Custom Pitch Stuff Grades

 
I've made an app allowing anyone to see what my models think of any hypothetical pitch.

Custom Pitch Stuff

I've written a lot on this blog about my pitch evaluation models, taking pitch characteristics and locations, without outcomes, to assess quality. You may have seen my ratings of MLB pitchers on my web app. My Stuff model doesn't look at where a pitch goes, it grades a pitch solely on its movement, speed and release point. So far I have only applied this model to real MLB pitches with publicly available data.

However, there is no reason that this can't be applied to any hypothetical pitch. More minor league and amateur pitchers than ever are analyzing their pitches by using tracking data to measure the velocity and movement they can impart on the ball. I've made an application which can take these inputs and evaluate a player's Stuff, along with showing how the Stuff Grade changes with the velocity and movement of the pitch.

How to use the app

I'm not an expert at making user interfaces, and I'm aware that the inputs to the app can be a little clunky and hard to understand, so here's a short checklist of the key things to do to measure pitch Stuff using the app. I'll go through each point in more detail below.

1. Are you measuring stuff for an existing MLB pitch or a hypothetical pitch?
2. Is the pitch their primary fastball?
3. Does the movement direction make sense for the pitcher handedness?

1. 

If you are measuring the stuff for an existing MLB pitch then things are a little easier. You just need to select the pitcher and pitch type from the third and fourth boxes and the sliders will automatically update to the average (median) qualities of that pitch from 2021.

If you're not using an existing pitch then you'll have to update the sliders manually. Firstly make sure that the correct pitch type and handedness are chosen in the top two boxes. The pitch types are as follows:

• FF = Four-Seam Fastball
• SI = Sinker
• FC = Cut Fastball
• SL = Slider
• CU = Curveball
• CH = Changeup
• FS = Splitter

Then the sliders should be easy to use. Speed and spin rate are in units of mph and rpm respectively. Pitch movement is in inches and is relative to a pitch with no spin (this is different to the total vertical movement as shown on baseballsavant.com which includes the effect of gravity). Release points and extension are where the ball leaves the hand. All these are relative to the catcher's perspective so negative horizontal movement / release point is armside for a RHP and gloveside for a LHP.

2.

If the pitch you are measuring is the primary fastball then leave the checkbox below the sliders ticked and press the "Go!" button.

If not then you will have to add in the primary fastball speed and movement characteristics. If you are using an existing pitcher then you can change the input to their primary fastball then click the "Update fastball characteristics.." button to automatically set these values. Otherwise just set the values to the average primary fastball of the pitcher whose Stuff you are measuring.

3.

Make sure the horizontal movement direction and release point are correctly aligned for the pitcher handedness and pitch type. Remember negative is armside for a RHP and negative for a LHP. Similarly positive is gloveside for a RHP and negative for a LHP.

Here are the example inputs for a Tyler Glasnow (RHP) 4-seam fastball:

Below are the inputs for a Tyler Glasnow curveball (RHP non-fastball):

Understanding the results

After clicking the "Go!" button the app will run a bunch of predictions. A loading screen should pop up for around 10 seconds but if it doesn't please don't click the "Go!" button a bunch of times or the app will run even more slowly processing all the requests!

The first result is the pitch grade. These grades are adjusted to the 20-80 scouting scale so that 50 is MLB average and every 10 points up or down represents one standard deviation. An easy way to understand this is that 50 = top 50%, 60 = top 15%, 70 = top 2.5%, and 80 = top few players in MLB.

The underlying metric for these Stuff Grades is the expected run value per swing which is based on predicted whiff rates, foul ball rates, and the types of batted balls that are expected to be produced, both in exit velocity and launch angle. The grades are normalised by pitch type, the average slider will get much better results per swing than the average sinker so the grade is always shown relative to the average Stuff for that particular pitch.  These grades do not take into account non-swing outcomes and also do not look at pitch locations. In reality these are much more important than raw Stuff, but are harder to measure in short sample sizes.

In addition the expected Whiff%, Groundball%, and HardHit% are included to give more context to why the model likes or doesn't like a particular pitch. These will also depend heavily on where the pitch is located in reality.

Below the main results are a series of graphs showing how the Stuff Grade, xWhiff%, xGB%, and xHH% vary when modifying pitch speed and movement. The first set show how varying the pitch speed and the total amount of movement affects the results. The second set show how changing the movement angle affects the results. These can be useful to find out what changes would improve my model's evaluation of a particular pitch's Stuff.

Downloading your results

After running a set of predictions, a box will appear where you can add a label to your pitch and save it. Then you can download the results as a csv file by clicking the button below.

You must also choose a unique identifier for your pitches and a label for individual pitches before saving and downloading.

FAQs:

• Why does the Stuff Grade change when I run with the same pitch multiple times?
• I simulate the same pitch many times in different contexts (various ball/strike counts and batter handedness). I also vary the pitch parameters slightly for each simulated pitch to represent the real life variation in pitch speed, spin and movement from pitch to pitch. This means that the results can sometimes vary slightly for the same pitch parameters. 
• Why are these Stuff Grades different to those on the MLB pitcher grade app?
• The MLB pitcher grade app uses all the pitches which were thrown by a pitcher which includes varying context such as the ball-strike count and the batter handedness. However, this app attempts to measure context neutral Stuff by simulating the pitch in a representative variety of counts and against a mix of LHB and RHB. This can explain small differences in Stuff Grade between the two apps.
• My results look very odd/wrong
• There are multiple reasons why the predictions may be not working correctly. Firstly make sure that you have the right pitch type, handedness, and movement directions. Use the example images above to see what the inputs should look like. If that is not the issue then the pitch may not lie within the usual bounds of existing MLB pitches, the model cannot extrapolate beyond the data which it has been trained on and so the predictions will only make sense if the pitch is already similar to other pitches at the major league level. 

If you are still not sure about anything then feel free to contact me at camerongrove96[at]gmail[dot]com.