The classes and methods associated with the implementation of rts, cts
and its in Splus and ts in R accomplish some of these objectives.
For many time series programs the availability of a window method
provided by those classes is the main method which is needed. However,
the time attributes of calculated objects are
often lost and programmers must re-assign time attributes to the resulting
object if they are to be retained. Historically this was done with
tsp, but this relies on a particular time representation and
would not work for other representations. In order to address this,
the
tframe(x) <- tframe(y)
to make the time frame of x the same as that of y,
without the need to
worry about what time representation is used in y.
In this assignment x and y need not be too similar (one
might be a univariate series while the other is a matrix or an array or
list of spatial or panel time series data), as long as they are similar
in the time dimension. For the case where tsp(x) <- tsp(y) would
make sense,
that is effectively what the above tframe assignment will do, and for
existing code, most of the conversion to these more robust methods is
accomplished simply by changing "tsp" to "tframe" and nrow() for a time
series matrix to periods().
The tframe assignment example above is accomplished by switching the dispatch so that it follows the classes of the tframe of y, rather than the classes of x as would normally be done for the above kind of assignment. Doing this in a generic way allows for the possibility of future classes of time representation. This is different from the way that rts, cts, its and ts are implemented, in the sense that it is the tframe of the data which is assigned a class indicating the time representation, not the data object itself.
The most general (last) class of the tframe should be "tframe". The method "is.tframe" checks if an object is a tframe, and the method "is.tframed" checks if an object has a tframe. In general, tframe methods act on the time frame (tframe) and tframed methods act on data which is tframed.
More specific methods can be defined for any special time representation (methods are defined in this library for tframes of class c("default", "tframe") which in Splus are the old style tsp convention for time). Also, there is a sketched implementaion for rts, cts, its, and a class and methods style implementation of tsp called tf. The tframe's specific classes are called rtstframe, ctstframe, and itstframe, to prevent confusion using inherit(). The R version includes an implementaion for the ts time series class with tframe's specific class is tstframe.
When implementing a new time series, suppose it is called zzz, then the tframe attibute of an object using this time frame should have class c("zzz", "tframe"). (Note zzz should be different from the class of the object itself.) The most important methods which need to be supported are tframe.zzz(), start.zzz(), end.zzz(), and periods.zzz(). While frequency.zzz() should not in theory be necessary, it makes porting code much easier. Other methods which should be supported are time.zzz(), checktframeConsistent.zzz(), tfTruncate.zzz(), tfExpand.zzz(), earliestStartIndex.zzz(), earliestEndIndex.zzz(), latestStartIndex.zzz(), and latestEndIndex.zzz().
The method tfwindow is used in this library and is typically just the same as window, but the new name has been used because of historical changes and bugs in window, and in order to support the argument "warn" to suppress messages (when objects are windowed unnecessarily).
One implication of the desire to be able to use a statement like tframe(x) <- tframe(y) is that the tframe should not indicate which dim of the data is the time dimension. In general this will have to be another attribute of the data, but the current convention of using the first dimension for matrix data and the length for vector data, makes it unnecessary to specify.
Operations which should be possible on tframed data:
In the time dimension - window (tfwindow), splice In other dimensions -tbind (like cbind) with shift to align the the tframe (and NA pad.start, pad.end, pad=TRUE/FALSE) - [ ] without losing the tframe. This is done with selectSeries().
The attribute "seriesNames" is also supported as a way to indicate the names of series in an object. This overlaps with the use of "names" and "dimnames[[2]]" used previously for series names in S, but seems necessary in order to have a more complete generic decompostion of the time dimension from the other dimensions.
Many of the functions in the library are not yet individually documented (and were it exists the documentation is a draft) however, the functions are all very short and can be examined easily. The code in the tests subdirectory provides a short set of tests and may serve as an example.
The main objects are: tfdata time series input and output data structure
tframe,
tframed,
tfwindow,
tfplot