I initially created Baseball Graphs to develop new and better ways of presenting baseball statistics. The 2003 season was my first full-season attempt, and it turned out pretty well. I now spend a lot of my time creating and formatting the baseball stats at The Hardball Times.
Along the way, I've also picked up the writing bug. I spend most of my energy writing for THT, but sometimes I like to write about more geeky stuff, baseball research and things like that. That's what this blog is for.
I'm very interested in graphics and other information display, so I'll probably post links to creative stuff like that. I also plan to write historical articles that complement the historical graphs. And I have inexplicably developed an interest in baseball win-related stats, such as Bill James's Win Shares and something called Win Probability Added. So you can expect a few posts about that stuff too.
If you're interested, I've created a complex spreadsheet full of WPA logic, which you can download at ftp.baseballgraphs.com/wpa. If you'd like to read more about Win Shares, try these links:
Win Shares Intro
I posted a review of a Boston/Tampa Bay game a couple of days ago at The Hardball Times. I also added the “P” of each situation to the graph, which might appeal a bit more to everyone’s graphical esthetics.
Also, there is a satirical take on Win Probability from Tim Keown at ESPN. I have a feeling he might be a contributor to the RedsZone Forum!
I love the P values as an additional series on the graph. I had played around with it on the Backe outing after I had already put the graph up, and then I haven’t had time to do WPA for the last week. It really is a nice addition with regard both to information and aesthetics (it reduces blank space).
Were the comments purposely closed on the new upload thread?
The addition of the HR/SO WPA change thing got me thinking again about how best to measure the importance of each plate appearance.
I shouldn’t be too difficult to figure out the probability of each common PA-result for each base-out state. Each PA-result would have a corresponding change in WPA. The expected value of all outcomes should be 0 (right?), but the standard deviation would tell you the degree to which that single PA affects the game. Instead of just using the extreme, you could use “all” the possible events to test the importance of a situation.
If I remember my basic stats correctly, the SD of a distribution is the sqrt of the sum of each possible event P(Xi)*(Xi-EV)^2.
Forgot to turn the comments on. Sorry. Though I am having a heck of a time with spam commenters.
I think your approach is sound, Sky. Got to admit that I don’t have the time or energy to tackle it, but it makes sense.
If you want to get real technical, the probability of expected outcomes would differ by run environment, right? That might be the “next phase” of your approach.
Yes, things should change by run environment. However, the only way that I’ve come up with to calculate the probabilities is empirically. How does this sound to people:
Using Retrosheet data, count the number of times each base-out situation follows each of the 24 base-out situations. This will allow me to calculate the probability of each transition. By combining these probabilities with the score and inning of a game, you can compute the WPA of each possible transition. Finally, you can compute the standard deviation of each outcome.
Now, I realize that the base-out transition probabilities depend on score and out, but I think that dividing up the data any finer would lead to sample size error, but i’m not sure. Also, which years do you guys recommend I use to collect data? I’m thinking 1990 on, but is there any commonly accepted start date of the common era? (Btw, where do I find the ‘93+ data files?)
-Sky
| <<Previous Article: Astros' WPA | Next Article: WPA 1.4>> |