We research approaches for predicting waiting times of customers within service systems. Our focus is on call centers in which accurate waiting time predictions may enable improved work-force management and lead to increased customer satisfaction.
Standard approaches for design and management of service systems that include customers, servers and queues rely on Queueing Theory (QT). In particular, QT is often used for predicting waiting times which is the focus of this work.
QT is often criticized since its formulas are based on fundamental assumptions regarding the model describing a given system — assumptions which are usually violated.
To overcome this issue, we hypothesize that additional external characteristics (“features”) may be required in order to provide an accurate prediction of the waiting time for many real-life systems. We further hypothesize that a Machine Learning (ML) model incorporating those characteristics will improve the performance of waiting time predictions. Consequently, in this research we explore the use of ML-based algorithms for waiting time prediction and investigate to what extent they overcome the limits of QT methods which are typically based on “simple model” assumptions. In a series of experiments, we use multiple ML- and QT-based models to predict the waiting times of customers within simple (“synthetic”) and real-life call centers.
We analyze the performance of the different models, in order to determine for which methods and under what conditions ML performance compares favorably with QT performance.