NHL Machine Learning Draft

The initial version of the predictions for the 2020 draft class are available. See home for the links.

As usual, re-draft results are posted to evaluate the algorithm. These predictions use the same model that's used for the current data, but the predictions are done on players for which we have an idea of where their careers are headed.

In summary, this release shows a decent improvement for forward predictions, but the model for defensemen remains a work in progress, even within the already narrow realm we're working in (PPG is not as aligned with 'success' as it is for forwards).

For forwards, the model appears to be 'correct' (within about .2 PPG) about 60% of the time, and 'not wrong' (within about .4 ppg) about 85% of the time. There are a lot of caveats to those metrics, so they just give an approximate idea of model performance. Model performance for defensemen is still unremarkable.

There have been many updates since the last release, including:

• Most notably, all data is now provided by Elite Prospects. This is a huge step forward in terms of completeness and richness of data. This also changes the Data policy.
• The results pages now show intermediate model probabilities. This is direct output from the model. A players PPG prediction is derived from these probabilities.
• Many algorithm updates; too many to list here (plus they might be too 'in the weeds' for this post). Many are related to the fact that we now have more solid data.
• Many cosmetic updates. Sortable columns in the data tables, permalinks, and a few other goodies.
• No more old model data, nobody cared and it was expensive to maintain. I may move back to a 'lite' version of this in the future.

I'm aiming for another release closer to the 2020 draft.

As usual, feel free to reach out with any questions on twitter.