Churchill Systems

 

Churchill White Paper

Demand Forecasting & The Seasonality Factor – Harve C. Light

 

Introduction To Demand Seasonality

Frequently, Churchill is asked to comment about the impact of Seasonality on Demand Forecasting.  This white paper overviews Seasonality, its impact on demand forecasting and the inclusion of Seasonality in demand forecasting methodologies.  Note:  This paper intended to be a basic explanation of the topic and not an academic-type paper.

 

What Is Seasonality?

Within the timeframe of any given period of time such as one year, merchandise (as well as selling locations) experience distinct periods of time, such as a few weeks, when one or more types of merchandise are more (or less) desired by the buying public when compared to an average week.  This repeatable increase or decrease in periodic consumer buying interest is defined as the merchandise or location’s “Seasonality” and therefore must be accounted for when forecasting demand for the affected merchandise or location.  To complicate matters, while the seasonal demand repeats over each year, etc. the actual seasonality periods can vary from year to year. (For example, Holidays)

 

What Is the Importance of Demand When Generating Forecasts?

Depending on the purpose of the forecasting, the importance of the Seasonality Factor can be very significant or very insignificant.  For example, in continuity item regular periods (non-promotional) forecasting, Churchill performance testing has shown Seasonality to be the most significant single source of forecasting error.  Poor Seasonal Indexes equals poor demand forecasts.  On the other hand, there are many examples of merchandise that are required by consumers throughout the year where other factors such as price and promotions are more significant that seasonality.

 

One final comment. Seasonality of demand can be a very “local” factor.  Churchill sees this situation whenever we build forecasting models for convenience or neighborhood stores.  For example, individual stores may serve specific ethnic communities who observe holidays, etc. specific to the community.

 

How Is The Seasonality Factor Expressed In Demand Forecasting Methodologies?

There are many techniques for quantifying the seasonality component for demand forecasting purposes.  One methodology that will demonstrate the principle is to utilize the “Relative Seasonality Index” methodology.  Conceptually this methodology is a two-step methodology where:

• Step One calculates the Average Periodic (weekly) Demand value as = Total Annual Demand / 52 Weekly Periods.
• Step Two calculates the Relative Seasonality Index values as = Weekly Total Demand / Average Periodic (Weekly) Demand.

The output values of this methodology form a relative seasonality index where (quantitatively) the Average Demand Week will have a value = 1.0 and Weeks with historically greater seasonal demand will have index values greater than 1.0 (for example 1.82) and Weeks with historically less seasonal demand will have index values of less than 1.0 (for example .86).

 

What Are Some Examples Of The Seasonality Index Being Utilized In Demand Forecasting?

Here are three examples.

 

Example #1 relates to the forecasting of regular (non-promotional) continuity merchandise demand.  Forecasting methodologies such as exponential smoothing, moving averaging, etc. first determine the underlying (nonseasonal) demand for the merchandise, location, etc. and then add the merchandise’s Seasonal Index values to include Seasonality for specific periods such as weeks, resulting in seasonalized demand forecasts for each week.

 

Example #2 is a causal forecasting example such as the forecasting of promotional price lift.  These forecasts can require the inclusion of the Seasonality Index values as Causal-based Forecasting Input values.  Inclusion of the Seasonality Factor values can be important because Seasonality can cause another causal factor, such as Price Discount on Holiday merchandise, to be more or less significant e.g. more intense before the Holiday periods (resulting in higher lifts) when compared to the exact same Price Discount promoted four months after the Holiday periods which generally result in lower promotional lift.

 

Example #3 is a limited life product forecasting example.  Seasonality can be an important factor during the limited lift season that the merchandise is offered for sale.  For example, Swimsuits tend to be more in demand during the early and warmer weeks of the season and less in demand during the later and perhaps colder weeks of the season.

 

How Does Churchill Generate Specific and Detailed Seasonal Indexes Values?

First, I have to mention that Churchill expends substantial energy effort developing the best Seasonal Indexes possible within the limitations of the customer’s historical sales data.  Second, Churchill utilizes a combination of human engineered data preparation combined with our proprietary clustering and profiling technology.  Our seasonality modeling tool is called Profile Cluster Builder™ (PCB).  An executive summary of our PCB tool is included as the last page of this paper.

 

In general terms, Churchill models Seasonal Indexes utilizing the following methodology.  As you read the steps of this methodology, please remember that, in most cases, we are building literally many thousands (or more) of seasonal indexes for the retailer.

1. Churchill starts by carefully defining the Merchandise Level and Location Level forecasting requirements that the Seasonal Indexes are to be built to support.
2. Since historical data accuracy generally limits the retailer’s historical data usability, Churchill must audit the historical data for the purpose of improving data quality. Over the years, Churchill has developed several data cleansing techniques that help with this data issue.
3. Next, since holidays and other seasonal events often change from year to year, the next step in this process is to organize the historical data from multiple historical years into one common format that represents Seasonality history properly aligned for the upcoming year to be forecasted.  This step not only includes the specific period (week) of the holiday, etc. but also the affected weeks leading up to and following the event period.
4. With the historical data aligned as best as possible, Churchill now turns to its Profile Cluster Builder tool to perform its “magic”. Note: While the following tasks are presented in a sequential format, in reality there is much overlap in these steps.
   1. Gathers the historical data from similar merchandise/locations into common groupings. Churchill utilizes degree of fit type metrics for the grouping metric.
   2. Isolates and separates individual merchandise and locations whose historical data patterns are considered to be too different to be utilized in profile building.
   3. From the qualified historical data, PCB builds demand profiles that will represent the clustered groupings.
   4. And this brings us back to the identified outliers. PCB, using a two-step process, places the outliers with the already built profiles that statistically are most appropriate.
5. After the PCB clusters and profiles have been built and tested, a series of “higher level” default profiles will need to be built and tested. These higher level default profiles will be utilized for forecasting seasonality in future new merchandise and/or locations that were not part of the original profile modeling process.
6. We now have a collection of seasonal profiles and related cross reference tables that can be utilized for application-specific demand forecasting. Before these Seasonal Profiles can be shipped to the retailer, two additional things must occur.
   1. The Seasonal Profiles need to be performance tested for the purpose at hand.
   2. Sometimes, the format of the Seasonal Profiles must be transformed to align with the retailer’s installed demand forecasting application.

In Closing

Because Churchill recognizes that Seasonality and the related Seasonal Indexes are a very important component in retail demand forecasting performance and conversely that lack of good Seasonal Indexes will be a major source of forecasting performance error, we take Seasonal Profiling very seriously.  One final comment.  Year to year, seasonality constantly changes.  For example, product assortments evolve, selling locations open and close, holiday dates change, leap year occurs every four years and retail calendars require adjustment.  Therefore, to achieve and then maintain high quality forecasts, please recognize that Seasonal Indexes require regular annual updating.