Change-Point Detection in Multivariate Time Series Considering Interdependencies in Financial Markets and Applications of the Tool

This thesis addresses the problem of detecting structural breakpoints (change points) in multivariate time series by leveraging graph-based representations that capture the interdependencies between financial assets. The primary objective of the research is to analyze existing methods and develop a unified approach for early detection of structural shifts, incorporating directional relationships between assets and macroeconomic factors. The thesis is structured as follows. Chapter 1 presents the problem formulation and provides a comprehensive literature review of change point detection (CPD) methods. It includes theoretical descriptions of key models such as Causal-RuLSIF, Recurve, and graph neural network (GNN)-based architectures. The chapter concludes with a formal statement of research hypotheses. Chapter 2 focuses on the domain context. The analysis is based on high-liquidity crypto assets and major macroeconomic indicators. The chapter describes the structure of the input time series, and the external variables used to account for environmental factors and enhance interpretability. Chapter 3 outlines the experimental setup and evaluation metrics for benchmarking model performance. Chapter 4 presents a detailed comparison of the proposed methods and discusses the empirical results. The developed tool demonstrates improved ability to detect structural changes compared to standard baselines, and shows practical potential in applications such as risk management, algorithmic trading, and analytics for decentralized finance (DeFi). Keywords: change point detection, multivariate time series, cryptocurrencies, graph-based models, RuLSIF, Recurve, PCMCI, Transfer Entropy, DeFi