Why did we need EBP?
In 2015 and 2016 I developed a formula designed to uncover the dependence of a baseball team’s run production on its On Base Percentage. I called it Expected Binomial Production, or EBP. In advance of publishing that, I wondered whether Sabermetricians and other fans of the mathematics of baseball may upon first seeing it, think “What in the world do we need yet another run estimator for? We’ve had good ones for decades – it’s a closed topic.”
While it is true that EBP provides a foundation that could lead to the most accurate of all run estimators, it was designed to serve a different purpose. It was designed to show how run production varies with OBP. No existing run estimator demonstrates an explicit dependence of run production on OBP, something EBP does. Nor can existing run estimators perform some of EBP’s other cool tricks, like predicting frequencies of innings with particular numbers of runs, and showing how run production would be different in a game of baseball played with 4-out innings, or some other number of outs per inning.
Could it have worked using what we already had?
It did occur to me later, however, that even though the formulas for existing run estimators don’t show an explicit relationship between run production and OBP, that they should do so implicitly. The key was realizing that the numbers that are put into those existing run estimators (walks, hits, etc.) will undergo the same kind of explosive growth as EBP does when OBP gets close to its maximum value of 1. This is the Fixed Outs Explosion that I refer to elsewhere. And because each of those inputs will undergo the Fixed Outs Explosion (combined with the fact that the formulas are somewhat proportional to these inputs), each of these run estimators will undergo the Fixed Outs Explosion as well. Even those with linear formulas will exhibit this nonlinear behavior. And that told me that, if I could feed in realistic input numbers that go along with all possible values of OBP (from 0 to 1), I might be able to get these run estimators to show me a reasonable relationship between run production and OBP.
This led me to wonder, how well will they do that? And how well do their results compare to EBP’s?
Let’s see if we can do this using what we already had
At first I tried coaxing the OBP dependence out of some of the main known run estimators by reworking their formulas to show an explicit dependence on OBP. It quickly became clear, however, that that approach could not work with most of them. I would have to mock up data that shows how each input used by these formulas – walks, hits, strike outs, double plays, stolen bases, and more – will vary with changes in OBP. Then I would run my mocked-up numbers through the run estimator formulas to plot curves showing what the p-dependence of these run estimators might look like.
I found it quite difficult to find formulas that produced mocked-up numbers that made sense. The hardest part, and the most critical part, was getting the dependence of plate appearances as a function of p correct. All the details of how I came up with this and the other formulas are spelled out in How I mocked up OBP dependence for basic offensive baseball statistics, for those who are interested. Here, I’ll just list a table displaying an abbreviated version of the resulting basic stats for the 1994 Cleveland Indians:
Important to note is that these simulated numbers do factor in outs and base advances on the basepaths. Total number of outs is assumed constant, so as outs made by batters goes down to zero as p goes up to 1, the number of outs made by baserunners goes up by the same amount. This prevents plate appearances – and correspondingly run production – from going up to infinity. The factor that decides how high it does go is the choice of what percentage of baserunners make outs. This choice was made by fitting a curve to a graph of team seasonal numbers for percentage of baserunners making outs plotted against p.
Taking these numbers and plugging them into the run estimator formulas gives an estimated p-dependence to each formula. These are just estimates and could be wrong – especially, at extreme values for p, strategy will change and with it the relative importances of power to on base percentage, and so therefore their relative levels of use. But as we will see, I have pretty good confidence in them.
Plots of the p-dependence of the run estimators
I’ll present these in three sets of charts, grouping run estimators by their functional forms: Linear, Nonlinear (OB*Advancement), and Hybrids of these two. For descriptions of these categories of estimators, and of each of the individual estimators themselves, see my page of Descriptions of baseball run estimators.
The Linear run estimators
Our first set of charts shows the run production that the following five linear run estimators give us when we provide them the input data we generated for the 1994 Cleveland Indians (and EBPt included for reference):
- Extrapolated Runs (XR)
- New Estimated Runs Produced (NERP)
- Estimated Runs Produced (ERP)
- Equivalent Average runs (EqR)
- Runs based on Weighted Runs Above Average (wR_PA)
Oh my. They all go negative between p=.12 and p=.19. That’s absurd, and it reaches absurdity for OBP’s not far from the ones you see against the best pitchers in the game. In other words, these go negative at what is possibly the expected OBP in some actual single MLB games. Also, their runs per plate appearance plots don’t have the upward curvature we expected – they’re straight lines, well *almost* perfectly straight lines, anyway. But that’s to be expected given that they’re linear formulas. Finally, at p=1, their runs-per-plate-appearance (and runs-per-base-reach) numbers are well below the .90 to .97 range we’re looking for.
The advantages of linear formulas include their simplicity of use, but you pay for that in having a relatively narrow range of on-base percentages in which they are accurate. Notice how all the curves bunch together in the p=.30 to p=.37 range within which almost all team seasonal values fall. Everybody’s right where it counts, but outside the team average values for p that we see in actual gameplay, linear run estimators appear to do poorly. And for individual matchups, we’ll see values of p outside that range.
