christophsax / seasonal Goto Github PK

R error handling instead of X13

x <- seas(AirPassengers, x11.type = "trend")
static(x, test = F)

with possible re-evaluation

do some speed tests: how much time uses reading spectrum data by default?

probably a x13 issue, it happens when seas is combined with transform.function = log.

in inspect, warnings should be returned immediately and not after quitting the function

• different series should be plotted
• in the revisions save all available output.

7.18 TRANSFORM

series { data = (879 899 462 670 985 973 ...)
          start = 1967.jan }
transform{data =(1 1.5.75 1 1...) mode = ratio
adjust = lom }

R-code:

# transform
tf <- rnorm(144)

AirPassengers.tf <- AirPassengers * tf
m <- seas(AirPassengers, 
          AirPassengers.tf
)

# transform back
final(m) / tf

The history spec leads to questionable figures. The following four models do not greatly differ but history treats them very different. It has something to do with the log, but also with x11/seats. The final series in the x$sae element ist also not identifiable in the mode. Need for inquiry. Delays history spec.

library(seasonal)
m1 <- seas(x = AirPassengers, arima.model = "(0 1 1)(0 1 1)", 
           regression.aictest = NULL, outlier = NULL, 
           transform.function = "log")

m2 <- seas(x = AirPassengers, arima.model = "(0 1 1)(0 1 1)", 
           regression.aictest = NULL, outlier = NULL, 
           transform.function = "none")

m3 <- seas(x = AirPassengers, arima.model = "(0 1 1)(0 1 1)", 
           regression.aictest = NULL, outlier = NULL, 
           transform.function = "log", x11 = list())

m4 <- seas(x = AirPassengers, arima.model = "(0 1 1)(0 1 1)", 
           regression.aictest = NULL, outlier = NULL, 
           transform.function = "none", x11 = list())

r1 <- revisions(m1)
r2 <- revisions(m2)
r3 <- revisions(m3)
r4 <- revisions(m4)

ts.plot(r1$sae, col = c("black", "red"))
ts.plot(r2$sae, col = c("black", "red"))
ts.plot(r3$sae, col = c("black", "red"))
ts.plot(r4$sae, col = c("black", "red"))

texreg S4 method for seasonal is implemented, but it is unclear how a method definition for an outside function should work.

Currently, x11regression and regression are mutually exclusive. That should be changed.

    series{ title = "Automobile Sales"
            file = "carsales.dat"
            start = 1975.Jan }
    transform{ function = log }
    regression{ variables = (const)
                user = (strike80 strike85 strike90)
                file = "strike.dat"
                format = "(3f12.0)" }
    arima{ model = (0 1 1)(0 1 1)12 }
    x11{  title = ("Car Sales in US"
                   "Adjust for strikes in 80, 85, 90")
          save = seasonal appendfcst = yes
          }
    x11regression{   variables = ( td easter[8] )  }

when solved, update wiki

This should also show the model:

> seas(AirPassengers, arima.model = c(2, 1, 1, 0, 2, 0))
Warning in seas(AirPassengers, arima.model = c(2, 1, 1, 0, 2, 0)) :
  Model used for SEATS decomposition is different:

a comprehensive list of useful statistics should be shown

• Information criterion (AICc and BIC)
• QS Statistic for seasonality
• Q Statistic for autocorrelation
• (Normality statistic)

probably also affects other non standard freq. needs testing

# bimonthly data
  require(tempdisagg)
  AirPassengersBM <- ta(AirPassengers, to = 6)
  m <- seas(AirPassengersBM, 
       regression.aictest = NULL,
       outlier.types = c("ao", "ls", "tc"),
       forecast.maxlead = 18
  )

final(m) should return bimonthly data

X-13 has support for internal missing values, could be included in seasonal

model selection

• avoid NA in selection
• automatic reestimation as an additional output

graphical outlier deselection?

New plots

• slidingspan
• history (delayed due to #5)
• tidy up (reidual plot and spectra probably not needed)

• analyze View() and Edit()
  • is there a way to change the defaults to make View more usable on all platforms?

content of series help with comments should be sufficient

one help page for plots, one for diagnostical statistics

with numeric model input, the nonseasonal part is used as the seasonal part:

m1 <- seas(AirPassengers, arima.model = c(2, 1, 1, 0, 0, 0))
m2 <- seas(AirPassengers, arima.model = "(2 1 1)(0 0 0)")
identical(spc(m1), spc(m2))

The underlying function as.character.arima is to blame:

> as.character.arima(c(3, 3, 2, 1, 3, 3))
[1] "(3 3 2)(3 3 2)"
> seasonal:::as.character.arima(c(2, 1, 1, 0, 0, 0))
[1] "(2 1 1)(2 1 1)"
> seasonal:::as.character.arima(c(2, 2, 1, 0, 0, 2))
[1] "(2 2 1)(2 2 1)"
> seasonal:::as.character.arima(c(2, 1, 1, 2, 0, 0))
[1] "(2 1 1)(2 1 1)"
> seasonal:::as.character.arima(c(2, 1, 1, 0, 1, 0))
[1] "(2 1 1)(2 1 1)"
> seasonal:::as.character.arima(c(2, 1, 1, 0, 1, 0))
[1] "(2 1 1)(2 1 1)"
> seasonal:::as.character.arima(c(2, 1, 1))
[1] "(2 1 1)"

seas(AirPassengers,
     transform.function = "log",
     arima.model = c(0, 1, 2, 0, 1, 1),
     x11 = list()
)

this should be possible as an alternative to

seas(AirPassengers,
     transform.function = "log",
     arima.model = "(0 1 2)(0 1 1)",
     x11 = list()
)

Add support for user defined variables

Add support for mts objects

For an unknown X-13-ARIMA SEATS related reason, the regressors in the x11regression spec do not apperar in the output.

    seas(AirPassengers, 
         x11regression.variables = c("td/1950.1/", "easter[8]", 
                                     "labor[10]", "thank[10]"),
         x11.seasonalma = "x11default",
         x11.sigmalim = c(1.8, 2.9),
         x11.appendfcst = "yes",
         )

when solved, update wiki

investigate whether the residual statistics in the summary are appropriate

suggestion by @mbannert:

use transform.function = "none" if series has negative values to avoid x13 error

use of defaults and the dots should be better explained.

otherwise, no change in usage

if x13 is installed in a folder with a space in the name, e.g c:\prsgram files\blabla, run_x13 does not work.

perhaps, the identify function could be used to unselect (or perhaps select) outliers in a plot

function

• applied to a "seas" object
• calls plot.seas
• uses identify to get the info from clicks
• adjust the model
• reestimate and replots it
• return static call

rather than explicitly import series from x-13, it would be more flexible to import all produced time series. Like that:

1. list all time series files produced by X13
2. import the time series that have been produced (read_series)
3. the main series probably still need some extra treatment, as they have different names in seats and x11
4. when saved, it is unclear how the series should be accessed.

installation, see readme (instead of 'included')

slidingspans recalls seas with activated slidingspans spec. returns object of class slidingspans.

print.slidingspans shows the output section.
plot.slidingspans gives a nice overview of the 4 (or more) spans. Perhaps also a second panel with max. deviations.

x11 or seas

• a list of user defined plot functions could be supplied
• perhaps the plot part could be based on a list of functions

set up a testing routine for new releases

• put all examples in a test file in inst/tests
• numerical comparison with some outputs (save in data)
• add x11 tests
• small function that performs all tests
• run function on different platforms