Quantitative portfolio management and financial decision-making with future generation of AI models

Abstract

This study investigates Large language models (LLMs) and agentic artificial intelligence (Agentic AI) for their efficacy within the finance domain, particularly focusing on asset allocation, portfolio management, and financial decision-making. Portfolio construction is a complex task that requires careful consideration of multiple asset classes, risk profiles, market conditions, and user-inclusive investment objectives. LLMs and pre-trained models often hallucinate due to their lack of access to real-time and updated information. Additionally, they are not particularly effective at mathematical reasoning. While LLMs exhibit remarkable capabilities in understanding financial terminology and reasoning over structured or unstructured data, they face critical limitations in their access to real-time financial information and market dynamics, which are essential for making informed investment decisions. To address these challenges, our research explores these problems through a three-phase approach: first, the application of advanced prompt engineering techniques; second, the implementation of retrieval-augmented generation (RAG); and finally, the development of a multi-agent system architecture. We show through this research (i) a strong contribution to the intersection of AI and financial decision-making; (ii) that LLMs and agents could act as catalysts for one another rather than relying solely on pre-trained LLMs. Therefore, enhanced performance and more complex execution can be achieved by integrating external agents; (iii) that optimizing the quality and reliability of model-generated responses is possible; (iv) improving the overall effectiveness of financial decision-making processes and portfolio management. As an evaluation mechanism we have done a case-study of comparison with Wealthsimple portfolio classifications by replicating their asset class structure appropriately through our agentic workflow.