The goal of this project is to apply data-based analysis to an unsolved and difficult problem, which I am intimately familiar with: evaluating ice hockey. Due to the reasons described below, I believe hockey analytics models require more feature engineering and domain knowledge than data analysis in many other sports. As someone who has played hockey my entire life, and paid attention to the team development aspects of the NHL, I feel I have acquired domain knowledge that I am excited to apply to this problem. Through this project, I hope to learn lessons about data visualization, analysis methods and techniques along the way, as well as revealing some interesting, and possibly unexpected, insights about the game I love.
Hockey is unique among popular professional sports due to a combination of several factors:
- There are relatively few stoppages or clearly demarcated events that can be used as data points
- Games are decided by relatively low scores (randomness significantly affects outcomes)
- The domain of legal player actions is huge
- Precise and complex manipulation of the playing object (puck)
- Few restrictions in allowable actions
- Continuous substitution of players during game play
- Players are constantly moving at relatively high speeds
Although these characteristics also apply to other sports, no other sport incorporates all of them. For example, both baseball and American football have continuous play of only a few seconds on average, whereas the time between whistles in hockey is normally on the order of minutes. Football (soccer) has nearly continuous play, but the rules and equipment limit the domain of possible actions players can take. I know relatively little about cricket, but I believe its format and restrictions are similar to baseball, at least regarding the characteristics listed.
Furthermore, compounding effects strongly impact team performance in hockey. For example, the way player skill sets complement each other has a significant effect on the impact of a line (the skaters of one team on the ice together), but analyzing this is confounded by the substitution of players every 30-40 seconds, including during game play. The number of player combinations on the ice together, and the frequency with which this changes, is greater than in any other sport. Assigning credit is also difficult in hockey. Events rarely come to fruition solely through the actions of a single player. In most cases, four to six players contribute to any one event. Even defining and separating events is challenging, due to the degree of flow in game play. The combination of these factors make hockey a uniquely challenging sport in which to gather useful data and to evaluate analytically.
The repo includes an environment.yml file, which may be used to create a conda environment containing the packages requiured for the project to run. Additionally, the repo includes a setup.py file which allows the project to be installed as a package.
To create the conda environment from the provided environment.yml file, simply run the follwing command from the directory in which the file is located. For further information, refer to the conda documentation regarding managing environments.
conda env create -f environment.yml
NOTE: The following has not yet been tested, as I am currently working on the project in editable mode (see below).
To install the project as a package in your current environment, activate the environment, cd to the Hockey-Analytics directory where the setup.py file is located, and then run the following command:
python setup.py install
NOTE: This is currently how I am running the project on my computer. Installing as described above will be tested later, although to my understanding that should already work. However, in that case, the package must be reinstalled each time a change is made to the project.
Alternatively, if you wish to contributue to or make changes to the project, you may install the package in development mode using the editable option, which is enabled by the setuptools library.
pip install --editable .
nhl_api (package): series of modules containing functions used to retrieve, parse, and reformat the data from NHL.com's API
api_scripts (directory): contains scripts used to pull game data from NHL.com's API. The data is retrieved in JSON format and parsed to python dictionaries before being stored as .csv files. Also contains a series of scripts used to clean and reformat the data, readying it to be stored in a database.
nhl_database (package): series of modules containing functions used to create a postgresql database in a Docker container, then populate it with the reformatted data from the NHL API
analyses (directory): collection of notebooks in which exploratory data analyses (EDAs) are performed on data retrieved from the database using SQL query functions defined in the nhl_database package
models (package): series of modules containing functions used to define the models which are the primary product of this project
TODO: model_output (directory): collection of scripts and notebooks in which the models are trained and tested
TODO: include commands required to pull and format data
Once the data is reformatted and stored as .csv files in the directory Hockery-Analytics/data/, run these commands from Hockey-Analytics/nhl_database/
docker compose up -d
python3 db_create_sqla_core.pyWhere the username, password and database name postgres environment variables are retrieved from a .env file. The username and password are important for connecting to the database through a database tool such as DBeaver. Since the database is hosted locally, and there is nothing sensitive on it, I have used a generic username and password combination.
The database schema is shown below. The box headings are table names, while the entries listed below are the primary keys and their datatypes.
These jupyter notebooks illustrate patterns and trends in the data, to identify interesting features and guide the model design process. Descriptions of particular analyses will be added here as they are completed.
TODO: Add description of the analysis
Note: Models have round ends, input features are in rectangles
flowchart BT;
A([Expected Goals]) --> B([Player Impact]);
C([In-Game Win Probability]) --> B;
B --> D([Team Success]);
X[shot data] --> A;
Y[game state, team strengths] --> C;
Z[player attributes, statistics] --> B;
In-Game Win Probability: Evaluates the probability a team will win a particular game given the current game state (score, time, etc...). The first version of the model is simply based on historical outcomes. Later, team strengths will be included as inputs to improve performance. Used to determine whether individual players normally impact the game during important moments, or contribute to scoring opportunities when the outcome of the game is no longer in question.
Expected goals (xG): Evaluates the offensive contributions of a player or team by aggregating the number and effectiveness of scoring opportunities they create. Shots are approximately 15 times more common than goals (i.e. "scoring events"). Therefore, using all shots to measure player or team contributions provides many more data points to train the model. Expected goals are used to help determine the impact of a player on a game, both offensively (xG created) and defensively (xG allowed), or the strength/style of a team.
Whether players or teams consistently outperform their expected goal totals is an indication of the individual player's or team's talent. Consistent differences between xG and actual goals also indicates that the model is not capturing all important information, and thus can be improved. Similarly, understanding the characteristics of which teams and players consistently outperform their xG totals may lead to a better understanding of what makes an individual player or team successful. Expected goal models are commonly used throughout hockey analytics. Therefore, I can compare the effectiveness of my model against other existing models, and adopt the publicly available xG statistics, if necessary.
Player impact: Evaluates the overall impact of a player on an individual game. This takes into account the player's average scoring/defensive contributions, as well as their tendency to contribute at important times. It is an open question in the analytics community what player attributes most strongly influence team success (i.e. winning games). Therefore, the player impact model is the key cog in this pipeline, and will likely evolve the most throughout the modelling process.
Team success: Evaluates the probability that a team will win individual games. This may be used to predict success across seasons by sequentially predicting all the games in a season and taking the cumulative result. On the other hand, a second model may be used to predict team success over longer time frames, based on team attributes. Both models will depend on the aggregation of individual player contributions, as well as how the mix of player attributes affects the overall team effectiveness. The cumulative model will also have a temporal aspect, accounting for recent success (momentum), travel, etc...