GitHub

Current Scope And Limitations

Epsilon is a public beta library for local, reproducible Bayesian marketing mix modelling in Julia. The current release is useful for small and medium local MMM workflows, especially when the modelling process needs to be explicit, scriptable, and inspectable.

The package is not trying to be a complete commercial MMM platform. Some areas are intentionally narrow while the core API settles.

What Is Supported Today

The strongest supported path is a time-series MMM run from a local configuration bundle:

config.yml
dataset.csv
holidays.csv

That workflow supports:

  • Turing/NUTS MCMC fitting for time-series models;
  • adstock and saturation transforms;
  • holidays, events, controls, Fourier seasonality, and trend terms where configured;
  • structured result folders with manifests, diagnostics, fit summaries, decomposition outputs, response curves, plots when available, and skipped stage markers;
  • blocked holdout validation for time-series models;
  • bounded lift-test and cost-per-target calibration for the maintained time-series MCMC path;
  • fixed-budget channel optimisation over fitted response surfaces.

Panel models are supported on a bounded surface. They are useful for declared panel dimensions and deterministic coordinate metadata, but they should not yet be treated as feature-equivalent to the time-series path.

What Is Intentionally Out Of Scope

The following are not part of the maintained public surface:

  • variational inference;
  • hosted dashboards, managed services, AI advisors, or background UI workflows;
  • panel holdout validation;
  • panel calibration;
  • free channel-by-panel optimisation;
  • fully automated prior-sensitivity refitting;
  • arbitrary future spend-path simulation;
  • portable binary interchange for Julia .jls artifacts.

Unsupported paths should fail clearly or write explicit skipped-stage markers when they are optional pipeline stages. Epsilon should not silently substitute a different statistical model.

Runtime Expectations

Epsilon uses Turing/NUTS MCMC for the maintained fitting path. This gives a transparent Bayesian workflow, but it is not the fastest possible execution strategy for every model shape. Larger runs can take material local compute time, especially when validation performs an additional refit.

Use runme.jl ... --quick or small sampler overrides when checking that a configuration bundle is wired correctly. Increase draws, tuning, and chains only when the goal is statistical inference rather than a workflow smoke check.

Runtime reduction remains an engineering priority, but speed work should not come at the cost of unclear model semantics.

Local Verification

Epsilon uses local verification scripts rather than hosted CI as the canonical release gate. For routine changes, prefer focused checks:

make format-check
make smoke
make test-file FILE=test/inference/recovery.jl

Before a release-facing tag or publication, run:

make check-release

See Supported Local Workflows for the runnable demo workflow and Support Boundaries for the detailed support matrix.