The Rfssa package provides the collections of necessary functions to implement functional singular spectrum analysis (FSSA)-based methods for analyzing univariate and multivariate functional time series (FTS). Univariate and multivariate FSSA are novel, non-parametric methods to perform decomposition and reconstruction of univariate and multivariate FTS respectively. In addition, the FSSA-based routines may be performed on FTS whose variables are observed over a one or two-dimensional domain. Finally, one may perform FSSA recurrent or vector forecasting of univariate or multivariate FTS observed over one-dimensional domains. Forecasting of FTS whose variables are observed over domains of dimension greater than one is under development.
The use of the package starts with the decomposition of functional time series (fts) objects using the fssa routine. Then a suitable grouping of the principal components is required for reconstruction (freconstruct) or forecasting (fforecast) which can be done heuristically by looking at the plots of the decomposition (plot). Once a suitable grouping is chosen, one may perform reconstruction where the sum of all the elements between the disjoint groups approximates the original FTS. One may also choose to perform forecasting after a grouping is chosen which returns future observations in each FTS specified by the groups.
This version of the package leverages a new S4 object for FTS objects (fts). The new object may be specified using a provided basis and grid, a requested basis and grid, or a mixture of provided and requested elements. We note that the FTS object may be univariate or multivariate and variables may be observed over one or two-dimensional domains. Validity checking of the S4 object constructor inputs was also added to help guide the user. The plotting of FTS objects (fts.plot) was also updated to allow the user to plot FTS variables observed over two-dimensional domains. Next, the FSSA routine (fssa) was updated to perform faster by leveraging the RSpectra and RcppEigen R packages, and the Eigen C++ package. We achieved a roughly 20 times speed up for certain data examples. We updated the plotting of fssa objects (fssa.plot) to allow for plotting of left singular functions that correspond with FTS variables observed over a two-dimensional domain. We updated FSSA reconstruction (freconstruct) to handle FTS whose variables are observed over one or two-dimensional domains. We also updated FTS arithmetic (such as FTS addition, FTS subtraction, etc.) to allow the user to perform scalar-FTS arithmetic on different variables of a multivariate FTS. In addition, we also now host the Callcenter, Jambi, and Montana datasets on GitHub to significantly decrease the size of the package. In order to load the data, one simply needs to use the load_github_data function. This same function can also be used to load data from any other public GitHub repository.
The first piece of new functionality that has been added is that the user may now specify univariate or multivariate FTS comprised of variables observed over one or two-dimensional domains. In addition, forecasting of univariate and multivariate FTS observed over one-dimensional domains by FSSA/MFSSA recurrent forecasting and FSSA/MFSSA vector forecasting has also been added. We have also added in a new data set (Montana) which provides the data for a multivariate FTS observed over different dimensional domains.
The reader should note that we do not utilize FTS plotting options in this README that are included in this update because of the large size of the resulting files. The reader should refer to the ‘help’ of fssa to see the same example but with the utilization of the new FTS plotting options.
You can install Rfssa from github with:
There are two basic examples which shows you how to use the basic fssa algorithm to analyze a univariate or multivariate functional time series: