Skip to content

Frequency too high for ets?

4 messages · rkevinburton at charter.net, Joshua Ulrich, Eric Zivot

rkevinburton at charter.net
# 
I have a time series that is basically weekly data for a year. So the frequency on the time-series is 52 (52 weeks (obeservations)/year). I have 3+ years of data. I am trying to fit a model to this data using ets in the forecast package (exponential smoothing) but I get:

Error in ets(.sublist$TimeSeries) : Frequency too high

I was looking in the documentation for 'ets' and there was no mention of the limits but apparently 52 is "too high".  Any suggestions?

Thank you.

Kevin
Joshua Ulrich
# 
Look at the source:
<snip>
debug: m <- frequency(y)
Browse[1]>
debug: if (m > 24) stop("Frequency too high")
Browse[1]>
Error in ets(x) : Frequency too high

--
http://quantemplation.blogspot.com
On Mon, Sep 22, 2008 at 12:15 PM, <rkevinburton at charter.net> wrote:
rkevinburton at charter.net
# 
Thank you. I did get a valiuable technique for debugging that I didn't know before from your comments.

So from your comments I gather the limit for exponential smoothing is 24. Not having much experience with exponential smoothing is it unreasonable to expect an algorithm to handle frequencies > 24? I started out with daily data (365 observations per year). I quickly realized that there were not any available fitting algorithms that could handle that degree of resolution. I tried splitting the year in half and splitting the year into quarters (91 observations per quarter). I finally found that if I have 52 observations per year I can use arima to fit and ARIMA model to my data. That seemed like a compromise that I could live with but it was still a compromise. Now if I want to fit an exponential smoothing model I need to further reduce the frequency by more than a half. Are there other packages that can handle this? I would like to be able to forecast down to any given week in the year if possible. In other words if I feed in something like week 48 I would like to be able to handle this. Reommendations?

Thank you.

Kevin
---- Josh Ulrich <josh.m.ulrich at gmail.com> wrote:
Eric Zivot
# 
This problem is just a limitation of representing your data as a ts object,
not a problem with any exponential smoothing algorithm. The ts class in R is
severely limited in the types of regularly spaced time series data that can
be represented (annual, quaterly or monthly). A simple fix is to redefine
your weekly time series so that it has a frequency of 1 instead of 52; that
is, disregard the fact that you have weekly data and just treat each week as
one observation. 

-----Original Message-----
From: r-sig-finance-bounces at stat.math.ethz.ch
[mailto:r-sig-finance-bounces at stat.math.ethz.ch] On Behalf Of
rkevinburton at charter.net
Sent: Monday, September 22, 2008 11:38 AM
To: Josh Ulrich
Cc: r-sig-finance at stat.math.ethz.ch
Subject: Re: [R-SIG-Finance] Frequency too high for ets?

Thank you. I did get a valiuable technique for debugging that I didn't know
before from your comments.

So from your comments I gather the limit for exponential smoothing is 24.
Not having much experience with exponential smoothing is it unreasonable to
expect an algorithm to handle frequencies > 24? I started out with daily
data (365 observations per year). I quickly realized that there were not any
available fitting algorithms that could handle that degree of resolution. I
tried splitting the year in half and splitting the year into quarters (91
observations per quarter). I finally found that if I have 52 observations
per year I can use arima to fit and ARIMA model to my data. That seemed like
a compromise that I could live with but it was still a compromise. Now if I
want to fit an exponential smoothing model I need to further reduce the
frequency by more than a half. Are there other packages that can handle
this? I would like to be able to forecast down to any given week in the year
if possible. In other words if I feed in something like week 48 I would like
to be able to !
 handle this. Reommendations?

Thank you.

Kevin
---- Josh Ulrich <josh.m.ulrich at gmail.com> wrote:
frequency on the time-series is 52 (52 weeks (obeservations)/year). I have
3+ years of data. I am trying to fit a model to this data using ets in the
forecast package (exponential smoothing) but I get:
the limits but apparently 52 is "too high".  Any suggestions?
_______________________________________________
R-SIG-Finance at stat.math.ethz.ch mailing list
https://stat.ethz.ch/mailman/listinfo/r-sig-finance
-- Subscriber-posting only.
-- If you want to post, subscribe first.