PuckVision Player Index (PVPI)
Technical Model Overview
Overview
Data-Driven Player Evaluation
The PuckVision Player Index (PVPI) is a data-driven player evaluation system that aims to distill raw NHL player production into a clear and accessible performance score. It assigns a rating on a 60–100 scale, enabling consistent comparisons across player roles, positions, and game situations.
The core goal of PVPI is to translate box score statistics and selected advanced metrics into an easy-to-consume numerical value that reflects a player's contribution to team success.
Core Model Structure
How PVPI Works
Game-Level Scoring
At its foundation, PVPI calculates a game-level score using a weighted formula based on event-level stats (goals, assists, expected goals, takeaways, giveaways, blocked shots, etc.).
Situational Weighting
Each game situation—5v5, 5v4, 4v5, and Other—is evaluated with distinct weights to reflect its strategic importance and frequency. Scoring frameworks are position-specific to capture role differences between forwards and defensemen.
Key Adjustments
Defensive Adjustment Rating (DAR)
A contextual bonus or penalty based on a player's expected goals against per 60 minutes.
Time on Ice Adjustment
High-minute players receive bonuses to reflect coach trust and physical toll.
Rating Conversion and Tiering
Performance Classification
Once a player's aggregate game score is computed, it is mapped to a 60–100 rating scale. This rating is then translated into one of several tiers. The tier boundaries are based on both natural gaps in the rating distribution and positional lineup logic.
Forward Tiers
Defenseman Tiers
Goalie Tiers
Tier sizes are constrained based on how many players can realistically fill those roles across 32 NHL teams (e.g., 32 number one defensemen, 96 top-6 forwards, etc.).
Feature Weight Derivation
Scientific Methodology
Logistic Regression Approach
The weights for each statistic (e.g., goals, xG, takeaways) were not selected arbitrarily. Instead, we began by building a game prediction model using logistic regression. The inputs were the PVPI scores of all players in each game. We then iteratively refined the statistical weights to maximize predictive accuracy.
This reverse-engineering approach ensures that each individual statistic's weight in PVPI contributes meaningfully to predicting team success.
Model Validation
Proven Performance
Predictive Performance
A simple logistic regression model using PVPI scores for all players in a game predicted the winning team with 65.3% accuracy on the 2024–25 test set. This performance significantly exceeds the 50% baseline of random chance, showing that PVPI captures real predictive signal related to on-ice outcomes.
Award Correlation (2023–24)
Hart Trophy (MVP)
4 of top 5 PVPI forwards:
Norris Trophy (Top Defenseman)
4 of top 5 PVPI defensemen:
Vezina Trophy (Top Goalie)
4 of top 5 PVPI goalies:
Strengths of PVPI
Transparent
Each stat has a visible, auditable weight.
Modular
Breakdown by situation and skill category (offense, defense, transition).
Award-Aligned
Closely matches voting results.
Predictive
Directly tied to game outcomes.
Normalized
Works across positions and team styles.
Limitations
Box Score-Based
Does not track intangible or unquantifiable impacts such as:
- Net-front presence
- Physical intimidation
- Leadership or communication
Example: Marcus Foligno (MIN) may appear as a bottom-line contributor, but provides veteran leadership and grit that aren't reflected.
Context-Free
Does not adjust for teammates or quality of competition. Linemate and deployment context are not currently incorporated.
Model Drift
Needs recalibration each season to reflect changes in league scoring rates and playing style.
Use Cases
Contract Modeling
Market value analysis and contract projections based on player performance.
Team Strength
Team strength modeling and win probability forecasts.
Trade Simulations
Trade simulations and armchair GM tools.
Scouting Enhancement
Player tier assignments and scouting enhancement.
Final Notes
PVPI is not designed to replace traditional scouting but to supplement it with rigorous, data-driven context. It transforms raw statistical outputs into intuitive ratings aligned with actual team value and league outcomes.
Through award validation and prediction modeling, PVPI has demonstrated strong correlation with real-world success. It is a cornerstone of modern hockey analytics in PuckVision's broader evaluation and decision-making framework.