[Waterstpub87]

I see the statistic XFIP being thrown around a lot on the boards, as a "true" measure of what a pitchers ERA should be. Statements like " with an ERA of 2 and an XFIP of 4, Pitcher x is due for a regression".

Up front, I don't like the idea of XFIP at all. Ground balls count as hits too, I don't think walks matter as much, and I don't think strikeouts are as important as people make them out to be.

I am open-minded, so I wanted to test how well XFIP was at predicted next month's ERA as a way to test.

The formula I used for XFIP was the one listed here : http://www.fangraphs.com/library/pitching/xfip/

xFIP = ((13*(Fly balls * lgHR/FB%))+(3*(BB+HBP))-(2*K))/IP + constant

I used the 2014 constant for Lghr/fb% of 9.50%. I used 3.10 as the constant in the second part of the equation. Using a sample size of 64 starting pitchers, selected based on having pitched over twenty innings per month from April to September, using 2014 season as the data for the test.

I tested two ways, termed short term reversion and long term reversion.

Short term reversion would be last months XFIP compared to this months ERA. So for June, it would be June ERA-May XFIP = Prediction Error

Long Term Reversion= Average of monthly values up to that point, so for June it would be June ERA-(April XFIP+MayXFIP)/2

Short Term Reversion Results
XFIP ERA
Average 1.325 1.444
Std Deviation .44 .589
Count 38 26

Average is average error. Std Deviation is the standard deviation of total average errors between pitchers, count if the number of pitchers that XFIP was more accurate on.

Technically, based on this data, they are not incorrect when they say it is more accurate at predicting future era. However, it only outperforms last months ERA by an average of .12, which is less then the deviation of errors between pitchers. I would be cautious relying on this to predict future era results, because the results of this test show it to be inconclusive.

Long Term Reversion
XFIP ERA
Average 1.25 1.30
Std Deviation .387 .43
Count 37 27

Again, technically, it is slightly better than ERA. However, it is the same situation, the gain is approximately .05 runs more accurate. The difference is even less in magnitude using the cumulative average.

Based on these results, I would be cautious about drawing conclusions of a pitchers future performance based on XFIP.

[evo34]

Well, predicting pitcher performance based on 1-2 months of anything is never advisable. If you wanted to use xFIP to predict ERA, though, you need to consider park effects and defense, neither of which are included (much) in xFIP. E.g., Colorado pitchers will always, over a large enough sample, under-perform their xFIPs (ERA will be higher) simply because Coors has a much higher hit rate and HR rate than the stat assumes.

[KingHutch]

Isn't SIERA a better predictor than xFIP?

[evo34]

None of them are designed to be pure ERA predictors, so you have to figure out what adjustments to make for park and defense. I prefer xFIP simply because it's pretty clear what it is incorporating and what it is not. I would imagine SIERA is the superior predictor in-season, if for some reason you were forced to use a single stat off the shelf without any adjustment.

[Waterstpub87]

Look, I think you misunderstand. I get why XFIP is bad, no defenses, park factors, excetra.

My point is that it is not anymore predictive then just using ERA itself in predicting future eras. Even if you adjusted it, and then made the same adjustments to era, that would likely hold. People tout XFIP like it is the be all and end all, pitchers regress or advance towards ,a true predictor of future results. Based on tests, it isn't.

[evo34]

"Even if you adjusted it, and then made the same adjustments to era, that would likely hold."

100% false.

[EXhoosier10]

Well, predicting pitcher performance based on 1-2 months of anything is never advisable. If you wanted to use xFIP to predict ERA, though, you need to consider park effects and defense, neither of which are included (much) in xFIP. E.g., Colorado pitchers will always, over a large enough sample, under-perform their xFIPs (ERA will be higher) simply because Coors has a much higher hit rate and HR rate than the stat assumes.

Seconded... your study is meaningless, just like 1 month of xfip is pretty meaningless. take pitchers with 20+ starts in a season, compare their first 10 game xfip to rest of season ERA vs 10 game ERA compared to rest of season ERA. you'll likely see much clearer results. If not, add more years. You will see it eventually

[Waterstpub87]

Seconded... your study is meaningless, just like 1 month of xfip is pretty meaningless. take pitchers with 20+ starts in a season, compare their first 10 game xfip to rest of season ERA vs 10 game ERA compared to rest of season ERA. you'll likely see much clearer results. If not, add more years. You will see it eventually

Exhoosier and Evo,

Appreciate the discussion.

Evo, If you tell me what to adjust the XFIP by to normalize it, happy to do so. Or, if you want to produce your own results (if this information is your betting secret sauce) We can compare those as well.

Exhooiser, fair enough point. To this, I measured the error going from July to September on a cumulative basis (XFIP up to that point to predict next month era). This would be approximately 15 starts.

The results are less clear. XFIP error to next month ERA is 1.33, ERA to next month era is 1.21.



How many years should I add? Have you tested it to find if the XFIP changes from year to year with pitchers? If so, how much?

[EXhoosier10]

2013 & 2014 correlation between 1h era and 2h ERA for qualified pitchers = .131
2013 & 2014 correlation between 1h xfip and 2h ERA for qualified pitchers = .253

count = 128

Can't think of why your "short term reversion" formula is showing that ERA has less of an average differenace than xfIP, but correlation si what you want to use, not this 'average difference' function