I want to understand the functioning of a tool. I've been told that the tool performs multiplicative decomposition of the time series before forecasting.
I've managed to decompose the series in R [decompose()] using multiplicative parameter, but can't run the series through Holt Winters(). Any idea/paper which can help me here?
Much thanks!
Hello, actually it's a problem to do it directly, however if you get an option in the treatment of the R module for such databases the NA option should serve (I haven't done it that way). Anyway, the other option is to try to extrapolate or intrapolate those missing values. There is a package that works and corrects that kind of things and the accuracy improves in some elements of ARIMA, which in my view is better technique than the Holt-Winters, I leave you the link:
https://cran.r-project.org/web/packages/imputeTS/imputeTS.pdf
I am afraid, contrarily to the suggestion of Carlos Andres Yanes Guerra, that the zeros in Vidisha Kanodia data are not missing data. If her data are monthly sales, for example, a zero observation means that they were no sales during that month. Perhaps the observations are counts. There are specific (and complex) methods for handling counts. There are also a few papers in the literature on data with zeroes. One way is to use an additive decomposition and an additive Holt-Winters method but I fear that the results will not be satisfactory, with many negative forecasts. Maybe a square root transformation can help. The forecasts on the transformed data will then have to be squared.
Thanks for your replies Guy Mélard and Carlos Andres Yanes Guerra . For the purpose of testing, I created some dummy time series with no zeros and ran multiplicative HoltWinters model on it. Then I added a constant (equal to 1/100th of the lowest non-zero value in the series) and ran the model again - I am getting very low deviation in the parameters obtained. Does this time series treatment seem well?
With pleasure. You say that the series has no zero. You should try with a series with zeros. Anyway it should depend on your data and the proportion of zeros.
- Badges
- Science topic
- Similar topics
- Biological Science
- Kinesiology
- Running