In config-i1/smooth: Forecasting Using State Space Models

knitr::opts_chunk$set(fig.width=6, fig.height=4, fig.path='Figs/', fig.show='hold',
                      warning=FALSE, message=FALSE)

SSARIMA stands for "State-space ARIMA" or "Several Seasonalities ARIMA". Both names show what happens in the heart of the function: it constructs ARIMA in a state-space form and allows to model several (actually more than several) seasonalities. ssarima() is a function included in smooth package. This vignette covers ssarima() and auto.ssarima() functions. For more details about the underlying model, read [@Svetunkov2019].

As usual, we will use data from Mcomp package, so it is advised to install it.

Let's load the necessary packages:

require(smooth)

The default call constructs ARIMA(0,1,1):

ssarima(AirPassengers, h=12, silent=FALSE)

Some more complicated model can be defined using parameter orders the following way:

ssarima(AirPassengers, orders=list(ar=c(0,1),i=c(1,0),ma=c(1,1)), lags=c(1,12), h=12)

This would construct seasonal ARIMA(0,1,1)(1,0,1)$_{12}$.

We could try selecting orders manually, but this can also be done automatically via auto.ssarima() function:

auto.ssarima(AirPassengers, h=12)

Automatic order selection in SSARIMA with optimised initials does not work well and in general is not recommended. This is partially because of the possible high number of parameters in some models and partially because of potential overfitting of first observations when non-zero order of AR is selected:

auto.ssarima(AirPassengers, h=12, initial="backcasting")
auto.ssarima(AirPassengers, h=12, initial="optimal")

As can be seen from the example above the model with optimal initials takes more time and we end up with a different model than in the case of backcasting.

A power of ssarima() function is that it can estimate SARIMA models with multiple seasonalities. For example, SARIMA(0,1,1)(0,0,1)_6(1,0,1)_12 model can be estimated the following way:

ssarima(AirPassengers, orders=list(ar=c(0,0,1),i=c(1,0,0),ma=c(1,1,1)), lags=c(1,6,12), h=12, silent=FALSE)

It probably does not make much sense for this type of data, it would make more sense on high frequency data (for example, taylor series from forecast package). However, keep in mind that multiple seasonal ARIMAs are very slow in estimation and are very capricious. So it is really hard to obtain an appropriate and efficient multiple seasonal ARIMA model. To tackle this issue, I've developed an alternative ARIMA model for multiple seasonalities, called msarima().

Now let's introduce some artificial exogenous variables:

x <- cbind(rnorm(length(AirPassengers),50,3),rnorm(length(AirPassengers),100,7))

If we save model:

ourModel <- auto.ssarima(AirPassengers, h=12, holdout=TRUE, xreg=x)

we can then reuse it:

ssarima(AirPassengers, model=ourModel, h=12, holdout=FALSE, xreg=x, interval=TRUE)

Finally, we can combine several SARIMA models:

ssarima(AirPassengers, h=12, holdout=FALSE, interval=TRUE, combine=TRUE)

MSARIMA

While SSARIMA is flexible, it is not fast. In fact, it cannot handle high frequency data well and most probably will take ages to estimate the parameter and produce forecasts. This is because of the transition matrix, which becomes huge in case of multiple seasonalities. The MSARIMA model (Multiple Seasonal ARIMA) is formulated in a different state-space form, which reduces the size of transition matrix, significantly reducing the computational time for cases with high frequency data.

There are auto.msarima() and msarima() function in the package, that do things similar to auto.ssarima() and ssarima(). Here's just one example of what can be done with it:

msarima(AirPassengers, orders=list(ar=c(0,0,1),i=c(1,0,0),ma=c(1,1,1)),lags=c(1,6,12),h=12, silent=FALSE)

The forecasts of the two models might differ due to the different state space form. The detailed explanation of MSARIMA is given in Chapter 9 of ADAM textbook.