Forecasting linehaul transit times & on time delivery probability using quantile regression forests

Abstract

with delays on the road and variabilities introduced by the major participants in the process, ie: distribution centers, drivers, etc. These sources of variability also make it difficult to measure the impact changes in transit time have on on-time performance. This paper focuses on trying to identify indicators of variability and incorporates them into quantile regression forest, a black box forecasting model, that will provide estimated scheduled transit times for a given probability of on-time arrival at the destination. With the use of Amazon's Q1 & Q2 2013 linehaul data, an analysis on performance trends based on length of haul were categorized to develop an understanding linehauls in North America. The outbound transportation team at Amazon faces the complex trade off between providing a sufficient amount of scheduled transit time to ensure ontime delivery to destination and the utilization rate of a truck. The ability to quantify how changes in scheduled transit time impact the performance of a particular linehaul allows transportation managers to assess this trade off. The paper explores a machine learning regression technique called quantile regression forests. The model was developed in R using the quantregforest package. It incorporates numerous factors about linehaul including: origin, destination, historical reporting on sources of late to arrivals, time to depart from origin and time of departure. The strengths of this black box model are in its ability to handle a large amount of data and continuously update its predicting structure to provide more accurate recommendations. Quantile regression forests also enable the user to specify the ontime performance percentage, p, that he/she wants the model to predict based on historical data. The final model at p = 95% provided a weight mean absolute percent error of 4.57% and a root mean square error of 2.22%. A four-week pilot was conducted to validate these predictions and the results are discussed.