Back to search

University of Thessaly (UTH)

Artificial intelligence and metaheuristic optimization algorithms for financial risk assessment and investment decision-making

Abstract

dc:description

The field of finance has long faced complex, difficult challenges. The present thesis, titled “Artificial Intelligence and Metaheuristic Optimization Algorithms for Financial Risk Estimation and Investment Decision-Making” studies and evaluates modern artificial intelligence and metaheuristic optimization algorithms and proposes hybrid models to address critical problems such as credit risk assessment in financial institutions, portfolio optimization, and stock closing-price prediction. It first presents the objectives and the research challenges the thesis seeks to address. This is followed by a literature review that highlights the growing research interest in these types of algorithms in recent years, both in the broader domain of Financial Science and, more specifically, in the area of portfolio optimization. Next, four applications are presented in detail in the fields of financial risk and investment decision-making. First, the problem of customer churn in financial institutions is examined, which is a demanding challenge not only in banking but across many industries. The performance of deep learning models is evaluated with respect to the choice of activation function and optimization algorithm. Different activation functions and optimization methods are investigated with the aim of assessing their performance, while the use of the APTx activation function is proposed. The respective activation function appears capable of delivering better results, and this work is the first to recommend its use in a practical application. In addition, given that such problems typically exhibit pronounced class imbalance, various resampling methods are evaluated. The second application explores another application of credit risk, the case of credit card default, an important challenge for financial institutions. For this purpose, an optimized LightGBM via the metaheuristic Pareto-like Sequential Sampling (PSS) algorithm is proposed. The performance of the proposed methodology compared against a popular Bayesian optimization method, the Tree-Structured Parzen Estimator (TPE). Furthermore, the use of Scrambled Halton Sequences (SHS) is proposed. As noted by the authors of PSS , this algorithm is compatible with any sampling method. The results showed that the proposed use of SHS provides more accurate results and better generalization performance compared to Monte Carlo and Latin Hypercube Sampling (LHS) methods. In this application, LightGBM’s built-in functionalities are also used to address class imbalance and high dimensionality. The results show that the proposed model achieves better performance compared to similar studies in the literature, without the need to use complex architectures. Subsequently, the use of metaheuristic algorithms for portfolio optimization is investigated. In particular, the Artificial Gorillas Troops Optimizer (AGTO) is proposed. Sixteen different metaheuristic algorithms are compared to assess their effectiveness in portfolio optimization—many of which have limited or no prior application in the literature—such as the Tuna Swarm Optimizer, the Artificial Rabbits Optimizer, and the Marine Predators Algorithm (MPA). Their performance is also examined with respect to different population sizes and numbers of training epochs. By optimizing a portfolioconsisting of the constituents of the Dow Jones Industrial Average (DJIA), the results showed that AGTO can provide the best performance in portfolio optimization. Finally, the hybrid model GJO-GRU|LSTM is proposed, namely an optimized hybrid deep neural network using the Golden Jackal Optimizer (GJO), for predicting stock closing prices. While many studies in the literature focus on tuning the hyperparameters of LSTM or GRU networks for time-series forecasting, the present application attempts to optimize both the parameters and the architecture itself: depending on the case, the network type (GRU or LSTM) and its corresponding parameters are selected dynamically. The performance of the proposed model is compared against sixteen deep-learning models, as well as GA-GRU|LSTM and PSO-GRU|LSTM, which denote optimized models based on Genetic Algorithms (GA) and Particle Swarm Optimization (PSO). The results indicated that GJO-GRU|LSTM delivers more accurate forecasts in the vast majority of cases. The thesis concludes with the main findings and a summary of its contributions. Its aim is to enrich applied research on Artificial Intelligence and metaheuristic optimization through models capable of supporting well-founded decisions on critical issues in finance. In recent years, these algorithms have been widely adopted across various fields, due to their ability to overcome the limitations of traditional models. We consider that the evaluation of the algorithms investigated, and the proposed approaches make a significant contribution to the field, deepening our understanding of their impact on financial science and enriching the existing literature both within finance and in the broader domains of artificial intelligence and metaheuristic optimization.

Degree

thesis:*
Grantor dc:publisher
University of Thessaly (UTH)
Year dc:date
2025

Author and committee

dc:creator, dc:contributor.*
Authors dc:creator
  • Gkonis, Vasileios
  • Γκόνης, Βασίλειος

Subjects

dc:subject × 16

Rights

Language dc:language
eng

Identifiers

dc:identifier.*
Identifier
10.12681/eadd/60621

Chain of custody

source
Harvested from
Greek National Archive of PhD Theses
Base URL
phdtheses.ekt.gr/eadd_oai/request
Last updated
2026-11-18
Source record
OAI-PMH GetRecord
citation

Gkonis, Vasileios; Γκόνης, Βασίλειος. Artificial intelligence and metaheuristic optimization algorithms for financial risk assessment and investment decision-making. University of Thessaly (UTH), 2025. http://hdl.handle.net/10442/hedi/60621