Where Am I?: An Introduction to ML Draft05.12.2019 : meta, update
Happy Mothers' Day!
The goal of this post is to introduce visitors to the project, and hopefully answer a few questions people may have.
The high-level goal of the project is simply to use statistical analysis and machine learning to predict a prospect's NHL career using data that's available prior to the NHL draft.
To drill down on this idea, here's the basic idea this site uses to predict the pro career of a prospect heading into the draft:
- Gather pre-draft statistics for a bunch of draft picks from previous seasons. More information on data collection can be found here: Disclaimer, Data Sources, Content Licensing
- Gather NHL statistics for those players. For players that are still active, this requires some projection or extrapolation to estimate their career statistics based on their to-date data. Each player's NHL career is then 'labeled' using some metric of interest, for example, NHL points per game.
- For each prospect eligible for the current NHL draft, find players from previous draft years who have similar pre-draft statistics or attributes, and use these players to predict a prospect's NHL career in terms of the label metric noted above. This is performed using mostly standard machine learning techniques.
- For each player from the 'old' drafts, use all the pre-draft data except that player's data to predict that player's NHL career. This prediction can then be checked against reality. This is the basic idea for cross-validation data, which is provided with most releases (e.g., Cross-validation data for defensemen), and case be used to evaluate and understand how the model is performing.
This project is an ongoing experiment, and so changes and/or differences in predictions between releases may seem awkward and non-sensical. It's a slowly-converging process. I try to do one release or update when midterm CSS rankings come out, and one when final CSS rankings are made available. I also do releases when the model or code has changed significantly.
I'm always interested in feedback and questions, so please feel free to reach out on twitter.