I am really glad I listend to this. I subscribed and currently in the process of listening to most of the past episodes. In my previous comment I mentioned a model I developed using a manual bayesian based approach for predicting team win probability based off of pinnicale's spreads. That didn't do all that well for a lot of reasons. One of which being that I approached bayesian statistics in an overly simplified way and another being that I picked a really hard target. What I got from this podcast was to try for lower hanging fruit and to not try and reinvent the wheel. In the past two days I was able to develop a model using pymc3 for predicting team and game totals that has a r2 score of roughly twice that of pinnicale and is actually profitable in back testing. I'm still fumbling my way around pymc3 but I can see myself doing some really cool stuff in the long term with it. Thanks again for the awesome and informative podcast!
At about 26:00 he mentions an example where someone did a really simple bayesian based approach for baseball using a team's win record and a pitcher's win record. I've done something similar in python for NBA using Pinnicale's spreads. I started off with a prior of the team's season win record and then used the spread from Pinnicale to update that prior. The end result was something very similar to Pinnicale's moneyline implied probabilities. That simple bayesian approach had an ROC AUC score of 68.1% and a Brier score of 22.6% whereas Pinnicale had 69.38% and 21.2% respectively during that sample training time period. While it was an interesting/fun learning experince it isn't very useful for betting NBA games as far as I can tell. I've been thinking about what else I could apply a similar concept to increase the calibration/profitability.
