Seasonal adjustment is a crucial step preprocessing the data for further research. Researcher however has a number of options for trend-cycle-seasonal decomposition. The most common (judging by the number of citations in empirical literature) rival seasonal decomposition methods are X-11(12)-ARIMA, Tramo/Seats (both implemented in Demetra+) and $R$'s stl. Seeking to avoid random choice between the above-mentioned decomposition techniques (or other simple methods like seasonal dummy variables) I would like to know a basic strategy that leads to choosing seasonal decomposition method effectively.
Several important subquestions (links to a discussion are welcome too) could be:
- What are the similarities and differences, strong and weak points of the methods? Are there any special cases when one method is more preferable than the others?
- Could you provide general guides to what is inside the black-box of different decomposition methods?
- Are there special tricks choosing the parameters for the methods (I am not always satisfied with the defaults,
stlfor example has many parameters to deal with, sometimes I feel I just don't know how to choose these ones in a right way).
- Is it possible to suggest some (statistical) criteria that the time series is seasonally adjusted efficiently (correlogram analysis, spectral density? small sample size criteria? robustness?).