Technical Overview


Consumption behavior patterns in brick and mortar stores often result from local factors that are not always available as structured dataset features. For example, the presence of a school near a store can affect one subset of products (e.g., products that children prefer), while a local product’s promotion by a nearby competitor can affect another subset of products (e.g., complementary or competing products). Accordingly, a store can share consumption patterns related to specific subsets of products with specific stores that are affected by similar hidden factors, even if these stores are different in size, format or region and do not necessarily belong to the same stores’ segment. Trying to explain such complex consumption behavior is extremely time consuming and practically impossible when many of these external and internal factors are hidden and dynamic. At the same time, with the right technology, the actual consumption behavior patterns that are reflected in point-of-sales data can be extracted and used to trigger granular and actionable recommendation.

CB4’s technology is based on proprietary data-compression algorithms that can automatically capture the local consumption patterns at each store, regardless of their affecting factors. Detected patterns are not bound by assumptions on how stores are segmented or clustered, or how various internal and external factors affect the consumption behavior locally.

CB4’s algorithms automatically detect similar consumption patterns among products and stores, regardless of whether those products were purchased together by the same customer (‘basket analysis’) in specific stores. The algorithms automatically identify the related ‘fuzzy-clustered’ patterns and define an exact sales benchmark for target products at each store. It then generates recommendations based on anomalies and unmet opportunities that are detected for specific products in specific stores.

Analysis Granularity Granular analysis applies to products in stores, generating recommendations at a SKU-in-store level Conventional statistical analysis often applies to aggregated data, e.g., category / brand / chain
Insights Descriptive and self-explanatory, presented in natural language Often a Black Box solution, difficult to interpret
Data Set Only 4 weeks of history required for analysis Usually 24-36 months of history required for analysis
Data Type Only sales data is needed, other data, such as inventory promotion and pricing can be easily integrated External data required, relying on time consuming IT projects. Also often require sensors/HW installation at stores
Model Automated machine learning and AI models with no need for continuous maintenance or upkeep Continuous investment and maintenance by data science and analysts
Dimensionality Multiple SKUs that can be analyzed in parallel Often one dimensional – each SKU is analyzed on its own

Algorithm Highlights

  • Applies advanced data enrichment and preprocessing methods for effortless automated handling of persistent databases.
  • Automatically generates and adaptively refines the analytic models without requiring data scientists to consistently update and validate the models
  • Data-agnostic, parameter-free algorithms that do not assume a conventional time-series model nor simplified auto-correlated processes.
  • Machine learning & AI algorithms that are not regression-based and can discover highly non-linear, conditional and state-dependent patterns.
  • Generates in-depth and accurate insights and recommendations at unparalleled speed and precision.
  • Unique modeling at the finest granular level enables a dynamic data drill-down to more detailed levels and a scalable, decentralized approach.