Time Series forecasting with NumPyro
summary
Implements custom probabilistic time‑series in NumPyro that actually recover unseen demand using censoring likelihoods (switching forecasts from flat to upward when capacity truncation is modelled), composes components via a reusable transition written for the scan function, and fits hierarchical pools that pull regional trends into sparse-product forecasts. Shows discrete alternatives (zero‑inflated negative binomial) and compares inference: full MCMC (NUTS) is accurate but slow, SVI vs NUTS trades bias for speed and scales far better for many series.
AI-written summary · about
chapters (17)
- 0:00Introduction
- 1:17Motivation for forecasting
- 3:04Classical methods and baselines
- 6:34Probabilistic forecasting with NumPyro
- 9:26Understanding scan in NumPyro
- 15:26Exponential smoothing in NumPyro
- 21:53ARMA models and transitions
- 23:21Handling intermittent demand
- 28:45Zero-inflated models
- 30:59Censored demand challenges
- 35:35Business versus model metrics
- 36:16Hierarchical modeling
- 41:12MCMC vs variational inference
- 42:38Pricing models and elasticity
- 45:51Extending to other frameworks
- 46:56Experimental state space models
- 48:28Q&A
related talks
29mProbabilistic demand forecasting at scale
41mReliably forecasting time-series in real-time
42mRobust Automated Forecasting in Python and R
40mMulti-Series Forecasting at Scale with StatsForecast
35mWhat we learned by running a large custom Bayesian...
01h 27mBayesian Time Series
42mRobust Automated Forecasting in Python and R
30mRealtime Time Series Anomaly Detection in Production