Easy Model Blending using Pseudo-Bayesian Model Averaging in Python
summary
Repository releases an open source BayesBlend package that fits candidate models with CmdStanPy/ArviZ, extracts out-of-sample likelihoods and posterior predictions, and fits pseudo‑Bayesian weights via hierarchical stacking. A 50‑insurer study (data from 1988–1997) shows stacked blends improve expected log predictive density versus individual models, that pseudo‑BMA is sensitive to outliers, and that blending multiple restricted models recovers time‑varying behavior — the library runs and the hierarchical blend wins on the validation set.
AI-written summary · about
chapters (14)
- 0:00Introduction
- 1:04Model blending matters
- 5:17Bas blend library
- 8:06Requirements for model stacking
- 10:57Addressing audience questions on model assumptions
- 13:44Mathematical foundation
- 15:14API demo with example
- 19:07Insurance loss forecasting
- 24:01Loss development modeling
- 28:38Candidate models for blending
- 30:25Blended model performance
- 31:51Blending methods comparison
- 33:17Blending approaches
- 34:39Summary
related talks
06mPitfalls of Full Bayesian Inference in Universal Probabilistic Programming
34mFast, beautiful and easy Bayesian modelling: Can you have it all?
01h 27mFlexible Statistical Modeling
38mUsable A/B testing – A Bayesian approach
42mSolving Real-World Business Problems with Bayesian Modeling
14mProbabilistic Programming in Python
31mBias Meets Bayes: A Bayesian Perspective on Improving Model Fairness
01h 24mBayesianism and Survival Analysis