Data-Driven Studies of Caller Behavior Under Call Center Innovations

Customer service call centers have recently used innovations such as new technology, new data sources, and new predictive analytics techniques to improve service quality, decrease operating costs, and increase sales revenue. In this dissertation we study how companies can exploit some of these innovations by performing three empirical studies of caller behavior under call center innovations. In the first chapter we perform an empirical study of a call center that uses callback technology. We formulate the parameters of a structural model of the caller decision-making process under callback technology and impute their underlying preferences from the data. We use these preferences to predict the performance measures of this call center under various callback policies, and make recommendations regarding how call centers should offer callbacks to improve service quality and increase system throughput. In the second chapter we study how call centers can improve their performance measures by using caller history data to predict caller abandonment and redialing behavior, and prioritize callers based on the predictions. We refer to these prioritization policies as history-based priority policies. To predict the performance measures of this call center under history-based priority policies, we formulate and estimate the parameters of mixed-logit models that rely on callers' history of waiting times, abandonment decisions, and redialing decisions to predict their current abandonment and redialing behavior. We use our models to predict call center performance measures under a variety of history-based priority policies and find that, relative to the first-come, first-served policy, certain history-based priority policies reduce average waiting times and/or increase opportunities for agents to sell additional products. In the third chapter we study how managers can use a variety of predictive analytics techniques and various sources of customer data to predict caller abandonment and redialing behavior. We train several machine learning methods to predict caller abandonment and redialing behavior using caller history data. We find that neural network configurations generate the most accurate predictions, but that logistic regression and decision trees provide nearly comparable prediction accuracy. We also find that caller history substantially increases the prediction accuracy of all the prediction methods.Doctor of Philosoph