Baseball is a game that relies heavily on statistics and data. Ever since the days of Henry Chadwick and his invention of the box score there has been a hardcore part of the baseball community that has, and continues to have, an insatiable appetite for data.
There have been three data revolutions in baseball, and I strongly believe we are on the cusp of a fourth. The first revolution was started by the incisive writing and astute analysis of Bill James, quickly followed by Pete Palmer and his tome, The Hidden Game of Baseball. The invention of Retrosheet by David Smith suddenly saw reams of data available to all—this was the second. The third has only really happened in recent years led by the ubiquity of cheap computing, free storage and the Internet, which has created a bustling online sabermetric community captained by the likes of Tom Tango and Mitchel Lichtman.
Today, we sit on the verge of the fourth, which is the complete digitization of baseball by snazzy computer software. The first incarnation of this is MLB’s Enhanced Gameday (EG) system. Enhanced Gameday is a pitcher focused tool that allows the speed and trajectory of every pitch to be precisely mapped from when in leaves the hurler’s hand to when it nestles in the catcher’s mitt. Whereas before we had to largely guess about pitch type and location we can now apply science to look at such things as pitcher release points, how well a hurler paints the corners of the zone, and how velocity and movement hold up over a season—it goes without saying that this revolutionizes pitching analysis.
At some point EG will be rolled out to hitters and fielders and will allow us to track the trajectory and hardness of every hit to every part of the park. That is the future; let’s get back to the present and the current incarnation of Enhanced Gameday.
About Enhanced Gameday
The system was first tested in the 2006 postseason and was supposed to have been rolled out to all parks for April 2007. Unfortunately this hasn’t happened, and at the moment only eight parks are covered: Anaheim, Chicago (AL), Seattle, Texas, San Diego, Atlanta, Boston and Los Angeles. A number of parameters are logged including: strike zone location, pitch release point, velocity release components, amount of break and release/plate speed among others.
This is stored as an XML file on the Gameday web server, which is part of mlb.com. These data are available to anyone to analyze and already a ton of great work has been done. Joe P Sheehan has used the data both here at THT and also in his regular weekly slot over at Baseball Analysts.
Blogs, such as Seth Billfer’s excellent Detroit Tiger Weblog also uses EG data to analyze pitcher performance. By browsing those articles it’s easy to see how these data will revolutionize our understanding of the game.
However, don’t get carried away just yet. There is one huge, gargantuan question that no-one seems to have answered but still needs asking, namely: Can we trust the data?
So, Can We Trust the Data?
Just because something is measured by computer software doesn’t mean it should automatically be accepted as the gospel. We must prove it. This is a question I have often mused since Enhanced Gameday’s inception.
A recent article at Seth Billfer’s blog persuaded me to lift myself out of my current state of paralysis and investigate. Take a look for yourself:
This shows the release point of Tigers pitcher Mike Maroth for two different games, both with EG systems installed. The orange dots are pitches thrown in Anaheim, the blue ones in Toronto. Notice something? No PhD is required to detect a systematic difference in release points between the two games. Why is this? Is the mound at a different height, or a different shape, or is EG somehow calibrated differently at different parks, or has Maroth just adjusted his mechanics?
We can eliminate (I’d imagine—perhaps Derek Zumsteg can tell us differently) differential mound height and shape. I refer you to rule 1.04:
The pitcher’s plate shall be 10 inches above the level of home plate. The degree of slope from a point 6 inches in front of the pitcher’s plate to a point six feet toward home plate shall be one inch to one foot, and such degree of slope shall be uniform.
That leaves two options: either EG has calibration issues or Maroth has tinkered with his mechanics between games.
The problem with Maroth is that he has only two outings in an EG park. To fully explore this issue we need to look at a pitcher, preferably a starter, who has had several outings in multiple EG parks. This will allow us to compare performance across both the same and different parks so we can wheedle out any data bias. Fortunately such a pitcher exists and his name is Kevin Millwood. Millwood has pitched in five EG parks: Texas (twice), Seattle, Anaheim, Boston, Toronto and Chicago.
We’ll look at a number of pitch spray charts cut by different EG parameters.
