Paul, I procrastinated too long to get that text file you posted with HOM ID's. Can you resend it to Rallymonkey (numeral for five) at comcast dot net?

Dan,

Where can I find your complete WARP file? I can only find the 1987-2005 file.


I can't find pitcher data. Is it there?

The Rosenheck DB will be up shortly (FTP client says 17 minutes). It's the older version though, but I'm not sure when I'll get to update it for his latest baserunning fielding adjustments (some time this weekend or possibly not until early next week) and didn't want to wait.

I'll post it in the main entry once it's up to date.

New, revised numbers coming soon! *does happy dance*

The new revised numbers are already out, Dan posted them like a year ago. I just need to incorporate them into my database, which has some good queries on all the data, etc. . . . that's what I meant . . .

No, I haven't published the pitcher data because it's so preliminary (and because I lost my spreadsheet until recently!). I can email it to you if you'd like, jimd, but it's based on a long-outdated version of BP's DERA at this point.

Cross-posting an example Larry Walker's 1997 season from his player thread, as a work-through example of how I calculate WARP for the 1987-2005 period:

Walker had 64 unintentional walks, 14 intentional walks, 14 hit by pitch, 109 singles, 46 doubles, 4 triples, 49 home runs, 4 sacrifice flies, 90 strikeouts, and 270 fielded outs. According to Baseball Prospectus, his baserunning was 2.5 runs above average, and he hit into 2.17 fewer double plays than a league average batter would have given his opportunities.

OK, let's bust out the eXtrapolated Runs estimator. UIBB and HBP are worth 1/3 of a run, IBB are 1/4. 1B are 1/2, 2B are .72, 3B 1.04, and HR 1.44. SF are 0.37, as are net DP (before counting the effect of the extra out they consume). To make total league XR equal total league runs scored for the 1997 NL, fielded outs are worth -.102 runs, and strikeouts are worth -.111 runs. So Walker produced (64+14)/3 + (14/4) + (109/2) + (.72*46) + (4*1.04) + (49*1.44) + (4*.37) - (90*.111) - (270*.102) + 2.5 + (2.17*.37) = 159.1 runs.

The average team in the 1997 NL had 4,171 batting outs. Walker consumed 4 SF + 90 K + 270 fielded out - 2.17 Net DP = 361.83 of them, leaving 3,809.17 for his teammates. The 1997 NL scored .1788 runs per batting out, and Coors Field had a park factor of 122, so an average lineup in Coors would generate .1788*1.22 = .2181 runs per out. .2181 runs per out times 3,809.17 outs yields 830.8 runs for Walker's teammates. Adding on his 159.1 runs means that an average 1997 NL lineup in Coors, with 664 PA replaced by Larry Walker, would have scored 989.9 runs.

Now, on to the defense. Walker spent time at three positions: first base, center field, and right field. In 25 innings at first base, all three metrics that show a statistically significant correlation to an average of PBP stats during the period they are available (TotalZone, Chris Dial's Runs Saved in playing time, and Simple Fielding Runs) find his fielding exactly average. In 13 innings in center field, both RSpt and TotalZone give him +1 (SFR is not available for the outfield before 2003), so he's credited with one run above average there. And in 1,235 innings in right field, RSpt has him at -2.4, while TotalZone has him at +4.1. The equation that provides the best fit to the PBP average in RF is .68*RSpt + .23*TZ, which comes out to -0.7. However, repeating this procedure on all the RF in the league produces an average that is slightly above 0; to zero it out, we have to subtract .0013 runs per inning, bringing his final RF range figure down to -2.3 runs. Finally, Sean Smith finds his arm to be 4.8 runs above average. I regress this figure 13% to the mean, because that provides the best fit to the Smith/UZRarm average for the years where both are available, reducing it to 4.2 runs above average. So his total defensive contribution is 1 - 2.3 + 4.2 = 2.9 runs above average.

A 1997 NL average team in Coors would score 746*1.22 = 910.1 runs. Walker's defense takes away 2.9 runs, leaving the Average Team Plus Walker's opponents with 907.2 runs.

A team scoring 989.9 runs and allowing 907.2 runs in 162 games has a Pythagenpat exponent of 2.016, meaning that the Average Team Plus Walker would win 88.1 games.

The 2005 standard deviation (which I use as a base) was 5.6% lower than the 1997 NL regression-projected standard deviation, so we pull Walker 5.6% back to the mean, down to 87.7 wins.

My methodology for determining replacement level (explained at length in my WARP thread) finds that an average team with a replacement player in Walker's playing time and mix of positions would win 80.0 games in a league with the 2005 standard deviation. Thus, Walker was 87.7-80 = 7.7 standard deviation-adjusted wins above replacement (WARP2).

