Estimating Hierarchical Demand in Retail Operations
HASTERAD method automatically determines product hierarchies.
1Need for Hierarchical Structures in Retail
Hierarchical structures based on product attributes are crucial for inventory and assortment optimization in retail. Traditional methods fail to automatically extract these structures from aggregate sales data. HASTERAD aims to fill this gap by using standard SKU-store-week scanner data. It analyzes price and promotion variations to identify substitution patterns between products.
2How HASTERAD Algorithm Works
HASTERAD recursively partitions the product space by comparing attribute-based substitution patterns. When tested on the IRI Marketing Data Set, it produced a stable and economically interpretable hierarchy. The algorithm delivers effective results even in multi-store panels where stockouts are unobservable. For single-store data, performance remains strong despite sparse data because stockouts are observable.
3Test Results on Synthetic Data
A three-step optimization process was implemented to measure HASTERAD's economic value. On synthetic data, the algorithm extracted a hierarchy, estimated demand parameters on this hierarchy, and optimized assortments. Results showed that HASTERAD's chosen assortments missed only 11% of the profit achievable by a perfect planner, even in challenging scenarios where traditional flat demand models missed nearly 100%.
4Importance of Model Selection
The key finding highlights the critical role of model selection. A simple hierarchy with the best in-sample fit explained observed sales as well as HASTERAD's hierarchy. However, assortment optimizations based on this hierarchy yielded significantly lower expected profits. This occurs because fit evaluates observed purchases, while optimization depends on predicting demand for unobserved assortments.
5Multi-Store vs Single-Store Scenarios
HASTERAD maintains strong performance even in multi-store panels where stockouts are unobservable. For single-store data, the algorithm remains effective despite sparse data because stockouts are observable. These findings are robust across various alternative specifications. The algorithm's adaptability to different data structures promises wide applications in retail.
6Limitations of Traditional Models
Traditional flat demand models can lead to nearly 100% profit loss in challenging scenarios. They fail to capture hierarchical relationships and substitution dynamics between products adequately. HASTERAD addresses this gap, enabling retailers to make smarter inventory and assortment decisions. This advantage becomes more pronounced in dynamic retail environments with frequent promotions and price changes.
7Economic Interpretability
HASTERAD's hierarchies are not only statistically robust but also economically interpretable. This makes it easier for retail managers to understand and incorporate the algorithm's outputs into decision-making. For example, substitution patterns, price elasticities, and promotion effects within specific product categories become clearly visible, enabling more informed strategic planning.
8Implications for Retail Industry
HASTERAD's success has the potential to transform data-driven decision-making in retail operations. The algorithm could significantly improve inventory management, assortment optimization, and promotion strategies for large retailers, becoming a key competitive differentiator. Future integration of HASTERAD-like methods into retail software is anticipated.