varmapack simulates and analyzes Gaussian VAR, VMA,
VARMA, and VARMAX time-series models. Simulated stationary series have
the correct distribution from the first returned term, without
discarding a burn-in segment. The package also provides model testcases,
theoretical and sample autocovariances, covariance-to-correlation
conversion, innovation-covariance recovery, spectral radii, and impulse
response functions.
The Getting Started with Varmapack vignette shows model construction, simulation, testcases, and analysis. The Mathematical Description of Varmapack vignette defines the supported models and describes the simulation method.
After installing the CRAN release, open the vignettes with
vignette("getting-started", package = "varmapack") and
vignette("mathematical-description", package = "varmapack").
The installed reference documentation is available through
help(package = "varmapack"). For information about the
underlying C library, see the C
README.
The package will be installable from CRAN when released:
install.packages("varmapack")For the development version, install the companion
randompack package first:
remotes::install_github("jonasson2/randompack", subdir = "r-package")
remotes::install_github("jonasson2/varmapack", subdir = "r-package")model <- varmapack_testcase("smallARMA1")
gamma <- model$acvf(maxlag = 2)
corr <- varmapack_cov2corr(gamma)
Sig <- varmapack_cov2sig(gamma[, , 1], A = model$A, B = model$B)varmapack_cov2corr() converts theoretical or sample
autocovariances to correlations by dividing each entry by the product of
the corresponding lag-zero marginal standard deviations. The returned
array has the same shape as the input. Lag-zero diagonal entries are
exactly one. Other entries are not clipped to [-1,1], so
correlations obtained from lag-corrected sample autocovariances may lie
outside that interval.
varmapack_cov2sig(cov, A = NULL, B = NULL) recovers
innovation covariance from a specified lag-zero model covariance and
fixed AR and MA coefficients. cov is a single
r by r matrix; only its lower triangle is
used. The returned matrix is symmetric, and inputs are unchanged. The AR
model must be stationary, and both covariances must be positive
semidefinite; singular covariances are allowed. Incompatible targets and
unreliable inverse problems raise errors.
rng <- randompack::randompack_rng()
rng$seed(123)
t <- varmapack_gamma_times(100, shape = 16, mean_gap = 1, rng = rng)
model <- varmapack_ou_model(c(1, 2), matrix(c(1, .5, .5, 1), 2))
X <- model$sim(t, nrep = 10, rng = rng) # 2 by 100 by 10The OU model simulates exact stationary Gaussian paths at supplied
times, without burn-in. cov is the stationary covariance,
not innovation covariance; it and the driving covariance must be
positive semidefinite. For diagonal drift, tau gives the
persistence times. For a full drift matrix, use
varmapack_ou_model(cov = cov, drift = K), with drift
-K %*% (X - mu). The autoregressive coefficient matrix over
a gap d is the matrix exponential of -K*d.
Gamma gaps have CV 1/sqrt(shape), without normalization of
the realized mean; shape = Inf gives equal spacing.
varmapack_slepian_model(theta = matrix(.5), cov = matrix(1.25))
uses the same sim interface for a Slepian-based model,
which reduces to MA(1) at unit spacing. Both models have an
acvf method for fractional lags; the OU model also has a
specrad method for the autoregressive matrix at a given
time gap.
An additional European
stock-index example fits a VAR(1) to the built-in
EuStockMarkets log returns and generates exact stationary
simulations from the fitted model.