Essays on customer- and firm-side learning in pricing optimization and campaign management
In the first essay, we study a firm's dynamic pricing problem in a continuous-time setting where multiple market segments learn the product quality from post-purchase customer opinions and exert heterogeneous levels of social influence on other segments. To represent segment belief evolution in a continuous-time model, we approximate individual-level interactions with statistical averages and propose a fluid Bayesian updating process, which allows one to circumvent an NP-Hard problem of pricing on social networks. We use optimal control theory to derive the optimality condition and demonstrate in numerical study how the firm can improve revenue by performing price differentiation according to segment influence and learning behaviour. In the second essay, we modify the social leaning model proposed above and apply it to estimate the market belief evolution about a movie's quality for 157 movies that release between October 2013 and June 2015. We propose a two-step maximum likelihood estimation procedure to address the parameter identification problem. In the first step, we estimate each movie's true quality and initial population size along with seasonality effect by using each movie's daily sales and the number of theatres that play the movie. In the second step, we estimate market-level learning speed and belief evolution for each movie's quality by using tweets as a proxy for market opinions. Finally, the model that incorporates social learning achieves a better accuracy of forecast on the 7-day per-theatre sales than otherwise. In the third essay, we study the effects of multi-channel marketing and customer borrowing activities on the conversion behaviour. We model the prospective customer's willingness to convert as a continuous latent variable, propose a state-space formulation, and address major heterogeneity and endogeneity concerns. The empirical findings show that there is a potential antagonistic effect between SMS and other channels, and a customer's existing borrowing relationships may not prevent her from establishing a new borrowing relationship. In the one-month ahead prediction, the model achieves a conversion rate of 17.6%, better than 9.55% achieved by the bank's targeting strategies. Finally, we demonstrate how the marketing department can potentially use Q-learning to identify the best action for each customer.