Quick Introduction to the tfarima Package

In tfarima: Transfer Function and ARIMA Models

knitr::opts_chunk$set(
  collapse = TRUE,
  comment = "#>",
  fig.width = 5,
  fig.height = 3.5,
  fig.align = "center",
  dpi = 120
)
library(tfarima)
data("BuildingMat")
data("seriesJ")
options(prompt = "R> ", continue = "+  ", width = 70, useFancyQuotes = FALSE)

Introduction

The tfarima package provides a comprehensive framework for specifying, estimating, and analyzing Transfer Function and ARIMA models in a unified and transparent way.\ It introduces several S3 classes such as lagpol, um, tfm, and ucarima that allow users to work directly with lag polynomials, univariate models, transfer function models, and UCARIMA decompositions, respectively.

This vignette offers a quick overview of the main features of the package, demonstrating how to define models, simulate data, estimate parameters, and perform diagnostic and forecasting analyses.

---

First-Order Autoregressive (Markov) Process

We begin with a simple AR(1) model to illustrate basic functionality.

# Creating an AR(1) process with mean 0.5 and innovation variance 1.5
ar1p <- um(ar = "1 - 0.8B", mu = 0.5, sig2 = 1.5)

# Displaying model properties
phi(ar1p)
roots(ar1p)
psi.weights(ar1p, 10)
autocorr(ar1p, 10)
spec(ar1p)[1:5, ]
display(ar1p)

Now we can simulate a realization and fit the same model back to the data.

set.seed(1234)
z <- sim(ar1p, n = 100)
ide(z)
ar1p.fit <- fit(ar1p, z)
diagchk(ar1p.fit)
p <- predict(ar1p.fit, n.ahead = 5)
plot(p, n.back = 20)

---

Lag Polynomial Objects (lagpol Class)

The lagpol class provides a convenient symbolic representation of lag polynomials, including seasonal and multiplicative structures.

lp1 <- lagpol(param = c(theta = 0.5), p = 2)
lp2 <- lagpol(param = c(Theta1 = 1.2, Theta2 = -0.9), s = 12)
lp3 <- lagpol(param = c(theta = 0.8, Theta = 0.9),
              coef = c("theta", "Theta", "-theta*Theta"),
              lags = c(1, 12, 13))
lp3

It also supports compact symbolic specifications through lagpol0():

i  <- lagpol0(list(1, c(1, 3), c(1, 12)), "i")
ma <- lagpol0("(1 - 0.8B)(1 - 0.7B3)(1 - 0.6B12)", "ma")
printLagpolList(i)
printLagpolList(ma)

---

Univariate Models (um Class)

The um class represents univariate ARIMA-type models, combining lag polynomials and parameters in a structured way.

ar1p <- um(ar = "(1 - 0.9B)")
ar1n <- um(ar = "(1 + 0.9B)")
ma1p <- um(ma = "(1 - 0.9B)")
ma1n <- um(ma = "(1 + 0.9B)")
ar2c <- um(ar = "(1 - 1.52B + 0.8B^2)")

display(list(ar1p, ar1n, ma1p, ma1n, ar2c),
        lag.max = 20, graphs = c("acf", "pacf", "spec"))

---

Example: Building Material Series

Here we identify and estimate a seasonal ARIMA model for the BuildingMat dataset.

Z <- BuildingMat
ide(Z, transf = list(
  list(bc = TRUE),
  list(bc = TRUE, S = 1),
  list(bc = TRUE, d = 1, D = 1), 
  list(bc = TRUE, d = 1, D = 1, i = lagpol(s = 3, coef = "1"))
))

um1 <- um(Z, bc = TRUE,
          i  = list(1, c(1, 3), c(1, 12)),
          ma = list(theta = 1, Theta = c(1, 3), THETA = c(1, 12)))
diagchk(um1)

p <- predict(um1, n.ahead = 12, i = c(1, 12))
plot(p, n.back = 72)

A UCARIMA decomposition can be obtained for the fitted model:

uca1 <- as.ucarima(um1, i = 3)
uc1 <- wkfilter(uca1)
plot(uc1, main = "")

---

Transfer Function Models (tfm Class)

Transfer function models allow us to model dynamic relationships between input and output time series.

Y <- seriesJ$Y - mean(seriesJ$Y)
X <- seriesJ$X - mean(seriesJ$X)

umx <- um(X, ar = 3)
umy <- fit(umx, Y)
pccf(X, Y, um.x = umx, lag.max = 16)

tfx  <- tfest(Y, X, delay = 3, p = 2, q = 2, um.x = umx, um.y = umy)
tfmy <- tfm(Y, inputs = tfx, noise = um(ar = 2))
tfmy

Another example with BJsales data:

Y <- window(BJsales, end = 140)
X <- window(BJsales.lead, end = 140)
umy <- um(Y, i = 1, ma = "1-0.620B", mu = 0.033, fit = FALSE)
umx <- um(X, i = 1, ma = "1 - 0.4483B", fit = FALSE)
tfx <- tf(X, delay = 3, w0 = 4.716, ar = "1-0.725B", um = umx)
tfmy <- tfm(output = Y, inputs = tfx, noise = umy, fit = FALSE)
plot(predict(tfmy, n.ahead = 10, ori = 89), n.back = 15, ylab = expression(Y[t]))

---

Outliers and Calendar Effects

The outliers() function can automatically detect additive outliers, level shifts, and calendar-related effects.

tfm1 <- outliers(um1, calendar = TRUE, easter = TRUE,
                 form = "wd", c = 4, p.value = 0.05, n.ahead = 24)
tfm1

---

UCARIMA Decompositions

Finally, the ucarima class provides a unified structure for unobserved components models.

trend <- um(i = 2, ma = c(theta = 1), sig2 = c(s2t = 0.1))
seas  <- um(i = "12", 
             ma = lagpol(coef = c("0.7944","0.6185","0.4699",
                                  "0.3465","0.2458","0.1658",
                                  "0.1041","0.0587","0.0275","0.0086")),
             sig2 = c(s2s = 0.01))
irreg <- um(sig2 = c(s2i = 1))

uca1 <- ucarima(Z, bc = TRUE,
                ucm = list(trend = trend, seas = seas, irreg = irreg))
rf1  <- as.um(uca1)
uca3 <- as.ucarima(rf1, i = 2, canonical = FALSE)
uca3

---

Summary

This vignette has illustrated how tfarima supports the entire workflow for ARIMA and transfer function modeling:\ from model specification (lagpol, um) and simulation, to estimation (fit), diagnostics (diagchk), forecasting (predict), and structural decomposition (ucarima).

For more formal details, theoretical background, and advanced examples, see the accompanying article and package documentation.

tfarima documentation built on Nov. 5, 2025, 7:43 p.m.