Bookmark and Share

Notice: On April 23, 2014, Statalist moved from an email list to a forum, based at

[Date Prev][Date Next][Thread Prev][Thread Next][Date Index][Thread Index]

Re: st: Filtering methods with short time series

From   Jorge Eduardo Pérez Pérez <[email protected]>
To   "[email protected]" <[email protected]>
Subject   Re: st: Filtering methods with short time series
Date   Mon, 31 Jan 2011 15:42:45 -0500

Try to avoid -cfitzrw-, I have compared the output from
implementations of the Christiano Fritzgerald filter in other software
to the Stata output and they are different. I already contacted the
authors about this, but I haven't received any response.

I don't understand why you argue that you need I(2) data to use the HP
filter. You can use it on series with a lower order of integration,
with the caveat that this might induce spurious autocorrelation on the
filtered series.

You might want to try the user written commands -bking- or -couliari-,
that implement different approximations of the band pass filter. The
latter estimates end points directly, thus avoiding the "future
information" problem you write about. -bking- will only produce the
filtered series, but you can obtain the cycle substracting the
filtered series from the original series. -couliari- will produce both
a trend and a cycle.

To improve the smoothing to better suit your needs, you might want to
try to vary the plo (s) and phi (e) options of bking(couliari), to
extract cycles of the frequency you desire.
Jorge Eduardo Pérez Pérez

On Mon, Jan 31, 2011 at 3:12 PM, kokootchke <[email protected]> wrote:
> Dear all:
> I am inspecting time series data of relatively short length (monthly, 2003-2009). The data are the amount of time spent by individuals on various activities on a daily basis (cooking, sleeping, etc.), they are aggregated data, and they suffer from many pitfalls, e.g., the series are short, they look very different depending on each activity--and definitely not I(2) to be able to use an HP filter.
> What I would like to do is to separate out any long-term trends from the short-term fluctuations, as usual. I have used the typical HP filter, calculated the optimal smoothing parameter (always something between 50,000 and 70,000 for my monthly series), and then use the -hprescott- command to generate the corresponding cycles and trends. This gives me some reasonable filtered data, with the expected phase shifts every now and then. Also, sometimes the cycles look way too much like the original series. Due to the nature of the data, I am certain that this is not the ideal filter in my case -- in fact, ideally, I would like to use an asymmetric filter that doesn't use future information to extract the trend.
> I have also used a Christiano-Fitzgerald (-cfitzrw-) setting the low/high bars for the bandpass at 24/72 months, respectively. The filtered data produced with this filter are very strange, though: many of the series look wavy, like a sine/cosine function, and the trends often look very different from the apparent trends in the original series. Also, it's the first time I use this -cfitzrw- command and I'd like to know if it can also produce both a cycle and a trend series, the way -hprescott- does.
> Could you provide any feedback as to how I can make these (or other) filters work better given my data?
> Thank you.Adrian
> *
> *   For searches and help try:
> *
> *
> *

*   For searches and help try:

© Copyright 1996–2018 StataCorp LLC   |   Terms of use   |   Privacy   |   Contact us   |   Site index