THE BOOK--Playing The Percentages In Baseball

Well this kinda takes the thunder out of my annual piece.

Nick, if you’ve done any actual work on it, your thoughts and results would be valuable.  As far as I can tell, Dr. Thomas only tracked two innings worth of data.

I think his paper has been discussed on this site previously:
http://www.acthomas.ca/papers/catcher-project-report.pdf

At the time he was looking for funding in order to proceed with any additional research.

Is this guy collecting data?  It says on the web site that this is a “test of catcher spotting ability?” I don’t get it.

In addition, I wrote this to him:

“You need to explain what the red and black squares on the catcher outline are.  Is the red supposed to be the strike zone? If yes, what is the black square?  How far from the red square is the black square?”

Tango, where did this come from?  Who is this guy?

It seems like they might just be testing these modes of input to see which is the best.  For one thing, they need to run the video twice, once for the location and another time for the actual location.  And I think it would help if it were in slow motion (not sure).

And can’t they use video image identification techniques to do this automatically?  It is pretty clear where the catcher’s glove is and where the ball ends up. Maybe it is trickier than I think, especially if the catcher moves his target for deception or the batter hits the ball.

MGL, if you read the link I posted, as well as the main http://www.acthomas.ca site, most of your questions will be answered.

And can’t they use video image identification techniques to do this automatically?  It is pretty clear where the catcher’s glove is and where the ball ends up. Maybe it is trickier than I think, especially if the catcher moves his target for deception or the batter hits the ball.

Nick can probably elaborate more than I can.  However, I didn’t notice a problem when the batter hit the ball, nor did I particularly notice the catcher moving for deception much if at all.  However, picking the catcher’s glove up out of his shin guards was sometimes pretty tricky to do.  They’re both dark colors.  If the angle of his glove was right, you get to see the outline of the glove, but if he turned it just so, it was dark on dark and very hard to pick up.  A human would have an easier time with that than a simple video ID technique.  Of course, with more money, you could make a more complex video ID technique that should be able to handle that.

Basically what I did was, using MLB.tv, track the approximate catchers glove position at the point of release (using the St. Louis home game cameras - perfect dead on angle, go Cards!), and used the pitch f/x location for where the ball ended up (so that won’t be affected by a batted ball).  It’s a pretty easy process once you get going (pause the video at the point of release, chart catcher location, rinse and repeat) and since there are really only a few different glove positions the catcher goes to, you don’t really have to worry about precision on each pitch. 

I’m sure that there is some way to automate the process a la Pitch f/x, however, I think a good solution would be to have the stringers enter the glove locations (of course this would be almost impossible with some of the non-centered cameras).

Nick - BIS is already tracking target position, presumably by having the stringers enter the glove location just as you suggest.  Since MLB Gameday stringers don’t view the play on camera entering target position would not be possible for them.

A Gage R&R seems like a perfect test to perform on these test methods and the stringers to tease out where and how much error can be attributed to the method and how much to the stringer.

Hi all, I’m the author of the original work (thanks for posting it, Tom!)

Generally: the purpose of the applet is to test the three particular modes of human-coded input, namely with an observer who’s watching a game in real time, like the color commentator might do. The real question of interest is: how many different trackers/stringers would it take to get an accurate measurement of location? That way, we’d know how to farm the job out on a larger scale if necessary.

Also, I don’t know how well an automated system will do the job, especially when there’s no glove flashed—pretty much any time a runner’s on second, the “obvious” information is hidden. Hence the human element.

Nick: If you’ve got actual data, you’re well ahead of anything I can possibly do right now with actual results—I’ve been waiting to do more tracking until the reliability question was answered. I look forward to hearing what you’ve got, especially if you’ve put time into thinking about the data collection problem.

Steve: The R&R is the kind of thing I’m trying to do with the completed data, multiple measurements on the same observations. One thing I really want to know (and hopefully this won’t spoil my results) is how much the “telestrator” aids (or hurts) the measurement process.

Andrew C. Thomas,

On your website the links to the two hockey papers are broken.  There are a lot of hockey fans that read this site, and personally I think all of them should read “The Impact of Puck Possession and Location on Ice Hockey Strategy”, that’s just good stuff.  I think that after reading it, and performing a bit of back-of-a-cigarette-pack math ... a lot of NHL coaches are going to seem less foolish.

I haven’t read “Inter-Arrival Times of Goals in Ice Hockey”, though a quick google shows that it is posted at hockeyanalyics.com, for anyone interested.

Fixed the links to the hockey papers. Thanks for the kudos!

BIS is already tracking target position, presumably by having the stringers enter the glove location just as you suggest.

Yes, they collect this data in the same way (and at the same time) as the pitch location data. I’m not comfortable with the accuracy of their pitch location data, though:

http://www.insidethebook.com/ee/index.php/site/comments/strikeout_rates_and_thats_all_there_is/