)

12 – Metrics Overview: Defensive Structure

3 Application: defensive positioning

1 – Modeling Defensive Positioning

2 – Modeling Quality vs Quantity

3 – Modeling 5v5 vs 5v4

3 – OAV_t Snapshots

4 – OAV_t Animations

4 Application: Zone Entries

1 – Modeling Zone Entries

2 – Visualizing Zone Entries

3 – Zone Entry Animations

Title

The Impact of Skating Speed and Style with Tracking Data

​

Authors

Jeff Goeree and Joe Gratz, Stathletes

​

Abstract

Skating speed and style are fundamental drivers of hockey performance, shaping both defensive resilience and offensive effectiveness. This study leverages advanced player and puck tracking data to investigate how skating attributes influence game outcomes across defensive and offensive phases.�Our dataset comprises 289 American Hockey League (AHL) games from the 2024–25 season (October 2024–April 2025). Each game contains approximately 85,000–105,000 puck positions and 800,000–1,000,000 player positions, enabling fine-grained spatial and temporal analysis. Importantly, the tracking data is aligned with event-level outcomes (e.g., shots, goals, passes, takeaways), allowing us to assess the consequences of skating behaviors on scoring opportunities.�First, we examine defensive skating styles, ranging from aggressive, high-pressure systems to passive containment and their impact on shot suppression and expected goals against (xGA) at both even strength (5v5) and during penalty kill (5v4). We assess whether aggressive defensive pressure reduces shot volume while concentrating risk into fewer but higher-quality chances and explore how these tradeoffs vary across game states.�Second, we investigate how player skating characteristics, including acceleration, top-end speed, and edge control, influence offensive zone entry success. We test whether superior skaters generate more clean entries, odd-man rushes, and rush-based expected goals, and whether these advantages translate into faster zone establishment on the power play.�This study highlights the indispensable role of tracking data in advancing hockey analytics and tactical understanding.

1 - Data Generation Process

AHL games (2024–25 season)

All data was collected using only:

• Broadcast Video
• Team Rosters

1 - Dataset Overview

~85k-105k

~800-1m

280

Puck positions per game�

Player positions per game

1 - Data Generation Process

2 - Tracking Demo

1 - Data Generation Process

2 - Tracking Demo

1 - Data Generation Process

2 - Tracking Demo

1 - Data Generation Process

3 - AI Models

1. Camera shot selection (remove commercials, stoppages, crowd shots)

2. Player detection (YOLOv12)�

3. Puck detection (YOLOv5)�

4. Player tracking (BoT-SORT)

5. Jersey number detection and recognition (CLIP4STR)�

6. Game clock OCR (Parseq)�

Rink Feature segmentation & Homography (Several Models)

1 - Data Generation Process

4 - From detections to reliable tracking data

• Post-processing phase to improve data after models are run
• Smooth and improve player tracking and puck tracking
• Interpolation for player and puck tracking gaps

​

• Raw detections are on a per-frame basis and can be noisy (missed frames, false positives, jitter).

​

​

• Kalman Filter integrates detections

over time → stable trajectories.

​

1 - Data Generation Process

5 - Puck and Player Coordinate Data Sample

2 - Feature engineering, metric derivation

1 - Signal smoothing & motion calculations

2 - Feature engineering, metric derivation

2 - Metrics Overview: Speed and Motion

Defining a “Speed Burst”

Summary of Metrics

• Distance (D)
• Ice Time (TOI)
• Energy (E) - Calories
• Power (P) - Cals/min
• Speed Bursts – sec,ft/sec

2 - Feature engineering, metric derivation

3 - Tools for measuring defensive structure

Region control (Voronoi)

• Visualize control regions for players
• Most likely to reach points in their zone first
• Useful partition for creating player metrics

Convex hull for defensive units

• Smallest convex polygon containing all defenders
• Expresses the structure of a defensive unit
• 3-5 players may form the boundary

2 - Feature engineering, metric derivation

4 - Gap Closure Metrics

We define a Gap to be the shortest distance from any defending player to the puck’s location:

We define Gap Closure as rate of change in Gap (closing or widening):

2 - Feature engineering, metric derivation

5 - Roaming Metrics

Roaming = 4.6ft

Stretch = 59%

Roaming = 5.2ft

Stretch = 8%

We define roaming (1) and Stretch (2) scores to defenders for a particular interval, based on variance of positions around the Center of Position (CoP):

