Search code examples
rtime-serieszoorollapplyrolling-computation

Building a rolling mean forecast

I have a series of intermittent demand called parts (sample below) and I want to develop a rolling mean forecast of a training set and a test set. My code is below as well. the series fitmean calculates the rolling mean, but there are two problems:

  1. It adds a 13th element when all I really want is to get a rolling mean of 12; and,
  2. The dates go from Jun 2016 to Jun 2017, so when I subtract testparts I only get the 6 values for Jan to Jun 2017.

Is there a way to (1) remove the 13th element at the end of fitmean, and (2) change the dates so they match with testparts?

Thank you.

library(forecast,zoo)
parts<-matrix(c(0,0,0,0,0,0,2,0,0,0,0,0,3,0,0,0,0,0,1,0,0,7,0,0),nrow=24,ncol=1)
parts<-ts(parts,f=12,start=c(2016,1))
maemean<-matrix(NA,nrow=12,ncol=1)
  trainparts<-window(parts,end=c(2016,12))
  testparts<-window(parts,start=c(2017,1),end=c(2017,12))
  fitmean<-round(rollapply(parts, width=12, by = 1, FUN = mean))
  maemean<-abs(fitmean-testparts)

Jan-16  0
Feb-16  0
Mar-16  0
Apr-16  0
May-16  0
Jun-16  0
Jul-16  2
Aug-16  0
Sep-16  0
Oct-16  0
Nov-16  0
Dec-16  0
Jan-17  3
Feb-17  0
Mar-17  0
Apr-17  0
May-17  0
Jun-17  0
Jul-17  1
Aug-17  0
Sep-17  0
Oct-17  7
Nov-17  0
Dec-17  0

Clarification:

The above list should break down to a training set from Jan-16 to Dec-16 and a test set from Jan-17 to Dec-17. What I want to do is use a rolling mean so the average of Jan-16 to Dec-16 (rounded, which is 0) becomes the forecast for Jan-17, and so on, i.e., Feb-16 to Jan-17, etc. The output should look like this

Jan-17  0
Feb-17  0
Mar-17  0
Apr-17  0
May-17  0
Jun-17  0
Jul-17  0
Aug-17  0
Sep-17  0
Oct-17  0
Nov-17  1
Dec-17  1

Unfortunately, I get this with 13 vice 12 elements.

     Jan Feb Mar Apr May Jun Jul Aug Sep Oct Nov Dec
2016                       0   0   0   0   0   0   0
2017   0   0   0   1   1   1
Solution

  • 1) width = list(...) Removing all the irrelevant code from the question and changing the rollapply line we have this where -seq(12) is a vector of offsets instructing rollapply to pass the first prior, second prior, ... twelfth prior values to mean at each point.

    library(zoo)

# test data
parts <- matrix(c(0, 0, 0, 0, 0, 0, 2, 0, 0, 0, 0, 0, 3, 0, 0, 0, 0, 0, 1, 0, 0, 7, 0, 0), 
  nrow = 24, ncol = 1)
parts <- ts(parts, freq = 12, start = c(2016, 1))

round(rollapply(parts, list(-seq(12)), FUN = mean))

    giving:

         Jan Feb Mar Apr May Jun Jul Aug Sep Oct Nov Dec
2017   0   0   0   0   0   0   0   0   0   0   1   1

    2) rollsumr Another approach would be to take a rolling sum of width 13 and then subtract off the current value and divide by 12:

    round((rollsumr(parts, 13) - parts) / 12)
##      Jan Feb Mar Apr May Jun Jul Aug Sep Oct Nov Dec
## 2017   0   0   0   0   0   0   0   0   0   0   1   1