Satlr - Thanks for following up - glad you deemed my earlier response useful. You raised a couple of good questions - let me try to tackle them one by one: "This was a question raised by one of my customer during demo. They expected that the forecasting line would have ups and downs exactly how the business would be in future." Yes, I understand that desire - unfortunately (or fortunately depending how you look at it) the future remains uncertain. All a statistical model can do is the suggest the most likely behavior based on the pattern in your data - the ups and downs are caused by random variation most likely. Forecasting is not about curve fitting - in fact it has been shown many times that models which fit the past extremely well, are typically poor in predictive the future. Some people refer to this as overfitting the past. Data = historic pattern + random variation - don't try to model randomness. "Understanding variation" is crucial - I'd like to refer you to this blog: Mike Gilliland - The Business Forecasting Deal where you will find lots of useful information about what can be accomplished by statistical forecasting. "But the forecast points in negative which concerns me." Remember that the statistical model does not know anything about the business context - it "only" looks at the data and tells you what is most likely going to happen. The fact that sales cannot go negative is something you know as a user. Statistical forecasting does not answer the question of "what should happen" - instead it focuses on the question of "what will happen most likely - based on past behavior"). I agree that it might be useful to have an option in VA which forces the forecast to be non-negative. In SAS Forecast Server you can specify such setting - but it is a post-processing activity. First you model in a unconstraint fashion - then your forecasts will be constrained by rules like: non-negative numbers only. "So I filtered Jan to see whether it gives me correct forecast and seems to be fine. But could you suggest me whether this approach is correct. The data point (Jan 2014) has 10 million USD which got omitted after which I am getting a smooth increasing trend. But is it good to do this way by omitting such a huge value data point by considering that as an anomaly?" This is a tricky question to answer without seeing the data at hand. Some general thoughts: modifying past values is usually not a good practice, after all this is what actually happened, right? However, you may want to clean data from extreme events - in particular if you know what caused them. Maybe some one-time sales event caused an extraordinary peek. Forecasting tries to predict "ordinary" behavior. Actually, what you would like to do is to flag these extreme values either as outliers or special events and model them using more complex techniques such as ARIMAX or UCM. Note that these are not trivial tasks. This is beyond the purpose of Visual Analytics in my honest opinion - for such advanced modeling questions you may want to have a look at specialized environments, such as SAS Forecast Server for example. Thanks, Udo
... View more
