I have a problem consists of forecasting the next 16 values for 250 time series of daily demands, Can I forecast just for one period ahead and then multiple it by 16? is it correct? I was wondering if you give me your suggestions.
$\begingroup$
$\endgroup$
4
-
$\begingroup$ The answer to your question is "No", unless your daily demands are all the same. IrishStat, below, lists a bunch of things that can affect time series. Also, when you say, "Can I forecast just for one period", what kind of method are you using to forecast? $\endgroup$– WayneCommented Jan 24, 2012 at 18:37
-
$\begingroup$ thank you so much for your answer, for example SES(single exponential smoothing) $\endgroup$– RojiCommented Jan 24, 2012 at 18:40
-
$\begingroup$ With an SES, I believe the prediction, no matter how many steps out, is a straight line, so your initial suggestion works. (Though, the prediction interval grows, so I wonder if the prediction is actually meaningful 16 values out.) I'd suggest that you need to consider the factors that Irish mentions and switch to at least double exponential (includes a trend) or triple (includes trend and seasonality, also known as Holt-Winters). Then there's ARIMA, State Space methods, ... $\endgroup$– WayneCommented Jan 24, 2012 at 19:00
-
$\begingroup$ :wayne I would never recommend assuming any particular class of an ARIMA model whose family includes all varieties of HW excepting the multiplicative variety but rather one should allow the data to speak to the model identification process. $\endgroup$– IrishStatCommented Jan 25, 2012 at 14:05
Add a comment
|
1 Answer
$\begingroup$
$\endgroup$
1
Only if your model is a random walk or a simple mean model ,otherwise you will have to forecast out 16 periods for each of your 250 time series. You might want to take into account some factors like 1) day-of-the-week-effects 2 ) auto-projective structure i.e. the impact of previous values on the forecast 3) the impact of events like holidays such as lead,contempraneous and lag effects 4) possible changes in the model parameters over time 5) possible changes in the variance of the errors over time 6) level shifts in your time series 7) local time trends in your time series 8) the impact of unusual values ( one time anomalies ) . If you deal corrrectly with these eight considerations you forecast might be useful.
-
$\begingroup$ thank you , that sounds interesting! I appreciate your help $\endgroup$– RojiCommented Jan 24, 2012 at 18:41