Periodicity of Weekly Zoo - R-help

Mon, Dec 3, 2012 5:30 PM #

Hi List,

I have weekly sales observations for several products drawn via ODBC.
Source data is available at
https://www.dropbox.com/s/78vxae5ic8tnutf/asr.csv.

This is retail sales data, so will contain seasonality and trend
information. I expect to see 52 or 53 observations per year, each
observation occuring on the same day of the week (Saturday). Ultimately
I'm looking to feed these series into forecasting models for demand
planning.

The data has issues with internal gaps, so while I've been able to
create a ts that appears to respect the frequency and period, I suspect
that a zoo is going to be a better data container. Unfortunately, I'm
not understanding the use of zoo() to describe frequency/period/deltat.

In the example below I use sales[,16] (aka $p) as it has several
periods (data between 2004 and 2012). I've tried using frequency=52, =7
and =1, but get the same result each time; every data point ends up in
cycle 1 and I don't have the periodicity needed to find seasonality.

Attaching package: 'zoo'

The following object(s) are masked from 'package:base':

as.Date, as.Date.numeric

+ sales$date_end, frequency = 52)

[1] 52

2004-08-14 2004-08-21 2004-08-28 2004-09-04 2004-09-11 2004-09-18 
         1          1          1          1          1          1 
2004-09-25 2004-10-02 2004-10-09 2004-10-16 2004-10-23 2004-10-30 
         1          1          1          1          1          1 
2004-11-06 2004-11-13 2004-11-20 2004-11-27 2004-12-04 2004-12-11 
         1          1          1          1          1          1 
2004-12-18 2004-12-25 2005-01-01 2005-01-08 2005-01-15 2005-01-22 
         1          1          1          1          1          1 
2005-01-29 2005-02-05 2005-02-12 2005-02-19 2005-02-26 2005-03-05 
         1          1          1          1          1          1

Doubtless it's some facile error that will make me feel sheepish, but
I've been staring at this for a bit now and just getting nowhere. Any
pointers would be greatly appreciated.

Thanks,

Andrew

Gabor Grothendieck

Mon, Dec 3, 2012 8:25 PM #

On Mon, Dec 3, 2012 at 8:30 PM, Andrew Freedman

<andrewf at hotsprings.com.au> wrote:

Hi List,

I have weekly sales observations for several products drawn via ODBC.
Source data is available at
https://www.dropbox.com/s/78vxae5ic8tnutf/asr.csv.

This is retail sales data, so will contain seasonality and trend
information. I expect to see 52 or 53 observations per year, each
observation occuring on the same day of the week (Saturday). Ultimately
I'm looking to feed these series into forecasting models for demand
planning.

The data has issues with internal gaps, so while I've been able to
create a ts that appears to respect the frequency and period, I suspect
that a zoo is going to be a better data container. Unfortunately, I'm
not understanding the use of zoo() to describe frequency/period/deltat.

In the example below I use sales[,16] (aka $p) as it has several
periods (data between 2004 and 2012). I've tried using frequency=52, =7
and =1, but get the same result each time; every data point ends up in
cycle 1 and I don't have the periodicity needed to find seasonality.

sales <- read.csv("asr.csv")
library(zoo)

Attaching package: 'zoo'

The following object(s) are masked from 'package:base':

as.Date, as.Date.numeric

sales.zoo <- zoo(subset(sales, select=c(2:length(sales))), order.by=

+ sales$date_end, frequency = 52)

sales.zoo.i <- na.approx(sales.zoo) # interpolate internal NA values
frequency(sales.zoo.i) # 52, which seems right

[1] 52

cycle(sales.zoo.i[1:20,16]) # everything is in the same cycle...

2004-08-14 2004-08-21 2004-08-28 2004-09-04 2004-09-11 2004-09-18
         1          1          1          1          1          1
2004-09-25 2004-10-02 2004-10-09 2004-10-16 2004-10-23 2004-10-30
         1          1          1          1          1          1
2004-11-06 2004-11-13 2004-11-20 2004-11-27 2004-12-04 2004-12-11
         1          1          1          1          1          1
2004-12-18 2004-12-25 2005-01-01 2005-01-08 2005-01-15 2005-01-22
         1          1          1          1          1          1
2005-01-29 2005-02-05 2005-02-12 2005-02-19 2005-02-26 2005-03-05
         1          1          1          1          1          1

Doubtless it's some facile error that will make me feel sheepish, but
I've been staring at this for a bit now and just getting nowhere. Any
pointers would be greatly appreciated.

A complete cycle is always represented by 1 time unit so if you wanted
a complete cycle to be a year then you would need to represent time in
years and fractions of a year, not as "Date" class.  That is how "ts"
class works too.

Since weeks don't evenly divide years you will have to approximate
this in order to have a frequency of 52.  There are many ways to do
this but below we drop week 00 in 53 week years so that there are 52
weeks in every year:  Years with 52 weeks don't have a week 00 so this
makes all years 52 weeks.

z <- read.zoo("asr.csv", sep = ",", header = TRUE)

# drop week "00"
z0 <- z[ format(time(z), "%W") != "00" ]
t0 <- time(z0)

# convert time to year + fraction
time(z0) <- as.numeric(format(t0, "%Y")) +
  (as.numeric(format(t0, "%W")) - 1) / 52

# convert to zooreg class (almost regularly spaced)
zr <- as.zooreg(z0)
frequency(zr) # 52
head(cycle(zr))



--
Statistics & Software Consulting
GKX Group, GKX Associates Inc.
tel: 1-877-GKX-GROUP
email: ggrothendieck at gmail.com