When a company introduces a new product, e.g., via online sales channels, there is often limited information available on the demand. In these situations, adopting a dynamic pric[1]ing strategy proves more advantageous in increasing revenues than setting a static price. While incorporating dynamic pricing, sellers are presented with the challenge of “earning and learning”, which involves finding the right balance between maximizing immediate revenue (earning) and enhancing long-term revenues by acquiring valuable insights from customer behavior (learning). This thesis explores the intersection of two pivotal research fields that constitute dynamic pricing, as suggested in [14]: statistical learning, which is applied to estimate demand function parameters, and price optimization, which seeks to dynamically adjust prices in response to real-time market conditions to maximize (ex[1]pected) revenues. Motivated by the transformative potential of dynamic pricing in various industries, in[1]cluding insurance, this study specifically compares three dynamic pricing models that incorporate demand covariates—variables that might predict demand—to assess their ef[1]fectiveness in a simulated market environment. The models that we study include the linear demand model, which employs Greedy It[1]erated Least Squares (GILS) to estimate the unknown demand model parameters; the binary model, which utilizes Maximum Likelihood Estimation (MLE) to efficiently han[1]dle binary outcome data; and the Generalized Linear Model (GLM), which adopts modern modeling techniques based on empirical data properties, moving away from rigid theoret[1]ical distributional assumptions. Additionally, the GLM integrates a Thompson Sampling step in the estimation procedure, enhancing parameter estimation accuracy by balancing exploration and exploitation in data-driven decision-making. Each demand model and its corresponding pricing strategy are evaluated both qualita[1]tively and quantitatively. The quantitative analysis is conducted using numerical simu[1]lations, and it measures performance by the regret, defined as the expected revenue loss compared to an oracle that knows the true underlying demand function [35]. Extensive simulations show that the generalized linear model is particularly effective in complex environments, where prices influence demand differently depending on specific demand covariates. In contrast, simpler models perform better (i.e., achieve smaller regret) in more predictable settings where these complexities are absent. This analysis highlights the importance of selecting an appropriate pricing model based on specific market dy[1]namics and underscores the critical role of dynamic pricing in modern online sales.     «

When a company introduces a new product, e.g., via online sales channels, there is often limited information available on the demand. In these situations, adopting a dynamic pric[1]ing strategy proves more advantageous in increasing revenues than setting a static price. While incorporating dynamic pricing, sellers are presented with the challenge of “earning and learning”, which involves finding the right balance between maximizing immediate revenue (earning) and enhancing long-term revenues by...     »