I prepared the following notes for a consulting client, and I thought they might be of interest to some other people too.
Let y_t denote the value of the time series at time t, and suppose we wish to fit a trend with correlated errors of the form y_t = f(t) + n_t, where f(t) represents the possibly nonlinear trend and n_t is an autocorrelated error process.
For example, if f(t) = \beta_0+\beta_1 t is a linear function, then we can simply set x_{1,t}=t and define y_t = \beta_0 + \beta_1x_{1,t} + n_t. In matrix form we can write \boldsymbol{y} = \beta_0 + \boldsymbol{X}\boldsymbol{\beta} + \boldsymbol{n}, where \boldsymbol{y}=[y_1,\dots,y_T]', \boldsymbol{n}=[n_1,\dots,n_T]', \boldsymbol{\beta}=[\beta_1] and \boldsymbol{X} = [x_{1,1},\dots,x_{1,T}]'. Note that I have left the intercept \beta_0 out of the vector \boldsymbol{\beta} so that the \boldsymbol{X} matrix matches the required xreg argument in auto.arima.
This model can be estimated by setting the xreg argument to be a matrix with one column:
\boldsymbol{X} = \left[\begin{array}{c}
1\\
2\\
3\\
4\\
\vdots\\
T
\end{array}\right]
x1 <- 1:length(y)
fit <- auto.arima(y, xreg=x1)The associated coefficient is the slope of the trend line.
Here is a simple example of a linear trend fitted to the Asian sheep data from the fpp package :
library(fpp)
T <- length(livestock)
x1 <- seq(T)
fit <- auto.arima(livestock, xreg=x1)
fc <- forecast(fit, xreg=T+seq(10))
b0 <- coef(fit)["intercept"]
b1 <- coef(fit)["x1"]
t <- seq(T+10)
trend <- ts(b0 + b1*t, start=start(livestock))
plot(fc, main="Linear trend with AR(1) errors")
lines(trend, col='red')
A more flexible approach is to use a piecewise linear trend which bends at some time. If the trend bends at time \tau, then it can be specified by including the following predictors in the model. \begin{align}
x_{1,t} &= t \\
x_{2,t} &= \begin{cases}
0 & t < \tau;\\
(t-\tau) & t \ge \tau.
\end{cases}
\end{align} In auto.arima, set xreg to be a matrix with two columns:
\boldsymbol{X} = \left[\begin{array}{ll}
1 & 0\\
2 & 0\\
3 & 0\\
4 & 0\\
\vdots\\
\tau & 0 \\
\tau+1 & 1\\
\tau+2 & 2\\
\vdots \\
T & T-\tau
\end{array}\right]
fit <- auto.arima(y, xreg=cbind(x1, pmax(0,x1-tau))If the associated coefficients of x_{1,t} and x_{2,t} are \beta_1 and \beta_2, then \beta_1 gives the slope of the trend before time \tau, while the slope of the line after time \tau is given by \beta_1+\beta_2.
This can be extended to allow any number of “bend points” known as knots. Just add additional columns with 0s before each knot, and values 1, 2, … after the knot.
Here is a piecewise linear trend fitted to the Asian sheep data with knots at years 1990 and 1992:
x2 <- pmax(0, x1-30)
x3 <- pmax(0, x1-32)
fit <- auto.arima(livestock, xreg=cbind(x1,x2,x3))
fc <- forecast(fit,
xreg=cbind(max(x1)+seq(10), max(x2)+seq(10), max(x3)+seq(10)))
b0 <- coef(fit)["intercept"]
b1 <- coef(fit)["x1"]
b2 <- coef(fit)["x2"]
b3 <- coef(fit)["x3"]
trend <- ts(b0 + b1*t + b2*pmax(0,t-30) + b3*pmax(0,t-32),
start=start(livestock))
plot(fc, main="Piecewise linear trend with AR(1) errors")
lines(trend, col='red')
If there is to be no trend before the first knot, but a piecewise linear trend thereafter, leave out the first column of the above matrix \boldsymbol{X}.
If there is to be a piecewise linear trend up to the last knot, but no trend thereafter, a slightly modified set up can be used. For one knot at time \tau, we can set \boldsymbol{X} = \left[\begin{array}{r} 1-\tau \\ 2-\tau \\ \vdots\\ -2\\ -1\\ 0 \\ 0 \\ \vdots \\ 0 \end{array}\right]
xreg <- pmin(0, x1-tau)where the first 0 in the column is in row \tau. Additional knots can be handled in the same way. For example, if there are two knots, then \beta_1+\beta_2 will be the slope of the trend up to the first knot, and \beta_2 will be the slope between the first and second knots.