Newsvendor Decisions with Two-Sided Learning

A substantial amount of work in operations management considers settings where a firm faces uncertain demand that depends on a parameter that is ex ante unknown, but can be learned by observing historical data. This work typically assumes that future demand is unaffected by the firm's learning process. However, in the new era of social media, it is increasingly the case that the information used by the firm to gauge future demand (e.g., sales and stockouts, consumer feedback) is now also observable to the consumers and may influence their purchase decisions. In this paper, we consider a newsvendor model where a product of ex ante unknown value is sold in an environment where learning is "two-sided," in that both the firm and the consumers learn the product's value over time by observing the same information. The analysis establishes a fundamental and consequential insight: when learning is two-sided, the value of information is often negative for the firm. As a result, we show that the firm should stock less in environments with two-sided learning as compared to settings with one-sided learning. Moreover, under certain conditions we identify, we show that the optimal quantity can be even lower than the critical-fractile policy, in stark contrast to the recurring prescription found in existing literature of "stocking more" in the presence of learning. Further results and numerical experiments suggest that the loss for the firm from failing to account for the two-sidedness of the learning process can be significant