Author: Jonah Hofeld – Applied Mathematics Major at Boston University and CornBelters Analytics Intern

1. Introduction

For over a decade, strikeouts have been rising at every level of baseball. At the major league level, whiff rates have risen from roughly 19%-21% (before Statcast revolution), to 25%-26% as of 2026. There is no doubt that this is a result of the analytical movement on the pitching side, but what about for the Prospect League?

Since the Prospect League has newly implemented Trackman in every stadium, we are now able to track the drivers of whiffs, and what specific profiles create whiffs for certain pitches. In this article, I utilize 2026 Prospect League Trackman data (from all games played so far) to build a logistic regression model that identifies the physical pitch characteristics that are most predictive of generating a whiff.

2. Defining a Whiff

A whiff in baseball is simply defined as a swing and a miss. From this, we can deduce a whiff rate or Whiff%, which is the amount of times a batter swings and misses divided by all swings. FanGraphs defines Whiff% as the inverse of Contact%. Another related statistic is SwStr%, defined as how often a pitcher generates a swing and miss per pitches he throws.

For this analysis, we focus on Whiff% since we want to understand what makes a pitch hard to hit when a hitter commits to swinging.

3. League-Wide Whiff Rates

Before jumping into the fancier techniques, it is always helpful to know the baseline. Across all 2026 Prospect League games in our dataset, the league-wide Whiff% is 25.2%. For pitch-by-pitch data, we see:

Pitch TypePitchesSwingsWhiffsWhiff%
Sweeper7102076933.3%
Slider4,8661,86261633.1%
Curveball2,56588229133.0%
Changeup1,91489026229.4%
Cutter84738711128.7%
4-seam16,5117,1671,54721.6%
Sinker1,0134598719.0%

You probably didn’t need this chart to know that breaking balls dominate the top of the whiff leaderboard. Sweepers, sliders, and curveballs all sit at 33% or above, which is roughly 8 points above the league average for whiffs. Changeups and cutters sit just above the league average, while four-seamers and sinkers trail since they are designed to be hit, and their whiff value comes from location and velocity rather than shape-induced confusion.

3. The Model

To identify what pitch characteristics drive whiffs, we fit a separate logistic regression model for each of the seven pitch types in our focus. Logistic regression is a supervised machine learning method, meaning it learns from labeled data. In this case, the model learns from swings that are a whiff (1) or a non-whiff (0). The model then outputs a probability between zero and one.

The core of logistic regression is the sigmoid function:

ℙ(Whiff)=11+e(−z)\mathbb{P}{(Whiff)} = \frac{1}{1 + e^(-z)}

This reads “the probability of a whiff is equal to the sigmoid function.” The sigmoid function takes any real number z and squishes it into the range (0,1). This allows us to interpret the output as a probability.

z itself is a linear combination of the input features and their learned weights:

z=β0+β1x1+β2x2+...+βnxnz = \beta_0 + \beta_1x_1 +\beta_2x_2 + …+ \beta_nx_n

Where each x is a pitch characteristic (velocity, induced vertical break, horizontal break, spin rate, etc) and each β is the coefficient that the model learned for that feature. A positive β means that as the feature increases, the probability of a whiff increases. A negative β means the opposite.

All features were standardized before fitting using the transformation:

xi∗=xi−μiσix^*_i = \frac{x_i – \mu_i}{\sigma_i}

Where

μi,σi,xi∗\mu_i, \newline \sigma_i, \newline x^*_i

are defined as the mean, the standard deviation, and the resulting feature, which follows a normal distribution with mean zero and standard deviation one:

xi∗∼𝒩(0,1)x^*_i \sim \mathcal{N}(0, 1)

for a feature i across all pitch types. This puts all features on the same scale so that the coefficients are directly comparable within each model. A coefficient of 0.3 on velocity means velocity has a stronger effect on whiff probability than a coefficient of 0.1 on spin rate.

To evaluate the model fit, we used AUC (Area Under the ROC Curve), which measures how well the model distinguishes whiffs from non-whiffs on a 0.5 (random) to 1.0 (perfect) scale:

Pitch TypeSwingsWhiff%AUC
Sweeper20733.3%0.772
Slider1,86133.1%0.743
Curveball88233.0%0.694
Changeup88929.5%0.692
4-seam7,16421.6%0.633
Cutter38728.7%0.632
Sinker45919.0%0.608

The sweeper model at 0.772 is the strongest of any pitch, meaning pitch characteristics are highly predictive of whiff outcomes. This makes sense because a pitch like a sweeper relies on a specific pitch shape to be effective. The slider follows at 0.743. Breaking balls as a group are seemingly the most model-predictable whiff pitches. Four-seams and cutters are much harder to predict from stuff alone since whiffs for these pitches are more location-dependent rather than shape-dependent. The sinker at 0.608 reflects the fact that sinkers are for generating ground balls as opposed to inducing whiffs.

