Development and Customization of an AI Assistant for Enhancing Portfolio Management and Investment Decision-Making

This thesis explores the development of an AI-based financial assistant using FinGPT, an open-source large language model designed for financial applications. The project addresses two key limitations of existing financial tools: the lack of transparency in proprietary systems and the insufficient domain adaptation of generic LLMs. By leveraging FinGPT, the assistant can be fine-tuned on curated financial datasets such as earnings reports, market data, and news articles, enabling the generation of structured, context-aware reports tailored to portfolio management and investment decision-making. The methodology combines data collection and preprocessing, fine-tuning with LoRA for efficiency, and regular retraining to ensure adaptability to changing market conditions. The system design includes modules for portfolio summaries, sector performance analyses, and country-level economic outlooks, with outputs validated through standard evaluation metrics and expert review. While challenges remain due to the “black-box” nature of neural networks, the open-source framework of FinGPT allows control over training data and periodic updates, ensuring greater transparency and trustworthiness compared to closed models. The expected contribution of this work is a prototype that demonstrates how financial LLMs can support investors and analysts by improving the accuracy, efficiency, and accessibility of financial reporting and decision support.