Is WPA Predictive for Batters?

One of the biggest complaints I see about WPA is that it’s not predictive. The mere mention of it’s non-predictability seems to be enough for many to write it off as a mere toy used by some of stats community.

So let’s see how it actually correlates from year to year compared to the stats we all know, like AVG, OBP, SLG, and OPS. I’ll throw in Batting Runs Above Average for fun too.

Looking at the r-squared from 2005 to 2006 for batters with over 300 plate appearances, here’s how WPA stacks up against the regulars:

AVG: .12
WPA: .27
BRAA: .35
OBP: .36
OPS: .36
SLG: .38

Here’s the same deal, 2004 to 2005.

AVG: .14
WPA: .24
OBP: .27
OPS: .30
BRAA: .31
SLG: .33

It’s true, WPA doesn’t correlate as well from year to year as OBP, SLG, or OPS, but it does have some correlation from year to year. In 2004, a players OBP was almost indicative of his 2005 OBP as his 2004 WPA was of his 2005 WPA. Yet, that wasn’t quite the case in 2005 to 2006. BRAA which is calculated by using Run Expectancy on a play-by-play basis (much like WPA uses Win Expectancy), holds its own against the regulars.

Anyway, the point is, let’s stop using the argument that WPA isn’t predictive as a crutch, because it does actually show some correlation from year to year.

Print This Post

David Appelman is the creator of FanGraphs.

15 Responses to “Is WPA Predictive for Batters?”

You can follow any responses to this entry through the RSS 2.0 feed.
  1. Pizza Cutter says:

    David, how far back do you have WPA data available? If you have several years worth, you might try something like an auto-regressive (AR1) covariance matrix to find out what the intra-class correlation is. (For those who don’t know, this is a multiple observation technique that allows a more complete picture of how much correlation there is from year to year on an individual level.) If you like, e-mail me (I believe you can see my e-mail address from my post.)

    Vote -1 Vote +1

  2. tangotiger says:

    And you will find that WPA/LI will have even more predicatability.

    However, the big question, does “clutch” or WPA minus WPA/LI have predicability? You’ll likely get an r under .10.

    Vote -1 Vote +1

  3. Erik says:

    Thanks David. I’ve been one of the guilty and have now seen the light.

    Vote -1 Vote +1

  4. WPA/LI for 2005 to 2006 was .36. For Clutch, it’s .01, as suspected.

    Vote -1 Vote +1

  5. tangotiger says:

    Great, so WPA/LI and OPS are equally predictive.

    Vote -1 Vote +1

  6. birtelcom says:

    I take it that the Clutch season-to-season correlation of .01 is significant evidence for the theory, in all of sabermetric history one of the most heavily debated, that the timing of hitting performance (to perform better in particularly advantageous moments) is not a “skill” in the sense that it is repeatable but rather is essentially random. It also suggests, I guess, that most of the season-to-season correlation for WPA, and its similarity to the correlation for OPS, is simply a reflection of the fact that most of WPA derives from overall hitting skill (reflected fairly well by OPS, pace tango), which we know is repeatable, and that only a small component of total WPA reflects timing issues. I still love WPAs, but not because they reflect some repeatable skill different from that already expressed by OPS and similar stats, but rather because they express important elements of the value of past player performance not already reflected in OPS and similar stats.

    Vote -1 Vote +1

  7. Chris Constancio says:

    Exactly – if you were to control for a few player ability measures, such as ability to get on base or hit extra-base hits, WPA would have no predictive usefulness.

    Vote -1 Vote +1

  8. tangotiger says:

    Right, the power of WPA is to tell you a particular story. At the end of the season, we’d be able to easily pick out the 20 PA that most influenced a real-life game. That’s its value.

    If you’re looking for “is he really better than that guy”, WPA has little, if any, competitive advantage over anything else out there.

    WPA should be treated for what it’s built for.

    Vote -1 Vote +1

  9. enoscountry says:

    The last three commenters get it right. In addition, WPA is so team dependent (not to mention inning dependent) that your going to get some significant correlation from year to year. Hitters on better teams, hitters who consistently hit higher in the order, are going to have higher WPA because the are more likely to hit in close games and hit later in the game.

    Vote -1 Vote +1

  10. dpm says:

    I don’t feel like I have enough information to know whether WPA is a great predictor of future success, but I do think it’s the best outcome measure I’ve seen for assessing current-season-to-date success. That is, it may not tell me if a player is lucky or good, but it tells me who got good results.

    If anyone else buys into that, here’s what I’d like to know. What’s the best prior-year predictor of current-year-to-date WPA? Is it prior-year WPA or is it one of the other measures? I don’t care that much how well last year’s OPS predicts this year’s OPS if last year’s OPS doesn’t predict this year’s WPA.

    Vote -1 Vote +1

  11. tangotiger says:

    There’s no question that prior year OPS and prior year WPA/LI will predict current year WPA the best.

    I agree with your assessment that we don’t care about OPS-to-OPS predictability, but rather what in the prior years predicts the next year’s WPA, since what we are after is wins.

    Vote -1 Vote +1

    • zacksf says:

      I believe the OPS part, but I am skeptical about the power of WPA/LI to predict next years WPA. All those divisions by small numbers worry me. Do you have any numbers for correlations that support that assertion?

      Vote -1 Vote +1

  12. dpm says:

    I think you’re probably right, but I’d still like to see the numbers. Maybe this is covered in an FAQ elsewhere, but is it possible to download a dataset that has each player’s WPA and some of the other key stats? I’d be happy to crunch the numbers myself if I had the data.

    Vote -1 Vote +1

    • zack says:

      @dpm (and TangoTiger)

      I realize this post is ancient, however, I think you are really asking the right questions and would love to know if there has been any follow up on this. (Particularly, what are the best predictors of WPA and related issues. Probably RE24 could also be in the discussion at this point.)

      Vote -1 Vote +1