【公开学术报告】Shall We Only Store Popular Products?

Shall We Only Store Popular Products? Warehouse Assortment Selection for E-Companies

Guest Speaker:     LIN Hongyuan (NUS)

Date & Time:        9:30-11:00 (Beijing Time), Tuesday 7th, Nov. 2023

Zoom Meeting:     91718585329（Password: 594265）

Click the Link:       https://zoom.us/j/91718585329

Problem definition: This paper studies the single-warehouse assortment selection problem that aims to minimize the order fulfillment cost under the cardinality constraint. We propose two fulfillment-related cost functions corresponding to spillover fulfillment and order-splitting, respectively. This problem includes the fill rate maximization problem as a special case. We show that although the objective function is submodular for a broad class of cost functions, the fill rate maximization problem with the largest order size being two is N Phard. Methodology: To make the problem tractable to solve, we formulate the general warehouse assortment problem under the two types of cost functions as mixed integer linear programs (MILPs). Furthermore, we propose a simple heuristic called the marginal choice indexing (MCI) policy that allows the warehouse to store the most popular products. This policy is easy to compute and hence is scalable to large-size problems. Although the performance of MCI can be arbitrarily bad in some extreme scenarios, we find a general condition under which it is optimal. This condition is satisfied by many classic discrete choice models and multi-purchase choice models. We also demonstrate by synthetic experiments that the average performance of the MCI policy is good, and more importantly, it is robust when the observed demand is noisy. Managerial implications: Through extensive numerical experiments on a real-world dataset from RiRiShun Logistics, we find that the MCI policy is surprisingly near-optimal in all the settings we tested. Simply applying the MCI policy, the fill rate is estimated to improve by 9.18% on average compared to the current practice for the local transfer centers (LTCs) on the training data set. More surprisingly, the MCI policy outperforms the MILP optimal solution in 14 out of 25 cases on the test data set, reiterating its robustness against the noise of demand estimation.

Key words: warehouse assortment selection, demand choice models, submodular functions, marginal choice probability.