THE BOOK--Playing The Percentages In Baseball

That is a great article.  Simple, accurate, easy for the average fan to understand.  Gets the point across quite well.  I like the part that says something like, “Intuition won’t allow you to figure out when to walk Bonds or not.” I say that all the time about most strategic decisions in baseball.  You would think that a reasonably smart manager or GM would say that to himself - “Gee, I don’t know the answer to that (whatever ‘that’ is), but I’m sure that we can pay someone to find out the answer.”

Thanks, Mitchel.  You’ve been extremely helpful with many of my stories in the past, and positive feedback from you means a lot.

Dan

Thump up to this article. I wonder why Barry himself seldom mentioned his ridiculous enormous walks are such huge contribution to Giants… maybe he considers sabermetrics as crap of outsiders, too…

This article illustrates a complete misunderstanding of what Win Expectancy means.  To take the Win Expectancy value that is based on the chance of scoring given a theoretical lineup of average players and claim that that value is a “mathematical certainty” of the chance of scoring when Bonds is batting is just idiotic.  Intuition would be better than this pseudo-mathematical “proof”.

You are right the word certainty cannot be used here.

However, it’s fairly easy to get a win probability table with any batter up, given the rest of the lineup is average. 

While I’m sure you know this, I’ll explain for the others benefit: you calculate the win probability of winning with an all-average lineup by figuring:
1. the frequency of each event (3% HR, 16% singles, 9% walks, etc)
2. for each event, the frequency they will transition into each of the 24 base out states (plus if any run scores)
3. for each game state (base/out, inning, new score differential), the win probability given an all-average lineup
Multiply 1x2x3, and figure the weighted average.

With Bonds at bat (as a PH), you simply change #1.  You could change #2 (maybe a Bonds double clears the bases more than average).

Now, since Bonds is not a PH, he’s got an effect in future innings as well, so #3 above (based on an all average rest-of-game lineup) won’t apply as strongly.

Peter are we on board with all this?  Or do you have an objection outside of this?

Your explanation is correct as stated but the problem is that the rest of the lineup is not “average” nor are the pitchers.  The rest of the SF lineup had a batting line of .261 BA,.316 OBP, and .402 SLG which would have put it 23d of the 30 teams.  Right handed pitchers walked Bonds at a 50% higher rate than left handed pitchers, appropriately since Bonds had a 50% higher home run rate against righties.  Although I haven’t calculated it, there is presumably also a selection bias of the quality of pitchers that chose to pitch to Bonds and those that chose to walk him.  So it is highly likely that if those pitchers who walked him had actually pitched to him that he would have performed at a rate much higher than his existing stats.

The bottom line is that Win Expectancy, although useful as a starting point for a generalized discussion of strategy, has no value in determining strategy for specific situations.  The factors are much too complex, and this is particularly apparent when trying to evaluate strategy for someone like Bonds, whose performance is at the extreme end of the spectrum. 

I like to see sabermetrics in the mainstream media as much as anyone.  But when a useful measure like WE is misused it makes us an easy target for critcism and it is too easy for our detractors to dismiss all sabermetric analysis.

Peter, I think we’re on the same page here. 

There’s no question that the batter/pitcher matchup has to be used (notably the handedness aspect).  So, if Bonds is a “true” +.020 wins / PA player in 2002-04 (in a random situation), that he’d be say +.016 against LHP and +.022 against RHP (or whatever it is).

And in higher leverage situations (say LI=3.0), that those numbers are roughly 3x higher (though again, you’d get better numbers knowing the exact game state, and his exact frequencies against each hand).

As for the overall team totals, you could summarize your position as to “what happens in a game where you have a .550 team playing against a .450 team?”.  It’s not just the Giants offense, but their pitching and fielding, the same for their opponents.

The question therefore to research is: how much of a difference in in-game strategy does all this make?  That is, if the Giants have a .450 (or .400 or .600) chance of winning to start the game, how does this alter the odds of whether to walk Bonds or not?  And, how much do the odds change once you make sure that everyone gets to come to bat one batter earlier?

These questions are easily answered with a sim.  I usually find that when MGL (sim-based) and I (simple Markov based) try to answer these questions, that we end up getting similar answers.

Your objections are legitimate.  Your conclusion (as would be mine or anyone else out there) however (intuition over simple math) is still debatable.

I should also note that the author used my charts here:
http://www.tangotiger.net/walkbondschart.html
http://www.tangotiger.net/walkbondschart2.html

Rather than say “walk/don’t walk”, and realizing that this is a simple model, I used 5 different categories:
http://www.tangotiger.net/walkbonds.html

The “Go With Gut” category I used with a certain margin of uncertainty.  That is, I think I used a +/- .020 win difference, that if something was within .020 wins (which is fairly sizable) that I allowed enough uncertainty (handedness, matchups, etc), that it’s not necessarily clear what to do.  But the “Walk Now!” and “Do NOT WALK” would have a gap that was wide enough (I forget what it was… at least .050 win difference, probably more), that it would be too hard to construct a scenario where you would want to do the opposite.

That said, that was done in 2004, and there are some things I’d want to do (a little) differently.

Just to be clear, I am not saying that there isn’t a mathematical analysis that would not be superior to intuition, I am just saying that the analysis as presented in the article had no justification that it was superior.

And I agree that a good simulation is necessary to get closer to a correct analysis.  But a good simulation is incredibly complex.  As I mentioned in my previous post, there is no assurance that Bonds would not perform at a much higher rate against those pitchers that are currently walking him if they chose to pitch to him instead.  I looked at the difference in overall quality of the pitchers that walked him and the pitchers that didn’t and it is statistically insignificant but that doesn’t mean that there aren’t meaningful differences in fly ball rate or pitch consistency that would be important.  What happens during a specific AB is very complex to model.  If you are using a simulation to predict the results of a game or an aggregate of ABs over the course of a year than you don’t need to include these individual factors because they tend to average out.  That is the reason that your Markov and MGL’s sim can agree much of the time.  But if you are looking to predict the results of a single AB or a small group of ABs then these factors become much more important and the sim has to be much more complex.

I think we are not only on the same page, but using the same words.

I agree that modelling matchups is complex.  For many years, I’ve been talking about the silliness of doing “park adjustments” that apply equally to all.  When Barry Bonds specifically faces Gregg Olson specifically at Giants Stadium specifically, whatever model you have, will come with a certain level of uncertainty.  Furthermore, a matchup is not a 1 event situation, but made up of mini-events (pitches), and therefore, you have a count to consider and model.  It is definitely a complex animal.

"What happens during a specific AB is very complex to model.”

No it isn’t.

“If you are using a simulation to predict the results of a game or an aggregate of ABs over the course of a year than you don’t need to include these individual factors because they tend to average out.  That is the reason that your Markov and MGL’s sim can agree much of the time.  But if you are looking to predict the results of a single AB or a small group of ABs then these factors become much more important and the sim has to be much more complex.”

That is crap.  Has someone invaded your body?

Tango you are being way too solicitous.  It is easy to figure out with a sim or without whether it is correct to IBB Bonds or not or too close to call.  Having those 3 categories is the key, because all of things that are difficult to model will fall into that slop in between the walk or don’t walk. It just depends on how big you want to make the “too close to call” category.  If you have a good model or sim, it does not need to be that big.  Intuition will get you nowhere (other than the obvious ones) unless you are some kind of savant, which is possible I suppose.

MGL - I’ve been following the discussion over in the Bochy thread and given what you have said there I don’t know what differences we have that could cause you to resort to the type of personal attack that you used here in your post #11.

In the Bochy thread you stated that your sim is “very complex” and includes factors such as weather, umpire bias, park effects, speed factors, etc.  In post #10 in this thread Tango correctly observes that an individual AB is made up of the mini-events of each pitch.  Surely you would concede that including all this information in your sim makes it much more complex than the type of Markov that you describe in your book that is based on league average values alone.  It is my understanding that the Win Expectancy values used in the above article were only based on such a generalized Markov.

So why is it wrong for me to conclude that a detailed sim such as yours will be more likely to agree with a generalized Markov when data is aggregated and less likely to agree on the results of a specific at bat?

For what its worth, I agree with your position in the Bochy thread.  The best way to evaluate strategic decisions is to develop the best sim possible and go with the results of that sim whatever they might be.  You obviously have developed a very complex and detailed sim and I’m willing to concede that it is probably the best sim being used to evaluate strategic decisions.  The really interesting discussions are going to occur when more people develop equally complex sims independently but use different assumptions than you do and come up with different results.