tempdisagg 1.0 (2020-02-07)

major changes

works now with most time series classes, as supported by the tsbox package.
disagregation is possible to all frequencies (e.g., monthly to daily). Disaggregation takes into account the calendar, e.g., the fact that February is shorter than other months. (#30)
new method: “fast”, a shortcut for chow-lin-fixed with fixed.rho = 0.99999. The method returns approximately the same results as “denton-cholette”, but is much faster. (#14)
new vignettes: intro to tempdisagg, disaggregation to high frequency

new methods: “dynamic-maxlog”, “dynamic-minrss”, “dynamic-fixed”, as described in Santos Silva and Cardoso, 2001. Many thanks to Tommaso Di Fonzo for providing a blueprint written in GAUSS.
updated documentation to include new methods.

better checks for non-time-series inputs. (https://github.com/christophsax/tempdisagg/issues/20)
added extensive numerical testing on travis.

ta() returns correct results if conversion is “last” or “first”, and the first or the last period is incomplete. (https://github.com/christophsax/tempdisagg/issues/22)

retropolation: ‘td’ will performs both extra- and retropolation if the high frequency series covers a larger time span than the low frequency series.
low frequency values are ignored if series is longer than high frequency series (with a warning).
suggestion to use ‘denton-cholette’ when the original ‘denton’ method is chosen.

Our R-Journal article on temporal disaggregation explains tempdisagg in more detail. Links are included in the package description, the help files and the README file.

predict method for ‘td’ is now different from fitted:
- $fitted.values of a ‘td’ object now containts the low-frequency fitted values of a regression or the low-frequency indicator in case of the Denton methods. The values are be accessed by fitted().
- The final high frequency series is now stored in $values. As before, these values are accessed by predict().
Package overview (?tempdisagg)
Demo (demo(tempdisagg))
argument ‘truncated.rho = 0’ instead of ‘no.neg = TRUE’. This allows for truncation values different from 0. Default behavior is the same as in 0.21.

in 0.21, ta() produced an error if less than a low-frequency unit was covered by high frequency data. Now it produces series containing only NA.
If a singular data matrix is entered, there is a new warning.

new methods available: “chow-lin-fixed” and “litterman-fixed”. Using the “fixed.rho” argument, an autoregressive parameter may be specified by the user.
interface changes: “chow-lin-maxlog-ecotrim” and “chow-lin-maxlog-quilis” are defined as new methods. No need for the old ‘vcov’ argument anymore.
new defaults: method = “chow-lin-maxlog”, neg.rho = FALSE with positive values for rho only, the chow-lin-maxlog method generally outperforms the other methods.
all relevant arguments are directly entered to td()
summary output: If neg.rho = FALSE and a negative rho is truncated to 0, and indicator is shown in the summary output.
non time-series mode: optionally, standard vectors can be used instead of time series. In this case, the frequency of low frequency variable is 1, while the fraction of the high frequency variable is specified by the ‘to’ argument
updated help files

td() is rewritten and has a clear structure now.
GLS Regressions are performed by the new CalcGLS() function, which uses QR-decomposition instead of matrix-inversion. This is faster and numerically stable. It resolves an issue wher large (or small) numbers have led to a ‘system is computationally singular’ error.