Hi all,
I have received many replies suggesting ways to solve my problem. Thank
you so much.
Actually, what I have done is first estimate a linear trend and remove it
from the original series. Concerning the varied period of seasonal effect,
I fit the detrended data against the number of days to the end of month
and remove this fitted value from the detrended series. I call this
deseasonalize if it works :P.
When processing the deseasonalized series, I found although the residuals
do not perform well but the lag-1 difference of the residuals are much
better. At least there is no obvious autocorrelations and partial
autocorrelations. ACF of the squared value of the differenced residual cut
off immediately after lag 2. Does that mean I need to fit a GARCH model?
Do you think this method makes any sense?
Many thanks,
LS
Dear allstats:
I am currently working on a daily time series with seasonal period of a
month. The problem is due to the existence of bank holidays and varied
number of working days in a calendar month, the length of the seasonal
period varies from 18 days to 24 days. I want to remove the seasonality
and fit a ARMA model of the residuals.
Does anybody know how to fit a model like this or any related articles
dealing with seasonal effects of different length?
Many thanks,
LS
|