University of West Attica
Machine learning and complex systems time series analysis in quantitative finance and risk management
Abstract
dc:descriptionQuantitative finance is a rapidly evolving scientific field that leverages methods from disciplines such as Mathematics, Physics, and Computer Science to understand and predict financial phenomena. At the same time, it is closely linked to financial risk management, providing tools for the assessment, monitoring, and prediction of uncertainty as well as extreme events in financial markets. This PhD thesis approaches quantitative finance from the perspective of complex systems and machine/deep learning. Although both of these scientific fields are relatively recent, they already possess a strong theoretical foundation and numerous applications in finance. However, despite their growing use, there has not yet been a systematic investigation of the potential integration of these areas for application in finance, particularly in risk management. A primary motivation of this thesis was to address this gap in the literature, shedding light on this relatively unexplored scientific field. More specifically, the main objective was the development of machine/deep learning models to forecast market volatility and, consequently, potential risk. These models were designed to incorporate as input information about market dynamics revealed through complexity-based methods. To achieve this goal, an extensive analysis of financial markets was conducted using quantitative techniques derived from the science of complex systems and econophysics. Before focusing exclusively on financial data, the feasibility of applying methods from different scientific disciplines to the study of extreme phenomena in complex systems was examined. In this context, financial techniques were applied to the study of space weather, specifically to the forecasting of magnetic storms, highlighting the importance of the interdisciplinary perspective. Subsequently, the study concentrated solely on financial data. Initially, the potential to detect critical dynamics in time series using the Method of Critical Fluctuations was investigated. While this method has been applied across various scientific fields, this was the first detailed and comprehensive analysis in finance. Next, the underlying mechanisms contributing to the formation of financial “bubbles” and market crashes were studied. In this context, an enhanced version of the Log-Periodic Power Law model was employed to explore the role of external factors in triggering the collapse of an already unstable financial system. At the same time, the dynamic complexity and fractal characteristics of financial time series during periods of high uncertainty were examined. Measures such as entropy, Fisher information, and multifractal analysis were applied to quantify levels of complexity and multifractality before and during the COVID-19 pandemic. As an extension of this analysis, one of the most fundamental issues in economics, which has preoccupied the scientific community for decades, was studied: the Efficient Market Hypothesis, formulated by Eugene F. Fama. In this context, the evolution of market efficiency levels and the predictability of various financial instruments were investigated. Moreover, this thesis proposed a novel analysis plane whose coordinates consist of Approximate Entropy and the multifractal spectrum, providing a tool to study the randomness levels of financial time series. This approach enabled the assessment of both randomness and structural complexity, offering deeper insights into market dynamics. Following this in-depth analysis of financial markets from the perspective of complex systems and a thorough examination of the information revealed by complexity measures, machine/deep learning models were developed that incorporate these complexity measures as features to forecast market volatility. It is noteworthy that this represents the first time such an analysis has been conducted for market volatility, as existing models typically rely on traditional financial techniques and historical price data. The main findings of this thesis can be summarized as follows: (i) The application of methods from different scientific disciplines to the analysis of complex systems can yield remarkable results. (ii) Methods based on the science of complex systems can significantly enhance the ability to predict extreme financial events, reveal key features of financial time series that facilitate the identification of periods of market inefficiency, and effectively measure the temporal evolution of market randomness. Consequently, there appear to be periods during which financial markets are, in principle, predictable, challenging the theoretical basis of the Efficient Market Hypothesis. (iii) The integration of complex systems science with machine/deep learning methods is feasible and capable of producing meaningful results. Specifically, incorporating features based on complexity measures improves the predictive performance of machine/deep learning models, demonstrating their ability to capture hidden market structures and nonlinear dependencies. This novel approach bridges the gap between econophysics and machine/deep learning, paving the way for more robust and adaptive financial forecasting models.
Degree
thesis:*- Grantor dc:publisher
- University of West Attica
- Year dc:date
- 2026
Author and committee
dc:creator, dc:contributor.*- Authors dc:creator
-
- Zitis, Pavlos
- Ζήτης, Παύλος
Subjects
dc:subject × 20- Ανάλυση χρονοσειρών
- Διαχείριση χρηματοοικονομικών κινδύνων
- Μηχανική μάθηση
- Οικονομική Φυσική
- Πολύπλοκα συστήματα
- Πρόβλεψη χρονοσειρών
- Ποσοτική χρηματοοικονομική
- Complex systems
- Econophysics
- Financial risk management
- Machine learning
- Time series analysis
- Time series forecasting
- Quantitative Finance
- Επιστήμη Ηλεκτρονικών Υπολογιστών και Πληροφορική
- Φυσικές Επιστήμες
- Τεχνητή νοημοσύνη
- Computer and Information Sciences
- Natural Sciences
- Artificial Intelligence
Rights
- Language dc:language
- eng
Identifiers
dc:identifier.*- Identifier
- 10.12681/eadd/60939