NHL Machine Learning Draft

Predictions for the 2019 draft class have been updated. Here are the predictions for forwards, and here are the predictions for defensemen. As usual, validation data is posted for previous years.

Since the 2019 draft has come and gone, the main goal of the re-prediction is to show progress on the model. Updates since the original predictions for the 2019 NHL draft include:

• The classifier output has been switched from a prediction of a player's number of games played in the NHL and the number of points scored to a single metric: points-per-game. This lends a little less insight to a player's career, but is technically easier to deal with in some ways, and also shows better model performance (see below). It's also a little more resistant to injury shortened careers and other oddities. In sum, I traded away some potential insight for model performance.
• Similar players are now computed. We also compute 'nearest superstar', which conceptually is what the player might look like if they became a top-flight player.
• Predictions for defensemen. The performance (with point-per-game output) was finally good enough to release, although it's still not great. The downside is that points-per-game tells even less of the story for defensemen than it does for forwards.

Specific data about the model performance for a given draft year (e.g.) is found at the bottom of each prediction page, but here's a short summary of how the model is doing:

• Forwards:

  • The model is 'exactly' right about 50% of the time. This means we can guess a forward's career PPG within about .15 ppg. 50% doesn't sound great (and it isn't), but keep in mind we're not flipping a coin.
  • The model is 'not wrong' about 72% of the time. This means we can guess a forward's career PPG within about .3 ppg.

• Defenseman:

  • The model is 'exactly' right about 45% of the time. This means we can guess a defensemen's career PPG within about .2 ppg.
  • The model is 'not wrong' about 65% of the time. This means we can guess a defensemen's career PPG within about .4 ppg.

Note that there's a lot of players we didn't take a guess at, due to missing data or other technical reasons, so the above figures would need some adjusting if we want to consider the entire draft class.

Overall, I'd say the results are decent but obviously there is much room for improvment. Looking at some of the misses, some defy explanation where others are more acceptable (e.g., player is very short or had a terrible junior season). There's a lot of other interesting ways to interpret some of the predictions, maybe I'll post about this in the future.

Since this is a 'spare time' project for me, the next major update will probably be for the 2020 draft class predictions.

Feel free to reach out on twitter.