DynCount 0.2.0
New features
- New observation family
family = "multinomial" for
choice counts: y is an n x K matrix, the row
totals are the known trials, and each non-baseline category has its own
latent additive-log-ratio (ALR) series
log(p_k / p_baseline) following the chosen latent dynamics
and innovation structure, with no parameters shared across categories.
The baseline is selected with the new baseline argument
(default "largest"). Per-category ALR offsets, forecasting
(forecast_trials = future totals, by default the last
non-zero row total), summary(), predict()
(including type = "prob"), forecast() and the
plot functions (one panel per category, category argument)
all support the new family; posterior draws carry a trailing category
dimension. With K = 2 the model coincides with the binomial
family.
- New simulator
simulate_dynamic_multinomial(). Its
baseline can be given as a column index or a category name,
and its offset as a scalar, a per-category vector or a
matrix.
- Forecasting is now post hoc:
forecast(fit, horizon = H)
forward-simulates the latent path from every stored posterior draw (with
increments from the fitted innovation structure) and draws responses
from the observation model. This targets the same posterior predictive
distribution as the previous in-sampler forecasts, but the horizon, the
forecast offset and the forecast trials can now be chosen after fitting.
forecast() and plot_forecast() gain
horizon, forecast_offset,
forecast_trials and seed arguments. Fitting
with horizon = H still stores an H-step
forecast in the fit.
- The draws of the stochastic-volatility parameters
(
sv_mu, sv_phi, sv_sigma) are now
stored and reported by summary(). The draws of the mixture
weights and component variances (mix_weight,
mix_var) and of the overall innovation scale
(scale) are stored as well.
- The example data sets gain a
date column (the Monday of
each ISO week).
Changes to defaults
- The default prior on the innovation variance is now the vague
InvGamma(0.01, 0.01) instead of
InvGamma(2.5, 0.5), which put almost no mass below
sigma = 0.1 and could force rough latent paths on smooth
series. Pass dynamic_prior(var_shape = 2.5, var_rate = 0.5)
to recover the old behaviour. See ?dynamic_prior for the
trade-offs.
- The relative variances of the
"mixture" components now
have their own hyperparameters mix_var_shape /
mix_var_rate (defaults 2.5 / 0.5,
i.e. unchanged) rather than reusing var_shape /
var_rate.
Breaking changes
draws$z and draws$sig2 now have one column
per observation. The initial latent state is stored separately as
draws$z0, and forecast states are only in
draws$forecast_z.draws$gate and draws$pi_open are
NULL for fits without zero inflation, and
draws$nu is NULL unless
innovations = "t".forecast() is now the generic from the generics package
(re-exported), so forecast(fit) also works when the
forecast or fable packages are attached.zero_inflation in fit_dynamic_model() must
be a single TRUE/FALSE; other values (e.g. a
probability, as in the simulators) are an error, as is combining
zero_inflation = TRUE with a different explicit
zeros.
Sampler
- The sampler works on a matrix of latent series; the Poisson and
binomial families are the single-column case.
- The latent states are updated in a checkerboard order. As the GMRF
precision is tridiagonal, the states with odd and with even time index
are conditionally independent given each other, so each half is updated
in one vectorised step with the same single-site adaptive Metropolis
moves. This makes fitting roughly 7 to 17 times faster.
- The Student-t degrees of freedom are updated with the mixing weights
integrated out (a partially collapsed Gibbs step), which improves their
effective sample size several-fold.
Minor improvements and fixes
- A
seed argument no longer changes the global random
number stream. The previous RNG state is restored after fitting,
forecasting and simulating.
- Arguments passed through
... to the plot functions
(e.g. main, xlab, ylim) now
override the defaults instead of causing an error.
- Warnings for arguments that are ignored (
trials or
forecast_trials for the Poisson family,
baseline for non-multinomial families, forecast inputs with
horizon = 0) and for forecasting a model with an offset
without a forecast_offset.
simulate_dynamic_binomial() validates
trials.
DynCount 0.1.0