with pm.Model() as model:
trace = pm.sample(draws=500)
az.summary(trace)
r_hat = 1.00 ess_bulk = 1842
divergences: 0
In partnership with
Live today · 5:00 PM ET
Most people run pm.sample() and hope.
When it works, it looks easy. When it doesn't, you're staring at a wall of warnings with no idea whether the problem is the sampler, the parameterization, or the model itself. This session opens the box.
Today · Wed 12 August · 5:00 PM ET
MCMC in Practice: Fitting and Checking Bayesian Models with PyMC
Dr. Chris Fonnesbeck · Core developer of PyMC. He wrote the library you're about to debug.
PyMCArviZNUTSDiagnostics
Try it — a real Metropolis sampler
accept — · draw 0
Drag the step size. Too small and every proposal is accepted but the chain crawls. Too large and almost everything is rejected, so it sticks in place. Somewhere in between you get the fuzzy caterpillar you actually want. Chris will show you how NUTS finds that setting for you — and what to do when it can't.
What you'll walk out with
- A real read on R-hat, ESS, MCSE and divergences — what each one is actually measuring, and which failure it catches
- The difference between a broken sampler and a broken model — and why a model can converge beautifully and still be wrong
- A Statcast pitch-velocity model taken through the full loop: fit it, break it, diagnose it, fix it
- An interactive marimo notebook with six explorers — rejection sampling, step-size tuning, the curse of dimensionality, NUTS vs Metropolis side by side
Live, with the notebook open
Real Q&A with a sports analytics expert.
Session 2 — Session 1 not required
Starts today · 5:00 PM ET
Save your seat.
Register below and your join link lands in your inbox straight away.
Can't make 5:00? Register anyway and we'll send you the recording.
No spam. Your join link and the recording.
Posit Managed Services Partner
20+ instructors from PyMC, Apache Arrow, Wharton, Columbia & pro sports
Taught by people who build the tools you use