Examples

Executed end-to-end and re-run in CI with reduced sampling settings.

• Univariate forecasting — weekly BART ridership with a random-walk local level, annual Fourier seasonality, and a Student-T likelihood; ADVI, CRPS evaluation, and a rolling-origin backtest.

• Hierarchical forecasting — hourly arrivals to one BART station from all 50 origins at once: batch dims, per-series levels and weekly seasonality, shared scales.

• Electricity demand with covariates — Victoria hourly demand with daily/weekly seasonality and a quadratic temperature response; forecasting with full-horizon covariates.

• Exponential smoothing in state-space form — a damped Holt-Winters single-source-of-error recursion written with pytensor.scan, fit with NUTS.

• Local level two ways — scan-based Markov latents vs. the pymc-extras statespace backend on the same model: posterior quality, runtime, and a shared backtest.

• Retail demand under stockouts — censored demand on the FreshRetailNet-50K panel: a hierarchical damped-trend model with a floored saturating availability factor, batched predictive sampling, and a counterfactual demand forecast at full availability.