2020.2.0 Model Release / 2020 NHL Draft Roundup

10.17.2020 : release, 2020

New predictions are posted for prospects that were eligible for the 2020 NHL Draft:

Obviously, the 2020 NHL Draft has come and gone, and as much as I would have liked to release this update prior to the draft, that just wasn't in the cards. The good thing is I get to write about the model output as compared to how the draft actually transpired.

The main improvement in the new model is that the input data obtained from Elite Prospects is supplemented with data scraped from CHL and USHL game summaries, for example, here's one. This allows us to use a variety of different statistics that aren't available from EP data, like primary assists, minor penalties, even-strength stats, and many more. This new data mainly contributed to the models marked 'CHL/USHL' above.

The current model release considers only prospects that were in their first year of draft eligibility for the 2020 NHL draft. To that, there is a bit of a balancing act between making predictions for as many players as possible, and maintaining model accuracy, and generally keeping things reasonable. For example, a CHL/USHL/J20/MHL/x/y model for defensemen was not provided for this release, simply because treating that diversity of leagues in the same model degrades accuracy. The underlying code can also handle overagers, but maybe that's a topic for another day.

With the logistics out of the way, let's move on to the some analysis:

High Value Picks:

There were some picks for whom the model predicts a better NHL career than their actual overall draft position might imply:

Defense is always trickier, but here's a potential value pick or two anyway:

There are many more picks of varying value. These were mostly obtained by sorting by the Miss Probability and First-Liner Probability columns in the model pages.

Low Value Picks:

There were also some players who the model ranked much lower than their actual draft position:

On balance, the model is quite good at calling negatives (relatively fewer false negatives than false positives, which makes sense since that is the nature of the NHL draft), but only time will tell.

Potentially Weak Defensemen Class:

I heard some discussion prior to the draft indicating that this year's defensemen class might not quite be up to snuff. Who knows whether or not that will turn out to be the case, but one way to at least get a grip on this is to compare the probabilities from the model output to those found in the cross-validation data (essentially probabilities from previous drafts).

Just looking at miss probabilities for the defensemen model posted above, it does appear that there aren't a whole lot of sure-things in this year's draft. There may be some inherent bias in the model that's responsible for this, but also, perhaps not, as this issue is not seen with this year's forward models.

Potential UDFA Finds:

Here's an interesting tweet:

Of course, points aren't everything. Here are some undrafted forwards for whom the model produced favorable probabilities:

For defensemen, things are a little foggier owing to the underlying fogginess of the model, but there are still some players who might be worth a shot:

These players were identified by sorting by 'Miss Prob.' on the model pages and finding the first few undrafted players.


Stay healthy, and feel free to reach out on twitter.