PA – How it should work (Time series analysis)

Dear readers,

Four weeks ago, I documented my first steps in PA here: http://scn.sap.com/community/predictive-analytics/blog/2015/11/23/my-first-steps-in-sap-pa–time-series-analysis
My questions (http://scn.sap.com/message/16376119 ) were all answered, unfortunately I did not have much time since then to continue a more detailed analysis.

One more fact contributed in neglecting that topic: The tool is just not handily enough for my purpose!

I don’t want to spend hours in finding out how to use those KxShell script language, so I got quite annoyed after repeating all the required single steps to predict the share values per single company.

At least, I decided to create second – and for the time being –  last blog about “PA – how it should work”:

Most important thing is: I want to analyze more sets of data (e.g. different companies) at once, uploaded from ONE single csv file!

That means, I would like to see an additional characteristic e.g. between Target and Weight, like “Category” which differentiates the individual data sets.

In my case it would be the WKN (ISIN) which organizes the split of companies to be analyzed:

The idea is that I want to be able to execute the steps like loading the Source data, loading the Description file (which is identical for all sets) and selecting my variables (also the same for all sets) only once.

I guess there are reasons I don’t understand yet, but in my personal opinion, the PA engine should be able to use the different data sets to recognize the overall trend, too. Not necessarily always, but e.g. on demand by another checkbox, for independent or inter-dependent analysis.

Finally of course, the Forecasts should be enhanced by the category, which enables a Filter to select one set after the other and analyze the outcome. E.g.:

In that way it would be possible to check the results for e.g. 10 companies within seconds. As of now repeating all steps over and over again to create the forecast graphs takes ~5 minutes for one. And this each time!

In addition I found two more “bugs”, for which I don’t have an explanation or don’t know the root cause?!

I saved a data model 4 weeks ago, and now wanted to create the chart again, but it looks like this:

Whereas 4 weeks ago it looked like this (the red line above is the green one below):

It seems PA recognizes the current date (Operating system date?) and finds 4 missing weeks from the source data to today and therefore does not work properly anymore. Could that be?

I just loaded the model and displayed the chart. The current version with outliers in every week never ever appeared before.

Then, two more things were strange…

I tried to use the PA test version two days before expiration date… SAP, tell me please what is this?

An information message which exceeds the monitor screen’s height, where you cannot see the bottom or any “OK”-button?

Second strange thing was that I could not load my files anymore.

Every time I tried to browse the Folder or the Data Set the whole program closed without any message. Neither warning nor error message.

I’m just not able to create a new time series analysis again. Only loading my old model still works.

Final summary: Even if – or maybe especially because – I’m from IT department and know how an “ergonomic” software should work, I’m quite disappointed about the features (I used version 2.2).

The quality of my predicted data result seems to vary tremendously. Either there are up to +/- 40% min/max values or sometimes also none.

The MAPE value however seems always good (~ 0,0xx) if I skip those variables which PA identifies as suspicious.

Of course I must admit that my evaluation is solely based on time series analysis, but in the current state I cannot see any added value which justifies the license costs.

Maybe I need to collect daily data for better results. But the effort here is huge, especially when prediction can only be done for one set after another. And I don’t want a day-based forecast but I want results and a trend (always!) for the next 12 to 24 weeks.

Thanks for reading,

Martin

PS: Feel free to share a document/blog post if you know one, about “Creating a KxShell script for dummies” that describes what I would like to have had with my “category” above.