Topics: Coronavirus | AI | Data Science | Deep Learning | Machine Learning | Python | R | Statistics

KDnuggets Home » News » 2020 » Mar » Opinions » Forecasting Stories: Is it seasonality or not? ( 20:n11 )

Forecasting Stories: Is it seasonality or not?


Kicking off with a series of forecasting stories, starting with seasonality and its business applications. This first article speaks of course corrections that were based on weather and calendar driven seasonality.



By Rajneet Kaur, Data Scientist

We all love stories, don't we? The only difference is that I like telling them too.

And hence I am going to share a series of business cases from my 5 years of forecasting experience, but as stories. We will be focused on applications, which become a price of cake, when we understand the role of each component of time-series. The first two of the series will be based on 'Seasonality', my favorite component, almost an obsession. So let's get started...

 

CASE 1: ISN'T IT OR IS IT SEASONALITY?

 
Let me take you to Aug 2014.

I am handling Netherlands country engagement for a top pharmaceutical client’s highest revenue(70%) generating drugs. Lets call them Camicade and Dimponi(name changed).

Sales for Camicade and Dimponi

Their approximate units sales are present in the graph.

As a young enthusiastic forecasting person, I notice that every 3rd month peaks – something that looks like seasonality.

Tom, the demand planner happily accepts the suggestion that we explicitly include this in our forecast.

The next peak is to happen in Oct 2014 as per this observation.

What happens next???

Now lets me take you to Sep 14. Both drugs peak! What is amiss?

No alt text provided for this image

The Answer Lies Here...

Yes, The Calendar

Now, For the Solution...

No alt text provided for this image

As Tom later explained, wholesalers placed orders on Mondays, and shipments occurred on Tuesdays. Months with 5 Tuesdays, would naturally peak!

This in my previous forecasting world, was called the famous 4-4-5 effect. Roughly every 3rd month had 5 order/shipment weeks.

LESSON LEARNT:

If the granularity of forecasting period(day/week/month/year) does not match the granularity at which business transactions occur(order/shipment/others), we might observe peaks, which is not actually seasonality!

Seasonality, is thus not just a correlation coefficient or an ACF number, there must be business logic behind seasonality!

 

CASE 2: WEATHER SEASONALITY IS FUNNY!

 
Herius(name changed) is a pharmaceutical drug, used to treat allergic rhinitis, and allergy due to high amount of pollen grain density in the air. Since the pollen season peaks in Oct-Nov, so does the sales of Herius

During a study in which we, we a team of 2 had to ascertain whether about 70 drugs had seasonality, and if so, what the seasonality period is, we found an astonishing fact. Over a few years, the seasonality period of Herius had reduced from 12(months) to 11(months)!!! This can be seen in the ACF plot in the diagram.

ACF Plot of seasonality

After all, this is weather seasonality and not calendar seasonality we are talking about. We also showed this to the country demand planner and he was both surprised and thankful! In the above case, the production, SCM and sales team could plan to be prepared for higher demand for Herius in October instead of November.

Seasonality can be tricky, even though by definition it should repeat after a fixed period.

However, of the 2 types of seasonality, weather seasonality can change course as the weather itself does. Should the 2nd type, calendar driven seasonality be more consistent?

Not necessarily. For instance Good Friday was on 30 Mar 2018 but 19 Apr 2019, hence both month and week being different.

Leveraging seasonality correctly is critical for a low error in Predictive analytics. It can also help drive effective business decisions and course corrections, as we will see in the next article. Would really like to hear back on your thoughts on the article.

In case you are interested in ways to structure projects in Analytics space, you could visit my article on Data Science Problem Solving.

 
Bio: Rajneet Kaur is a passionate and result-driven marketing and data science executive with 6 years of business and technology experience.

Original. Reposted with permission.

Related:


Sign Up

By subscribing you accept KDnuggets Privacy Policy