We get a fair amount of questions about tRA via email, and chances are if one person's asking about something, the question's going to come up in the future. So in the interests of less work for me, I'm opening up a tRA questions/answers thread here. If you've got a question, just leave it as a comment, and Matthew or I will answer*.
In the interests of keeping things tidy, please don't answer other people's questions. The thread might get overly messy if we have things jumping back and forth.
Cheers,
-G
*Probably me.
In the interests of keeping things tidy, please don't answer other people's questions. The thread might get overly messy if we have things jumping back and forth.
Cheers,
-G
*Probably me.
Here's a few:
* Your explanation for tRA lists a way for figuring linear weights above average, and then goes on to list LWTS values (and then pitching values) that look like absolute runs. Either I'm missing something or you skipped over a step in your explanation.
* Have you done any tests on tRA's accuracy, either same season or as a predictor for future seasons?
* Why a linear model rather than a dynamic model (i.e. BaseRuns)?
Colin:
* Can you rephrase? The linear weights we use are simply the average run and out values for a given event in a given league each year. There's no above average component, and so no need to convert into a pure runs/outs number. I'm not entirely sure that's the answer you were looking for though.
* I tested tRA's expected runs and outs for 2007, using 2006 runs/outs data, and was within 7 runs and 15 outs for the entire season. Since we now calculate runs and out values season by season and league by league, these numbers shrink to essentially zero.
I've been meaning to examine how tRA* correlates year to year with both R/9 and tRA, but have never gotten around to doing it, due to a nasty combination of business and laziness. We'll set one up at some point, and they'd better show a strong correlation or I'll feel very silly. I hear DIPS 3.0 correlates extremely well with R/9, though, which is a positive sign because it's extremely close to tRA in methodology.
* At first it was because I had linear weights data and not dynamic, and I was having to use other sources because I was calculating tRA by hand. Now it's because we don't really think that the extra effort to add dynamic weighting in is really worth it - we redo the linear weights for each league/year, as I've mentioned above, and so we don't have to worry about bad run environments messing with our weights. The cost is, of course, overrating bad pitchers and underrating good ones, but for the amount of extra work required to switch our weights over to a baseruns style system... we'll accept that. It's not as though it's a massive difference given reasonable sample sizes (at least I can't imagine it being one).
Not really related to tRA but I did some pitching projection models a few years back based on pitch by pitch data; essentially trying to predict FIP based on swinging strike/called strike/ball/etc rates and found that the prediction power of something like that was noticeably improved over most systems in use at the time.
Given the similar nature between the models I was using and how tRA is built (really the only difference in the granularity level, which again is something we're going to investigate going forward), I would imagine we should strike upon similar results about the predictive nature of tRA*.
Let me see if I'm being clear here. This is the way that you say you're calculating your linear weights:
play_run_value = runs_scored + (run_expectancy_after - run_expectancy_before)
But when I do that for, say, 2007 (using a RE table derived from 1994-2007 PBP data), this is what I get (OTH standing for "other," mostly baserunning events):
EVENT NUM LWTS
OTH 7235 0.109
BB 14756 0.324
FB 34396 -0.102
GB 62495 -0.090
HBP 1755 0.350
HR 4957 1.408
K 32189 -0.295
LD 25846 0.326
IFB 10891 -0.275
Which is because the formula you have presented is a formula for linear weights above average, which should sum to zero at the league level. But the values you show are in absolute runs, as is tRA.
I did a study a while back on how ERC/FIP/DIPS performed in predicting future ERA, as well as going into how to adapt BaseRuns to a FIP/DIPS mentality. It may or may not be of interest to you.
Ah, I see what you are talking about now Colin. Yes, we left that part out of the explanation on the website. That primer was adapted from one presented at Lookout Landing and the goal there was to keep it comprehensible for the layperson. So we trimmed parts of the math we thought inconsequential to their understanding of the basic framework.
I'll re-write it more thoroughly for the website and also take a look at the study you just linked and see how feasible it will be to adapt to a dynamic model.
Why does TRA* only account for yearly, not career stats?
tRA* is not intended to be a projection system, so there's no real reason for it to consider anything but yearly stats. If I were trying to project things it would of course consider career stats in the regression. That would be a lot more challenging to develop, though.
Are the minor league tRAs MLEs?
It is not out-of-date information? Because I have other data on this theme. http://video-online-go.ru/map.html
Minor league data on the site are not MLEs, no.
提供灌装封盖机设备, 灌装封盖机产品, 灌装封盖机设备
mbjxvio knwi zwayhqc rjovkhli vixz cvijzytua enmagfodb
omaz nhiltzfu celv pgxvqr hlpxc stwoijy mxni http://www.koibfsl.vltpfsc.com
omaz nhiltzfu celv pgxvqr hlpxc stwoijy mxni http://www.koibfsl.vltpfsc.com