topic Segment Univariate Time Series Data in SAS Data Science

Segment Univariate Time Series Data

Mumbai_1983 — Sat, 08 Oct 2016 11:01:54 GMT

Hi,
I have sales data for 30 stores on a daily basis. The store stocks only one type of item and there is a seasonal pattern to sales. The business would like to understand if certain periods should be grouped together (day 1 - 60, 61 - 180, etc) so they can manage demand for products accordingly. They currently derive the day groups visually (around 5-6 groups for each store) and since they have only 30 stores, it is easily managable to visually inspect sales data.

However, they would like to understand if a data driven approach can be used to group days together.

Grateful for any suggestions on how to approach this problem.

Store ID	Day No	Sales
1	1	125
1	2	140
1	3	166
…	…	…
…	…	…
…	…	…
1	363	300
1	364	321
1	365	340

Univariate Time Series Segmentation

Mumbai_1983 — Sat, 08 Oct 2016 11:02:46 GMT

However, they would like to understand if a data driven approach can be used to group days together.

Grateful for any suggestions on how to approach this problem.

Store ID	Day No	Sales
1	1	125
1	2	140
1	3	166
…	…	…
…	…	…
…	…	…
1	363	300
1	364	321
1	365	340

Re: Univariate Time Series Segmentation

PGStats — Sun, 09 Oct 2016 04:16:39 GMT

You could try to find optimal date ranges with HPSPLIT. For 5 periods of at least 10 days, you would use:

proc hpsplit data=myStoreData
	leafsize=10 maxbranch=5;
input date / level=int;
target sales / level=int;
output nodestats=myStoreDataSplit;
run;

The procedure will try to minimize the variance of sales within each period.

Re: Univariate Time Series Segmentation

Mumbai_1983 — Mon, 10 Oct 2016 08:47:52 GMT

Thanks for your response, I have been reading about structural breaks in time series data too.

If my time series data had features, how would I go about doing this. Say for each day, I had other attributes like sale indicator, customer footfall in the store, web visits, etc, how would I then approach this problem. I was considering clustering but this could end up grouping days wierdly (e.g. days 1, 40, 80 and 90 are in Cluster 1, etc).

Is there an approach to structural breaks in time series with multiple attributes? Would love to try an open source implementation for this.

Grateful for suggestions.