Subscribe to DSC Newsletter

I have never been formally trained on how to deal with seasonality. But I wanted to take a moment to share my perspective based on experience, which I hope readers will find fairly straightforward. Some people use sales revenues in order to evaluate seasonal differences. I find it more desirable to analyze units sold if possible. A price increase resulting in slightly higher revenues does not in itself represent increased demand. Nor should discounted prices leading to reduced revenues necessarily be regarded as reduced demand. Below I present some fictitious data. I offer this data as a controlled sample. I would expect an objective person to say, "This product is selling poorly." Although the data is clearly fabricated, I will make some cosmetic adjustments in a moment to obscure the fine details.

To make the data more realistic, I applied a 50 percent random fluctuation across the entire sample. Based on my experience, I would say that the resulting pattern shown below remains unrealistically coherent. In real life, seasonality is literally "seasonal": there are distinct periods of increase and decline. However, the absence of these periods does not negate the usefulness of the general approach presented here.

Although it is theoretically possible to evaluate daily data on a daily basis, I consider it more likely for an organization to assess change on a weekly or monthly basis. Nonetheless, it is not the periodicity itself that is important but rather the need to filter or even out fluctuations. This evening out can occur by using weekly or monthly totals; or one can be daring and use moving averages. (These options all seem reasonable to me. A bit depends on the sort of visibility that the organization hopes to gain.) In the next image, I use to 5-day moving average, which of course is similar to using 5-day totals divided by 5.

Up to this point, I haven't actually addressed the issue of seasonality. What I suggest in order to evaluate seasonality is to use a moving ratio of the current period over the previous period: e.g. June 1 to June 5 of this year compared to June 1 to June 5 of the last year. Using this technique, if randomness is evident, over time the distribution would be back and forth around 1. On the other hand, if seasonality is evident but there has been little change, the distribution would be around 1. In the original "Bogus Baseline Sales Data" illustration, it seems like sales are headed for the gutter. However, according to the moving ratio, a slight recovery is indicated near the end of the sampling period as shown below.

There is some interest in prediction in data science. A projection from the underlying sales data is possible ignoring seasonal shifts. However, a reasonable seasonal projection can be obtained using a segment of the moving ratio. Extract the equation, and apply it to past data after extending.

If exactness is highly desirable, or the consequences associated with inexactness are extreme, then huge fluctuations in the ratio would be problematic. I suggest that the best approach in this case would be to apply a pattern recognition technique from an adaptive algorithm to determine the "best bet." It remains a bet, though. I personally avoid gambling in high-risk scenarios.

Seasonality Challenge

Seasonality represents an important "concept" that should always be considered among data scientists. Simply saying that sales demonstrate seasonal shifts and then not examining the reasons indicates either ambivalence or apathy - possibly absence of motivation - often lack of resources. Are sales literally affected by "seasons": changes in outdoor temperature, lighting, and precipitation? How does one know if discounts, promotions, and loyalty incentives are working? Sales represent outcomes - not antecedents. Consequently, every opportunity should be taken to develop an understanding of contributing factors. In a laboratory using the "scientific method," one would attempt to control potential determinants; this is a mostly deductive process. If the antecedents are uncontrolled and the business setting is quite complex, discovery becomes an issue of induction and inductive methodologies. (This does not negate the need for scientific confirmation if at all feasible.)

On the assumption of a closed system (no market expansion), clear improvements in seasonal sales might mean that competitors are struggling. It could also mean that existing consumers want more of the product. If the system is open (still expanding) and sales are increasing, presumably demand might be related to the larger market - because there are more people although not necessarily because all persons wants more. (If two screwdrivers are purchased, it might be because one person likes screwdrivers enough to buy two; or there might two people that each want a screwdriver. I'm uncertain if that makes sense. It's for a different blog to cover.) Suffice it to say that it is possible to monitor seasonality without understanding any underlying antecedents. To monitor something because it is important and yet make little effort to understand it is illogical: ensuring the persistence of the latter delegitimizes the need for the former.

Views: 2489

Comment

You need to be a member of Data Science Central to add comments!

Join Data Science Central

Follow Us

Videos

  • Add Videos
  • View All

Resources

© 2017   Data Science Central   Powered by

Badges  |  Report an Issue  |  Privacy Policy  |  Terms of Service