Upgrading My Individual Pitch Result Metric

On a personal level, the All-Star break can be declared a success as I’ve made major improvements to my pitch result evaluator, pERA. I was supposed to do dive into it last season, but I spent most of the time dealing with the league’s new rules so this update got pushed off until now. I planned on adding Ball Percentage (Ball%), Called Strikes (CStr%), and StatCast batted ball information. I felt each add would provide a clearer picture of the pitcher’s pitches. I eventually found out I was double counting the same information with Ball% and CStr% and needed to remove one.

The key element of pERA is to grade out each pitch and have a way to value a pitcher who adds or losses a pitch. If the pitcher throws an ineffective change and ditches it, how should the pitcher’s results change? To answer the question, I created pERA

It was almost seven years ago when I created the initial formula that included:

• Swinging Strike Rate: Ability of a pitch to generate swings-and-misses.
• Groundball Rate: If the pitcher is on either end of the groundball spectrum, they can outperform their ERA estimators by generating fieldable contact.
• Walk Rate: I included the walk rate once all the individual pitches were added up. This was a major issue that elevated several pitches that were never near the strike zone.

The overall results have stood the test of time but needed some fine-tuning. Here are the improvements I planned on incorporating.

• Called plus Swinging-Strike Rate: All strikes are the same and this improvement catches the starters who paint the corners. The idea came from the work of Alex Fast at Pitcher List.
• Fieldable Contact (<-5 degrees, >40 degrees): This addition puts all the fieldable batted balls into one bucket
• Ball%: This addition was huge and I’m a little disappointed I didn’t think of it when I created the initial formula. Some recent work with Ball% was the light bulb I needed to include it.

I had about 95% of this article’s first draft done but found out that I was double counting the same items, Ball% and CStr%. Control pitchers, who could paint the zone, got called strikes and fewer balls. Their fastballs looked elite but weren’t. Called Strike Rate got cut and I just went with Swinging Strike Rate.

With the two additions, I collected a dataset to transform the raw data into something useful. I used pitchers who threw at least 80 IP from 2015 to 2023 (1237 in the sample). I was limited to those seasons because of StatCast’s availability.

With the dataset, I created the following equations.

• K% = SwStr% * 2.1
• BB% = .6846*Ball%-.1695
• ERA = -10.55 * (K%-BB%) + 5.6144
• Weak contact adjustment [46% Weak% is where there is no adjustment] = -3.4904 * Weak% + 1.6142 The weak contact allows a single overall value to be generated without the addition of an individual pitch score because popups and groundball are combined. I might move to a polynomial equation where values on the extreme are weighed more (-297.5 * Weak%^3+418.5 *Weak%^2-198.4*Weak%+31.695).

With the math and background out of the way, it’s time for idiot (failed the first time) checks. Here are the top-10 starters (min 60 IP) and relievers (min 20 IP). Additionally, the top grades for each pitch (link to the full sheet).

Overall pERA

 

Pitcher ERA (min 300 pitches)

Some observations.

• Robert Stephenson (top reliever): The biggest disconnect between his results is the Ball%. Usually, a Ball% of 31% (140 balls in 447pitches) would point to a 1.6 BB/9, not Stephenson’s 4.1 BB/9.
• Lance Lynn shows his face a few times. The biggest issues with Lynn are his curve (5.41 pERA) and change (4.91 pERA). Both have a high Ball% over 50%.
• Few, if any, sinkers are effective compared to other pitches
• Splitters: Just seven guys who feature the pitch but the pitch is effective for all of them. Underutilized option.

Overall, I’m happy with how everything turned out. I feel I’m about 95% finished with a few places to clean up around the edges. In addition to cleaning up the data, here are some areas I could improve on:

• As mentioned above, I could use a different equation for the batted ball adjustment.
• Regress the pERA values. In small samples, extremes show up, so it’s nice to incorporate some regression. One item I am going to immediately is a regress is the batted ball data. With fewer samples than strikes and balls, the batted ball takes longer to stabilize.
• I want/need to dive into a pitcher’s Hard Hit Rate (Exit Velocity based). Every time I’ve tried to incorporate it, I can’t find any correlation once the fieldable data (launch angle) is included.
• I want to run some year-to-year correlations and comparisons. The correlations compare to xFIP and SIERA, but I want a final equation before running all these tests and publishing the results.
• The next step and it will be a major one going forward is how much to weigh a pitcher’s STUFFF (pitch shape-based projections) with the actual results. Should STUFFF be used for three games? A 50/50 mix after five games? This step would then determine how much to adjust season-long projections.
• The pERA values for right and left-handed hitters. How pitchers handle certain hitter-handedness can give an idea of when a pitcher might struggle against certain lineups.

Well, that’s it for this article. What I’d like is any feedback on how it might be improved. I need to nail down these pitch results values before moving to the next step of combining them with STUFFF values.