[thomaspinder]

Implements spec #393: the two root fixes behind pymc-labs/pymc-marketing#3293 (review items P1.1 and P1.2), for release 0.1.4.

Closes #393
Closes #394
Closes #395

Per-equation exog_prior_scale (#394)

VAR.exog_prior_scale takes a float or a tuple with one entry per endogenous equation, in endog_names order, mirroring MinnesotaPrior.own_lag_mean. A field validator requires every entry to be positive and finite, so the scalar and the tuple are validated the same way (a scalar inf, which gt=0 let through, is now rejected). The tuple's length is checked where the exogenous prior is built, against the number of endogenous variables. Row i of the prior sd is scale_i * sigma_i / s_j; the scalar path gives the same numbers as before. fit, prior_predictive and build_in_model all go through that construction. ADR-0012 records the amendment; the docstring drops the stale "applies only to fit".

Truncated own lag for one latent series with lags=1 (#395)

For exactly one latent endogenous series with lag order 1 the latent block is the own lag B[0, 0] and the stationary region is |B[0, 0]| < 1, so the constraint is now exact and smooth: latent_own_lag is the own lag's Minnesota Normal truncated to (−1, 1) under PyMC's interval transform, B_free is a flat Normal over every other lag coefficient, and B is a Deterministic assembled from the two under its existing name and (var, coeff) dims. No Potential and no initial value are registered in that case; every other case (no latent series, lags > 1, several latent series) is unchanged. The posterior is the same truncated prior the Potential imposed, up to its normalising constant.

Consequences: the own lag starts at 0 whatever its prior mean, so the host model is clonable for any own-lag mean, Impulso's default of 1 included; jitter stays inside the region; prior predictive draws of the own lag are truncated, so prior-predictive latent paths are stationary; no eig in the graph, so the JAX backend differentiates the model for this case (#378 stays open for the Potential cases). On the slow one-latent sampling setup under the default own-lag mean of 1, the Potential gave 268–300 divergent draws of 400 on three seeds; the truncated own lag gives none, with R-hat at most 1.011 on the own lag.

_register_coefficients owns B and the latent-block constraint in both forms, written as a free block plus a constrained block so that a general parametrisation of the latent block (#397) replaces the constrained block without a second rewrite. ADR-0017 records the decision; ADR-0016's initial-value, Potential and bounded-own-lag bullets point at it; the glossary entry for the stationarity constraint says how it is enforced in each case.

Commits

Tests first (strict xfail), then the implementation that unmarks them, per ticket: four commits.

Verification

• uv run python -m pytest -m "not slow" -q, ruff check, ruff format --check src tests, ty check: clean on the merged branch; all slow tests in tests/test_build_in_model.py pass.
• Known, pre-existing on main: ruff format --check . fails on 11 Markdown files (ruff 0.17.2 formats Python blocks in Markdown); none are touched here. The existing JAX test for two latent series with lags=2 fails on main too when JAX is installed (JAX samplers fail on models with latent series: the latent_stationarity gradient goes through eig #378); CI skips it.

🤖 Generated with Claude Code

[codecov]

Codecov Report

✅ All modified and coverable lines are covered by tests.

📢 Thoughts on this report? Let us know!