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 (Svetunkov and Boylan 2019).
As usual, we will use data from Mcomp
package, so it is
advised to install it.
Let’s load the necessary packages:
The default call constructs ARIMA(0,1,1):
## Time elapsed: 0.01 seconds
## Model estimated: ARIMA(0,1,1)
## Matrix of MA terms:
## Lag 1
## MA(1) 0.4066
## Initial values were produced using backcasting.
##
## Loss function type: likelihood; Loss function value: 700.9086
## Error standard deviation: 31.6734
## Sample size: 144
## Number of estimated parameters: 2
## Number of degrees of freedom: 142
## Information criteria:
## AIC AICc BIC BICc
## 1405.817 1405.902 1411.757 1411.968
Some more complicated model can be defined using parameter
orders
the following way:
## Time elapsed: 0.03 seconds
## Model estimated: SARIMA(0,1,1)[1](1,0,1)[12]
## Matrix of AR terms:
## Lag 12
## AR(1) 1
## Matrix of MA terms:
## Lag 1 Lag 12
## MA(1) -0.31 0.1851
## Initial values were produced using backcasting.
##
## Loss function type: likelihood; Loss function value: 558.2012
## Error standard deviation: 11.8407
## Sample size: 144
## Number of estimated parameters: 4
## Number of degrees of freedom: 140
## Information criteria:
## AIC AICc BIC BICc
## 1124.402 1124.690 1136.282 1136.997
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:
## Time elapsed: 1.76 seconds
## Model estimated: SARIMA(0,1,3)[1](0,1,0)[12]
## Matrix of MA terms:
## Lag 1
## MA(1) -0.3587
## MA(2) 0.1056
## MA(3) -0.2230
## Initial values were produced using backcasting.
##
## Loss function type: likelihood; Loss function value: 549.9204
## Error standard deviation: 11.179
## Sample size: 144
## Number of estimated parameters: 4
## Number of degrees of freedom: 140
## Information criteria:
## AIC AICc BIC BICc
## 1107.841 1108.129 1119.720 1120.435
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:
## Time elapsed: 1.68 seconds
## Model estimated: SARIMA(0,1,3)[1](0,1,0)[12]
## Matrix of MA terms:
## Lag 1
## MA(1) -0.3587
## MA(2) 0.1056
## MA(3) -0.2230
## Initial values were produced using backcasting.
##
## Loss function type: likelihood; Loss function value: 549.9204
## Error standard deviation: 11.179
## Sample size: 144
## Number of estimated parameters: 4
## Number of degrees of freedom: 140
## Information criteria:
## AIC AICc BIC BICc
## 1107.841 1108.129 1119.720 1120.435
## Time elapsed: 15.13 seconds
## Model estimated: SARIMA(0,1,3)[1](0,1,0)[12] with drift
## Matrix of MA terms:
## Lag 1
## MA(1) -0.3645
## MA(2) 0.1120
## MA(3) -0.2297
## Constant value is: 0.2186
## Initial values were optimised.
##
## Loss function type: likelihood; Loss function value: 549.269
## Error standard deviation: 11.7305
## Sample size: 144
## Number of estimated parameters: 18
## Number of degrees of freedom: 126
## Information criteria:
## AIC AICc BIC BICc
## 1134.538 1140.010 1187.995 1201.592
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:
If we save model:
we can then reuse it:
## Time elapsed: 0.07 seconds
## Model estimated: SARIMAX(0,1,3)[1](0,1,0)[12] with drift
## Matrix of MA terms:
## Lag 1
## MA(1) -0.2593
## MA(2) 0.1397
## MA(3) -0.2587
## Constant value is: 0.3841
## Initial values were provided by user.
## Xreg coefficients were estimated in a normal style
##
## Loss function type: likelihood; Loss function value: 551.0437
## Error standard deviation: 11.1478
## Sample size: 144
## Number of estimated parameters: 1
## Number of provided parameters: 19
## Number of degrees of freedom: 143
## Information criteria:
## AIC AICc BIC BICc
## 1104.087 1104.116 1107.057 1107.127
##
## 95% parametric prediction interval was constructed
Finally, we can combine several SARIMA models:
## Time elapsed: 0.01 seconds
## Model estimated: ARIMA(0,1,1)
## Matrix of MA terms:
## Lag 1
## MA(1) 0.4066
## Initial values were produced using backcasting.
##
## Loss function type: likelihood; Loss function value: 700.9086
## Error standard deviation: 31.6734
## Sample size: 144
## Number of estimated parameters: 2
## Number of degrees of freedom: 142
## Information criteria:
## AIC AICc BIC BICc
## 1405.817 1405.902 1411.757 1411.968
##
## 95% parametric prediction interval was constructed
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)
## Time elapsed: 0.26 seconds
## Model estimated using msarima() function: SARIMA(0,1,1)[1](0,0,1)[6](1,0,1)[12]
## With optimal initialisation
## Distribution assumed in the model: Normal
## Loss function type: likelihood; Loss function value: 583.1003
## ARMA parameters of the model:
## Lag 12
## AR(1) 0.9997
## Lag 1 Lag 6 Lag 12
## MA(1) -0.0744 -0.1244 -0.0084
##
## Sample size: 144
## Number of estimated parameters: 24
## Number of degrees of freedom: 120
## Information criteria:
## AIC AICc BIC BICc
## 1214.201 1224.285 1285.476 1310.534
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.