Improving Truck Driver Turnover Predictions With Survival Stacking and Reference-Dependent Covariates

Abstract

The U.S. motor carrier industry has suffered from high driver turnover rates for decades. The need is high for developing techniques that enable carriers to predict the turnover risk (quit probability) of each driver for each time period, as they can provide timely warnings to carriers regarding which drivers are about to quit, allowing carriers to take necessary preventive actions. Despite the strong practitioner interests, development of such techniques has received limited attention from scholars. The existing approach, moreover, uses traditional survival models that typically assume smooth and monotonic hazard functions, which may not properly capture the drivers' complex decision dynamics. We present a more flexible approach to the individual-level, time-specific predictions of driver turnover risk by utilizing two concepts that are new to the literature, namely survival stacking and loss aversion. We theoretically explain why using these concepts can improve the accuracy of turnover predictions and empirically examine, using the data obtained from three different motor carriers, if and to what extent these concepts can enhance performance. Results indicate that these concepts are useful for improving driver turnover predictions. The approach presented in this paper was recently adapted by a U.S. motor carrier as a decision support system.