4

I am looking for intermittent demand time series, i.e., demands that have "many" zeros. Such series commonly occur in the retail business for single SKUs in single locations, or as the demand for spare parts. For example:

intermittent demand example

So far, I have only found two time series in the tsintermittent package for R (one of them plotted above), but I'd really like to have many, many series to do some serious testing. (For instance, the M3 competition had 3003 time series.) Rob Hyndman's Time Series Data Library hangs on me, but there are no retail time series there, either, and there likely is nothing under "Industry".

Can anyone point me to a good source?

albert
  • 11,885
  • 4
  • 30
  • 57
Stephan Kolassa
  • 231
  • 1
  • 9
  • 1
    Incidentally, it seems like a [tag:time-series] tag would make sense, but I don't have the rep yet to create a new tag. There are about six questions that could usefully be tagged [tag:time-series]. – Stephan Kolassa Jan 15 '15 at 19:40
  • 1
    If the time-series is sparse, the data may be distributed as an event list (one record per occurance, with the time of when it happened), rather than as a time series. – Joe Jan 16 '15 at 18:42
  • I think the search terms you are looking for are "burst", "anomaly" or "outlier" detection. The many zeros you mention could just be the difference from the mean. – philshem Jan 17 '15 at 15:50
  • I added the 'time-series' tag to this question and the others you linked. Thanks for bringing it up! – philshem Jan 17 '15 at 15:52
  • Would synthetic data be acceptable, or do you specifically want real, historical examples? In the real world, I've seen this type of series looking at a single business's job postings over time, but I'm afraid I don't know of any public postings data off the top of my head. – J. Miller Feb 06 '15 at 21:55
  • @J.Miller: I'm really looking for real data. There are ways of generating this kind of data, but I don't like assessing my ideas on synthetic data - you always get out what you put in, and it's rather dubious whether this generalizes to the real world. – Stephan Kolassa Feb 07 '15 at 08:35
  • Did anyone get such data in the meantime? I am also looking for intermittent demand time series for my research. Unfortunately, I could not find any yet. Thanks a lot for any support! – Andi Apr 08 '19 at 09:34
  • @StephanKolassa - I know this was many years ago but I am wondering if you have found many of such series or built a compilation of your own at this point. I am becoming interested in this problem. Also, wondering if you have any suggestions for modeling aside from Croston's method and its variants. – RDizzl3 Feb 18 '20 at 00:42
  • @RDizzl3: unfortunately, I haven't found anything, so I'm working off our internal datasets, which is a problem for reproducibility. I have hopes for the upcoming M5 forecasting competition, which will feature over 43,000 time series on product-location-day granularity from Walmart - some of these should be intermittent. Finally, this article in the IJF from 2016 encapsulates my (still) current thinking. I don't like Croston much, because IMO we need density forecasts. – Stephan Kolassa Feb 18 '20 at 10:02

0 Answers0