fdlm_factor: Sample the dynamic factors

In drkowal/FDLM: A Bayesian Multivariate Functional Dynamic Linear Model

Description

Sample the dynamic factors (latent state space variables) using the simulation smoothing algorithm of Koopman and Durbin (2001), implemented in the KFAS package.

Usage

fdlm_factor(Y, sigma_et, Wt, Fmat = NULL, YF = NULL, Gt = NULL,
  W0 = NULL, kfas_model = NULL, useFastImpute = FALSE)

Arguments

Y	the T x m data observation matrix, where T is the number of time points and m is the number of observation points (NAs allowed)
sigma_et	T- or 1-dimensional vector of observation error standard deviation(s)
Wt	K x K matrix or K x K x T array of evolution error covariances
Fmat	m x K matrix of FLCs; only needed for useFastImpute = FALSE
YF	T x K matrix of data Y projected onto FLCs, Y%*%Fmat, which takes the place of Y; only needed for useFastImpute = TRUE
Gt	K x K evolution matrix; if NULL, set as identity (for random walk)
W0	K x K matrix of initial evolution error covariances; if NULL, set to diag(10^-4, K)
kfas_model	SSModel object from KFAS package; if NULL, construct model w/in the sampler (might be slower!)
useFastImpute	logical; when TRUE, use imputation/projection scheme for the dynamic factors; otherwise use full state space model for factors (slower)

Value

The T x K matrix of dynamic factors, Beta.

Note

The sampler has two options: useFastImpute = TRUE, in which the response is YF = Y%*%Fmat (T x K) and the observation matrix is the identity (K x K); and useFastImpute = FALSE, in which the response is Y (T x m) and the observation matrix is Fmat (m x K). Recall that typically K < < m, so useFastImpute = TRUE is often much faster.

Examples

# Read in the yield curve data:
data("US_Yields")

# Restrict to dates since 2006:
Y = Y[which(dates > as.Date("2006-01-01")),];

# This is a simple (yet meaningless) example:
K = 3
Beta = fdlm_factor(Y, sigma_et = 1, Wt = diag(K), Fmat = diag(ncol(Y))[,1:K])

drkowal/FDLM documentation built on May 20, 2019, 5:20 p.m.