I can email it to you if you'd like, jimd,

Thanks. It would be much appreciated.

but it's based on a long-outdated version of BP's DERA at this point.

I want to incorporate your numbers into my system for next year's election.
I need pitcher's data, even if it is somewhat out-of-date.
Whenever you update it, let us all know (via this page).

I also updated my email address (which had gone stale).

Thanks,
--JimD

Joe Dimino has requested that I update my WARP numbers to include the six seasons since I developed the stat. (Wow, it's been that long?!?!) The trickiest thing would be calculating FWAA. The problem is that back in the good old days you had play-by-play defensive stats from both databases, BIS and STATS, which to my knowledge are equally good. Now that MGL has switched to BIS, though, the only remaining stat I know of tapping the STATS database is Chris Dial's RSpt, which is based on brute Zone Rating rather than the more detailed STATS numbers. So what's the best way to approach this? I suppose I should do a historical regression of RSpt and Dewan on the STATS UZR/Dewan/PMR hybrid I used to use, and then use that equation except replacing Dewan with 50% Dewan and 50% BIS UZR? That will lead to more tightly clustered, lower-stdev FWAA for recent seasons than I had for 2003-05, however...

If you're changing FWAA methodology I would suggest changing the numbers for all applicable years. It might be time for an revised version, not just an extension of the previous version.

Is there any financial incentive for BBTF to host your WAR database? It seems like the WAR calculators drive plenty of page hits.

There is no consistent FWAA methodology, DL from MN--or rather, there is, which is to make the best use of the available data. To rehash, I arbitrarily defined an average of, if I recall correctly, 45% UZR, 30% Dewan, and 25% PMR as the "true" fielding measure for 2003-05. (I figured that the STATS and BIS databases were equal, but since I had two good BIS metrics and only one good STATS metric I'd collectively give BIS a 55% weighting, with a slight preference for Dewan over PMR). I the ran regressions on the other defensive stats--RSpt, SFR, TZ, DRA, FRAA, and FWS--against this value at each position for the 2003-05 period. That let me know how much to weight each stat and how much to regress to the mean at each position for pre-2003 seasons in which some or all of the preferred PBP metrics are not available.

Strangely, with MGL's switch to BIS and the death (to my knowledge) of PMR, we now know less about defense than we did in the 2003-08 or 09 "golden age." Keeping a consistent FWAA methodology would simply entail repeating the 1987-2003 process for recent seasons to come up with a best guess of what the 45/30/25 weighted average would be likely to tell us. However, I'm afraid this is also impossible, as I don't have databases of both STATS-based and BIS-based UZR for the same season to see how they correlate. I'm not quite sure how to get around that one...

I'm happy to put it up as many places as possible. If anyone at BBTF wants to host it I would be delighted for them to do so! Speaking of which, I have to get the link at Tango's site back up...

If anyone wants Dan's numbers in a manipulable format, I've got them in an MS Access DB that works great for me.

I've also written these queries:

position crosstab - lists players and their WARP by position

League leader WAR by season

MVPs

Pennants Added (career total)

Total MVPs - (Honus Wagner wins with 13. Barry had 10, Hornsby and Ruth 9, Mantle 8, Cobb, Mays, ARod 7 - through 2005).

Seasonal WAR for everyone since 1893.

Also have the data table for all of Dan's WAR by component (with my war credit add for certain players) and the master table from the Lahman DB in there as well and an extra column designating HoMers (hasn't been updated for awhile though).

Coincidentally, my reasons for leaving DanR WARP behind are touched on exactly by the sudden interest in FWAA. Basically, I'm not a fan of 'averaging' fielding results from a variety of systems. I'm not a fan of some of the systems used, so I think averaging them in fact mixes bad data with good data.

Fra paolo, do you have issues with Dewan's Plus/Minus, Pinto's Probabilistic Model of Range, or MGL's UZR? If not, then you should be fine with my 1987-2005 numbers, which are a statistical approximation of a weighted average of those stats. Pre-'87, I have to get around to doing studies on DRA, TZ and SFR alone for the Retrosheet era.

I don't see how anyone could have much confidence in any of the fielding metrics. Averaging them seems quite reasonable to me.

Fra paolo, do you have issues with Dewan's Plus/Minus, Pinto's Probabilistic Model of Range, or MGL's UZR?

Yes. It's been ages since I read how PMR is constructed, but I am not a particular fan of either Dewan's or MGL's metrics.

