DRC+ really isn’t any good at predicting next year’s wOBA for team switchers


The decent performance I get for DRC+ projecting low-PA players is from assigning pitchers terrible DRC+s no matter how well they hit.  The rest of the post is fine, but this is all even more bullshit than I realized at the time of writing.



Required knowledge: wOBA and DRC+

Part 2: The DRC+ team-switcher claim is utter statistical malpractice

TL;DR: raw DRC+ is a little better overall than projecting everybody to be league average, but actually worse than that for team-switchers. Best-regressed DRC+ has about a 2.5-point MAE improvement over “everybody hits league average” for switchers and a 1 point MAE improvement over best-regressed wOBA. Regressed DRC+ has a huge advantage projecting very-low-PA seasons and starts losing to regressed wOBA around the 300 PA mark.

Having already demonstrated and explained why DRC+ is structurally unfit for use in WAR/BWARP, the purpose of this next experiment was to test the claims here and here that DRC+ was something special when it came to projecting next year’s wOBA for team-switchers. It fails that test convincingly, but a little regression work gives a decent projection for players that *don’t* switch.  Unfortunately for DRC+, that prediction is only marginally better overall than the same methodology using wOBA, and only has a real advantage at the low-PA end.  Regressed wOBA rules the high-PA end.

Methodology Overview

To test their claim, and to account for leaguewide wOBA changing every year, I normalized every batter-season’s wOBA onto a 100 scale by taking (batter wOBA)/(league average wOBA for that season) * 100.  I’ll call that wOBA% from now on.  Normalization to wOBA% makes sense because many of the factors that influence leaguewide wOBA changes in the upcoming year, from the changing strike zone to the changing baseball itself are not something DRC+ tries to predict- or ever should try to predict. Using wOBA% instead of raw wOBA removes a good deal of nonsense noise at no cost.

Team switchers were not explicitly defined, but since the test sample must be composed of players who had PAs in consecutive seasons, I’m defining a team switcher as anybody who didn’t appear entirely for the same team in both years (e.g. half a season for team A followed by 1.5 seasons for team B is a team switcher).

I also normalized DRC+ to DRC+% similarly since it was coming out a bit under 100, but since DRC+ is on the run scale, I used 100*Sqrt(Max(DRC+,0)/100) to put it on the same scale as wOBA.  Seasons from 2010-2018 were used, although 2010 was only used to project 2011.  Every pair of consecutive seasons where a player had at least 1 PA was eligible.

MAEs were calculated weighted by the harmonic mean of the PAs in year T and year T+1.  Best regressions were determined the same way (using all pairs of seasons, not just switchers).  Since MAEs are calculated in wOBA%, I multiplied by 3.20 (i.e. a wOBA of .320) to put them back on the wOBA-points scale to report.

Tests and Results

The first thing I tried was simply using year T DRC+% as the projection for year T+1 wOBA% and benchmarking that against an “everybody 100 wOBA%” (LgAvg) projection and restricting it to pairs of qualified seasons (500+ PA in each).  DRC+ had an MAE of 25.1 points of wOBA against LgAvg’s 34.3 overall, but 26.5 vs 26.1 on team switchers.

In the attempt to see if the signal could be improved, I regressed year T DRC+% with league average (100 wOBA%) PAs and the minimum weighted MAE came with adding 89 average PAs.  Doing the same optimization with year T wOBA% came up with adding 332 average PAs.  For reporting purposes, I broke the players up into 3 groups

  1. 0-99 PAs in year T, which is just enough to capture all pitchers (2016 Bumgarner, 97 PA) as well as a bunch of callups and fill-ins who aren’t really MLB quality.
  2. 400+ PAs in year T, which is all full-time players and primary sides of platoons, etc.  That number is kind of arbitrary, but it’s a little over 50% of the average PA per position, assuming some PHing, and moving it around 25 PAs isn’t really going to affect the big picture analysis anyway
  3. 100-399 PAs in year T to cover everybody else.


This is a sample report.

Table 1.  wOBA MAEs, 400+ PA in both seasons.  LgAvg=100

Min 400 PA both seasons raw DRC+% LgAvg regd DRC+% regd wOBA% year T+1 wOBA% year T wOBA% N
all 26.0 32.3 25.1 24.5 106.6 107.5 1152
switch 28.4 26.7 26.9 25.8 103.6 105.3 303
same 25.2 34.3 24.4 24.1 107.7 108.2 849

Any PA cutoff biases the sample, but using a PA cutoff in both seasons is especially bad form because it excludes players who would have reached the cutoff in year T+1 if they hadn’t been benched for sucking.  Even with artificially tight performance constraints, the regressions are virtually useless for team switchers- only 1 point of MAE improvement for wOBA and nothing for DRC+. To avoid the extra bias problem, future results will include all (1+ PA) year T+1 seasons.