Take 1: Speed & Location
First up is a plot of all of Millwood’s pitches by velocity and location. Each color point represents a different speed range:
Yup, that feels like the Millwood we know and love—a low 90s fastball pitcher who sprays it around a bit. That’s the first test passed. Now look at the same picture but split by game. Is there any appearance of bias?
Emphatically not. Although he seems better able to locate the strike zone in some games that in others, the picture is random. So far, so good.
Take 2: Release Point
Looking at release points should be more informative as it is where Seth Billfer detected bias. Here is Millwood’s release point chart for the five different parks. Each colour marker represents a different game.
There it is again. Different parks show consistently different release points! Is this a mechanical issue? Is it an issue with EG? Am I missing something? Let’s have a think:
- The main outliers are Safeco and Rogers Centre, although a gradation exists across all parks. The difference between the extremes is almost 2/3 of a foot, about 20cm—that is a significant number
- Each set of points would trace a similar shaped “arc” (the line of best fit would be parallel). This implies it isn’t a mechanics issue because if it was I’d suggest we’d see a more erratic distribution, whereas we are seeing a consistently different release point
- There is no apparent difference in release points for the two home games (the red and green markers). Again indication that it isn’t an issue of changing mechanics
Take 3: Release and Plate Speed
Another test of accuracy is to look at release and plate speed for each EG park. Below is a plot of pitch speed as it leaves the hurler’s hand (y-axis) versus when it crosses the plate (x-axis):
Again we see the same systematic bias that we saw for the release points. Now, we might expect to see some difference because of park effects. But is this trend related to park (atmosphere would be the largest influence I’d imagine)? Or is it an issue with the EG set-up? Or something else?
Here are some thoughts and observations:
- Both home appearances (red and green markers) have a similar distribution, which shows inter-game consistency (variations aren’t a result of how a hurler feels on that particular day)
- There is a systematic difference in start speeds, which is surprising as park effects should have much less (or no) influence on a pitcher’s release velocity. It suggests either measurement error or that the reading is taken at different points (or measurement bias/error)
- The line of best fit through each set of points has the same gradient. If atmospherics was a consideration we’d expect to see different slopes as some parks cause pitches tail off more than other parks do (see below).
As we said earlier the red and green lines are on top of each other as they represent the same park, but there is a clear difference in the positioning (but not slope) of the other lines.
Take 4: Velocity Components
A further test is to look at velocity components on release. In the chart below the Y-component is the velocity in feet per second in the forward direction, the Z-component is velocity in the vertical direction, and the X-component is velocity in the horizontal direction.
Interestingly on these charts the bias is much harder to detect but it is there if you hunt. Our interest is mostly in the Y-component and if you study the right graph you can see the blue and pink markers are consistently lower than the red and green (ie, higher forward speed as the axis is inverted). It isn’t convincing but it supports test 3 and indicates that the problem mainly lies with the forward velocity component (as you’d expect). This, again, suggests measurement error—perhaps the point of measurement is a couple of feet later, which would account for the difference.
Take 5: Break
Finally let’s look at break. The parameters plotted are pfx_x and pfx_z. These are the X and Z components of break and represent the difference in movement between the actual break and the expected break given no spin. Our hypothesis is that this should be mostly random as the measurement is park agnostic (the atmosphere is more or less a constant in both measurements).
Again, the picture is fuzzy although one can start to imagine that the different games form marginally separate clusters (the light blue dots seem to be a shade to the left, on average, of the dark blue dots). This lends support to the calibration theory but it by no means concrete evidence.
It is difficult to draw 100% firm conclusions but based on Millwood’s exploits in five different parks I’m reasonably comfortable asserting that there seems to be speed and location bias between the parks. This is a worry as one of the attractions of computer based data is that measurements are made more precisely. However, MLB is only in the roll-out phase on EG so it isn’t surprising that there are some consistency issues. I only hope these are resolved once every park is on EG.
Finally, if you think you have better explanations for some of the phenomena observed then don’t be shy and
References & Resources
A big thanks to Joe P Sheehan who helped me understand some of the EG parameters and also some of the graphings.