Thomas Koch, Massachusetts Institute of Technology

Abstract

This talk presents research conducted at MIT on modeling maritime logistics and intermodal container flows, as well as related work performed as an ORISE Research Fellow supporting the US Department of Transportation FLOW initiative. Port operations are vital in global supply chains, with U.S. ports handling 70% of the nation's trade by weight and 40% by value. Given its importance, accurate demand forecasting for ports is necessary for efficient planning and operational improvements. Machine learning (ML) models, especially Long Short-Term Memory (LSTM) neural networks, have emerged as powerful tools for time series forecasting in this domain. While most prior studies focused on single-port forecasting, this study explores the application of a sequence-to-sequence LSTM model for multivariate forecasting across multiple ports and commodities. Using data from the U.S. Census Bureau (2003-2024), the study trained and validated a LSTM model to predict container throughput at eight major U.S. ports, covering 776 port-commodity combinations. The model outperformed traditional methods like linear regression, random forest, and gradient boosting, reducing forecasting errors by 6% overall. It also proved effective as multi-period forecasting technique, providing accurate predictions up to three months in advance.

Bio:

Thomas Koch is a Postdoctoral Associate at the MIT Center for Transportation & Logistics, where he works with the U.S. DOT on the FLOW project to enhance freight and maritime efficiency. Combining a Ph.D. in multimodal transport with expertise in high-performance computing and predictive modeling, he develops scalable, data-driven solutions for complex supply chain challenges. His work focuses on integrating real-time analytics and geospatial data to improve visibility and resilience across global transportation networks.