Welcome to NHL-ML-Draft
Blog -- 2019 Predictions Release (version 2019.0.0)
2019-05-16 00:00:00 -- release, 2019
The initial version of the predictions for the 2019 draft class are posted here. As noted in the introductory post, only forwards are being predicted at this time.
As usual, 'ex-post-facto' results are also posted. These predictions use the same model that's used for the 2019 data, but the predictions are done on players for which we have an idea of where their careers are headed. Here are links to the data:
The combined miss rate (i.e., incorrect predictions) for the ex-post-facto data is about 37%. This means that the current model seems to able to get at least a cursory read on a player's career about 63% of the time. This is also only considering players we took a shot on. For some players, we admitted up front we didn't have enough data. The results aren't excellent, but they also aren't bad, I'd say. I think things look a little brighter if you consider these predictions are being made on data only, and a quite meager dataset at that (more on this later). Obviously, this is all still an experiment!
Interestingly, to my eye, the model appears to be a bit more pessimistic than the last release (version 0.2.0). That is to say, this model has more false negatives compared to 0.2.0, and fewer false positives.
The model seems to just prefer Kappo Kakko over Jack Hughes, which seems to be a developing story over the past few months. The model also thinks highly of Alex Turcotte. I'll be posting specifics about a few other predictions over the next few weeks.
Now that the original predictions have been made, the plan for the next year is approximately this:
Final summary: Predictions are out. They're probably less than excellent. I'm happy the project is at the point where I can come forward with something before draft day.
As usual, feel free to reach out with any questions on twitter.