FAQ

Who are you?

I'm Patrick, I'm currently a data scientist, and formerly an actuarial fellow, and a professional poker player. One way or another, I've made my living with data for 15 years.

This blog has an instagram where I post more / different content - @basedonactualmath

What is this blog about?

Right now all my content will be college football themed. Ranking the teams through various means, trying to predict each week's games, trying to understand each path each team is taking through this season.

In the past, I've written about other things. Politics and sports are two of my interests that lend themselves well to data analysis, and I've written about them both. I built models and in-game win probability graphs for the last two world cups. I built and published a model for the 2012 and 2016 presidential elections, I got 50 states right in 2012; less so in 2016, and I've built models for local races here in WA.

Views expressed are my own and do not reflect the views of my employer etc.

So you have predictions for every college football game?

Yes

What would happen if I used your odds to make bets in Vegas?

You're welcome to do so. In the past my model's predictions have been sliiightly better than predictions being made by the bookies in Vegas.

However, when you bet in Vegas you start at a disadvantage. To make an even money bet, you have to put up $11 and Vegas only puts up $10. $1 may not seems like a lot, but it is, it's 10% of the bet! So you have to be 10% better than Vegas just to break even. I don't think the model as-is is 10% better than Vegas, but maybe you want to combine the model with your intuition and give it a try? Go for it! YOLO etc., please report back and let me know how you wind up.

Also, sites that sell you winning sportsbetting picks are rip-offs.

Your rankings look very weird to me, for example you have Team X above Team Y, what's up with that?

My rankings are predictive - in short, my model would favor every team on the list against the teams ranked below it (on a neutral field), and the model adjust teams ratings when their on field performance was different than predicted. For example, even though Notre Dame did win against Virginia Tech this year, they barely won, and Virginia Tech has been weak, the model expected better from Notre Dame, and hence it revised their rating down, despite the win.

The model's weekly correlation with Vegas lines is typically 0.96, this week (week 11, 2019) it happens to be 0.97.

This approach is rather different from either the CFP or the AP voters (which leads to some surprising rankings); the AP / CFP ranking algorithm isn't... the most clear. But, they aren't as predictive as my rankings. For example, in the marquee match-up of the year, LSU is a 7 point underdog @ Alabama, despite being the higher rated team.

What's unique about your model?

I mean, I think the whole thing is awesome. But forced to choose... The number of moving parts under the hood are quite simple - yet powerful. The bowl forecasts (with %s) isn't something I've seen somewhere else, usually people just guess, and the CFP module is something I've developed that's a combination of data and heuristics, using trial and error trying to model things the CFP committee claims to care about vs. the things they do care about.

So far this year my model has correctly predicted the winner in 81% of games.

Why is your photo of you playing the piano?

Kind of hard to take a good photo of me doing data science. What am I gonna do? Screen cap my SQL?

No comments:

Post a Comment