(1) roaming tells us how defenders pursue attackers vs defending specific areas

(2) Stretch tells us how defenders are patrolling back & forth in specific lanes

2 - Feature engineering, metric derivation

6 - beyond Voronoi: Region Control

2 - Feature engineering, metric derivation

7 - visualizing region control

2 - Feature engineering, metric derivation

8 - pass completion % Model

Shallow neural network model trained on a fusion of Stathletes event data and puck/player tracking data

Pass completion % per x,y coordinate based on the current puck’s location, attacking/defending players, pass/carry/shooting lanes, and game state

2 - Feature engineering, metric derivation

9 - Visualizing pass completion % Model

2 - Feature engineering, metric derivation

10 - Action value models

Expected Possession Value xPV% per action based on the current puck’s location, attacking/defending players, pass/carry/shooting lanes, and game state

Shallow neural network model trained on a fusion of Stathletes event data and puck/player tracking data

2 - Feature engineering, metric derivation

11 - Optimal Action value (OAVt)

Possession as a Markov Decision Process (MDP):

S = State (current state)

A = Actions (to take from current state)

P = Transition Probability (to new states)

R = Rewards (of an action in a state)

2 - Feature engineering, metric derivation

12 - Metrics Overview: Defensive Structure

3 - Application: Defensive positioning

1 - Modeling Defensive Positioning

We train a gradient boosting model on Offensive zone possession Windows to explore the relationships between defensive attributes and scoring chance rates (shots), and scoring chance quality (xG per shot)

Interpreting Results

• Optimal Action Value carries the most weight, being a comprehensive metric

​

• Condensed defensive units (low hull area) are strong indicators of quality chances against

​

• High stretch Scores (defending in lanes or ellipses) are strong indicators of quality chances against

​

• Region control remains important (quality and area) regardless of optimal action value

​

• Calories burned reduces xg rate, while distance increases it – suggesting it may be optimal to either pursue aggressively or conserve energy, at the right moments, compared to passively chasing the puck carrier

3 - Application: Defensive positioning

2 - Modeling quality vs quantity

Feature Importance for Shot Quality (xG) at 5on5

Feature Importance for Shot Quantity (Shots) at 5on5

Comparing Shot Quality (left) vs Quantity (right)

Interpreting Results

• Gap is significantly more important in reducing shot quantity or volume, but its effect on quality is limited
• Region control is significant for suppressing shot quality but not quantity
• Distance and Calories are both less significant for shot quantity

3 - Application: Defensive positioning

3 - Modeling 5v5 vs 5v4

Feature Importance for Shot Quality (xG) at 5on5

Feature Importance for Shot Quality (xG) at 5on4

Comparing Shot Quality 5on5 (left) vs Shot Quality 5on4 (right)

Interpreting Results

• OAV (Optimal Action Value) is the dominant feature on the PK or 5v4
• implies factors like passing lanes and positioning are the driving factors in suppressing chances against
• Gap Closure sees its highest impact on the PK
• Stretch score also sees its highest impact on the PK, suggesting less elliptical formations may be optimal

3 - Application: Defensive positioning

4 - OAVt Snapshots

3 - Application: Defensive positioning

4 - OAVt Snapshots

4 - Application: Zone entries

1 - Modeling zone entries

Feature Importance for Rush Shot Quality (xG) at 5on5 (left) vs 5v4 (right)

We train a gradient boosting model on Zone entry windows to explore the relationships between defensive attributes and scoring chance rates (shots), and scoring chance quality (xG per shot)

Interpreting Results

• We see an increased importance in both Gap and Gap Closure on Zone Entries
• Stretched defenses imply odd-man rushes or bad positioning and have a strong impact at both 5v5 and 5v4
• Calories Burned sees its highest score yet at 5v4 for rush chances, showing the impact of speed upon entry

4 - Application: Zone entries

2 - visualizing zone entries

Note: Entry isn’t clean but due to high speed, low entry angle, and poor gap closure the entry is enough to create chaos in the defensive zone

Entry speed: 30 ft/sec

Entry Angle: 15 deg from center

Entry Gap: 6ft

Entry Gap Closure: -7ft/sec

​

-> Entry leads to a chance with >80% xG

4 - Application: Zone entries

2 - visualizing zone entries