Thomas Koch, PhD
Postdoctoral Associate
- Cambridge MA UNITED STATES
Dr. Koch leverages high-performance computing to enhance visibility and efficiency in maritime shipping and logistics.
Media
Social
Biography
Thomas holds a Ph.D. from Vrije Universiteit Amsterdam, where his research examined multimodal route choice after the opening of a metro line in Amsterdam. His expertise lies in algorithm development, large-scale data systems, and predictive modeling, applying advanced computational methods to optimize freight movement and supply chain resilience.
By integrating diverse data sources, Thomas develops scalable solutions that support data-driven decision-making in transportation networks. He is passionate about bridging cutting-edge technology with real-world logistics challenges, driving innovation in freight and intermodal transportation.
Research Focus
Intelligent & Autonomous Supply Chains
AI-driven, data-centric, and increasingly autonomous decision systems for supply chain planning and execution.
Future Logistics & Infrastructure Systems
Shaping freight, mobility, and infrastructure systems that enable global supply networks.
Critical Systems & Supply Network Resilience
Ensuring continuity, preparedness, and mission assurance under disruption and geopolitical volatility.
Areas of Expertise
Education
Vrije Universiteit Amsterdam (VU Amsterdam)
PhD
2022
Vrije Universiteit Amsterdam (VU Amsterdam)
MS
Computer Science
2017
Vrije Universiteit Amsterdam (VU Amsterdam)
BSc
Computer Science
2014
Languages
- English
Research Papers
Forecasting traffic flow by vehicle category on a major highway impacted by road maintenance works
Transportation Research ProcediaEline A Belt, Lili Yao, Xiaoyue Li, Nan Wang, Varun Shekhar, Thomas Koch, Elenna R Dugundji
2025-01-01
In this paper we look at the forecasting of traffic flow on a major highway in the Netherlands impacted by road maintenance works, examining the effects of lane closures on intensities per vehicle category. We apply several forecasting methodologies such as Prophet, Harmonic Regression, Seasonal Autoregressive (SAR), and Seasonal Autoregressive Integrated Moving Average (SARIMA) and compare them against a seasonal naive baseline model. We observe that SARIMA performs better than other models across all forecasting metrics for all sensors. This is mainly because of its capability of capturing linear trends and seasonality. There is also an opportunity to further improve the forecast accuracy of the SARIMA model by incorporating holiday information as seen in the Prophet model.
From Dock to Destination: Toward an End-to-End Simulation Study
Procedia Computer ScienceKevin Power, Yassine Lahlou-Kamal, Nikolay Aristov, Elenna Dugundji, Thomas Koch
2025-01-01
This study uses a discrete-event simulation model, built with open-source software, to analyze import container flows at the Port of New York/New Jersey. The model integrates input and parameter distributions derived from extensive data analysis of publicly available import records, enhanced by machine learning techniques, including Natural Language Processing for commodity classification using unstructured shipping manifest product descriptions. Initial results demonstrate the effectiveness of Gaussian Kernel Density Estimate (KDE) and Fourier models in representing container dwell times, reducing mean absolute error compared to normal distribution by up to 39.5% for dry containers and 24.8% for reefers. A fine-tuned BERT model achieves over 80% accuracy in commodity classification to the four-digit HS code level, enabling improved input data structuring for simulation.
Analysis of the impact of policy measures on parking behavior using interpretable time series models
Journal of Transport and Land UseElisabeth Fokker, Elenna Dugundji, Thomas Koch
2024-01-01
Growing awareness of the environmental impact of abundant parking has led to recent measures focused on decreasing car use in urban areas. This paper employs interpretable time series models to analyze the effects of these measures on parking demand. The study utilizes a dataset of more than 22 million parking transactions from 3,594 on-street selling points and 8 park-and-ride (P&R) locations in Amsterdam. Three models with external regressors, namely, Error Trend Seasonality (ETSX) models, Seasonal Autoregressive Integrated Moving Average (SARIMAX) models, and Interpretable Multi-Variate Long Short-Term Memory (IMV-LSTM) models, are compared against a Seasonal Naïve benchmark model. The ETSX model achieved the lowest error values, as indicated by both RMSE and SMAPE.