arviz-devs/arviz

> The backend-agnostic diagnostics and visualization layer of the Python Bayesian stack — if your sampler emits posterior draws, ArviZ is how you check them.

GitHub repo · Official website · License: Apache-2.0

Overview

ArviZ ("AR-vees") is a Python package for exploratory analysis of Bayesian models: posterior summaries, MCMC convergence diagnostics, model comparison, and posterior/prior predictive checking[^1]. Its core design decision is to be sampler-agnostic: PyMC, Stan (CmdStanPy/PyStan), NumPyro, Pyro, emcee, and TensorFlow Probability all convert their output into ArviZ's InferenceData container, so the same plot_trace, summary, loo, and rhat calls work regardless of which probabilistic programming library produced the draws.

The star count (~1.8k) badly understates its reach. ArviZ is a hard dependency of PyMC — pm.sample() returns an InferenceData object directly — and the de facto standard post-sampling toolkit across the Python PPL ecosystem. It is a NumFOCUS-affiliated community project with institutional backing from Aalto University and FCAI, and Aki Vehtari (of Stan and the R-hat/PSIS-LOO papers) is among its authors, which is why its diagnostics track the current statistical literature rather than textbook-era defaults[^2].

The defining tension in 2026 is the 1.0 modular split: the monolithic arviz package is being replaced by arviz-base (data + conversion), arviz-stats (diagnostics), and arviz-plots (visualization), installable today via pip install "arviz[preview]"[^3]. The main repo now largely carries documentation and the meta-package (GitHub's language detection reports "TeX" for this reason), while active development happens in the sub-package repos. Users face a real migration, with a dedicated guide[^4].

Getting Started

pip install arviz                  # stable 0.x API
pip install "arviz[preview]"       # 1.0 modular preview (arviz-base/stats/plots)

import arviz as az

# Bundled example: Eight Schools posterior with sample_stats + log_likelihood
idata = az.load_arviz_data("centered_eight")

az.summary(idata, var_names=["mu", "tau"])
# mean, sd, hdi_3%, hdi_97%, mcse, ess_bulk, ess_tail, r_hat per variable

az.plot_trace(idata, var_names=["mu", "tau"])   # convergence at a glance
az.plot_posterior(idata, var_names=["mu"])      # marginal + HDI
az.loo(idata)                                   # PSIS-LOO cross-validation

With PyMC there is no conversion step (pm.sample() returns InferenceData); with other samplers use az.from_cmdstanpy(...), az.from_numpyro(...), etc.

Architecture / How It Works

Everything revolves around InferenceData: a container of labeled xarray Dataset groups following a published schema[^5] — posterior, prior, posterior_predictive, log_likelihood, sample_stats, observed_data, and others. Dimensions are conventionally (chain, draw, *shape). Because groups are xarray Datasets, coordinate-aware indexing, broadcasting, and out-of-core computation via dask come along for free, and the whole object serializes to NetCDF or Zarr. The schema is deliberately cross-language: ArviZ.jl reads and writes the same layout from Julia.

On top of the container sit three concerns, which the 1.0 split makes literal package boundaries:

Conversion (arviz-base): from_* adapters per PPL plus

convert_to_inference_data for raw dicts/arrays. This is where coupling to each sampler's internals lives, and it is the layer that breaks when a PPL changes its output format.

Stats (arviz-stats): rank-normalized split R-hat, bulk/tail effective

sample size, and MCSE per Vehtari et al. 2021[^2]; PSIS-LOO and WAIC for model comparison[^6]; HDI computation; compare() for ranking models.

Plots (arviz-plots): trace, forest, pair, rank, energy, PPC, ESS

evolution plots. Legacy ArviZ supports matplotlib and bokeh backends; the new arviz-plots adds plotly and rebuilds plots on a composable faceting layer instead of monolithic functions.

Production Notes

Memory is the main operational hazard. InferenceData holds all draws

in RAM; a posterior_predictive or log_likelihood group on a model with large observed data multiplies size fast (chains × draws × observations × 8 bytes). Thin, subset with xarray selection, or use dask-backed arrays.

az.loo needs a stored pointwise log_likelihood group (PyMC: opt in

via pm.sample(idata_kwargs={"log_likelihood": True})). If PSIS Pareto-k diagnostics exceed 0.7, the LOO estimate is unreliable and ArviZ warns — those warnings flag influential observations, do not suppress them[^6].

