Link Prediction In A Temporal Money Flow Network Using Graph Neural Networks

Abstract

Many real-world phenomenon can be described as a graph that changes over time. Predicting what changes are likely to happen next is a common desire that machine learning pursues to answer. More recently, Graph Neural Networks (GNNs) have emerged as a powerful tool when answering graph related challenges, including link prediction. In this thesis, GNNs are used to study, how the investors of the Helsinki Stock Exchange fund their stock purchases by selling other stocks. This dynamic can be viewed as a flow of money from one stock to another, and can be presented as a network. The goal of this thesis is to establish the base predictability of the money flows for future research with a suitable GNN-based implementation.

The experimental results show that there is predictability in the money flow network that is captured by the methods used. The results are verified against a naive baseline. The findings justify further research on predicting the dynamics of the money flows, and offer themselves as a new baseline for future models.