In this thesis, we study two data-driven dynamic pricing problems and present the results in two essays. In the first essay, we study the problem of personalized price promotions in a market for storable goods. Rich data containing fine-grained information on consumer behavior allows retailers to identify heterogeneous consumer preferences and time-varying latent states, such as inventory, in markets for storable goods. As a result, retailers often resort to personalized dynamic price promotions to maximize sales. However, the traditional estimate-then-optimize approach may not fully exploit such rich data due to well-known problems such as model misspecification and variable estimation error. Instead, we develop a scalable offline reinforcement learning approach to derive promotion strategies directly from the data. Specifically, based on business insights and domain knowledge, we propose an appropriate approximation to the state-action value function and learn it directly from the data. Synthetic and natural data experiments show that our approach outperforms the conventional method. The result shows that our approach is more robust to the problems that have plagued the conventional two-stage approach. Meanwhile, our approach can reap most of the potential benefits of personalization. Our work provides an alternative path to data-driven personalized advertising by learning the strategies directly from the data, thereby mitigating the drawbacks of the conventional two-stage approach.

In the second essay considers a dynamic pricing problem for multiple room types in the hospitality industry. Based on empirical evidence from actual hotel data, we find that customers' booking behavior exhibits marked heterogeneity across dates of arrival (DoA). Thus, despite the abundance of historical booking data, data on the target DoA is relatively scarce. Building on the approach outlined in the previous essay, we present two transfer-offline reinforcement learning approaches aimed at exploiting additional information from other DoA data to improve the effectiveness of the dynamic pricing strategy for the target DoA. The experimental results show that both approaches pay off when transferring information from other DoAs. The corresponding pricing strategy generates about 15% additional revenue compared to learning only from sparse data. Moreover, the first approach proves to be more effective when accessing sparse target DoA data, while the second approach is required when encountering a completely new target DoA. Our proposed approach addresses the performance degradation of the data-driven approach when the amount of data is insufficient and provides valuable insights into the dynamic pricing problem of room types in the hospitality industry.