Wednesday, October 06, 2010

Division Series, October 7th Games

October 7th Simulations, Division Series Odds. ..

Visitors Home Probable Pitching Matchup Favorite Vegas Win Prob Simulator Win Prob Actual
Braves Giants D.Lowe vs T.Lincecum Giants 59.84% 62.27% SF 1-0
Rangers Rays C.Wilson vs J.Shields Rays 53.38% 48.86% TEX 6-0
Yankees Twins A.Pettitte vs C.Pavano Yankees 52.72% 52.57% NYA 5-2

Current ROI: 3.54%

MLB Scoreboard


Josh Engleman said...

I'm a little surprised by the Rangers/Rays prediction. How do your player projections handle CJ Wilson? I would've thought the Rays would be a slight favorite.

Xeifrank said...

Pitching stats/projections are heavily weighted by FIP. But yes, the Rays are a slight favorite.

Interesting matchup between these two teams for my simulator because during the regular season my simulator had an awful ROI of -9.8% on all bets made on Rays games, but on the other hand it had an amazing ROI of +20.8% on all Rangers games that it bet on.

These playoff games are way too small of a sample size to take away much from the results, but it is an interesting matchup for my simulator nonetheless.

Josh Engleman said...

Sorry, I should have been more specific. I was wondering how you handle Wilson's stats in regards to his old bullpen numbers? Also, have you looked to see if something like xFIP could be more accurate? Sorry if I'm bombarding you with questions. I'm very interested in everything you are doing and have wanted to set up something like this for myself for a long time now. I saw the screenshot you posted a while back of your sim. Looks incredible. I'm jealous.

Xeifrank said...

Josh, those are all great questions and actually questions that others like you have asked me before. This offseason I will attempt to test and compare a version that uses xFIP instead of FIP and try weighting between the two, ie - 70% xFIP, 30% FIP trying to find a more accurate way of projecting pitcher true talent levels. The reason I haven't done it is because I don't have all the data for it yet, but will hopefully get it shortly. The data is out there, just takes some work to download, format it and write the code to use it.

As far as C.J. Wilson's bullpen stats go, I ignore them. I know there is probably a solid method for converting from bullpen numbers to starter numbers, but I have chosen not to add that complexity.

I try to automate as much as I can, but I do know there are cases where I have to manually override the pitcher or hitter projections in a few select cases due to complexity issues. For example I do not have a MLE formula, I instead estimate rookie players as 0, 0.5, 1.0 etc... WAR players and give them stats for a player of that level until they have developed a larger sample size.

Through backtesting, using in and out of sample techniques I have come up with what I feel like is the best "one size fits all" method for projecting a players true talent level. The 5/4/3 + Regression method we are all familiar with has proven to be a loser in this arena. The weighting method I use is proprietary but it is nothing at all like the 5/4/3+R weighting for either pitchers or hitters.