08 Aug Time Series Data Science
As I look over the abstracts of the 50 some planned presentations for the upcoming FARCON conference sponsored by MinneAnalytics, I’m struck by the fact that there will not be much content on the subject Rob Beachy and I have been asked to cover in Axiom’s presentation time slot.
That subject is Time Series data science and as I muse over the uniqueness of our subject matter, I’ve spent some time trying to figure out why Time Series considerations are so often overlooked in traditional business intelligence analytics projects. My personal experience working with retail clients and product marketing goes back some 20+ years and I’ve learned that looking at data as a time series has many benefits.
In my presentation I plan to identify the unique characteristics of time series as they differ from the more easily understood multi-regression and summary statistics of random event driven data sets. At Axiom we have the methods to support, efficiently examine, and cross-correlate the time series which represent the data in retail and financial econometrics.
In the BI community, we are often so focused on making sense of huge amounts of Big Data and encompassing total analysis of each individual transaction that we don’t have time to realize that the summarized data at every point in the hierarchy is often connected in time to other points of each constituent series. Time Series processing of Big Data sets requires summarization in each level of hierarchy. While there are tools available as functions within SAS, STATA, SPSS and R only a handful of companies have the methods and software to deliver cumulative summarization of Big Data events on a time continuum. While this may not seem important, consider that the perfectly constructed event driven attribution model probably operates differently during the pre-Easter period than it does during mid-July. Likewise when Easter shifts from March to April, without considering this shift in a time series model, the attribution model should probably change as well. In the RFM client metrics we should likely associate the summarized Recency summary statistics with traffic surges during holiday seasons.
And another subject that is near and dear to us at Axiom is the use of government data to independently benchmark retail performance against econometric factors like employment, personal income, productivity, consumer expenditures, weather data and other performance benchmarks. As an example we work with specific data series that address variations in annual monthly consumer spending among the 300 some product categories that are covered by two different data series respectively from Bureau of Labor Statistics and the Bureau of Economic Analysis.
Localizing this econometric and weather data is another area of specialization at Axiom. While It’s one thing to blame the weather for poor performance at the national level, too often management loses sight of the fact that the weather is not the same in all areas of the country and in fact may only be affecting sales in a region of the country where they have limited regional sales. For that reason it is important to model and measure weather against sales in each region of the country and subsequently aggregate the summaries to higher levels using hierarchy matrices.
Hope to see you all at FARCON on the Wednesday the 12th.
No Comments