Whiff + (Heat+ update)
About a month ago, I released my first writeup on my initial stuff model Heat+. Since then, in between class and spring break, I've made some major upgrades and fundamental changes to the model that have greatly improved whiff rate prediction power. I've also renamed the model, from Heat+ to Whiff+ because I think that sounds better and is more descriptive of what the model tracks. In terms of future changes, I would like to implement a location aspect to my model, as fastballs at the top of the zone are much more effective.
Fundamental Change
The biggest change made comes from what outcome the model is being trained to look for. Stuff+, Heat+ and PitchingBot are trained on Run Value, meaning it looks at the true outcome of the pitch. Whiff+ is trained Whiff Probability, meaning it looks at whether or not the batter made contact with the pitch. This removes a lot of external factors and focuses more on the physical qualities of the pitch and its ability to miss bats. This process is not necessarily better than Run Value training, but answers a different question.
Whiff+ has an R^2 value of .478 when looking at same season whiff rates, indicating 47.8% of the variation in whiff rates can be explained by Whiff+. When looking at predictive rates, 2024 Whiff+ had an R^2 of .435 when looking at 2025 whiff rates, indicating 43.5% of the variation in Y2 whiff rates can be explained by Y1 Whiff+.
As can be seen, Whiff+ is considerably better than Stuff+ at predicting whiff rates and FIP across Y1 and Y2. It is important to note Whiff+ is still only trained on fastballs, I intent to train it on other pitches at some point down the line.
Payton Tolle stands out as a young prospect for my Red Sox. I'm excited to see how we can develop his secondaries and improve his command. Mason Miller is understandably near the top, while Edwin Uceta benefits from a great Tampa Bay coaching staff. Elvis Alvarado catches my eye as an impact bullpen arm for a promising Athletics team. Taylor Rashi stands out as a pitcher who's effectiveness is highly influenced by a unicorn trait, in this case his extreme arm angle at 72.7.
Leaderboards
Payton Tolle stands out as a young prospect for my Red Sox. I'm excited to see how we can develop his secondaries and improve his command. Mason Miller is understandably near the top, while Edwin Uceta benefits from a great Tampa Bay coaching staff. Elvis Alvarado catches my eye as an impact bullpen arm for a promising Athletics team. Taylor Rashi stands out as a pitcher who's effectiveness is highly influenced by a unicorn trait, in this case his extreme arm angle at 72.7.
Feature Importance
Here's the feature importance map for this first version of Whiff+.
The biggest difference is perceived velo, which is just taking extension into account when looking at pitch velocity. Arm angle interactions were removed, but release height interactions were added and IVB takes on a greater role. Batter handedness was removed as well to make the model more pitch-quality dependent. This is something I might mess around with in the future.
Hyperparameters
Whiff+ has a max_depth of 6, compared to Heat+'s 3. The added depth allows for better training of the model, as it is able to ask more questions before making a prediction. ETA has been increased from .03 to .05, allowing for faster learning.
Stickiness and Stabilization
Stickiness and stabilization stayed pretty much the same, here are the graphs for those who care.





Comments
Post a Comment