ARTIFICIAL INTELLIGENCE IN FINANCIAL TECHNOLOGY: METHODS, APPLICATIONS, AND CURRENT DEVELOPMENTS

Abstract

This study systematises modern methods of generative artificial intelligence, specifically large language models (LLMs), and analyses approaches for their application in the financial technology sector. It provides a summary of existing strategies for using LLMs in FinTech, including zero-shot, few-shot, fine-tuning, Retrieval-Augmented Generation (RAG), and training models from scratch. A comparative analysis of their cost and implementation complexity was performed, identifying the most suitable LLM integration options depending on the application task. The paper presents an algorithm for automatic investment portfolio rebalancing, which combines classical Markowitz Portfolio Theory (MPT), price forecasting using LSTM networks, and technical analysis signals. An extended version of the rebalancing algorithm is proposed, supplementing traditional quantitative methods with two LLM components: a market sentiment analysis module and a financial statement processing module. Integrating these components enables the processing of unstructured data, such as financial news, social media posts, and quarterly or annual corporate reports. Using such data significantly expands the input datasets for price forecasting models, which can improve the quality of investment decisions. Based on the analysed scientific publications, it is shown that combining technical and fundamental financial indicators with market sentiment assessment helps to increase the accuracy of price forecasting for financial instruments. The paper demonstrates the potential for using the proposed investment portfolio rebalancing method in automated financial advisory systems (Robo-Advisors). The main limitations of the study are highlighted, in particular the need to test the rebalancing algorithm in practice using real market data. Directions for further research are identified, relating to the experimental testing of the proposed model on historical data from various periods and the subsequent optimisation of LLM components based on the results of the experiments.