Spin Doctor BB

@903124S I think the flaw with looking at only consecutive pitch pairs is that any effect of a pitch 2+ pitches back is completely ignored. I do think more recent pitches are more important, but I don’t think you NEED a tunnel with the previous pitch for tunneling to work

My tweet yesterday shown a narrow sense of tunneling which is between consecutive pitch pair though I've tried on wider sense of tunneling which is it synergy of pitches in a pitcher arsenal. x.com/903124S/status… 2/n

Upon discussion with @srbrown70 here I'd briefly mention why I think a broader sense of pitch tunnel could also not be a thing though I'm very welcome people to post rebuttal since just a rough approach. 1/n

@903124S @albertbayes @tangotiger But then also that poses the question of whether the balance of the importance of intended location/tunneling is dependent on the pitcher’s command

@903124S @albertbayes @tangotiger Even just anecdotally from my experience with MLB The Show most top players do use some form of tunneling while also having the fastball/whatever regular pitch in the corners/on the edge

Why pitch tunneling don't work using Physics of baseball Few years ago I've found pitch tunneling which means two pitches head to two location after appear in similar location don't really work base on Statcast data. Now I'd try to see a physical explaination of it 1/n

@903124S @srbrown70 I do think I probably need to control for count (and maybe number of previous swings, bc it does affect results), but my model only looks at 1) current trajectory of pitch, 2) distances between trajectories, and 3) whether the hitter swung at previous pitches

@srbrown70 @SpinDoctor_BB Yeah I've read multiple works trying to use time series e.g. LSTM or transfromer to expalin it. Still imo you still need to separate other pitches in the arsenal since due to regression to the mean other pitches would definitely impact run value without consider tunneling

@903124S @albertbayes @tangotiger role of the pitch. If you look at only one pitch, then it makes sense to throw it only in the corners, but if you have maybe a string of 5 consecutive pitches, then it may make sense to have some pitches that coax called strikes/chases. And that’s where I think tunneling is

@903124S @albertbayes @tangotiger I think if you’re solely trying to maximize run value effect, it makes sense to only throw your best pitch in the corner of the zone. Of course, that’s not the most viable strategy, and I do think if you deviate from that, it’s a sort of give and take situation depending on the

@srbrown70 @903124S I’d argue that there’s 3 types of tunneling: 1) pitches within a plate appearance 2) pitches on average in an arsenal (basically arsenal effects in BPro, Mix+) 3. the hitter uses representative heuristics to guess current pitch based on all pitches a hitter has ever seen

@903124S @SpinDoctor_BB see second photo for caption which clearly ran way too long

Here is a link to the slideshow I presented yesterday at Saberseminar 2025. I hope to release a writeup of what was covered in the presentation + some extra bits of research sometime soon on my substack. Very, very fun experience docs.google.com/presentation/d…

Unbelievable presentation by @SpinDoctor_BB evaluating the effect of pitch tunnels on swing and contact probability. His modeling and insights were next-level for such a raw analytical topic.

haven't done much baseball research recently (will start again soon) anyways excited to announce that I will be taking my talents to UCLA as a Math of Computation major for the next 4 years

thinking about evaluating batter approach/selectiveness by comparing the density plot/KDE of top 50% EV batted balls and the density plot of predicted (if the pitch was a fastball) locations

been busy doing stuff (college applications) but thinking about how pitching is probably still stuff/tunnel dependent and how the number of pitches prevents degradation of stuff (allowing for good performance multiple times through the order)

@_holmj_ this is really really useful (and hopefully will save me the 20 minutes required to download data from pybaseball)

Just updated the statcast_pitches package to include new arm angle data and all other new columns! Download as much or as little baseball savant pitch data from 2015-present in 3-45 seconds. github.com/Jensen-holm/st…

sneak peek 2 from tunnelling article: in order to quantify tunnelling, you need some kind of model to know just based on plain singular pitch data how good a pitch is. and after some headbashing, i threw together a xgbclassifier pitch quality model (here's 2023's best pitches)

Here are the feature importances of these models, key note is that Xs/Zs close to the beginning/end of the 15 ft path are important, as well as player_name (introduced as a categorical variable, so within pitcher effects do exist)

sneak peeks from tunnelling article: a hitter (or theoretical machine learning powered hitter) can predict final plate x (r2 = .78) and final plate z (r2 = .94) just from ball flight positions within the first 15 feet decently well (30 feet was too well)

@MaxwellResnick League wide, but I’m including the pitcher as a variable (so it hopefully weighs the pitcher’s pitches as more important, currently running code to get feature importances on this)

@_holmj_ see if pitches prior to previous pitch have some impact, albeit a deafened impact. anyways very good work

@_holmj_ very interesting, i'm looking through some of your code and a lot of the ideas seem very similar to stuff in my jupyter notebook i'm also looking at sequential pairings but defining the tunneling as integral of euclidean distance and then final euclidean distance, but I want to

if you have any tunnelling/pitch synergy articles/tweets/anything that you really like, let me know so I can reference it in my tunnelling article

I created a Bluesky account if you want to follow me there article coming soon