Sunday, March 3, 2024

Factors that might influence the difference between ERA and FIP

My last post mentioned that Aaron Bummer had a 6.79 ERA last year while his FIP ERA was 3.58 for a differential of -3.21 (FIP - ERA). That was the largest absolute differential last year for any pitcher with 50+ IP and it was 0.64 larger than the next largest.

So to look at what might explain why ERA differs from FIP (fielding independent ERA estimated using SOs, BBs and HRs), I ran a regression with FIP - ERA as the dependent variable and the following three independent variables:

SLG Diff (a pitcher's SLG allowed with runners on base minus the SLG they allowed with no runners on)
BAbip (the batting average a pitcher allows on balls in play, so it is dependent on how good his fielders are)
BQS/9 (Bequeathed runners that scored per 9 IP).

Bequeathed runners represents the number of runners left on base by a pitcher when that pitcher leaves the game. Any bequeathed runner who scores an earned run after a pitcher has left the game will be counted against that pitcher's ERA (from mlb.com https://www.mlb.com/glossary/advanced-stats/bequeathed-runners).
 
I used SLG Diff because some pitchers might have gotten hit pretty hard when they had runners on base, making their ERA higher than what we might otherwise expect based on their overall numbers.

I used BAbip because this is not controlled very much by the pitcher. A guy can have a low FIP but if his fielders can't catch the ball, his ERA will be high.

I used BQS/9 because a pitcher cannot control what happens after he leaves the game. Some pitchers get lucky and their bullpen bails them out. For others, it is the opposite.

I looked at all the guys who had 100+ IP last year. All data came from Baseball Reference and Stathead. There were 127 pitchers.

Here is the regression equation:

FIPERADIFF = .789*BQS/9 + 13.63*BAbip + 2.5*SLGDiff - 4.32

r-squared = .639, so 63.9% of the variance in the dependent variable is explained by the model.

standard error = .33

Here are the t-values for the three independent variables:
 
BQS/9)  6.7 (The p-value is < .00001)
BAbip)  11.6 (The p-value is < .00001)
SLG Diff)  5.04 (The p-value is < .00001)

The r-squared seems fairly high but it still means that 36.1% of the variation in FIPERADIFF is not explained.

The standard error seems high. I wish it was lower. The average absolute differential was about .44.

The t-values are all pretty high so each independent variable is significant. I used a website that converts t-values into p-values.

There may be some other variable that I should include. Maybe I could find the estimated FIPERADIFF for each guy and look at the 10 or so guys with the biggest differences between the estimated value and the actual (FIP - ERA). Maybe something that would be obvious to include would pop up.

No comments:

Post a Comment