All -
I am working on a forecasting model for a time series that does not have a seasonality to it. In other areas I have used ARIMA for those with single seasons, TBATS for those with multiple seasons. However, now, I am working with a time series that follows no seasonality. I am hoping to do some basic forecasting. I am dealing with about 1.5 years worth of data - taken daily. The plot of the time series looks as follows (I cannot post images yet, due to my status on StackOverflow, but this is a link to the image:
http://107.170.210.195/Rplot.png
Now, given that this is not seasonal, I build my time series this way:
incidentBacklogRolling12DailyTS <- ts(incidentBacklogRolling12Daily$Count, start=c(2014,327), frequency=365.25)
This may or may not be the correct way of doing this, but I am following my intuition here... Then I have tried various forecasting methods - obviously ARIMA and TBATS are out. However I tried ETS (which seems promising) and STL. Both results in sort of a "straight line" extending from the last point. Now, I know my time series does seem flat at the end, but there is a bit of fluctuation. I also experimented with HoltWinters, which produced an extremely accurate fit, but the forecast was, again, a straight line, this time sloping downwards. I was hoping there might be a good model out there for a non-seasonal series such as this one.
Does anyone have a suggestion for a model (and possibly paramters for that model) to try. Maybe what I am using is trying to hard - maybe there is a simple solution for this I am not seeing. Thanks for any suggestions!
In researching this over the past week or so, I wanted to post my solution that I arrived at for anyone looking into this in the future. I found some material posted by Rob Hyndman (researcher in this area from Melbourne). I found that in one of his lectures he recommends using the auto.arima() function for items such as these. I sent. Dr. Hyndman a note and he suggested the use of ETS for this. Being that there is no detectable, underlying pattern or seasonality, you'll want to forecast off of the most recent data, which you can do with methods like AIRMA and ETS. I tried them both out and was able to get acceptable results based on the data. Thanks for all who read this, like I said, I just wanted to provide my solution for any others that might be having the same issue.