The 0.x → 1.0 migration is a real break, not a rename: plotting

functions move to arviz_plots with different signatures and a backend-neutral API. Pin arviz<1 if you cannot absorb churn yet, and budget migration time using the official guide[^4].

Version coupling with PyMC. Because PyMC depends on ArviZ, upgrading

one can force the other; resolver conflicts in pinned environments are a recurring annoyance.

Diagnostics defaults have evolved. R-hat and ESS follow the

rank-normalized definitions since roughly 2019/2020, so thresholds differ from older literature (aim for R-hat < 1.01, not < 1.1)[^2]; comparing against numbers from pre-2019 estimators is apples to oranges.

Maintenance is healthy: pushes within the last month (as of mid-2026),

~106 open issues, active multi-repo dev organization.

When to Use / When Not

Use when:

You run MCMC or variational inference in any Python PPL and need

convergence checking, posterior visualization, or model comparison.

You want one storage format (InferenceData/NetCDF) for posterior results

shared across teams, languages (Julia), or archival.

You are comparing models fit in different PPLs — ArviZ is the common ground.

Avoid when:

You work primarily in R — the native Stan-ecosystem tooling is stronger there.
You need quick corner plots of raw emcee-style chains and nothing else —

a full InferenceData round-trip is overkill.

Your workflow is point estimates from scikit-learn-style APIs; ArviZ

assumes posterior draws exist.

Alternatives

stan-dev/bayesplot — the R equivalent for Stan-centric plotting; use it

when your modeling happens in R/brms/rstanarm.

stan-dev/posterior — R package for draws containers and the same modern

R-hat/ESS diagnostics; use for R-side storage and summaries.

mjskay/tidybayes — use when you want tidyverse/ggplot2-grammar posterior

workflows instead of prebuilt plot functions.

dfm/corner.py — use for lightweight corner plots straight from array-shaped

chains (astronomy/emcee tradition) without the InferenceData machinery.

arviz-devs/ArviZ.jl — the Julia interface to the same InferenceData schema,

for Turing.jl/Stan.jl users.

History

| Version | Date | Notes | |---------|------|-------| | — | 2015-07 | Repository created; grew out of shared plotting/diagnostic code in the PyMC and Stan communities. | | 0.1–0.3 | 2018 | First PyPI releases; xarray-backed InferenceData design established. | | 0.x | 2019-01 | JOSS paper (Kumar, Carroll, Hartikainen, Martin)[^1]; adoption as PyMC's return type followed. | | 0.6+ | 2019–2020 | Bokeh added as a second plotting backend; rank-normalized R-hat/ESS become the defaults[^2]. | | 1.0 preview | 2025–2026 | Modular split into arviz-base / arviz-stats / arviz-plots with plotly support; second JOSS paper[^3]; arviz[preview] install path and migration guide[^4]. |

References

[^1]: Kumar et al., "ArviZ a unified library for exploratory analysis of Bayesian models in Python", JOSS 2019. https://doi.org/10.21105/joss.01143 [^2]: Vehtari, Gelman, Simpson, Carpenter, Bürkner, "Rank-Normalization, Folding, and Localization: An Improved R-hat for Assessing Convergence of MCMC", Bayesian Analysis 2021. https://doi.org/10.1214/20-BA1221 [^3]: Martin et al., "ArviZ: a modular and flexible library for exploratory analysis of Bayesian models", JOSS 2026. https://doi.org/10.21105/joss.09889 [^4]: ArviZ migration guide (0.x → 1.0). https://python.arviz.org/en/latest/user_guide/migration_guide.html [^5]: InferenceData schema specification. https://python.arviz.org/en/latest/schema/schema.html [^6]: Vehtari, Gelman, Gabry, "Practical Bayesian model evaluation using leave-one-out cross-validation and WAIC", Statistics and Computing 2017. https://doi.org/10.1007/s11222-016-9696-4

Wiki: arviz-devs/arviz