Leader track 🎲 · six sessions, no code
For whoever signs off a number from small data. What a prior costs, why eighteen observations give you a distribution and not a point, how to borrow strength across units, how to read a posterior in a meeting, the argument about priors the field has not settled, and what to require in writing.
What a prior costs
A prior is a number of pseudo-observations. At Northgate the sensible one was confidently wrong; at Westfield a wrong one still moved the answer by 0.026 at n = 900.
🔍 Leader session 2Small data, honest uncertainty
Eighteen observations cannot give you a point. What they can give you is a distribution, and the sentence that goes on the slide.
🏪 Leader session 3Borrowing strength
Twelve small outlets know more together than any one alone. The pooling ladder, 0.054 to 0.031 to 0.025, and the six outlets it made worse.
📊 Leader session 4Reading a posterior
P(B beats A), expected loss, and why 0.915 was not enough to stop. Credible against confidence, in one sentence.
⚖️ Leader session 5The prior argument
Subjective, reference, weakly informative. The field has not settled it, and this session does not pretend to.
✍️ Leader session 6Commissioning a Bayesian answer
What to require in writing: the prior as a pseudo-count, the interval kind and level, the decision rule, the check that was run.
Practitioner track 🧪 · ten sessions, hands on
For the people who build the models. Conjugate updates by hand, the posterior as a distribution, MCMC and its diagnostics, PyMC 6 with nutpie and DataTree, hierarchical models, Bayesian A/B by expected loss, predictive checks, the executive sentence, and a capstone bench scored against the truth.
Beta-Binomial by hand
Conjugacy: a plus y, b plus n minus y. The posterior mean as a weighted average, and why that makes the prior legible.
🎚️ Practitioner session 2Four priors
Flat, weak, informed, wrong on the same eighteen customers. Pseudo-observations, and the interval that misses.
📈 Practitioner session 3The posterior is a distribution
Grid, then sampling. Summaries as quantiles, and why 89 percent equal-tailed is the default now.
🔗 Practitioner session 4MCMC and its diagnostics
Why sample at all. Chains, R-hat, effective sample size, divergences, and the rank plot that replaced the trace.
🐍 Practitioner session 5PyMC, the first model
PyMC 6.3, nutpie by default, a DataTree instead of InferenceData. Westfield end to end, as the tools are in 2026.
🧩 Practitioner session 6Hierarchical models
Partial pooling with a known truth. kappa as learned pooling strength, and where shrinkage costs you.
⚡ Practitioner session 7Bayesian A/B decisions
P(B beats A), expected loss, ROPE. The rule stops at n = 100 and is wrong 2 times in 200, counted.
🔬 Practitioner session 8Checking the model
Prior and posterior predictive checks. What a check can catch, what it cannot, and where LOO fits.
🗣️ Practitioner session 9Saying it to executives
The sentence, the interval with its kind and level, the probability of a difference, and the stake.
🧪 Practitioner session 10The pooling bench
Three live benches on a world whose truth is written down. Priors, pooling, and the A/B rule, all scored.
The whole thing on one page
Every idea in the course and the session that owns it. Click any node to jump there.
Three ways through
Everyone starts at session 1 of their track. These are the shortest useful routes if you cannot take all sixteen.
What is real here
Nothing on the benches is a stored result.
| The world | Twelve outlets from a model written out in assets/bayes-live.js, fixed seed, every true rate visible |
| Every posterior | An exact conjugate update, summarised by 20,000 draws at the ArviZ 1.x default of 89 percent equal-tailed |
| Partial pooling | An empirical-Bayes fit that maximises the marginal likelihood across outlets, so the pooling strength is learned, not set |
| The decision rule | Expected loss over 20,000 paired draws, and a wrong-stop rate counted over 200 replays rather than promised |