4. Features Generating Whiffs by Pitch Type

The coefficient plots below show the standardized coefficients for each pitch type and how they affect whiffs. Red bars show features that are positively associated with whiffs while blue bars show features negatively associated with whiffs. The length of each bar represents how strongly that feature influences whiff probability relative to others.

Four-Seam Fastball: As the coefficient plot shows, vertical approach angle is the dominant whiff driver for the four-seam with plate height as a distant second. This tells us that a release point, vertical break, and pitch location are all important for an effective fastball since these features influence VAA. We can also see that a four-seam up in the zone is more likely to induce a whiff.

Slider: The slider’s coefficient plot shows a clean story with IVB and velocity as the top two drivers followed by extension and spin rate. This tells us that the most effective sliders in the Prospect League combine depth with velocity. Unlike the curveball where break is the dominating feature, the slider rewards a more complete profile with sharp downward action and velocity. Additionally, the strong negative coefficient on VAA shows that steeper sliders suppress whiffs.

Curveball: The curveball’s plot shows induced vertical break as the clear leader with spin rate, velocity, horizontal break, and extension contributing positively behind it. This shows that elite curveballs in the Prospect League are characterized by depth, high spin rate, and velocity. The negative coefficients show plate height as the dominant suppressor, meaning curveballs left up in the zone are much less effective at generating whiffs. This makes complete sense because it is easier for hitters to track a curveball (hanger) that doesn’t finish down.

Changeup: Induced vertical break is again a driver for whiffs with velocity and horizontal break following. This is defendable since effective changeups have either good sink, good run, or both. Velocity as a driver reflects the fact that a changeup is meant to appear as a fastball, and a slower changeup may be easier to recognize out of the hand. The biggest suppressor is again plate height, showing an elevated changeup as a very hittable pitch.

Cutter: The cutter plot is defined by VAA at the top, similar to the four-seam. Extension and spin rate follow as drivers, meaning cutters thrown with high spin and closer to the plate generate more deception and therefore more whiffs. Since the cutter is essentially a hybrid between a slider and a fastball, horizontal break and velocity work against it because too much of either takes away from the tight late cutting action that makes the cutter effective.

Sinker: For the sinker, we can see that plate height seems to be the biggest driver with induced vertical break following closely. This means a sinker with ride at the top of the zone generates the most whiffs. VAA as the biggest suppressor makes sense because a flatter approach angle works against the sinking movement.

Sweeper: On the positive side, the plot shows VAA, IVB, spin rate, and extension as contributing to more whiffs. This suggests the best whiff-generating sweeper combines a flat approach angle with strong shape and deception. On the negative side, we see plate height and plate side as the biggest suppressors. This works since we want the sweeper to start in the zone and finish down and away.

5. Limitations

Trackman Pitch Tagging: While all pitches in the dataset are manually tagged by CornBelters analytics intern Brady Goodman using pitch feature data, we don’t know every pitchers arsenal and therefore can’t correctly tag every single pitch correctly, since features vary by pitcher. As an example, a pitcher’s sweeper might have a slider-like profile or a pitcher’s natural spin might be lower, causing a misclassification. Misclassified pitches pollute each pitch type’s model with observations that don’t belong, which can distort the coefficient estimates. Therefore, any findings should be interpreted with this in mind, especially for pitches with less distinct profiles.

Sample Size: Since Prospect League play began roughly 2 weeks ago, certain pitch types have relatively thin samples like sweeper (n= 207) and cutter (n=387). While logistic regression is fairly robust to small samples compared to more complex models, coefficients on minority pitch types should be treated with more caution then those on four-seams (n= 7,164) or sliders (n= 1,861). As the 2026 season progrosses and our data set grows, these models will become more reliable.

Count Context: It is important to note that whiff rates vary substantially by count. A pitcher in an 0-2 count can generate a whiff on a dirt ball but if this pitch was thrown in a 2-0 count, it would likely be taken by the hitter. This model pools all counts together, meaning coefficients reflect average effects across full count distribution rather than isolating whiff generation from count-level data.

5. Conclusion

Across seven pitch types and nearly 12,000 swings so far in the 2026 Prospect League, we are able to draw some notable inferences. Breaking balls: sweepers, sliders, and curveballs are the most reliable whiff generators, and their effectiveness is driven by depth and spin. Changeups work when they have sink and substantial velocity to keep hitters guessing. Fastballs generate whiffs at lower rates overall, but elevated four-seams with flat approach are most dangerous. Sinkers are ground ball pitches, and the model reflects that.

On the pitcher development side, we see that whiff generation is strongly correlated with the physical characteristics of a pitch. These characteristics: height, velocity, spin, etc, reveal predictions on whether a hitter is likely to make contact or not. In a developmental league where pitchers are constantly refining their arsenals, understanding the drivers for whiff generation gives both players and coaches a better target to work toward.

Sources

Posted in 2026 Kernelytics Blog, Uncategorized

Leave a Comment