THE BOOK--Playing The Percentages In Baseball

Thanks for the link, Tango. I should note that that graph is a straight average, and today I’m going to post a graph using the delta method, which will be much more helpful in showing how stuff changes by pitch thrown.

I have inning and pitch count data, but I’ve yet to write the code for time through the order. Shouldn’t be difficult (unless a pitcher faces a batter multiple times in the same inning).

As Jeremy says, he did not use the delta method, so that the graph is a mish-mash of different pitchers at the various intervals.  I’m not sure I see the point at all in that graph.  Plus, to tell you the truth, I really could not make heads or tails of the entire article itself.  Jeremy usually does great work, but I’m not sure about this one.  Looking forward to the “re-do.”

MGL, I think your projections of my work are overly optimistic. You’re not properly regressing to the mean. Anyhow, I got some utility out of the graph. But I’m having trouble with the re-do. I think I need reliable pitch classifications to take another step.

I think using pitch count here might be more appropriate than times through the order, just because Jeremy is trying to isolate how the pitcher’s “stuff”, as measured by Pitch F/X, changes as the game wears on from other things that affect pitch value (like the batter having seen the pitcher already in the game).  The number of times through the order has a big effect on the average value of a pitch or PA, but probably mostly for reasons outside of sudden changes in the velocity and/or spin of pitches.  If the question is how much of the difference is a pitcher’s stuff getting worse deeper into a game as opposed to other factors, pitch count seems the most apparent way to track that to me.  I would think the biggest issue with times through the order would be that that variable could relate directly to pitch selection (like Jeremy starts to get into at the end), so maybe it is an issue because of that.

Since the graph of true RV has definite peaks and valleys when plotted against pitch count that show where, on average, the order turns over, and the stuff RV has no such peaks and valleys, it seems to imply that the changes in a pitcher’s stuff over a game are relatively smooth, and don’t have sudden jumps as the order turns over.  Even if the actual RV of a pitch is more affected by times through the order than by pitch count, it doesn’t appear that that is true when you only look at how good a pitcher’s stuff is. 

I actually like the graph as is.  It shows the actual observed RV of pitches, which includes things like the effects of number of times through the order, and then it shows how much a pitcher’s stuff degrades as the game goes on.  This let’s you see how the effects of a pitcher’s stuff getting worse over a game compare to the effects of everything else that affects the RV of a pitch.  When I look at the graph, I get out of it that a pitcher’s stuff might get a bit worse deeper into the game, but that that is only a very small part of his results getting worse.  It tells me that a pitcher’s stuff getting worse is generally not the main culprit as to why his results get worse the longer he pitches in a game.

I’d like to see this done only with starters.  As noted, the first part of the graph is dragged down by dumping all the relievers in the first so many pitches.  I think looking only at starters in the graph will give a better idea of whether there is a gradual decline in stuff from early in the game to later in the game or if the graph is really flat and the slope is an artifact of relievers dropping out of the sample as the pitch count rises.  I think that change, and probably breaking this down by pitch type, would be the most things I’d most like to see added to this type of study.

The advantages to using times through the order

1. Quality if batter faced varies by position in the batting order. Evaluating first time faced 1-9 vs 2nd time faced 1-9, etc will control for much of this, except for PH for P late in game

2. Batter familiarity with the pitcher, not only will the pitcher be doing poorer as he tires, but when the batter gets his 2nd and 3rd look at the same pitcher he has a better knowledge base of that pitcher on that day. I believe this was studied in ‘The Book’.

3. To counteract #2, pitchers vary their pitch selection based on times through the order. That many pitchers go away from the fastball the 2nd time through is not just because the pitcher is saving his strength, but that he wants to give the batter a different look.

So, times through the order can create a distinct logical set of data that number of pitches does not. Also, by aggregating the data into larger buckets it’s probably easier to identify the trends.

The pitch selection problem could be an issue, but breaking it down by pitch type would address that.

Since you can see the distinct crests and troughs and sudden changes in the true RV line on the graph as the order turns over, but not in the stuff RV line, it doesn’t seem like there is much of an effect on the stuff RV from the spot in the order or the number of times through the order (beyond just tiring as pitch count rises).  If you are measuring the true RV, then you would need to account for those things, and they clearly have an effect on the graph against pitch count, but that doesn’t seem to be the case for stuff RV.  Since all Jeremy is measuring is the quality of stuff from the Pitch F/X data and doesn’t care about the end result, the quality of batter faced doesn’t really matter, unless pitchers deliberately throw worse stuff to the bottom of the order.  Given that there are no crests or troughs in the stuff RV that coincide with the true RV, that doesn’t seem to be the case.