The Stability of MNL-Based Demand under Dynamic Customer Substitution and its Algorithmic Implications

Abstract

We study the dynamic assortment planning problem under the widely-utilized Multinomial Logit choice model (MNL). In this single-period assortment optimization and inventory management problem, the retailer jointly decides on an assortment, i.e., a subset of products to be offered, as well as on the inventory levels of these products, aiming to maximize the expected revenue subject to a capacity constraint on the total number of units stocked. The demand process is formed by a stochastic stream of arriving customers, who dynamically substitute between products according to the MNL model. This modeling approach has motivated a growing line of research on joint assortment and inventory optimization, initiated by the seminal papers of Bassok et al. (1999) and Mahajan and van Ryzin (2001). The currently best-known provably-good approximation in the dynamic setting considered, recently devised by Aouad et al. (2018b), leads to an expected revenue of at least 0.139 times the optimum under increasing-failure rate demand distributions, far from being satisfactory in practical revenue management applications. In this paper, we establish novel stochastic inequalities showing that, for any given inventory levels, the expected demand of each offered product is stable under basic algorithmic operations, such as scaling the MNL preference weights and shifting inventory across certain products. By exploiting this newly-gained understanding, we devise the first approximation scheme for dynamic assortment planning under the MNL model, allowing one to efficiently compute inventory levels that approach the optimal expected revenue within any degree of accuracy. Our approximation scheme is employed in extensive computational experiments to concurrently measure the performance of various algorithmic practices proposed in earlier literature. These experiments provide further insights regarding the value of dynamic substitution models, in comparison to simple inventory models that overlook stock-out effects, and shed light on their real-life deployability.