THE BOOK--Playing The Percentages In Baseball

I just got a note that the team reaches base on error 3 times a game, with about 2 baserunning outs per game, along with 1 WP per game.  I’ll have to rework everything, since this will clearly impact what I’ve just done.

I certainly am not surprised that a softball league with those odd stats, plus the RBOE activity, does not conform to BsR. Yes, BsR is a “fundamental” type formula, but it certainly carries MLB assumptions about baserunner movement. The weights in the B factor are empirical, not theoretical. I mean, if you applied those B weights to little league, I am sure it would be off by a mile.

The surprise, for me, is that anyone would think that those weights are somehow inherent or universal. It’s certainly possible that the “structure” of the formula is universal, given the same rules of play. But the weights certainly are not. And that applies to the triples problem. The linear B factor is also just a model of a more complex function. If someone wants to solve that problem, that’s great. But then, how will those changes apply to a softball or little league?

I think the more useful information to take back to the league from this analysis is the value of an out.

At -0.84 runs, it is imperative to get any out possible on defense and avoid outs on offense.  I tried to stress to this an over 30 league I played in.  It had a high run environment (about 7 runs per game per team) and giving up bases and not gettings outs were huge.

Not being real conversant in baseruns - how important is the 3B factor?  In most softball leagues, even more than MLB, the triple is rare due to short fences.  Is the frequency of an event refelected in the linear weight - or equivalently will the low frequency not affect comparisons to actual data if the weight is appreciably “off”?

Re: Dufman, it’s impossible to teach non-mathematical types to appreciate the value of outs.  You will still see them running themselves out of inning after inning, or trying hopelessly to get the lead runner out, even in a 25-21 slugfest.

I’m under the impression that many of these statistics break down in extreme environments, especially if they use ratios or don’t adjust for context.

I re-ran using the new data:
1 - Reached on error was added in (3.4 per game!)
2 - Sac flies were included (I’m assuming the “sac” category supplied was SF)
3 - WP, PB, BK (1 per game)
4 - NO fudge factor to the “B” component

Results?  BaseRuns, using “MLB assumptions” predicted 678 runs in 54 games.  Actual? 668.  Bingo!

The LWTS values are: .73, 1.07, 1.51, 1.49, .54 for the basic hitting categories.  RBOE, WP are .75, .29.  Run value of the out is -.89.

The triples value is of course problematic, but no so egregious this time.

The most important aspect is the out value, which is enormous, as we know.  Batting outs, and even worse baserunning outs, are terrible prices to pay.

I think the players know this, since only 20% of the hits are for extrabases (MLB is 30%), with more triples than HR. 

Later this afternoon, I’ll produce the LWTS values using a simple-Markov process.  This won’t have an issue with the triples.

This reminds me of something similar did w/ Wiffleball leagues whose stats I could pull off of the web.  I pulled random leauges off the net, based on which stats were available.  I was mostly looking for traditional wiffleball, a 2 person game, with no base-running, 1 defender, and fixed advancement (1B = 1 base advanced, 2B = 2 bases, etc.). 

Several leagues I looked at (sorry, dont have R/G available right now) fit best to OPS and basic RC.  I didn’t try very hard to fit them with Base-runs compnents, but found OPS and RC more accurate (without modifying the B component, so I probably could have improved the accuracy)

I do remember another league (again, don’t know R/G off top of my head), which had baserunning, 3 outs per inning, multiple defenders, and multiple batters, and it was a nearly exact fit plugging in the basic Base-Runs formula with no modifications, more-so than OPS or RC.

I think I saved the info to a spreadsheet, I’ll have to look when I get home.  I wasn’t so interested in modelling a specific league, as much as I was testing the fixed advancement, which I assumed would make wiffle ball easier to model.  Run Creation in traditional wiffle should model easily and perfectly in a simulator because of the fixed advancement, all you should need is 1B%, 2B%, 3B%, HR%, BB%,outs per inning.

A simple-Markov approach gives us these run values, with the MLB version in parenthesis:
1b 0.85 (0.49)
2b 1.12 (0.77)
3b 1.30 (1.07)
hr 1.52 (1.48)
bb 0.73 (0.37)

In both cases, the assumption is that there are no outs on base, nor any basestealing, and that the “runner advancement on hits and outs” is the typical MLB patterns, and the the frequency of each event is independent of the base/out state.  The model is 100% pure math, meaning, it’s “perfect” under these assumptions.

As we can see, the gain in singles, doubles, and walks is virtually the same between the two leagues, 0.35-0.36 runs.  The gain in triples is only 0.23 runs, while the gain in HR is .04 runs.  What this means is that we’re still in the upward slope of the marginal run model.

Remember, at some point, all the run values will converge toward 1.000, meaning that the run values of the extra base hits will be lower, the higher the run environment.  We haven’t reached that point here.

So, if you want to use some LWTS values for your softball league, where the OBP is .500, and 20% of your hits are for extra bases, use these above values.  Set the run value of the out so that your totals come out to zero.

(Note: the simple-Markov approach estimated 13% more runs than actually scored for the softball league.  The reason is because of the baserunner outs not being part of the model.)

So, BaseRuns does a great job here, just blowing it on the triples.