Expected by Whom? A Skill-Adjusted Expected Goals Model for NHL Shooters and Goaltenders

J.T.P. Noel · 2025

Read Expected by Whom? A Skill-Adjusted Expected Goals Model for NHL Shooters and Goaltenders by J.T.P. Noel (2025) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2511-07703.txt.

Introduction

J. Noel builds an expected goals model for ice hockey that credits skill, not just shot location. Standard models rate a scoring chance by where the shot came from. Shooters and goaltenders, though, differ sharply, and the same chance means different things in different hands.

A light gradient boosted model uses league tracking data with features for overall, locational, and situational skill for both shooter and goalie. Shot binning and similarity distances shape those features. Separate datasets by skill pairing feed baseline comparisons. Coaches and analysts gain chance ratings that respect who shoots and who guards the net.

Catalog record

CREATORJ.T.P. Noel
YEAR2025
COLLECTIONPAPERS
FORMATTXT
FILE SIZE34 KB
VERSION1.0
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2511-07703

Browse the full library · About the Tower of Babel

Full text

The opening of the work, from the library's hosted text:

Expected by Whom? A Skill-Adjusted Expected Goals Model for NHL Shooters and Goaltenders J.T.P. Noel Abstract This study outlines a light gradient boosted model aimed at predicting shot outcomes in the NHL. The model uses the NHL’s spatiotemporal data to account for both the skill of shooters and goaltenders. This approach involves isolating and engineering features for different aspects of shooter and goaltender skill. These aspects include the overall skill, the locational skill, which is engineered using a shot binning technique previously outlined by Shuckers and Curro, and the situational skill, which is engineered using Gower distance. Three separate datasets were created based on the skill of the shooter and goaltender. For each, a baseline model was created in order to compare and contrast its performance with the skill-adjusted model. The results seen in this study show performance increases for the skill-adjusted model over the baseline model in log loss, brier scores, and area under the ROC curve. These performance increases have a high of 5% and outperform previous works, which have attempted to account only for player skill. This highlights the importance of accounting for both player and goaltender skill, while also accounting for different aspects of their skill. In future works, a skill-adjusted expected goals model could benefit models interested in predicting other aspects of the game, such as scoring leaders or individual game outcomes. Keywords: Expected Goals, Machine Learning, NHL, Sports Analytics 1 Introduction In recent years, analytics has changed the way teams, fans and players approach and understand the game of ice hockey [ 16 ] . This new age of analytics has ushered in several new performance indicators that can improve team evaluation, player evaluation, and the understanding of what ultimately leads to a team winning or losing. One of the more popular performance indicators to come from this era is expected goals (xG) [ 11 ] . Expected goals is a measurement of shot quality that typically uses spatiotemporal data and machine learning (ML) to assign a likelihood that a given shot attempt will result in a goal. Hockey’s introduction to expected goals is often attributed to Alan Ryder, who in 2004 proposed a method to account for shot quality in the NHL. Ryder’s work opens with a phrase that has become synonymous with expected goals “Not all shots on goal are created equal” [ 18 ] . This phrase perfectly encapsulates the concept behind expected goals, rather than looking at how many shots a given team or player has taken, it is more important to look at the quality of those shots. Since Ryder’s paper, expected goals have evolved and there have been several NHL expected goals models put forth which typically rely on machine learning to value a given shot. However, it should be noted that many of these models exist outside the academic sphere [ 24 , 25 , 14 ] . Expected goals models are not limited to only the NHL and hockey, as the metric is arguably more popular in European football/Soccer. Thus, there are a host of expected goals models for the sport [ 8 , 17 , 10 ] . While expected goals have been accepted as a good indicator of future goal-scoring in the NHL [ 24 ] , the performance indicator often struggles to account for players who are either skilled or unskilled shooters. There have been several proposed adjustments to account for shooter skill such as the model developed by Dawson Sprigings and Asmae Toumi [ 24 ] or the model created by Harry Shomer [ 22 ] . Both of these approaches resulted in no gains in model performance or very minimal gains. While both of these models provided valuable insight into the problem of shooter skill in the NHL, they both did not account for the skill of the goaltender. They also fail to explore any subsection of skill (i.e. a player may be more skilled when shooting from a certain location or in a certain situation). This is an area of study that this research aims to further the understanding of. The approach outlined in this paper accounts for the skill of both the shooter and goaltender, while also accounting for three separate subsections of skill, these subsets being overall skill, locational skill, and situational skill. The assumption here is that by accounting for different subsections of both shooter and goaltender skill, a skill-adjusted expected goals model will have more predictive power than a baseline model that does not account for skill. The models in this paper are engineered using an architecture similar to Shomer’s [ 22 ] . In this architecture, two expected goals models are stacked such that the second model uses the xG values obtained in the first model to generate features. However, it differs from the aforementioned approach as rather than only using the last two years to account for talent, all available 5v5 shots from 2010-2022 are used to judge a shooter’s or goaltender’s skill, these shots are then linearly weighted in order of occurrence to better exemplify the player’s skill over their entire career. Another distinguishing factor is that this approach creates models for different skill brackets of players and goaltenders. This action was taken as the factors that lead do a goal could be different based on the skill of the individuals involved. The use of this approach should create a clearer picture of how skill is related to shot outcome, how skill-based expected goals models correlate with a shooter or goaltender’s actual performance, and what this means for other types of prediction models in the NHL. This paper should also serve as one of the only comprehensive outlines of an NHL expected goals model in academic literature. 2 Literature Review Early work in the area of expected goals in hockey can be seen in work by Alan Ryder, who proposed a shot quality metric in 2004 [ 18 ] . Ryder’s approach made use of data from the 2002-03 NHL season. Ryder used several different situational factors affecting each shot such as the distance, shot type, and whether the shot was a rebound to determine the likelihood that a given shot would become a goal. At this same time, in the world of football/soccer. Ensum, Pollard, and Taylor used logistic regression to study the factors that affected the likelihood of a shot becoming a goal during the 2002 World Cup [ 5 ] . They found factors that are still used in models today such as distance and angle. The work of Ryder [ 18 ] , as well as Ensum, Pollard, and Taylor [ 5 ] laid the foundations for what would become the modern expected goals model by identifying factors that increased the likelihood of scoring. In 2012, Brian Macdonald presented a paper on expected goals in the NHL at the MIT SLOAN sports analytics conference [ 14 ] . This paper made use of situational factors along with a ridge regression model to create an adjusted plus-minus metric. This metric was then used to measure the positive or negative impact a player had for their team when they were on the ice. For instance, a player would have a positive

Read the complete text