Aristotle University Of Thessaloniki (AUTH)
Machine learning methods for the analysis of financial data
Abstract
dc:descriptionThe financial industry is a major adopter of emerging information technology, with recent advancements in Artificial Intelligence (AI) transforming the banking and investing sectors. Nevertheless, evidence suggests that standard Machine Learning (ML) methods require significant adaptation to be effective in finance, due to limited high-quality data, which tends to be noisy and non-stationary. This thesis advances financial ML by contributing new methods to address relevant problems in three distinct areas. These contributions advance ML theory and are sufficiently general to be applied across domains. Advancing the research area of Learning to Rank (LtR), the first contribution comprises new ranking methods for stock selection that can support portfolio management. While traditional ranking methods recommend candidates solely based on high expected rank, the proposed method Stochastic-Aware Bootstrap Ensemble Ranking (SABER) also quantifies the uncertainty associated with each rank, enabling the discovery of portfolios with improved return-to-risk ratios. Furthermore, the enhancement Merged Bootstrap Selection (MBS) supports variable-size portfolios, which consistently outperform traditional fixed-size portfolios. This research work demonstrates an effective financial application of LtR, with attention to predictive performance and computational complexity. Contributing to the research area of Ensemble Learning, the second research work proposes a finance-inspired estimator weighting method for Tree-Based Ensemble (TBE) models applied to Random Forest (RF). Traditional weight estimation methods rely exclusively on the individual performance of estimators, which introduces the risk of allocating most weight to highly correlated estimators, thereby reducing the variety of the ensemble, a critical factor in its effectiveness. The proposed method Markowitz Random Forest (MRF) addresses this limitation by considering the variance-covariance matrix of estimator errors in a quadratic optimization task resembling portfolio optimization. The result is a novel weight allocation method that assigns weight to high-performing estimators while discouraging the concentration of weight on highly correlated ones. The third contribution is a new approach for macro-level sentiment analysis within the research area of Natural Language Processing (NLP). Sentiment analysis of financial news is an effective strategy that powers trading and investing systems worldwide. Although the majority of retail investors prefer to invest in sectors or industries, related work has focused primarily on micro-level sentiment analyses regarding how specific news items may affect individual stocks. To address this issue, the proposed method, Sector-level Sentiment Analysis (SLSA), integrates two Supervised Learning (SL) models for the simultaneous prediction of both relevant sector and sentiment polarity of news, with the aggregation of scores revealing the prevailing sentiment of each sector. vii Additionally, this work explores a semi-supervised mechanism for relabeling likely mislabeled neutral news. In conclusion, the thesis provides additional discussion regarding the contributions in the three research areas. Furthermore, it also outlines relevant future work directions within the scope of financial ML.
Degree
thesis:*- Grantor dc:publisher
- Aristotle University Of Thessaloniki (AUTH)
- Year dc:date
- 2025
Author and committee
dc:creator, dc:contributor.*- Authors dc:creator
-
- Kouloumpris, Eleftherios
- Κουλούμπρης, Ελευθέριος
Subjects
dc:subject × 16- Μάθηση Κατάταξης
- Συλλογική Μάθηση
- Επεξεργασία φυσικής γλώσσας
- Βαθιά μάθηση
- Learning to Rank
- Ensemble learning
- Natural language processing
- Deep learning
- Επιστήμη Ηλεκτρονικών Υπολογιστών και Πληροφορική
- Φυσικές Επιστήμες
- Τεχνητή νοημοσύνη
- Διεπιστημονικές εφαρμογές επιστήμης ηλεκτρονικών υπολογιστών
- Computer and Information Sciences
- Natural Sciences
- Artificial Intelligence
- Computer Science Interdisciplinary Applications
Rights
- Language dc:language
- eng
Identifiers
dc:identifier.*- Identifier
- 10.12681/eadd/61230