Package {spFFBS}

Title: Spatiotemporal Propagation for Multivariate Bayesian Dynamic Learning
Version: 1.0-1
Maintainer: Luca Presicce <l.presicce@campus.unimib.it>
Description: Implementation of the Forward Filtering Backward Sampling (FFBS) algorithm with Dynamic Bayesian Predictive Stacking (DYNBPS) integration for multivariate spatiotemporal models, as introduced in "Adaptive Markovian Spatiotemporal Transfer Learning in Multivariate Bayesian Modeling" (Presicce and Banerjee, 2026+) <doi:10.48550/arXiv.2602.08544>. This methodology enables efficient Bayesian multivariate spatiotemporal modeling, utilizing dynamic predictive stacking to improve inference across multivariate time series of spatial datasets. The core functions leverage 'C++' for high-performance computation, making the framework well-suited for large-scale spatiotemporal data analysis in parallel computing environments.
LinkingTo: Rcpp, RcppArmadillo
Imports: spBPS, Rcpp (≥ 1.1.1), abind
Suggests: doParallel, mniw, MBA, ggplot2, patchwork, reshape2, knitr, rmarkdown
License: GPL (≥ 3)
Encoding: UTF-8
RoxygenNote: 7.3.3
VignetteBuilder: knitr
URL: https://lucapresicce.github.io/spFFBS/
NeedsCompilation: yes
Packaged: 2026-07-15 17:16:35 UTC; presi
Author: Luca Presicce ORCID iD [aut, cre]
Repository: CRAN
Date/Publication: 2026-07-15 19:20:02 UTC

spFFBS: Spatiotemporal Propagation for Multivariate Bayesian Dynamic Learning

Description

logo

Implementation of the Forward Filtering Backward Sampling (FFBS) algorithm with Dynamic Bayesian Predictive Stacking (DYNBPS) integration for multivariate spatiotemporal models, as introduced in "Adaptive Markovian Spatiotemporal Transfer Learning in Multivariate Bayesian Modeling" (Presicce and Banerjee, 2026+) doi:10.48550/arXiv.2602.08544. This methodology enables efficient Bayesian multivariate spatiotemporal modeling, utilizing dynamic predictive stacking to improve inference across multivariate time series of spatial datasets. The core functions leverage 'C++' for high-performance computation, making the framework well-suited for large-scale spatiotemporal data analysis in parallel computing environments.

Author(s)

Maintainer: Luca Presicce l.presicce@campus.unimib.it (ORCID)

See Also

Useful links:

Compute dynamic Bayesian predictive stacking weights (v2)

Description

Calls compute_Wt_cpp_v2() directly - no foreach overhead.

Usage

compute_Wt(out_FF, tau, phi, n_threads = 1L, verbose = TRUE)

Arguments

out_FF

Output of spFF3_v2.

tau

Numeric vector of tau grid values (for column names only).

phi

Numeric vector of phi grid values (for column names only).

n_threads

Number of OpenMP threads (default 1).

verbose

Logical; print progress? Default TRUE.

Value

n x J weight matrix.

Convert a full prior covariance matrix to block form for spFFBS

Description

Convert a full prior covariance matrix to block form for spFFBS

Usage

make_prior(m0, C0, nu0, Psi0, p)

Arguments

m0

(p+n) x q prior mean matrix.

C0

(p+n) x (p+n) prior covariance matrix (full).

nu0

Prior degrees of freedom (scalar).

Psi0

q x q prior scale matrix.

p

Number of regression coefficients (integer).

Value

A list with elements m, C_bb, C_bw, C_ww, nu, Psi ready to pass as prior to spFFBS.

spFFBS: Spatiotemporal Bayesian Pipeline (optimised)

Description

Optimised wrapper for the spFFBS v2 C++ backend. Runs forward filtering, dynamic stacking weight computation, and optionally backward sampling, temporal forecasting, and spatial interpolation.

Usage

spFFBS(
  Y,
  X,
  D,
  grid = list(tau = NULL, phi = NULL),
  prior,
  G_beta = NULL,
  rho = 1,
  do_BS = FALSE,
  do_forecast = FALSE,
  do_spatial = FALSE,
  L = 200L,
  tnew = NULL,
  X_all = NULL,
  spatial = NULL,
  n_threads = 1L,
  verbose = TRUE
)

Arguments

Y

Response array, n x q x T.

X

Covariate array, n x p x T. Replaces P.

D

n x n spatial distance matrix.

grid

List with elements tau and phi (numeric vectors).

prior

Block-form prior list; see make_prior. Elements: m ((p+n)xq), C_bb (pxp), C_bw (pxn), C_ww (nxn), nu (scalar), Psi (qxq). If the old-style full C is supplied instead of C_bb/C_bw/C_ww, the wrapper auto-converts using make_prior().

G_beta

pxp upper-left block of the system matrix G. Default NULL = identity.

rho

Scalar multiplier for the lower-right nxn identity block of G. Default 1.

do_BS

Logical; run backward sampling? Default FALSE.

do_forecast

Logical; run temporal forecast? Default FALSE.

do_spatial

Logical; run spatial interpolation? Default FALSE.

L

Number of posterior samples. Default 200.

tnew

Forecast horizon (required when do_forecast = TRUE).

X_all

nxpx(T+tnew) covariate array for forecasting (required when do_forecast = TRUE).

spatial

List for spatial interpolation: list(crd, crdtilde, Xtilde, t). Xtilde should be a uxp matrix (covariates at prediction sites).

n_threads

Number of OpenMP threads. Default 1.

verbose

Logical; print progress? Default TRUE.

Value

A list with (depending on flags):

FF

Forward filter output from spFF3_v2.

Wi

nxJ dynamic weight matrix.

Wglobal

Jx1 global (averaged) weight vector.

BS

(if do_BS) T-list of nxqxL sample cubes.

forecast

(if do_forecast) List with Y_pred array of dimensions nxqxLxT.

spatial

(if do_spatial) Named list of L-lists of prediction matrices, one per requested time.

Differences from spFFBS (v1)

Examples


set.seed(42)
n <- 50; T <- 10; p <- 2; q <- 2

coords <- matrix(runif(n * 2), ncol = 2)
D      <- as.matrix(dist(coords))
Y      <- array(rnorm(n * q * T), dim = c(n, q, T))
X      <- array(rnorm(n * p * T), dim = c(n, p, T))

prior  <- make_prior(
  m0   = matrix(0, n + p, q),
  C0   = diag(n + p),
  nu0  = 3,
  Psi0 = diag(q),
  p    = p
)

res <- spFFBS(
  Y = Y, X = X, D = D,
  grid  = list(tau = c(0.8, 0.9), phi = c(1, 2)),
  prior = prior,
  L = 50
)