Fitting Bayesian VHAR of SSVS Prior — bvhar

This function fits BVAR(p) with stochastic search variable selection (SSVS) prior.

Usage

bvhar_ssvs(
  y,
  har = c(5, 22),
  num_chains = 1,
  num_iter = 1000,
  num_burn = floor(num_iter/2),
  thinning = 1,
  bayes_spec = choose_ssvs(y = y, ord = har, type = "VHAR", param = c(0.1, 10),
    include_mean = include_mean, gamma_param = c(0.01, 0.01), mean_non = 0, sd_non = 0.1),
  init_spec = init_ssvs(type = "auto"),
  include_mean = TRUE,
  minnesota = c("no", "short", "longrun"),
  verbose = FALSE,
  num_thread = 1
)

# S3 method for class 'bvharssvs'
print(x, digits = max(3L, getOption("digits") - 3L), ...)

# S3 method for class 'bvharssvs'
knit_print(x, ...)

Arguments

y: Time series data of which columns indicate the variables
har: Numeric vector for weekly and monthly order. By default, c(5, 22).
num_chains: Number of MCMC chains
num_iter: MCMC iteration number
num_burn: Number of warm-up (burn-in). Half of the iteration is the default choice.
thinning: Thinning every thinning-th iteration
bayes_spec: A SSVS model specification by set_ssvs(). By default, use a default semiautomatic approach choose_ssvs().
init_spec: SSVS initialization specification by init_ssvs(). By default, use OLS for coefficient and cholesky factor while 1 for dummies.
include_mean: Add constant term (Default: TRUE) or not (FALSE)
minnesota: Apply cross-variable shrinkage structure (Minnesota-way). Two type: short type and longrun type. By default, no.
verbose: Print the progress bar in the console. By default, FALSE.
num_thread: Number of threads
x: bvharssvs object
digits: digit option to print
...: not used

Value

bvhar_ssvs returns an object named bvharssvs class. It is a list with the following components:

coefficients: Posterior mean of VAR coefficients.
chol_posterior: Posterior mean of cholesky factor matrix
covmat: Posterior mean of covariance matrix
omega_posterior: Posterior mean of omega
pip: Posterior inclusion probability
param: posterior::draws_df with every variable: alpha, eta, psi, omega, and gamma
param_names: Name of every parameter.
df: Numer of Coefficients: 3m + 1 or 3m
p: 3 (The number of terms. It contains this element for usage in other functions.)
week: Order for weekly term
month: Order for monthly term
m: Dimension of the data
obs: Sample size used when training = totobs - p
totobs: Total number of the observation
call: Matched call
process: Description of the model, e.g. VHAR_SSVS
type: include constant term (const) or not (none)
spec: SSVS specification defined by set_ssvs()
init: Initial specification defined by init_ssvs()
chain: The numer of chains
iter: Total iterations
burn: Burn-in
thin: Thinning
group: Indicators for group.
num_group: Number of groups.
HARtrans: VHAR linear transformation matrix
y0: $Y_0$
design: $X_0$
y: Raw input

Details

SSVS prior gives prior to parameters $\alpha = vec(A)$ (VAR coefficient) and $\Sigma_e^{-1} = \Psi \Psi^T$ (residual covariance).

$$\alpha_j \mid \gamma_j \sim (1 - \gamma_j) N(0, \kappa_{0j}^2) + \gamma_j N(0, \kappa_{1j}^2)$$ $$\gamma_j \sim Bernoulli(q_j)$$

and for upper triangular matrix $\Psi$,

$$\psi_{jj}^2 \sim Gamma(shape = a_j, rate = b_j)$$ $$\psi_{ij} \mid w_{ij} \sim (1 - w_{ij}) N(0, \kappa_{0,ij}^2) + w_{ij} N(0, \kappa_{1,ij}^2)$$ $$w_{ij} \sim Bernoulli(q_{ij})$$

Gibbs sampler is used for the estimation.

References

Kim, Y. G., and Baek, C. (n.d.). Working paper.