This happens sometimes, I’m not sure why…

The problem with seasonal models on intermittent data is exactly what you say – very low number of observations. So the default univariate models will struggle with fitting anything meaningful with seasonality. John Boylan, Stephan Kolassa and I are currently working on one of the potential solutions for this problem (using vector models instead of univariate ones), but this is an ongoing research, in an alpha stage…

One of the potential solutions for now is to use dummy variables for seasons instead (for the cases, where intermittent!=0) – this might work slightly better. You can use xreg parameter in es() for that. Please, let me know if this works.

I restarted my R session and now it works – I’m not sure what happened.

The results are, however, not as great as I had hoped. The main issue is that the data is seasonal, but I can only estimate a non-seasonal model due to too few non-zero values. I guess this is due to the fact that the underlying model tries to estimate seasonality for all of the year, and I only have data for around 14 weeks during the winter?

Do you have any ideas for how to proceed? I haven’t come across anything solving this problem yet, although I feel it should be quite a common problem in retailing.

In general this thing that you use:
es(data, “MNN”, intermittent = dummy_vector, h = h)
should work. If it doesn’t then it’s possible that you found a bug. Can you, please, report it here: https://github.com/config-i1/smooth/issues – so we can figure out what’s happening?

Hi Ivan,

Thanks for your answer. I do happen to know when the sales start and end, so I think the first approach would make most sense.

What do you mean exactly when you say to provide the dummy vector to the intermittent parameter? I tried fitting a model like this:
es(data, “MNN”, intermittent = dummy_vector, h = h), and got an error (i also tried fitting other types than MNN).

Hi Casper,

I don’t have a specific model for that yet, and I have not covered some specifics of the functions of smooth in order to give the detailed answer. But if you know when the sales start and when they end (so you know specific values for o_t), then you can provide those values as a vector of dummies in the intermittent parameter. The other option would be to model time varying probability based on ETS(M,N,M) model for demand occurrence (meaning that there are periods, when probability is very high and the others, when it’s very low), but this relies on logistic model that has not been covered yet, and has its own limitations. I will discuss it in the next post.

Cheers,
Ivan