Pitcher Win Value Correlations
Now that we’ve worked through the logistics of the pitcher win value formula that we recently added here on FanGraphs, I figured it was time to answer the important question – how predictive are they? Since pitchers have historically been evaluated by ERA, the belief has been that they are wildly inconsistent from year to year. However, we know that ERA includes a bunch of non-pitcher variables, and since we’re using FIP, we shouldn’t have to worry about the variation in those external forces.
So, how well does a pitcher’s Win Value correlate from one year to the next? Better than I expected, honestly. I took all pitchers with at least 10 IP in 2004/2005, 2005/2006, 2006/2007, and 2007/2008, and found the following year to year correlations:
2004 to 2005: .62
2005 to 2006: .69
2006 to 2007: .67
2007 to 2008: .55
That’s not bad at all. ERA, for example, has a year to year correlation to itself of around .4. Clearly, the inputs of FIP are more stable than the inputs of ERA, but we knew that already. However, since FIP is a rate stat and Win Value is a counting stat, I’m a bit surprised at how well the win values hold up, since it requires a combination of similar performance and playing time.
As time goes on, I’m sure we’ll improve the formula and push the year to year correlations higher, but as it stands, Pitcher Win Values 1.0 do a pretty nice job of predicting pitcher value from one year to the next.
Print This Post
Those correlations look quite good to me. I don’t think you’re going to get things a whole lot better on y-t-y correlations for one single stat on a pitcher’s overall value.
Personally, I think the only thing worth adding into the win values at some point is GB%.
I need to brush up on my correlation inputs a bit but isn’t 10IP a really small amount to use?
I was going to say the exact same thing, Graham.
What’s y-t-y correlation for FIP on its own? How about IP?
I am curious about FIP y-t-y correlation as well…
Dave,
Did you separate relievers from starters when determining these correlations? If not, I wonder whether the strength of the correlation has to do just with play time correlations between these two roles. It seems odd to me that play time + FIP (the ingredients in win values) should exceed FIP alone in reliability.
Is it just me, or does the 10 IP cutoff for the sample group seem very low. I would think that you would want it quite a bit higher as that amount of work would not prove to be statistically significant. It would also be interesting to see the results if the starters are separated from the relievers.
Hi Dave, I’ve really enjoyed your work here and at USSM. Been reading for about a year but this is my first time commenting. Thanks for explaining FIP and Win Values, a great set of posts.
I was wondering when you were going to get to predictive power of FIP generated win values on future performance – the crucial question. But showing that past FIP correlates to future FIP is sort of skirting the question. To be valuable, win values must predict future, well, wins. Of course Wins (with a capital W) are full of noise and only loosely correlated to performance. But for comparing across different methods of Win values, that’s OK, especially with large sample sizes.
I think it would be very powerful to show a positive correlation for something like the following (for starting pitchers): FIP on the Y-axis and # of games started won divided by teams total wins on the X-axis. Then ask does FIP show a stronger correlation then other measures of performance, such as ERA or K/9.
Something like this needs to be done (probably already has?) to show that Win values actually lead to wins for the team! Would love to hear your thoughts.