Table 2. wOBA MAEs, 400+ PA in year T, any PA year T+1.  LgAvg=100

Min 400 PA year T raw DRC+% LgAvg regd DRC+% regd wOBA% year t+1 wOBA% year T wOBA% N
all 27.6 32.6 26.6 26.3 105.0 106.6 1597
switch 30.3 28.8 29.0 28.2 101.0 103.9 473
same 26.4 34.2 25.6 25.5 106.6 107.7 1124

The bias in Table 1 is apparent now.  This population is simply worse to begin with in year T (marginal hitters were more likely to suck and get benched in year T+1 and not show up in Table 1) and dropped off more to year T+1.  Back to the post topic, neither DRC+ nor wOBA are any good for switchers, wOBA is a bit ahead of DRC+, and the projection for same-team players is a clear improvement on LgAvg.

Table 3.  100-399+ PA in year T, any PA year T+1.  LgAvg=100

100-399 PA year T raw DRC+% LgAvg regd DRC+% regd wOBA% year t+1 wOBA% year T wOBA% N
all 36.8 36.2 34.7 34.5 98.5 97.8 1574
switch 38.5 38.3 36.3 36.5 94.3 95.2 627
same 35.7 34.9 33.6 33.2 101.4 99.5 947

The league average errors are a good bit worse and now DRC+ and wOBA are pretty useless for everything, offering at best a 2-point improvement over LgAvg.  Also, the quality of the players here is clearly worse because….. better players get more PAs and make it into Table 2 instead.

Table 4. wOBA MAEs, 1-99+ PA in year T, any PA year T+1.  LgAvg=100

1-99 PA year T raw DRC+% LgAvg regd DRC+% regd wOBA% year t+1 wOBA% year T wOBA% N
all 110.2 102.2 62.9 92.5 84.3 69.8 2409
switch 91.4 96.7 64.3 87.7 73.3 69.8 841
same 120.3 105.2 62.2 95.1 90.2 69.7 1568

And this is interesting. Garbage hitters, giant MAEs, and regressed DRC+ winning by a mile for a change.  The other interesting thing here is that the players teams keep improve *a ton* and the ones they let go keep being godawful at the plate.  Somebody should look into that in more detail.

Seeing that the 100-399 PA group at least resembled MLB-quality hitters, albeit not the good ones, and that the 1-99 PA group was an abomination at the plate (it did include all the pitchers), I wondered what would happen if I cheated a little and tried to optimize on the 100+ PA group instead of everybody.  That group looks like

Table 5.  wOBA MAEs, 100+ PA in year T, any PA year T+1.  LgAvg=100

Min 100 PA year T raw DRC+% LgAvg regd DRC+% regd wOBA% year t+1 wOBA% year T wOBA% N
all 30.4 33.7 29.1 28.8 102.7 103.9 3171
switch 33.3 32.2 31.6 31.2 98.6 100.8 1100
same 28.9 34.5 27.7 27.5 104.9 105.6 2071

Again, useless for switchers, solid improvement for the ones who stayed.  Based on this, I decided to reoptimize based on a LgAvg of 103 using only players with 100+ PA year T just to see what would happen.

Trying a different league average

This is starting to look down the rabbit hole of regressing to talent (more PA is a proxy for more talent as we’ve seen) instead of to pure league average, but let’s see what happens.  The regression amounts came out to 243 added PA for DRC+ and 410 added PA for wOBA.  Doing that came up with

Table 6.  wOBA MAEs, 100+ PA in year T, any PA year T+1.  LgAvg=103

Min 100 PA year T raw DRC+% LgAvg regd DRC+% regd wOBA% year t+1 wOBA% year T wOBA% N
all 30.4 33.2 28.7 29.1 102.7 103.9 3171
switch 33.3 33.7 31.3 31.9 98.6 100.8 1100
same 28.9 32.9 27.2 27.6 104.9 105.6 2071

Well, that’s not an auspicious start, marginally helping the stayers (the switchers were closer to 100, so regressing towards 103 isn’t any help).  Let’s see if there’s any benefit in either group individually.

Table 7.  wOBA MAEs, 100-399 PA in year T, any PA year T+1.  LgAvg=103

100-399 PA year T raw DRC+% LgAvg regd DRC+% regd wOBA% year t+1 wOBA% year T wOBA% N
all 36.8 38.7 34.5 35.6 98.5 97.8 1574
switch 38.5 41.9 36.7 38.0 94.3 95.2 627
same 35.7 36.6 33.0 33.9 101.4 99.5 947

Well, that was a disaster for these guys. The top line using 100 LgAvg was 36.8 / 36.2 / 34.7 / 34.5 before, and regressing them further from their talent shockingly didn’t do any favors.  Was this made up for by the full-time players?

