Baseball Prospectus | Lies, Damned Lies: Estimating Pitch Counts

Nate Silver, April 23, 2003

Silicone. Margarine. O'Doul's. Why fool around with watered-down imitations when you've got the real thing ready and available?

Rightly or wrongly, a lot of attention has been focused on pitch counts in the past several years. That's partly because of the efforts of people like Rob Neyer, Keith Woolner, and Will Carroll, not to mention those coaches, executives and agents who understand the importance of protecting their golden-armed investments. Pitch counts have become easy to take for granted because pitch count data is more readily available now than it ever was in the past. These days, just about any self-respecting box score lists pitch counts alongside the rest of a pitcher's line, a far cry from the dirty newsprint days of yore, when pitch count references were about as common as mentions of Reality TV or the Information Superhighway.

But what about when you don't have pitch count information available? Like, say, you're at a ballgame, and wondering whether Dusty Baker should send Kerry Wood out for another inning? Or you're perusing through minor league stats? Or you're looking at old boxes on Retrosheet, which wonderful as they might be (this, folks, was the first game I ever attended), don't contain any information on pitch counts?

Well, it turns out that it's not that difficult to make a reasonable guess at pitch counts based on other information that's much easier to come by. Looking at a complete set of data from the 2001 and 2002 seasons as provided by Keith Woolner, I ran a simple linear regression of pitches thrown against various other characteristics of a pitcher's stat line. Here was the formula that I came up with:

Implied pitch count (IPC) = (3.17 * BFP) + (3.44 * BB) + (1.53 * K)
Before we go any further, let me disclaim that I'm not the first person to try and skin this cat. In particular, Boyd Nation has done some wonderful work in estimating pitch counts for college pitchers based on a similar approach. I'm publishing my results anyway because: