SBG: GH introduces himself below, but I wanted to say that I always enjoyed his Timberwolves blog. I'm glad to have him come aboard.
Hi all! I figure I should introduce myself for those of you who aren't familiar with me. I am GH (short for Greek House) of the now defunct "GH and Petey's Timberwolves Blog". SBG offered me the opportunity to write on his blog here a year ago, but it was just before I moved to Iowa. Once I got here, I had no real access to MN sports--a situation that I have since rectified.
I was reading ubelmann's fine post on baserunning the other night and felt like commenting. However, I couldn't find a way to post a short comment without writing what would amount to a short essay on the subject. I couldn't stop thinking about the subject, so I decided to actually try and work out some of the numbers and see what happened. I came up with a general formula which tells you when it's a good idea to try and take an extra base. Some of the numbers were familiar and some were surprising. I ended up writing a short research paper on the subject. Here is the link for those of you that are interested in the analysis.
The most surprising thing of all was that with 2 outs and a man on third, that runner should try to score as long as he will be safe a mere 35% of the time! I also did some analysis of the Lew Ford play from the other night, which I imagine is what ignited the idea for the original article. The gist of the analysis is that Lew Ford should attempt to score on that play if he can do so with less (probably a lot less) than a 41% success rate. I try to keep the article objective and not introduce opinions into it, but based on the analysis, Lew was absolutely correct to try and score on that play. I would have to watch the video again, but I believe that he's safe there probably around 50% of the time. --GH
GH: I look forward to reading the full paper. Quick correction: I think you meant "less than a 41 percent failure rate" above (or perhaps more than a 41 percent success rate; one of the two).
Oops...I don't think I said what I meant to say...and what I meant to say is a bit confusing.
What I had intended to say was that 41% was an upper bound for the minimum success rate needed for him to attempt to score. I don't have an actual value for what is needed for this particular situation. So for instance, the actual success rate required in this situation might be 35%. If Lew can score 35% (or more) of the time, then it's beneficial for him to attempt to score. Hopefully this makes some sense.
I have nothing of substance to add other than compliments on the pun-tastic (perhaps unwittingly so?) article title. Why, just this past week, when Bartlett got picked off, when Cuddyer got thrown out at third base, and when Ford got nailed at home by a mile, I yelled the title of this article out loud each time!
Hehe...no I didn't intend it initially, but did realize the pun after I came up with the title. As I said in the article, I don't think the decision to send Ford was really that bad here. I think the moral of the story here is that MLB teams need to be more aggressive taking the extra base when that extra base is home, but less aggressive when it's any other base.
Nice work, GreekHouse. I especially like your typesetting. Do you have a background in math, stats, or physics?
I think that your breakeven probability is a really good way to deal with early game situations where you generally just want to maximize the total number of runs scored.
One way to deal with the "Lew Ford Fiasco" and the importance of Ford's run (since it was late in a one-run game) is to check out Tangotiger's run frequency matrix. The data's old, but like you mention, we're just looking for a rough handle on the data anyway. With runners on 1st and 3rd (where Ford and Casilla would have been if they'd been more conservative) and two outs, teams failed to score 71.5% of the time. So if Ford and Ullger figured that they could score more than 28.5% of the time, you could consider that a good decision.
In his career, Tyner has 25.7% H/PA average, which is actually pretty close to the 28.5% you get from the average data. Tyner is mainly a bad hitter because he has zero power, and this was a situation where power wasn't necessarily very important, so I don't think that the Tyner adjustment would be very large (if you are only interested in whether or not Ford would score.)
Hey ubelmann,
I actually have a Master's degree in computer science and am currently a math PhD student at Iowa State (which is why I moved to Iowa).
I like your way of dealing with this situation. If you wanted to be really accurate, you could look at the win rates of teams when they're ahead by 1 run, 2 runs, etc. and use these to weigh the value of each further run that is scored that inning. Then, you could use the run frequency matrix to determine a probability that takes these weights into account and use that to help come up with a more accurate probability. In general though, those extra runs won't really add much since they are already relatively infrequent and by assigning them a lower weight than the first run, you reduce their contribution to the overall equation even further. So I think just looking at Ford's run as being the only important run in this situation would give a pretty good estimate.
For these methods, I don't believe that an accurate number is really that important anyway since the players and coaches won't be able to judge these things with any degree of accuracy anyway. The important thing for them to know is what the approximate probabilities are so they can make better judgments on the field.
I actually have a Master’s degree in computer science and am currently a math PhD student at Iowa State (which is why I moved to Iowa).
Did you start writing stuff in TeX in CS, or only after you got to math?
The important thing for them to know is what the approximate probabilities are so they can make better judgments on the field.
Right. I think a lot of people might be a bit surprised that a 25%-35% success rate is a good enough success rate in this case.
For these methods, I don’t believe that an accurate number is really that important anyway since the players and coaches won’t be able to judge these things with any degree of accuracy anyway.
This is true for a lot of things, but I wonder if they could do pretty well as far as stealing bases and tagging up for extra bases goes. It seems to me that every mananger/base coach should know how long it takes a baserunner to go 90 feet (and what the fluctuations around that time are.) Then, you could have scouts time throws from different outfielders, and figure out what the mean and standard deviations are there. (And you can keep track of how often their throws are on line.) With catchers, you can check what I believe is called their "pop time"--the time from when the ball is in their glove to when they release it--and the time it takes them to throw down to second base. With pitchers, you can check their delivery time to the plate.
Personally, I don't have the resources to collect/keep track of all that data, but it seems like a major league baseball team could. That should give them a pretty good idea which runners can steal on which catchers and which baserunners should be allowed to tag up for home on which outfielders.
Did you start writing stuff in TeX in CS, or only after you got to math?
I learned TeX during my last year of grad school in CS. My advisor wanted me to learn it before that, but I was too lazy. Eventually I decided that it would be nice to have type written copies of all my work and so I decided to learn TeX. I'm pretty sure that I'm proficient enough at it now that I can write up an assignment as quickly using TeX as I could if I did it by hand. Plus, it looks a lot nicer and gives me a nice electronic copy of everything I do.
This is true for a lot of things, but I wonder if they could do pretty well as far as stealing bases and tagging up for extra bases goes. It seems to me that every mananger/base coach should know how long it takes a baserunner to go 90 feet (and what the fluctuations around that time are.) Then, you could have scouts time throws from different outfielders, and figure out what the mean and standard deviations are there.
This might be possible, but you might have to make some attempt to categorize outfielders based on what sort of arm you think they have in order to get a large enough sample to make accurate assessments. Of course, there would still be some guess work since the base coach would have still have to estimate the depth of the outfielder and make further inferences based on the manner in which he catches it (charging the ball, running backwards, diving, etc.). You could probably do this an get a general idea of how deep an outfielder needs to be based on these things in order to attempt a score and modify this depth based on the particular runners and outfielders involved.
It certainly seems like a worthwhile project if you're an MLB team though. In particular, it would have a lot of value to a team like the Twins that doesn't have a lot of money since they need to squeeze every ounce of value out of everything that they have.
the problem is that in order to get meaningful data, you'd have to either actually run on the OFers, or at least credibly bluff, in order to get them to throw the way they would in the strategic situation of interest.
plus, as you note, the added variable of "what vector is the OFer on when he catches the ball?" that's a lot of explanatory factors to account for before making a split-second decision as a 3b coach. Nowhere near as simple as stealing second based on a pitcher's time-to-plate.
the problem is that in order to get meaningful data, you’d have to either actually run on the OFers, or at least credibly bluff, in order to get them to throw the way they would in the strategic situation of interest.
Or you can let players on other teams do that for you.
plus, as you note, the added variable of “what vector is the OFer on when he catches the ball?†that’s a lot of explanatory factors to account for before making a split-second decision as a 3b coach. Nowhere near as simple as stealing second based on a pitcher’s time-to-plate.
While not as simple as stealing second base, you should more or less be able to establish a lower bound for an outfielder's time to home plate, and generally just give the third base coach a more quantitative idea of just how good each outfielder is at getting the ball to the catcher.
by the way, GH, a hearty welcome. What this place needs is more BS
Of the half-baked variety, for sure.
One small quibble with the paper, GH. And no, I don't have a solution.
The 2003 run-expectation table is for observed data. That is, it includes strategic behaviors based on expectations. You treat that data as though it were independent of the risk/reward calculations of interest. But it in fact includes risky behaviors by the actors in 2003. Some 2003 runners trying to score from 1st, 2nd, 3rd were thrown out at the plate or another base, whereas others were not. If NONE had engaged in risk-taking behaviors, both the conditional outs and the conditional runs would have been different, but it is hard to predict how those numbers would have changed.
like I said, a small quibble. worth a footnote at best.
This is a good point. However, the general theory (from section 2 of the paper) is correct. The expectations do exist in all cases. If you could some how model each particular situation, you could use linear programming to find the ideal strategy. Short of that, you could do a lot of estimating to get a reasonable upper bound for your probability and only run in situations where you exceed that upper bound.