What a day. It was a long one, but I have rehabilitated my model and it should be good to go from here on out. I went back to January 27 and ran 10 days worth of model runs, one day at a time, to build a training database of roughly 400 games. The process took several hours as I pretty much had to perform my daily model routine 10 times, which included 2 Saturdays, that have a ton of games to run.
All games going back to January 27 on the model plays page now reflect the updated fixes. On the morning of February 14, the “Last 14” row in the model results SHOULD reflect 14 days of fixed model results.
Today’s plays listed on the site for February 11 have been updated to what I would have played had my model been correct this morning, instead of what I actually did. At the time of writing this post, those old plays were 7-6 for +0.4 units. Since I don’t want to make it look like I am running away from these plays, here are the old plays I actually put money on this morning:
And at the time of writing this post, my “new” plays are also 7-6, but for +1.4 units thanks to a +120 money line hit on Longwood.
Notably from the current 10 day training database was the fact the machine learning could not find a correlation between my models and under plays, so don’t expect any under plays for a while. Over plays however do show a correlation, and the 2 overs it picked today both won. The p-values on the machine learning models for ATS and ML plays are also much more improved from what they were with a much stronger correlation.
Another thing I did today was re-work the Blender. I had wanted to do this but was going to wait until the offseason because I felt my training database I had built up was more valuable, but since I had to scrap it today I went and made those changes. Interested to see if the changes I made to the Blender was for the better.
Hopefully these fixes lead to a strong finish to the basketball season.