Evolution of NHLseasonML...

• Re-write the code to be more "harnessable" - something that can easily built into a script and run multiple times with different hyper-/parameters.

** The database queries are still hard-coded

** COMPLETE - After the data is pulled, the remaining steps should be harnessable

• COMPLETE - Use some more/different data - include position, age and draft position information, for example.

• Include player ID as categorical data? I don't think this makes sense!

** Still remains in the flow to make it easy to attribute predictions to players

• Determine if using 3 years of lag is best.

• Does missing data affect the results?

** Preprocessing modified to assign ignore data values AFTER scaling, as it should have been ** I modified the performance quantification to handle missing data properly

• Does one need to use different models for different player archetypes? Can one model predict a grinder's and a sniper's performance?

• Can more than one responding variable be predicted at the same time? - Predict multiple (all?) stats at once.

• How can the design of the neural network be improved?

• COMPLETE - The final predictions should be probabilistic.

** Now that the training and prediction portion of the flow is harnessable, probabilistic predictions are possible.

• When it comes to predicting total points, is it better to predict goals and assists separately, then combine them? Or can a points prediction do a better job?