Practice · Baseball
Six thousand pitches from a Statcast sample, each with its pitch type and release speed, for the full model-checking loop.
One row per pitch: the type code and the release speed. Three pitch types, three very different speed distributions, and a mixture that a single-scale model gets wrong in a way the sampler diagnostics cannot see. It is the dataset the MCMC session uses to let each failure decide the next revision.
pitch_type | the pitch type code, for example FF, SL, CH |
release_speed | mph at release |
6,000 rows. Statcast, via Baseball Savant and pybaseball; a sample from one two-week window of a season, prepared for the MCMC in Practice session. For practice; cite the original source in anything you publish.
Compute the mean and standard deviation of release speed for each pitch type, and draw the three distributions on one plot. Which pair overlaps most?
Model speed with an intercept plus a dummy variable for every pitch type, and sample it. The chains will tell you something is wrong before the estimates do. Name the problem, fix the parameterization, and sample again.
Give each pitch type its own residual scale and run a posterior predictive check against the single-scale model. Then compare the two with PSIS-LOO. Does the check and the comparison agree, and what would you still want to add?
No answers are published here. Work the question, then watch the session and compare your route with the instructor's.
MCMC in Practice. Every dataset in the practice library names its session the same way.
First name and email. One message when a session recording, a skill, a cheat sheet or a posting worth your time arrives here, and no other mail. Replies reach a person.