Philosophically, I think the Wins Shares system is constructed correctly, as it starts at the team level and works its way down to the players. The systems you name, if I am recalling correctly, build up from individual plays, although there may be subsequent adjustments to match team total data. Plus Colin Wyers' work on observer bias has undermined the trust we can put in these metrics. While he overstates the significance of his discoveries, I don't doubt that the impact is sufficient for us to add a note of doubt in the reliability of zone-based metrics. Fundamentally, we are dealing with zones that are too small, and with data that starts as too individualized.

In fact, I am a bigger fan of Humphrys' work than any zone-based metric. We know who fielded the balls. We only have to worry about where the balls that didn't get fielded went.

EDIT: I also rather like Dial's 'brute' use of zone rating, because an elegant side-effect of his system, which I'm not sure he has noticed (I have tried to tell him, but I have trouble conveying my meaning), appears to be to expand the zones. So I think his numbers are better than Dewan's or MGL's.

fra paolo, that doesn't mean you'd endorse Fielding Win Shares, does it? With its artificial cap and limiting of 1B to a range of about +/- 4 runs?

Dewan is not a zone rating.

He looks at each ball and compares whether or not the play was made to the league average. It is not like zone rating in terms of making plays in and outside the assigned zone and it's artificial penalties.

I got Wizardry for Christmas, just started to read it. Think you absolutely need to look at it.

-- MWE

I have a copy, just have to get through it. That said, I've been in regular touch with Michael for years, while he was developing DRA and I was developing my WARP, so I'm pretty familiar with it.

MWE, thanks for mentioning Wizardry!

David, some things discussed in Wizardry that might be relevant to the discussion in this thread.

Agreed that we lost a great deal when we lost STATS-based UZR. Given all the troubles with batted ball data, highlighted I believe for the first time in my August 2007 articles at THT, it is crazy that we all rely on one non-open-source database for batted ball data ratings. See also my article in the 2012 Hardball Times Annual.

If you are able to do some studies of DRA and TZ versus UZR, you should know that TZ shares two major biases that will result in TZ correlating 'better' with UZR. The biases involve overrating (underrating) bad (good) fielders playing next to good (bad) fielders, and overrating the impact of errors and error-avoidance. Also, in my rush to get the book done, I did not take full advantage of play-by-play data for seasons since the 1950s. I'm working on new numbers that pick up on using Sean's idea of using a pitcher's career ground out to fly out to estimate fly balls and ground balls with less distortions based on the relative quality of a team's outfielders and infielders each year, while avoiding several distortions that TZ reintroduces. This is all discussed in chapter three.

I think that in time DRA/Wizardry will shift the consensus career valuations for a couple of dozen historically significant players by between five and fifteen career WAR. Jeter in particular is grotesquely overrated as a fielder by UZR and TZ. The version of DRA in Wizardry suggests he is a best a borderline Hall of Merit player; new version of DRA developed using post 1988 Retrosheet data show unambiguously that he is not a Hall of Merit player.

Jeter in particular is grotesquely overrated as a fielder by UZR and TZ. The version of DRA in Wizardry suggests he is a best a borderline Hall of Merit player; new version of DRA developed using post 1988 Retrosheet data show unambiguously that he is not a Hall of Merit player.

That is very interesting. He is remarkably poor.

[Dewan] looks at each ball and compares whether or not the play was made to the league average.

I understand that Plus/Minus divides the fields into zones. That's why I used the term 'zone-based' rather than zone rating.

fra paolo, that doesn't mean you'd endorse Fielding Win Shares, does it? With its artificial cap and limiting of 1B to a range of about +/- 4 runs?

It depends what one means by 'endorse'. I find Fielding Win Shares useful, and knowing those issues you mention one can accommodate occasions when they might be a factor.

I've argued this in HoM threads before, but my impression of fielding is that one is mainly interested in extreme cases. Fielding differences are mostly a very narrow band, which exacerbates the effect of being exceptionally good or bad. A bad glove really hurts, relatively, and a good glove really helps.

I read 'Wizardry' in the summer. It definitely is a must-read for anyone who hangs out here.

That is very interesting. He is remarkably poor.

Yes, but his poor ratings are exacerbated by a relatively unusual ball-in-play distribution and defensive alignment.

The Yankees as a team, from 1998-2002, had fewer balls hit into play on the left side of the diamond, and more balls hit into play on the right side of the diamond, than you would expect (even when you take into account the distribution of left/right hitting against them). The Yankees therefore tended to overshift both their infielders and their outfielders toward the right side of the diamond, even in the late innings - Brosius played closer to the SS/3B hole, Jeter cheated up the middle, Bernie played more toward right-center. The upshot of this alignment is that Yankee fielders tended to be penalized more heavily for hits allowed in every system, because those hits allowed were usually in areas of the field where other teams would be making plays. The rewards that Yankee fielders got for making plays, on balance, didn't counter that effect primarily because they had fewer balls in play than did most teams (they had a lot of strikeout pitchers). So while the Yankees generally had decent team defense, their key up-the-middle fielders tended to look bad in most defensive systems.

