bmgarch() now accepts backend = "cmdstanr" in addition to the default
"rstan". Both MCMC (NUTS) and variational Bayes (sampling_algorithm = "VB")
are supported with the new backend.
summary(), print(), plot(), forecast(), fitted(), loo(), and
model_weights() all work with cmdstanr-fitted objects.
Stan function a_b_scale_jacobian renamed to a_b_scale_log_jac to avoid a
future name collision with Stan 2.39's reserved _jacobian suffix (affects
inst/stan/functions/jacobian.stan).
The iterations argument now defaults to 2000 for MCMC and 30000 for VB
when left unspecified. Previously the single default of 2000 was too low for
ADVI gradient ascent.
The backend field is stored in every bmgarch object so that loo()
refits and other operations that call back into bmgarch() automatically use
the same backend.
Threading defaults: cmdstanr models compiled with stan_threads = TRUE now
default to threads_per_chain = 1 (MCMC) and threads = 1 (VB). Users can
override via ....
Fixed ensemble covariance underestimation in .weighted_samples(). The
between-model variance term required by the law of total variance,
Σ_m w_m (μ̂_m − μ̂_mix)(μ̂_m − μ̂_mix)ᵀ, is now added to the
weighted combination of H and H_forecasted. This corrects the
ensemble posterior predictive covariance for both backcasts and forecasts
whenever models with differing mean structures are combined. The fix is a
no-op for single-model objects.
model_weights() and loo() now work with cmdstanr objects: log_lik
extraction and chain/iteration metadata no longer rely on rstan S4 slot
access (@sim).
Weighted forecasts (forecast(..., weights = mw)) now work with cmdstanr
backends via a new .extract_param_list() helper that replicates
rstan::extract() output for cmdstanr draw objects.
Forecast draws passed to rstan::gqs() are now correctly extracted via
posterior::as_draws_matrix() for cmdstanr fits, fixing a "no dimnames"
error.
tests/testthat/test-backends.R covering summary, forecast, and
model_weights for all three combinations: rstan/MCMC, cmdstanr/MCMC, and
cmdstanr/VB on a CCC model.rstan 2.26.0 array syntax.VAR argument for a VAR(1) model.VB as a sampling algorithm option. bmgarch now takes VB and the standard MCMC argument in sampling_algorithm. VB is inherited from rstan and is still experimental - use with caution.loo function now takes m-ahead argument for the "backward" approximation allowing one to tune the forcast to arbitrary steps ahead.
Fixed package dependency versions.
Initial CRAN release