\( \newcommand{\mathbbm}[1]{\boldsymbol{\mathbf{#1}}} \)

19.2 ARIMA

The smooth package also includes two distinct ARIMA functions, which are implemented in the SSOE framework, and have slightly different purposes.

19.2.1 Multiple Seasonal ARIMA, msarima()

The msarima() function implements the multiple seasonal ARIMA model as discussed in Chapter 9. In fact, starting from smooth v3.2.0, it is just a wrapper of adam(), similarly to how es() is its wrapper for the ETS counterpart. Being a wrapper, the function does not support all the features of ADAM, but still has the following functionality (similar to the es()):

  1. Explanatory variables in the ARIMAX model (Chapter 10);
  2. Different loss functions, including the custom one (Chapter 11);
  3. Different initialisation methods (Section 11.4);
  4. Diagnostic plots (Chapter 14);
  5. Orders and explanatory variables selection (Chapter 15);
  6. Covariance matrix of parameters and related methods for reapplication of the ARIMA model (Chapter 16);
  7. Data generation from an estimated model (Section 16.1);
  8. All types of prediction intervals and other related aspects (Chapter 18).

Being a part of ADAM, the function is tailored for multiple seasonality and works efficiently in this case, especially when compared with other ARIMA implementations in R. It only supports the Normal distribution and focuses on the location model for continuous data (non-intermittent). Furthermore, similar to the es() function, it is agnostic of the provided data and will work with any type of univariate object (vector, ts, zoo, etc). Finally, the explanatory variables can be provided via the xreg parameter, similar to how it is done in the arima() function from the stats package.

The selection of the ARIMA orders is implemented via the auto.msarima() function, which has a slightly different algorithm than the one explained in Section 15.2. It does selection based on an information criteria (IC), checking sequentially the orders of I, then MA and the AR for all the orders up until the ones provided by the user. So, for example, the following code:

auto.msarima(y, orders=list(ar=c(3,3),

will test different orders of SARIMA up to SARIMA(3,2,3)(3,1,3)\(_{12}\). If the parameter fast is TRUE then the function will skip orders that do not lead to the reduction of IC (e.g. if MA(3) does not improve the IC, it will not test MA(2) and MA(1)). As a final step, the function will compare the model with constant with the model without it and select the most appropriate one. Finally, the auto.msarima() also supports combination of forecasts from ARIMA models from the same pool as in the case of order selection described above.

19.2.2 State space ARIMA, ssarima()

The ssarima() function implements ARIMA in a different form, the one explained in Chapter 11 of Hyndman et al. (2008) and then discussed in Svetunkov and Boylan (2020). It can be considered a conventional SSOE state space ARIMA. The main difference with MSARIMA is in its architecture, i.e. the measurement vector \(\mathbf{w}\) and the transition matrix \(\mathbf{F}\) differ between the two implementations. The measurement vector is always of the style: \[\begin{equation} \mathbf{w} = \begin{pmatrix} 1 \\ 0 \\ \vdots \\ 0 \end{pmatrix} , \tag{19.4} \end{equation}\] while the size of the transition matrix is \(K^2\), where \(K\) is the maximum order of the ARI or MA polynomial. This contrasts with the ADAM ARIMA (and consequently msarima()), where the size of the matrix equals to the number of elements in the polynomial equation. So, for example, SARIMA(1,1,2)(0,1,0)\(_12\) can be expanded into the following equation (based on a similar model discussed in Section 9.1): \[\begin{equation*} {y}_{t} = (1+\phi_1) {y}_{t-1} -\phi_1 {y}_{t-2} + {y}_{t-12} -(1+\phi_1) {y}_{t-13} + \phi_1 {y}_{t-14} + \theta_1 \epsilon_{t-1} + \theta_2 \epsilon_{t-2} + \epsilon_t . \end{equation*}\] Based on this, in the case of SSARIMA, the size of the transition matrix will be \(14 \times 14\), while in the case of MSARIMA, the matrix will be \(5 \times 5\). This difference in complexity pays of for MSARIMA when a multiple frequency model is needed, because the size of the transition matrix will impact the computational speed. Summarising all of this, ssarima() is not suitable for multiple frequencies, because it relies on the conventional state space architecture. Still, it has an advantage: it is in general easier to initialise than the ADAM ARIMA, because the vector of initials equals to \(K\), while in case of the latter it is up to \(\frac{K(K+1)}{2}\).

Finally, the auto.ssarima() function implements the order selection for state space ARIMA based on the algorithm explained in Svetunkov and Boylan (2020).


• Hyndman, R.J., Koehler, A.B., Ord, J.K., Snyder, R.D., 2008. Forecasting with Exponential Smoothing: The State Space Approach. Springer Berlin Heidelberg. https://doi.org/10.1007/978-3-540-71918-2
• Svetunkov, I., Boylan, J.E., 2020. State-space ARIMA for Supply-chain Forecasting. International Journal of Production Research. 58, 818–827. https://doi.org/10.1080/00207543.2019.1600764