Statistical Simulations

[TooLiveBrew]

Just a simple query from a dummy: When doing things like calculating win probability and simulating series/games/seasons, are teams matched against each other utilizing current batting, baserunning, pitching, and fielding stats? Or is it done more generically? I understand playing home/away is most likely weighted, but are things like how the teams have been playing lately (ie the last 10 games, or 5, or 17, etc.) factored in, too? I understand this may be quite complicated, but I'm very curious as to how teams are 'simmed' against each other.

[igor67]

If you are simulating 10,000 or more seasons you can essentially infer that the model will be rather straight forward. I think in Russ's case he just used pythagorean record in concert with an adjustment for home and road games. The thing is if you do that for so many seasons all the other potential effects show up in the final number for the most part, so you get a pretty fair estimation.

[rluzinski]

The way I do it is very generic:

Runs scored/against -- > expected win percentage -- > compare 2 team's win percentage = win probability --- > adjust for home road.

There's a lot of problems with this approach. Change in personel throughout the season is a biggy. Also, the using this approach for 1 game or one series ignores what part of each team's starting rotation is scheduled to start. That said, this approach gives some reasonable resuts.

but are things like how the teams have been playing lately (ie the last 10 games, or 5, or 17, etc.) factored in, too?

No and beyond changes to personel, they probably shouldn't be. Team slumps and streaks can be almost completely be attributed to random lumpings of wins and loses, as I tryed to show in this article:

Teams Streaks

Is it possible something else is also contributing to a streak? Sure but without knowing for sure, one must assume the most likely cause, luck.

A "real" simulation would take into account everything and would be more accurate for sure. Still, for as simple as an approach I take, it works pretty well.