Forecasting without sequences: graph representations for dynamic systems in finance and beyond using GNNs

Forecasting complex systems has traditionally been framed as a sequence learning problem, assuming that past trajectories contain sufficient information to predict future states. However, in many real-world domains—including financial markets—this assumption fails due to stochasticity, structural shifts, and hidden dynamics. This dissertation proposes a conceptual shift: to view forecasting not as extrapolating sequences, but as learning from structure. The work is divided into three major parts. First, we introduce the Sparse Spatio-Temporal Neural Network (sSTNN), a scalable architecture for large-domain forecasting tasks. While sSTNN performs well on smooth systems like ocean temperature, its failure on deforestation forecasting tasks exposes the limits of sequence-based approaches when faced with abrupt, structurally-driven dynamics. Second, focusing on financial markets, we demonstrate that meaningful structural information persists even at extremely short intraday timescales. Using advanced community detection methods on financial correlation networks, we show that sectoral organization can be reliably recovered from minutes or hours of return data. These findings suggest that market behavior is better captured through evolving relational structures rather than isolated time series. Finally, building on these insights, we develop a predictive framework using Message Passing Graph Neural Networks (MPGNNs) trained on financial graphs. Our models, incorporating both node and edge features, forecast future index movements with statistically significant accuracy across multiple temporal resolutions. Crucially, the success of these graph-based models is not solely explained by conventional sector clustering, indicating that deeper latent topological signals underlie market dynamics. Together, these contributions argue for a structural paradigm in forecasting complex systems, where topology and relational information serve as the primary modeling substrate. Beyond financial markets, the methods and insights developed here lay the groundwork for applying graph-based forecasting to a broader range of dynamic, structurally complex domains.

Name thesis:degree_name

Ph.D.

Level thesis:degree_level

Dissertation

Discipline thesis:degree_discipline

Informatics

Grantor

University of Illinois Urbana-Champaign

Year dc:date

2025

Author dc:creator

• Bracht, Eamon

Contributors dc:contributor

• Brunner, Robert J.
• Brunner, Robert J
• McMullin, Jeff
• Mendoza, Kim
• Carrasco Kind, Matias

• Gnn
• Forecasting
• Time Series
• Deforestation
• Stock Market
• Equity Markets
• Community Detection
• Graph Clustering
• Clustering
• Mpgnn

Statement dc:rights

• Copyright 2025 Eamon Bracht

Language dc:language

en, eng

Handle dc:identifier

https://hdl.handle.net/2142/130048

OAI record oai:identifier

https://www.ideals.illinois.edu/oai-pmh?verb=GetRecord&metadataPrefix=etdms&identifier=oai%3Awww.ideals.illinois.edu%3A2142%2F130048