Projecting innings pitched for individual games

Whether you’re trying to select a pitcher for a daily salary cap format or evaluating which pitcher to use in a traditional weekly league, it would be really useful to be able to accurately forecast the number of innings a pitcher will last on any given day. Not only would that give you some idea of how much impact his performance that day will have on his rate statistics, but it also is part of the equation for determining how many strikeouts to expect and how likely he is to get the decision in the game. It’s also a topic that I’ve never seen addressed in any detail.

What makes projecting innings pitched for an individual game interesting is that expected performance will have an impact on expected innings pitched. It’s easy to mistakenly think that because most pitchers average somewhere between 5.5 and seven innings pitched per game, it wouldn’t be worth the trouble of trying to forecast innings pitched. But that’s misleading, because it averages out performance in a variety of situations. Taking all of those into account, we might be looking at a range of expected performance that goes from around four innings pitched at the low end (a bad pitcher against a good offense in a hitter’s park) to maybe eight innings pitched at the high end (a good pitcher with a high pitch count limit against a weak offense in a pitcher’s park). That makes the attempt to forecast innings pitched seem much more worthwhile!

In this article, I’m going to build a simple model for forecasting or projecting how many innings a pitcher will last in any given game. I’ll keep things simple, focusing on the impact of the opposing offense on innings pitched. I hope to follow it up in the future with another article that tests the effectiveness of the model. Depending on the results, maybe I’ll tackle some of the other factors that impact innings pitched too.

The majority of the time, a pitcher will come out of the game for one of two reasons. Either he’s removed because he’s pitching badly, or he’s removed because he’s reached some sort of pitch count or innings pitched limit determined by the team. There are other reasons such as injury or being removed for a pinch hitter, but these generally play a relatively minor role. When looking at a pitcher’s average innings pitched in past games, we’re looking at data that takes both major factors into account. It reflects both the effectiveness (and efficiency) of the pitcher in past games, as well as any pitch count or innings pitched limits he’s been subject to.

Likewise, when looking at the average innings pitched by starting pitchers against the opposing team, we’re looking at data that reflect all aspects of how the hitters on that team have impacted opposing pitchers’ innings pitched. While it won’t be as precise as it would if we created a model based on the specific hitters in today’s lineup, we’re trying to build a simple, usable model…not a perfect simulation.

What I’m going to propose is laughably simple, but I suspect that for most situations, it’s going to be good enough to have value for many people. I’ll need three data inputs: average innings pitched for the league (NL or AL), average innings pitched of starting pitchers against the opposing offense, and average innings pitched for the starting pitcher. Basically, I’m going to create an adjustment factor by dividing the average innings pitched of starting pitchers against the opposing offense divided by the league-average innings pitched. This value (which will average one) will be used to multiply the average innings pitched of the starting pitcher being evaluated. That’s the IP forecast for today’s game.

Yes, there are all sorts of other factors that would make this a more accurate model:
-Which players are in today’s lineup?
-What are the effects of the park?
-What are the effects of being home or away?
-Should we use regression to the mean on the pitcher’s average innings pitched?
-Should we weigh the pitcher’s recent performance more heavily?

However, I believe that even the incredibly simple model I’ve described would provide value. For my next article (which should be in about a month), I’ll do some back-testing comparing the accuracy of using this model to forecast innings pitched rather than simply using the pitcher’s average innings pitched without taking the opponent into account.