I'm dealing with a system which monitors and records a time series (half hourly) which I plan to use to build a double seasonal time model (if possible using something that already exists, such as tbats from R's forecast).

My issues is related to the fact that I don't plan to build this model very often. Let's say I have three weeks of data which allowed me to take into account both the daily and weekly seasonalities. Two months later, it's a Tuesday, it's 10AM and I want to make a forecast for 12AM. Do I need to use all the data from the moment I made the model to this current date, or is there a way I can use a shorter period (such as the current day data) ? Speed is paramount and if I need to make a forecast every half an hour, I would really like only using a reduced period of time to make the forecast.

I'm basically looking for a method for dealing with growing time series as most literature deals with a static data sets.

  • $\begingroup$ It depends on the process which you observe. If it is "stable" using current day data is perfectly ok. I would run a simulation to determine how much forecast accuracy depends on the sample length. $\endgroup$ – mpiktas Feb 13 '14 at 14:45
  • $\begingroup$ The only solution which seems doable to me (in R's context) would be to save the three weeks period used to build the model, and to concatenate it with the time series beginning on the first day of the current week (even if this week is months away from when the model was built). I guess that to increase the stability, I could consider rebuilding the model every week from the past three weeks ? $\endgroup$ – wrousseau Feb 13 '14 at 14:53
  • $\begingroup$ How much time it takes to build model one time? For forecasting purposes it is always best to build the model with the most recent data. However since you have lots of data, I suggest doing various forecasting scenarios and then compare them. Then you will have a feel what is the cost of each scenario. $\endgroup$ – mpiktas Feb 13 '14 at 14:57
  • $\begingroup$ I'm looking at only 20 seconds so far for building the model using two weeks of half-hourly data. This is on a generated time series with daily and weekly patterns, with added normal noise. This is acceptable but I'm guessing I will need a longer training period or a higher frequency to obtain good results with real data. I will update this once I have established the proper scenarios. $\endgroup$ – wrousseau Feb 17 '14 at 11:48

Your Answer

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

Browse other questions tagged or ask your own question.