Specifies an ISSM model prior to estimation with option for automatic model selection.
issm_modelspec(
y,
auto = FALSE,
slope = TRUE,
slope_damped = FALSE,
seasonal = FALSE,
seasonal_frequency = 1,
seasonal_harmonics = NULL,
ar = 0,
ma = 0,
xreg = NULL,
variance = "constant",
garch_order = c(1, 1),
lambda = NULL,
lower = 0,
upper = 1,
distribution = "norm",
sampling = NULL,
init_garch = "unconditional",
sample_n = 10,
top_n = 1,
...
)An object of class “tsissm.spec” or “tsissm.autospec”.
an xts vector.
whether to use automatic model selection.
(Logical) slope component. If “auto” is TRUE, then this can be a vector of size 2 with TRUE and FALSE.
(Logical) slope dampening component. If “auto” is TRUE, then this can be a vector of size 2 with TRUE and FALSE.
(Logical) seasonal component(s).
vector of numeric seasonal frequencies (can be fractional).
the number of harmonics per seasonal frequency for the trigonometric seasonality. If “auto” is TRUE, this can be a list with slots for each seasonal frequency listing the sequence of harmonics to test for each.
AR order.
MA order.
an xts matrix of external regressors.
either “constant” or “dynamic”. In the latter case a GARCH model will be used. For the automatic selection case, both can be provided as a vector.
the order of the GARCH model (no automatic selection of GARCH order allowed).
the Box Cox lambda. If not NULL (no transformation), then either a numeric value or NA denoting automatic estimation.
lower bound for the transformation (defaults to 0).
upper bound for the transformation (defaults to 1.5).
a choice of the Normal (“norm”), Student (“std”) or Johnson's SU (“jsu”) distributions. There is no choice for selecting multiple choices for automatic selection.
(optional) sampling frequency of the dataset. If NULL, will try to identify from the timestamps of y. This is useful for plotting and extending the timestamps in the prediction horizon.
GARCH variance initialization method with options to use the “unconditional” variance or a “sample” of length “sample_n”.
the sample length to use if choosing to initialize the GARCH variance using the “sample” method.
how many models to return from the top when using automatic model selection. If this is equal to 1 then the best selected model based on lowest AIC will be returned else a list of the top_n estimated models (based on AIC).
not used.
The specification object holds the information and data which is then passed to the maximum likelihood estimation routines. Depending on whether automatic selection is chosen, it will dispatch to the appropriate estimation routine.
The specification performs some sanity checks on the arguments provided and sets up the required state space matrices and parameters which are used in the estimation stage.
De Livera, Alysha M and Hyndman, Rob J and Snyder, Ralph D, 2011, Forecasting time series with complex seasonal patterns using exponential smoothing, Journal of the American Statistical Association, 106(496), 1513--1527.