The Nonlinear run estimators
The next set of charts shows the run production that the following four nonlinear run estimators give us for the 1994 Cleveland Indians, with EBPt included for reference.
- Runs Created, and Runs Created – stolen bases (RC & RCsb)
- Runs Created – technical (RCtech)
- Total Offensive Productivity (TOP)
These include the original Runs Created, which set down the basic principle of multiplying a base-reaching rate by a base-advancement rate. All the others in this category are basically increasingly complex variations of this principle. (I’ve never seen TOP referenced outside of the original author’s web page about it, but it seemed intriuging so I thought I’d include it.)
Note that I show one line for RC and RCsb because their plots are pretty much identical (so are those of EBPt and EBPf, by the way, although those do have a very slightly visible difference). Here are the plots:
Well now, that’s much better. The runs per PA plots have a nice upward curve, as expected, and start at 0 when p=0, as they should. For the runs per base reach plots, at least we’re in the realm of the possible now at p=0, even though they’re all well below our expectation of .103 or so (the Cleveland Indians’ home run fraction). For these to be correct in a real-world situation, our assumption that the rate of home runs relative to other kinds of base reaches stays the same would have to be wrong. Instead of staying the same, it would have to decrease as on base percentage decreases, mostly disappearing in the case of TOP (from .103 to .032), and entirely disappearing in the case of all the versions of Runs Created. That’s entirely plausible, considering that a dropping OBP probably corresponds with a lot less good contact relative to bad contact, so a lot fewer balls hit a long distance relative to the number hit a short distance. Then again, this would be offset to some extent by the fact that teams would focus more on trying to hit home runs, as that’s pretty much the only way they can score anymore. I don’t think it would drop this much; in my judgement, these numbers are too low.
At p=1, however, no judgement is required to fault the results – they’re impossible. You can’t have more runs than base reaches, and certainly no more runs than plate appearances, so these curves should never go higher than 1. Yet they’re all peaking at higher than 1.1. This absurdity doesn’t happen on every team’s plot. It seems to be worse for teams that hit a high percentage of home runs. In the half dozen or so teams I’ve plotted, at least half of them share this flaw for these nonlinear estimators.
The Hybrid run estimators
Now on to the last two estimators we’ll look at – the hybrids:
- Runs Created – 2002 (RC2002)
- Base runs (BaseRun)
Both of these formulas are the sum of a linear portion (in the style of the linear run estimators) and a nonlinear product (in the style of the base-reaching rate times base-advancement rate nonlinear estimators). However, they go about this in very different ways, with very different results:
RC2002 is like the worst of both worlds, lacking the formulaic simplicity of the linear run estimators, while behaving just like them in its p-dependence, including taking negative run values for low p.
But look at Base runs. At p=0, the runs per base reach plot takes the value of the team’s home run fraction, right near where we expected it to be. At p=1, the value is .93 – comfortably within the .90 to .97 range we expect. It does so for every team I’ve looked at thus far. It’s been said of Base runs that it models the reality of the run-scoring process significantly better than any other run estimator. These results lend support to such praise.
So in retrospect, did we need EBP?
So we’ve found out that, after all, there was one existing run estimator (Base Runs) that could have been used to examine the OBP-dependence of run production. Which brings us back to the question stated at the beginning of this article, “What in the world do we need yet another run estimator for?” One could argue that because Base Runs appears to do a pretty good job of showing that OBP dependence, that EBP is superfluous, and didn’t need to be made. I have a few responses to that.
- Coaxing that OBP-dependence out of Base Runs was about as difficult as was creating EBP.
- At the time I only knew of the linear run estimators, and it didn’t seem to me that their OBP dependences, if even possible to find, would have the right shape (and they don’t).
- There are different shapes to the OBP-dependence curves for EBP and Base Runs that may carry some significance.
- EBP has plenty of room for improvement, and shows promise for paving the way to the creation of the most accurate run estimator yet.
- EBP may provide new insights on the connections between individual offensive events and run production.
- EBP still has those other cool tricks up its sleeve that run estimators cannot do.
About the shapes of these plots
There are some interesting things to note on the shapes of these curves. One is that every linear run estimator appears as an almost perfectly straight line on the runs per plate appearance charts, and curved on runs per base reach; while the nonlinear estimators appear very close to being straight lines on the runs per base reach chart, and curved on the runs per plate appearance chart. These aren’t quite straight lines on the runs per base reach chart, however. They tend to have a slight upward curvature for lower p values, and a slight downward curvature for upper p values. In the case of RCtech, it’s upward throughout the whole range for every team I’ve checked. In the case of TOP, sometimes it’s all upward; in the case of Base runs, for some teams it’s all upward, for some it’s all downward, and for some it’s up then down. But in every case, the curvature is so slight that to the eye these look pretty close to straight lines.
EBP, by comparison, has a very pronounced upward curvature followed by a very pronounced downward curvature for every team. Because EBP is the only one of these specifically derived based on p, I suspect the real-life p-dependence of run production to follow EBP’s curvature, but to have values in the high-p range closer to those of Base runs.