Abstract
This paper focuses on establishing a demand forecasting model to optimize product assortments from a set of SKUs in the same category, aiming for revenue maximization. The model considers consumer preferences at the attribute level and the possibility of substitution between different attributes. By decomposing products into specific attributes and multiplying their effects, the model provides a granular framework for prediction. An empirical analysis was conducted using e-commerce sales data for top beverage categories. A hypothetical retail store was modeled to predict sales volumes and total revenue for various SKUs.
Introduction: The Challenge of Assortment Optimization
Retailers face a constant "Tetris" challenge: limited shelf space necessitates strategic decisions on which Stock Keeping Units (SKUs) to carry to maximize profit. Traditional forecasting often treats SKUs in isolation, ignoring two critical realities of consumer behavior:
- Attribute Preference: Consumers choose combinations of brand, flavor, and size rather than just a "product."
- Substitution: If a preferred item is missing, customers often switch to a similar alternative rather than abandoning the purchase.
This paper addresses "category localization"—adjusting inventory to local tastes—while accounting for the complex ripple effects of product substitution.
Methodology: Attribute-Level Decomposition
The proposed demand estimation model relies on attribute-level decomposition. Instead of treating each SKU as a unique entity, it is broken down into constituent parts (e.g., Brand + Taste + Package).
Multiplicative Preference and Joint Attributes
The probability of a consumer preferring a specific SKU is calculated by multiplying the preference ratios of its individual attributes. A key finding is that attributes are rarely independent. For instance, brand preference often shifts depending on the package size. To address this, the researchers combined "Brand" and "Package" into a single joint attribute to improve model accuracy.
Substitution and Purchase Probability
The model calculates the likelihood ($\pi_{ij}$) that a consumer will switch from product $i$ to product $j$ based on attribute similarity. In the beverage case study, the authors assumed the substitution probability for taste was zero; consumers are more likely to switch brands or sizes within a flavor category than to switch flavors entirely. The total purchase probability for SKU $j$ is the sum of its "first-choice" demand and the "substitution" demand redirected from unavailable products.
Empirical Analysis and Results
The model was tested using sales data from Tmall for two major brands, Masterkong and Uni-president, covering 14 tastes and 2 package sizes (56 total SKUs). A hypothetical "Store S" carrying 36 of these SKUs was simulated.
The simulation yielded the following key results:
- Demand Capture: The 36 selected SKUs captured 87.90% of the total category demand.
- Volume and Revenue: The total estimated demand was 2,564 units, resulting in a predicted revenue of 11,176 yuan.
- Granularity: The model successfully generated specific "Predicted Demand" figures for every individual SKU, allowing for precise inventory planning.
Conclusion and Future Work
While the model provides a robust framework for assortment optimization, several challenges remain. Environmental factors like seasonality and promotions can skew preferences. Additionally, the computational complexity increases exponentially with the number of SKUs. Future research should incorporate time variables and Bayesian distributions to account for the dynamic and evolving nature of consumer tastes.
Summary updated