Can you advise some simple, robust maybe not perfectly accurate method for the automatic time series forecast. It should consider seasonality and trend and works both for stationary and non stationary data. Some data are seasonal, and some not, the season period is very different. It may not model residual or cyclic patterns. Just some approximation that is pleasant for the user eye and looks like reasonable forcast.

I have tried seasonal ARIMA, ETS, and looks like it not working even reasonable good for general case and I should fine tune it manually for a lot of cases. That looks like a huge of work.

  • 1
    $\begingroup$ Well, triple exponential smoothing/ETS would be my method of choice. It's really the simplest thing you can do while including trend and seasonality. If this is not sufficient, maybe you could edit your question to include an example where it does not work satisfactorily? It may well be that your series simply exhibit a lot of noise and are not very forecastable. $\endgroup$ Jan 26 '17 at 17:45
  • $\begingroup$ What is the horizon you are looking to forecast over? If you are looking "just a few" time-points ahead maybe just carrying over the deseasonalised last time-point is a competitive result. (And yeah, it is dead simple.) $\endgroup$
    – usεr11852
    Jan 26 '17 at 23:54
  • $\begingroup$ The simplest method for time series analysis would be simple moving average and weighted moving average. More info can be found here en.wikipedia.org/wiki/Moving_average $\endgroup$
    – prashanth
    Feb 1 '17 at 17:02

You might want to closely look at Forecasting weekly demand: based on ACF and PACF, is ARIMA appropriate? for some hints as to how to automate time series modelling. Imitation is the sincerest form of flattery thus you might try your hand at imitating AUTOBOX (which I have helped to develop) if you have time/patience available. No one model form works for all time series as you have painfully found out.

The data is the boss and the best or even reasonable model form is individually derived from the data via an iterative process yielding a white noise error term with all model coefficients statistically significant. This is not obtainable directly by closed form procedures. To facilitate quick'n easy your software selection should have macro options bypassing steps.

When buying a suit or getting a pair of glasses no one suit/glass prescription fits all.The suit has to be customized /fit to your physical measurements/data and your prescription to your eyes/data.


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.