Abstract
For the purposes of monthly time series analysis the original data are assumed to be additively composed of a trend-cycle, a seasonal, a calender-effect, and a residual component, the last of which may include a few extreme values. The first mentioned two components are estimated by moving filter applications derived from approximating functions by a regression approach. The appropriateness of the filters is judged and controlled by their transforms into the frequency domain. While model 3 of this procedure was designed series-specifically, the improved model 4 provides adaptive and end-stable results by a uniform set of filters to both components where general optimization aspects had already been included. The mathematical models used, their specifications and development, and some features of the new model like practical simplicity and definiteness are indicated. A hint is given to the integrated identification of extreme values and estimation of calendar-effects which in future will both be performed before the eventual seasonal adjustment.
Get full access to this article
View all access options for this article.