Handling Unobserved Substitutions among Alternatives through Choice Models: Applications for Assortment Planning and Optimization in Retail

Assortment planning helps retailers determine what products to stock and how much of each product to stock so that the retailers’ objectives related to profits and sales are met. Empirical evidence suggests that good assortment planning is of paramount importance to retailers. Significant financial benefits attributable to improved assortment plans have been reported in extant literature. However, there still is a gap between research and practice and a lot of opportunities exist for improving the retailer assortments, especially in estimating demand and substitutions between products. The focus of my dissertation research is on building a good product level demand forecasting and substitution probability estimation approach that is applicable on large industry scale problems, and then, combining the estimates with resource and service level constraints to optimize a retailer’s assortment. Assortment planning relies heavily on estimates of the future demand of individual products and substitution probabilities (The substitution probability between two products a and b is defined as the probability that a customer arriving with the intention of purchasing product a ends up purchasing product b if product a were not available). Stock-outs and the resultant customer substitution behavior introduce a set of challenges to both product level demand forecasting and the estimation of substitution rates, a topic that is often discussed but not well addressed in the literature. As a part of my finished research, I have developed the Reference Point Logit (RPL) model and the True Demand Estimator (TDE) that address the dual issues of estimation of substitution rates and demand estimation when there are unobserved substitutions among choice alternatives and I show the efficiency and effectiveness of the proposed model in this dissertation. Enhancements in the solution approaches that make it applicable to a wider range of problems have also been presented. Next, given the estimates of the true demand and the substitution probabilities between products, I have developed a mixed integer programming based optimization model that incorporates capacity and quality of service constraints in decision making. I focus on a single time period static assortment optimization problem in which both deterministic and stochastic demand are considered. I compare the results obtained by my estimation and optimization models to various methods proposed in academia and those used in the industry, and showcase the importance of (1) accuracy of the product level demand forecasts; (2) using the right substitution probability; (3) incorporation of quality of service related goals and resource constraints in assortment optimization. The case of assortment optimization in the salty snack category has been considered throughout this dissertation for illustration