I've been looking at ways to solve, especially with a focus on over/unders. Basing the calculation on negative binominal didn't work, a grid search for the best parameters is still ongoing. And we do get some interesting probabilities too.

Really fun stuff!

Playing around a bit on visualizing predictions. One thing that's been pretty challenging is the over under variation. Turns out, every hockey game lands at 6, so it's real easy to have a model just predict 6 for every game.

An example of the type of analysis I'm now producing. I'm not sure about that win probability with an upset rate of 40% or so. But it looks good! More to come around testing and experimentation.

For today's focus, I wanted to continue pushing on the implementation from the rest of the article - Claude made that pretty quick. I thought it was interesting to look at prediction accuracy with bias of future data (top part) versus only using historical data (bottom section)

So for a first draft, got the following! 3 main differences, will be interesting to see how those shake out.

Odds from Will Hill (for ease of use and speed, will hopefully find a market number programmatically soon).

Happy to take any and all feedback! More to certainly come

Had a chance to rewrite JobDescriptionAI.com, which I'm super proud of! AI is getting better writing HTML and CSS. I think it looks great

As part of that, I've been working on new tools for context management - including a SQL helper that analyzes tables and lets you pick which ones to send to the LLM

Working on Bluesky Analytics, and it's coming together! Very excited.