Re-adopt pymc-marketing transform backends once it supports PyMC 6

[drbenvincent]

Summary

As part of #212 (PyMC 6 / ArviZ 1 upgrade + conda removal), the pymc-marketing dependency was dropped because no released version supports PyMC 6 / ArviZ 1 (pymc-marketing currently pins pymc<6 and arviz<1). The two transform kernels pathmc used — geometric adstock and logistic saturation — are now vendored as plain pytensor implementations in pathmc/transforms.py (_geometric_adstock, _logistic_saturation).

Once pymc-marketing ships PyMC 6 / ArviZ 1 support, decide whether to delegate the built-in transforms back to pymc_marketing.mmm.transformers or keep the vendored kernels.

Upstream tracking: pymc-labs/pymc-marketing#2486 (compatibility issue) and pymc-labs/pymc-marketing#2616 (migration PR). Expected upstream ETA: end of July 2026.

Proposed work

• Watch for a pymc-marketing release with pymc>=6 support.
• Evaluate whether delegating buys anything beyond the vendored ~30 lines (e.g. additional adstock/saturation variants worth exposing through the transform registry).
• If re-adopting: restore the dependency in pyproject.toml, reinstate the xtensor bridging helpers, and remove the vendored kernels. If not: close this issue and keep the vendored implementations as the permanent home.
• Either way, verify numerical equivalence on the transform gate tests (tests/test_transforms_*.py) and refresh any affected docs freeze caches.

Notes

This is not a release blocker for 0.1.0 — the vendored kernels are tested and behaviorally equivalent (truncated-at-l_max geometric adstock, same logistic saturation formula).

[drbenvincent]

Block check date: 2026-08-07

Determination: still blocked. The blocker has not been resolved under the issue stated condition: pymc-marketing has not yet shipped a released version with PyMC 6 / ArviZ 1 support.

What I checked:

• pathmc issue Re-adopt pymc-marketing transform backends once it supports PyMC 6 #214 is OPEN and still carries the exact blocked? label.
• Upstream compatibility issue PyMC 6 / ArviZ 1 / PyTensor 3 compatibility pymc-marketing#2486 is still OPEN.
• Upstream migration PR chore: migrate to pymc>=6.0.0, arviz>=1.1.0, pytensor>=3.0.0 pymc-marketing#2616 merged on 2026-06-26 into v1.0.0, and upstream main now declares pymc>=6.0.0 and arviz>=1.2.0.
• The latest published upstream release is https://github.com/pymc-labs/pymc-marketing/releases/tag/0.19.4, published 2026-05-06; no compatible v2 release is published.
• I checked Re-adopt pymc-marketing transform backends once it supports PyMC 6 #214 comments, timeline, and linked work, including pathmc PR Drop conda dev env, adopt uv, upgrade to PyMC 6 / ArviZ 1 #218; there is no released upstream compatibility result that would authorize re-adopting the backend.

No code, worktree, branch, or PR was created because the issue remains blocked.

[drbenvincent]

Unblocked

pymc-marketing now pins pymc>=6.0.0,<6.1.0 and arviz>=1.2.0,<2.0 (see ~/git/pymc-marketing/pyproject.toml). The blocker described in this issue is resolved.

Next step

Re-evaluate delegation vs keeping vendored kernels. Beyond numerical parity, the main reason to re-adopt is exposing additional MMM transform variants (Weibull/delayed adstock, Michaelis–Menten saturation, etc.) through pathmc's transform registry without vendoring each one.

Related context: synthetic-data / MMM stress-testing work identified this as a gap while scoping pathmc as a generative backend for marketing DAGs.

[drbenvincent]

Follow-up tracked in #436 — PR #435 landed the interim hybrid (delegation infrastructure + vendored fallbacks + new MMM transforms), but full removal of vendored kernels remains blocked on pymc-marketing relaxing its pytensor pin.