tsoutliers(y, xreg = NULL, cval = NULL, delta = 0.7, n.start = 50,
types = c("AO", "LS", "TC"),
maxit = 1, maxit.iloop = 4, cval.reduce = 0.14286,
remove.method = c("en-masse", "bottom-up", "linear-regression"),
remove.cval = NULL,
tsmethod = c("auto.arima", "arima", "stsm"),
args.tsmethod = NULL, args.tsmodel = NULL, logfile = NULL)
tsoutliers0(x, xreg = NULL, cval = 3.5, delta = 0.7, n.start = 50,
types = c("AO", "LS", "TC"), maxit.iloop = 4,
remove.method = c("en-masse", "bottom-up", "linear-regression"),
remove.cval = NULL,
tsmethod = c("auto.arima", "arima", "stsm"), args.tsmethod = NULL,
args.tsmodel = NULL, logfile = NULL)stsm object.y.tsmethod = "stsm"."IO"), additive outliers ("AO"),
level shifts ("LS"), temporary changes ("TC")
andlocate.outliers.cval is reduced
if the procedure is run on the adjusted series, if maxit > 1.remove.outliers.cval. See details.args.tsmethod are referred to.tsmethod.stsm.model. Only for tsmethod = "stsm"NULL. It is the path to the file where
tracking information is printed. Ignored if NULL.tsoutliers."AO" additive outliers, "LS" level shifts,
and "TC" temporary changes are selected;
"IO" innovative outliers and "SLS" seasonal level shifts
can also be selected.tsoutliers0 is mostly a wrapper function around the functions
locate.outliers and remove.outliers.
tsoutliers iterates around tsoutliers0 first for the original series
and then for the adjusted series. The process stops if no additional outliers
are found in the current iteration or if maxit iterations are reached.
tsoutliers0 is an auxiliar function (it is the workhorse for tsoutliers
but it is not intended to be called directly by the user,
use tsoutliers(maxit = 1, ...) instead.
tsoutliers0 does not check the arguments since they are assumed to be passed
already checked by tsoutliers; the default value for cval is not based
on the sample size.
For the time being, tsoutliers0 is exported in the NAMESPACE since it
is convenient for debugging.
If no value is specified for argument cval a default value based on the sample size
is used. Let $n$ be the number of observations. If $n \le 50$ then cval is set
equal to $3.0$; If $n \ge 450$ then cval is set equal to $4.0$;
otherwise cval is set equal to $3 + 0.0025 * (n - 50)$.
If tsmethod is NULL, the following default arguments are used in the
function selected in tsmethod:
tsmethod = "auto.arima": allowdrift = FALSE, ic = "bic";
tsmethod = "arima" = order = c(0, 1, 1) seasonal = list(order = c(0, 1, 1));
tsmethod = "stsm" = stsm.method = "maxlik.fd.scoring", step = NULL,
information = "expected".
If args.tsmethod is NULL, the following lists are used by default,
respectively for each method:
auto.arima: list(allowdrift = FALSE, ic = "bic");
arima: list(order = c(0, 1, 1), seasonal = list(order = c(0, 1, 1))):
stsm: list(stsm.method = "maxlik.fd.scoring", step = NULL, information = "expected")).
xreg must be a matrix with time series attributes, tsp,
that must be the same same as tsp(x). Column names are also compulsory.
If there is only one regressor it may still have non-null dimension, i.e.
it must be a one-column matrix.
The external regressors (if any) should be defined in the argument xreg.
However, they may be also defined as an element in args.tsmethod since this list
is passed to function that fits the model. The function tsoutliers
deals with this possibility and returns a warning if "xreg" is defined
twice with different values. No checks are done in tsoutliers0.
If tsmethod = "stsm", the structural time series model can be defined
through argument args.tsmodel, which is passed to function
stsm.model.
By default, the local level model is chosen for series with one sample per unit
of time (e.g. annual data) and the basic structural model for seasonal data;
ssd = TRUE and sgfc = TRUE.
If maxit = 1 the procedure is run only once on the original series.
If maxit > 1 the procedure is run iteratively, first for the original series
and then for the adjusted series. The critical value used for the adjusted series
may be reduced by the factor cval.reduce, equal to $0.14286$ by default.
The new critical value is defined as $cval * (1 - cval.reduce)$.
By default, the same critical value is used in the first stage of the procedure
(location of outliers) and in the second stage (remove outliers).
Under the framework of structural time series models I noticed that
the default critical value based on the sample size is too high, since all the
potential outliers located in the first stage were discarded in the second stage
(even in simulated series with known location of outliers).
In order to investigate this issue, the argument remove.cval has been added.
In this way a different critical value can be used in the second stage.
Alternatively, the argument remove.cval could be omitted and simply choose
a lower critical value, cval, to be used in both stages.
However, using the argument remove.cval is more convenient since it avoids
locating too many outliers in the first stage.remove.cval is not affected by cval.reduce.
Gómez, V. and Maravall, A. (1996).
Programs TRAMO and SEATS. Instructions for the user.
Banco de España, Servicio de Estudios. Working paper number 9628.
Gómez, V. and Taguas, D. (1995).
Detección y Corrección Automática de Outliers con TRAMO:
Una Aplicación al IPC de Bienes Industriales no Energéticos.
Ministerio de Economía y Hacienda. Document number D-95006.
locate.outliers, remove.outliers,
plot.tsoutliers, print.tsoutliers.data("hicp")
tsoutliers(y = log(hicp[[1]]))Run the code above in your browser using DataLab