I remember many years ago that Bill James essentially wrote that when you develop a simulation, by definition the constraints (or lack thereof) that the author places on it will drive the results. While he wrote these thoughts in the context of determining the best team of all time, I believe that it could also apply to this discussion.

George Brett was a great hitter who had a 1980 season that was statistically an outlier for the rest of his career. The "rules" that one sets in trying to assess his true level of ability and simulate results will drive the outcome. It's impossible to have a simulation in a vacuum. I'm certain that he was both better (age 27 is often the best season for hitters when you exclude the impact of steroids) and luckier (higher BABIP) in 1980 than he was in other years.

I'm less convinced that '41 TW was as much of an outlier as was Brett in '80. TW had a higher BA for his career and had several other seasons of hitting over .340, including .388 in 1957 at age 39.

Posted by contrarian23 on 4/29/2016 9:38:00 PM (view original):
Well, the posts in this thread have suddenly become entirely unproductive.

I appreciate the earlier discussion. Thanks to those of you contributed.

FWIW, I might have more productive comments to make once I mull this over some more.

As the Go Go's would put it "Pay no mind to what they say Miguel Dilone hey hey"

I write this late Sunday night, read this thread late last week,and liked the philosophical aspects of skill v. luck. I recently drafted 98 Shane Spencer and was pondering a related philosophical issue. Prior versions of WIS projected 98 Spencer out for a full season of 373 AVG and 900 SLG, because the algorithm didn't consider PA/162. Insane. Was he really that skilled? No. He was lucky for 73 PA.

But.... Yes, because for 73 AB in 1998, Shane Spencer got fastballs just where he wanted but he was looking for them and got solid wood on the ball. Kinda just like George Brett did for 117 games in 1980 to hit 390.

I understand the reasons behind the statement that Brett wasn't a 390 hitter (much less a 454 hitter) but it's still a strange statement. I'm not a 390 hitter. You're probably not either. But George Brett was in 1980... for 117 games. He wasn't as tough as Nap and Ty in 1910 duking it out on the last day of the season, but they had softer balls, worse mitts, and few non-white athletes. They also didn't hit 390 that year. But in 1980, George Brett was a 390 hitter for 117 games. For sure, lots of skill AND luck went into it, his eyes and swing were on, the pitching and defense in the AL Central that year was just so, he had the right amount of pine tar on his bat, (and if you want to get philosophical, the wind eddies caused by butterflies flying out windows in rural Kansas (i.e. field conditions)).

To break it down to skill vs. luck, loses the moment. Brett needed skill and luck to hit 390 in 1980. To assume he hit 390 due to luck or a statistical anomaly and he was really only a 335 hitter assumes too much. How do we know he didn't have the skill that year to hit 420 but got unlucky and only hit 390?

Good question contrarian . All I can say for sure at this point on the subject is that if you draft him on a what if team, and you don't want to hit him 9th or get some 70-80% fatigue, you're gonna need someone at 3B off the bench.

I totally agree with you legaldh, I may have been overthinking the issue in my earlier posts. Plus it was probably bias on my part because as a Tigers fan I f-ing hated Brett.

I've reread the thread, including my own comments, and I've had more thoughts over the years, and want to add thoughts now...

I think a simulation should be able to respect both the fact that Brett is a .390 hitter and the fact that that is an outlying year. In other words, there should probably be some "credit" for the actual result even if we use an ability curve similar to what you had earlier in the thread.

But I think the important thing to bring up, that hasn't been brought up yet, is that there's another underlying assumption that simulations are making that I'm not sure is true: that batting average quality is linear. I am not at all convinced that it requires the same improvement (in real life) for a .335 hitter to become a .345 hitter as it does a .390 hitter to become a .400 hitter; in fact, I think the latter is MUCH harder, in fact, it is probably exponentially harder.

So if the complaint is "a .390 hitter can just as easily become a .450 hitter as a .330 hitter in a simulation, and that first one isn't realistic at all", well I think that's an error of the simulation. Without really any hard numbers to back it up, if we start from .390, I will sort of pull out of my *** estimate that it should be about as easy to get to .330 on the low end as it is to get to just .405 on the high end. That is, if the probability of a .330 or worse season is 10%, then even a mere .405 or better should be that same 10%, even if the "average" season is .390; I should see a lot of .391, .392 type seasons on the "high" end of the spectrum simply because it's that much harder to do.

That being said, the reason simulations don't do this is that there's not really a good way to accomplish this without building in consideration of past results, in the simulation itself, which is not something that simulations are good at doing or like doing.

The general statistical sense for an "unusual" event is a 5% tail unless you're in a field that needs a tighter definition for some reason, so I'll use that here. I would say that we should feel free to start with an ability curve, but as we do it, make sure that the actual numbers are not actually "unusual" events. Thus, Brett's 1980 season should start in simulation at no worse a number than that keeps the probability of a .390 season at at least greater than 5%. (As I said, the 5% isn't necessarily the only option for our tail, but we should figure out what we want it to be and use that for all players.) The exact place we'll need to start Brett, then, will determine on how we set the exponential curve. We might need to set it closer to .390 than you've set up here so far if we use this exponential curve.

Note that the exponential curve is not really just the bell curve. Doing that we'd just need to set it no more than 2 standard deviations away from the mean... but to even get that normal distribution we more or less need a linear dynamic, exactly what I want to get rid of here.

It's also possible, of course, that everything I said is simply too hard to implement. If we don't want to do this, than I think the previous pages cover the material pretty well.