Learning Bayesian Statistics #152 A Bayesian decision theory workflow, with Daniel Saunders
24 snips
Feb 26, 2026 Daniel Saunders, a philosopher-turned-Bayesian practitioner who builds decision-theory workflows and PyMC/PyTensor tooling. He discusses separating beliefs from utilities for team workflows. He explains shifting evaluation from accuracy to business value and demonstrates vectorized posterior optimization with PyTensor for pricing and profit. He covers risk-averse utility transforms and practical safeguards for industrial decision making.
AI Snips
Chapters
Books
Transcript
Episode notes
Separate Decision Theory From Modeling
- Decision theory should be treated as a separate but complementary workflow to model building.
- Daniel argues teams benefit from separating probability (posterior) from utility so modelers and decision-makers can work independently and then combine results.
Prioritize Decision Value Over Accuracy Metrics
- Evaluate models by the value of decisions they enable rather than by generic accuracy metrics.
- Daniel recommends quantifying the cost/benefit of added complexity (e.g., extra features or compute time) in dollars to justify modeling choices.
Use PyTensor To Glue Models And Optimizers
- Use PyTensor's compute graph to separate model parts and compose optimization objectives without rewriting code.
- PyTensor lets modelers expose price→sales chunks while optimizers add profit/unit operations and compile efficient callables for SciPy.
