Date of Award
Doctor of Philosophy (PhD)
Operations & Information Management
Marshall L. Fisher
This dissertation studies empirical operational problems in retail sales and service systems through three essays. We are in the middle of a remarkable rise in data analytics as available data, as well as the capability of artificial intelligence grows exponentially. This dissertation demonstrates how we can use traditional econometrics and cutting-edge machine learning models to provide data analytics for retail operations management. The first essay studies the value of unstructured social media text data in forecasting future fashion demands several months out, at a granular style-color level. Using recent advancement in natural language processing and machine learning techniques, we show that unstructured social media text data, after proper processing, has significant predictive in forecasting color demands, weeks or even months into the future. The predictive power measured by the improvement of the out-of-sample mean absolute deviation ranges from 18% to 25% over different manufacturing lead times and is robust across different geographic markets. Finally, we explore the mechanism and show that the predictive power comes from social media data capturing the fashion retail cycle, fashion influence theories, and marketing efforts. The second essay studies what factors and how they affect workforce productivity in service systems. Using a detailed time-stamped data, a novel framework, and a bivariate Probit model to contend econometric difficulties arising from workers’ making endogenous discretionary decisions, we provide a complete picture on how different factors affect instantaneous productivity collectively while explicitly controlling for the endogenous discretionary decisions. We find that peer effects have a negative impact on productivity with workers’ discretionary decisions serving as mediator. Scheduled interruptions and workers’ decisions to multitasking also have negative impacts while decisions to take breaks and task-switching improve productivity by 54% and 25% respectively. The third essay proposes a novel methodology to contend a challenging task – optimizing long-term cumulative reward in a partially observed high-order Markov decision process. We combine machine learning and approximate dynamic programming techniques to develop a tandem neural network structure. The methodology is applied to a real-world customer relationship dataset and outperforms extant models in the field in terms of both the predictive power and optimization performance.
Fu, Youran, "Essays On Empirical Operations Management In Retail Sales And Service" (2018). Publicly Accessible Penn Dissertations. 3053.