Jeter IS a below-average fielder; even cheating up the middle he didn't make a lot of plays going toward second (which has always been the rap against him). But what struck me when I actually had a chance to look at the Yankee BIP data was that the Yankees really "didn't" allow an unusually high number of hits up the middle - and that they DID allow an unusually high number of hits on Jeter's other side. That's when I started watching video and realizing what was going on.

-- MWE

The upshot of this alignment is that Yankee fielders tended to be penalized more heavily for hits allowed in every system, because those hits allowed were usually in areas of the field where other teams would be making plays.

More grist for my sceptical-of-zone-based-systems mill.

More grist for my sceptical-of-zone-based-systems mill.

Well, I don't want to overstate the case here - the net impact could not have been nearly large enough to raise Jeter to the level of Gold Glove quality defense, or even to the level of being above average defensively. There WERE plays not being made that an average shortstop would have made, even given Jeter's positioning - the "past a diving Jeter" meme isn't just an invention. But it strikes me as peculiar that a defense that was generally above-average overall could have fielders at key defensive positions that were as bad as the systems were showing Jeter and Bernie Williams to be.

-- MWE

Mike,

I may be able to report numbers in a day or two summarizing Yankee defense using updated DRA for all the Jeter years except 2000-02. The Yankees during that span were on average well below average on fielding. They had terrific pitching staffs.

Mike,

Also, do we really think the Yankees had this unique shift in batted ball distributions throughout Jeter's career? I did a new WOWY analysis for Jeter's pitchers, split by batter handedness. This would control for any bizarre but stable batted ball distributions. He was close to -500 plays on ground balls.

I think that in time DRA/Wizardry will shift the consensus career valuations for a couple of dozen historically significant players by between five and fifteen career WAR.

Who are the other players whom this will most affect?

Sweatpants, have you picked up Wizardry? It just made a very nice list: http://math-blog.com/2011/12/16/interesting-mathematics-books-2011/

the net impact could not have been nearly large enough to raise Jeter to the level of Gold Glove quality

Of course. But some of what you discover here would be less of a problem if the zones were not too small. And working from raw data like RZR's OOZ would probably be more useful in assessing the actual quality of defence of a team like these Yankees.

From 1998-2000 the Yankees overall were quite good at preventing hits on balls in play; in 2001-2002, not so much, but overall for the period were generally above average (even discounting 1998, which was a season for the ages in all respects including defensively). Those were the only five seasons for which I had good-quality location data for BIP for the Yankees, so I don't know the extent to which the right-side skew has changed.

One issue with a WoWY analysis on Jeter (and for most regular players) is that except for 2003 (and last season) he hasn't missed much time at SS, and many of the innings that he has missed come near the end of blowouts, where pitching patterns and BIP distributions can be very different. You're talking, in a number of cases, pretty small sizes of "Wo" - Jeter played all but about 40 of the innings that Clemens pitched for the Yankees between 1998 and 2002, for example - and it's not clear that the "Wo" sample of innings is representative of the much larger "WY" sample.

-- MWE

The Jeter WOWY would compare Jeter against the scores of shortstops who also fielded behind Clemens when the latter was not pitching for the Yankees. Also, the WOWY I used eliminated from the sample instances in which the Jeter and non-Jeter samples were too small. I forget right now the formula I used, but it involved comparing two binomial means. Also, the WOWY results were consistent with the basic new DRA based described in The 2012 Hardball Times Annual. We are talking a decade and a half of numbingly consistent seasons.

Just wanted to bump this . . . right now I'm just doing this to estimate post 2005 . . .

Run an individualized regression for each player. Compare his career DanR WAR through 2005, season by season, with BPro and B-R WAR. I suppose I could add fangraphs WAR as well.

Use the resulting individualized regression equation to estimate DanR WAR from 2006-present.

Obviously this has some pretty major flaws and will become less useful over time, as players have fewer pre-2005 years.

It would be better to do this for each component (batting, fielding, base-running), but haven't really seen the need yet. Plus I'd need to figure out the SFrac (fraction of a season played) and replacement level by each player. And not all systems break out the components like Dan does, some including fielding and replacement level together, etc..

Joe, I can email you the master spreadsheet and you should be able to figure it out and input all the data yourself if you'd like.

That would be great Dan, thanks!

Will you also be available for consulting services, should I need them (rather likely). :-)

Yes, of course. Remind me your email?

Hey Dan, sorry I missed your request . . . it's in your g mail chat window . . .