I have a large set of relatively simple time series with very similar behaviour, on which I would like to do short-term forecasts. These series are non-aligned, and at one moment in time, only a small fixed number of series is considered. After that, the series are dropped and never become relevant again.

An example to clarify:

At 12:00, three series [A, B, C] are considered, and each is forecasted using an instance of the generic model.

At 12:05, three series [B, C, D] are considered, and each is forecasted with its instance of the generic model. D being a series never seen by the model before. It is a very similar series to the previously seen, and I would expect the model to generalize and provide a reasonable forecast.

I am essentially looking for a model that I could fit on multiple time series and then create instances with data from different series and get forecasts.

What models could suit this problem well? How can I fit/train the model on all the training series evenly to not re-fit just for the last observed series?


1 Answer 1


I suspect that exponential smoothing models would do what you want, but I am not sure exactly what you are asking. They are very easy to run and forget data past a specific point - well they weight data heavily depending on what you tell it.

  • $\begingroup$ Thanks for the suggestion. I think I found what I am looking for, the key was redefining the problem as a regression problem. The similar nature of the series allows even with the time series information "lost" to provide decent results and generalize well over multiple series. $\endgroup$
    – holyfish
    Mar 4, 2021 at 10:57

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.