Demand Estimation Under Uncertain Consideration Sets
研究了考虑-选择(CTC)模型的统计性质，提出EM估计方法，并在合成数据和两个真实数据集上测试，发现CTC模型在训练数据有噪声或训练与测试数据不对称时预测精度优于经典随机效用模型。
In “Demand Estimation Under Uncertain Consideration Sets,” Jagabathula, Mitrofanov, and Vulcano investigate statistical properties of the consider-then-choose (CTC) models, which gained recent attention in the operations literature as an alternative to the classical random utility (RUM) models. The general class of CTC models is defined by a general joint distribution over ranking lists and consideration sets. Starting from the important result that the CTC and RUM classes are equivalent in terms of explanatory power, the authors characterize conditions under which CTC models become identified. Then, they propose expectation-maximization (EM) methods to solve the related estimation problem for different subclasses of CTC models, building from the provably convergent outer-approximation algorithm. Finally, subclasses of CTC models are tested on a synthetic data set and on two real data sets: one from a grocery chain and one from a peer-to-peer (P2P) car sharing platform. The results are consistent in assessing that CTC models tend to dominate RUM models with respect to prediction accuracy when the training data are noisy (i.e., transaction records do not necessarily reflect the physical inventory records) and when there is significant asymmetry between the training data set and the testing data set. These conditions are naturally verified in P2P sharing platforms and in retailers working on long-term forecasts (e.g., semester long) or geographical aggregate forecasts (e.g., forecasts at the distribution center level).