Forecasting CPG Sales on the Amazon Channel Behemoth

Consumer Goods industry is still adapting to the new rules of the game that are being set by Amazon and other e-commerce players. CPG companies' online channel is rapidly growing with 43% of CPG’s revenue growth already being driven by ecommerce and online sales expected to double in next five years.

One of the biggest challenges that the industry faces today is the poor forecast accuracy with respect to Amazon’s purchase orders. Leading players operate at less than 50% forecast accuracy for this channel.

Amazon’s own business practices complicate the matter.  Although CPG companies identify their products by UPCs, products offered in Amazon are identified by Amazon Standard Identification Numbers (ASINs) which are 10 alphanumerical characters unique to each Amazon item page in Amazon.

Since Amazon is an open marketplace, each item (i.e. ASIN) could be offered by different sellers at different price points and it is very likely that Amazon itself is one of the sellers amongst other third-party sellers.

While there might be multiple sellers for each ASIN, at each customer visit only one of the sellers' offer is linked to the buy button, or in Amazon terminology, only one seller shows in the buy box, and other sellers of the same ASIN will lose buy box on that visit.

Amazon uses an undisclosed classification for determining which seller can be in the buy box. However, there are three main factors that can determine which seller can be in the buy box:

  • The current offered price of an item
  • The stock availability at the seller
  • The seller reputation and competency: sellers that are selling items as Fulfilled By Amazon (FBA) are preferred to Merchant Fulfilled Network (MFN).

To reduce operating costs, Amazon doesn’t always keep a large inventory volume. This further complicates the matter as the reordering interval is short. For fast-moving items, Amazon could be submitting a purchase order twice a week, adding serious challenges to suppliers.

Unlike brick and mortar retailers, Amazon has some unique data sets that they share with their suppliers for each product on a weekly basis, including:

  • Glance views: which represents how many times each ASIN page has been viewed in a week
  • Unique visitors: the number of unique visitors who viewed each ASIN page in a week
  • Total number of customer reviews per ASIN updated weekly
  • The average customer review ranking per ASIN updated weekly
  • The ranking of the ASIN page unit sale and sales amount compared with other ASINs within the same Amazon category or subcategory
  • LBB (Lost Buy Box): the number of times that Amazon has lost the buy box to a third-party seller because of the pricing
  • Rep OOS: The Number of times that Amazon has lost the buy box because a replenishable item has been out of stock
  • Sellable and unsellable on-hand units: sellable and unsellable stock per ASIN per week

These above-mentioned fields can be used to quantify demand drivers and constraints on historical sales. For instance, the number of glance views can reflect the sponsoring of some items during the paid promotion programs to bring the item on top of search lists. The changes in sub-category sales ranking can reveal interesting dynamics between the promotion strategies of the ASIN supplier and the sellers of other competing products in the same sub-category. Another rather unique element to Amazon measurement is LBB which can be used to estimate third-party pricing strategies as a valuable factor in promotion planning.

For these reasons many CPG companies struggle to accurately forecast their sales on Amazon. However, AI/ML techniques can be utilized to obtain a high forecast accuracy, and it is very different than doing it for brick & mortar retailers / traditional offline channel. Given that Amazon already accounts for 39% of CPG ecommerce revenue, it’s critical to get it right and it deserves a new scientific approach that can unlock tremendous value.