What we learned by running a large custom Bayesian...
summary
Kolonial.no builds a custom Bayesian demand-forecasting pipeline in PyMC3 that models customer habit segments and transition-based state space dynamics, outputs prediction intervals (1st–99th percentiles) and yields measurable planning efficiency; initial MCMC samplers take ~15 hours and later variational inference cuts training to ~25 minutes for a model with ~thousand random variables. It works in production but monitoring, convergence diagnostics, brittle model wiring and holiday/regime shifts remain real operational problems.
AI-written summary · about
chapters (10)
- 0:00Bayesian forecasting at Kolonial
- 2:14Forecasting grocery demand
- 5:01Bayesian models for demand
- 8:27Custom models with small data
- 12:04Orders per customer with PyMC3
- 15:41Gains and efficiency improvements
- 18:32Challenges in large Bayesian models
- 21:41Optimizing model runtime
- 23:52Evaluating prediction intervals
- 26:44Q&A
related talks
34mFast, beautiful and easy Bayesian modelling: Can you have it all?
42mSolving Real-World Business Problems with Bayesian Modeling
01h 26mProbabilistic Programming and Bayesian Computing with PyMC
42mBayesian Non-parametric Models for Data Science using PyMC3
01h 27mBayesian Time Series
39mProbablistic Programming Data Science with PyMC3
44mAn introduction to PyMC3
36mClusterf*ck: A Practical Guide to Bayesian Hierarchical Modeling in PyMC3