Table 8.  wOBA MAEs, 400+ PA in year T, any PA year T+1.  LgAvg=103

400+ PA year T raw DRC+% LgAvg regd DRC+% regd wOBA% year t+1 wOBA% year T wOBA% N
all 27.6 30.8 26.1 26.3 105.0 106.6 1597
switch 30.3 29.1 28.3 28.5 101.0 103.9 473
same 26.4 31.5 25.2 25.4 106.6 107.7 1124

Not really.  This is a super marginal improvement over the previous top line of 27.6 / 32.6 / 26.6 / 26.3.  Only the LgAvg projection really benefits at all, and that’s not what we’re interested in.  Changing league average a little and optimizing over only MLB-quality hitters doesn’t seem to really accomplish anything great for the DRC+ or wOBA regressions.


There’s little-to-nothing to BP’s claim that DRC+ is something special for team switchers.  Raw DRC+% is worse than league average, and the MAE for best-regressed DRC+ is about 2 points better than league average overall, and that entire benefit is from the low-PA end.  It projects team-switching full-time players worse than assuming they’re league average.  However, for the really low-PA players, it is more accurate raw and much more accurate regressed than league average.

Likewise there’s also absolutely nothing to the claim that DRC+ is a significant improvement over wOBA for predicting year T+1 wOBA for switchers- the gap is actually *smaller* for switchers.  1.1 points of MAE for switchers and 1.5 points for stayers.  The best regression of DRC+ absolutely does shine in the very-low-PA group, but it’s also not good in the full-time player category.  Regressed DRC+ and regressed wOBA actually do make fairly decent, much better than league average predictions for full-time players who *stay*, for whatever an untested model fit to in-sample data is worth.


C’mon Man- Baseball Prospectus DRC+ Edition

Required knowledge: A couple of “advanced” baseball stats.  If you know BABIP, wRC+, and WAR, you shouldn’t have any trouble here.  If you know box score stats, you should be able to get the gist.

Baseball Prospectus recently introduced its Deserved Runs Created offensive metric that purports to isolate player contribution to PA outcomes instead of just tallying up the PA outcomes, and they’re using that number as an offensive input into their version of WAR.  On top of that, they’re pushing out articles trying to retcon the 2012 Trout vs. Cabrera “debate” in favor of Cabrera and trying to give Graig Nettles 15 more wins out of thin air. They appear to be quite serious and all-in on this concept as a more accurate measure of value.  It’s not.

The exact workings of the model are opaque, but there’s enough description of the basic concept and the gigantic biases are so obvious that I feel comfortable describing it in broad strokes. Instead of measuring actual PA outcomes (like OPS/wOBA/wRC+/etc) or being a competitive forecasting system (Steamer/ZIPS/PECOTA), it’s effectively just a shitty forecast based on one hitter-season of data at a time****.

It weights the more reliable (K/BB/HR) components more and the less reliable (BABIP) components less like projections do, but because it’s wearing blinders and can’t see more than one season at a time, it NEVER FUCKING LEARNS**** that some players really do have outlier BABIP skill and keeps over-regressing them year after year.  This is methodologically fatal.  It’s impossible to salvage a one-year-of-stats-regressed framework.  It might work as a career thing, but then year X WAR would change based on year X+1 performance.

Addendum for clarity: If DRC+ regresses each season as though that’s all the information it knows, then adds those regressed seasons up to determine career value, that is *NOT* the same as correctly regressing the total career.  If, for example, BABIP skill got regressed 50% each year, then DRC+ would effectively regress the final career value 50% as well (as the result of adding up 50%-regressed seasons), even though the proper regression after 8000 PAs is much, much less.  This is why the entire DRC+ concept and the other similarly constructed regressed-season BP metrics are broken beyond all repair.  /addendum


****The description is vague enough that it might actually use multiple years and slowly learn over a player’s career, but it definitely doesn’t understand that a career of outlier skill means that the outlier skill (likely) existed the whole time it was presenting, so the general problem of over-regressing year after year would still apply, just more to the earlier years. Trout has 7 full years and he’s still being underrated by 18, 18, and 11 points the last 3 years compared to wRC+ and 17 points over his whole career.

DRC+ loves good hitters with terrible BABIPs and particularly ones with bad BABIPs and lots of HRs.  Graig Nettles and his career .245 +/- .005 BABIP / 390 HRs looks great to DRC+ (120 vs 111 wRC+, +14.7 wins at the plate), as do Mark McGwire (164 vs 157, +8.5 wins), Harmon Killebrew (150 vs 142, +16.2 wins), Ernie Banks (129 vs 118, +20.8 wins), etc.  Guys who beat the hell out of the ball and run average-ish BABIPs are rated similarly to wRC+, Barry Bonds (175 vs 173), Hank Aaron (150 vs 153), Willie Mays (150 vs 154), Albert Pujols (147 vs 146), etc.

The flip side of that is that DRC+ really, really hates low-ISO/high BABIP quality hitters.  It underrates Tony Gwynn (119 vs 132, -12.9 wins) because it can’t figure out that the 8-time batting champ can hit. In addition, it hates Roberto Alomar (110 vs 118, -10.4 wins) Derek Jeter (105 vs 119, -17.9 wins), Rod Carew (112 v 132, -18.7 wins), etc.  This is simply absurd.

C’mon man.