TLDR: InvestAlign is a new framework that trains Large Language Models (LLMs) to understand investor decision-making, especially under herd behavior, by creating high-quality training data from theoretical solutions to simpler investment problems. This method overcomes the scarcity of real-user data, leading to faster model convergence and better alignment with human investment patterns in both simple and complex scenarios.
Large Language Models (LLMs) are becoming increasingly important in various fields, and finance is no exception. However, a significant hurdle in applying LLMs to complex areas like investor decision-making, especially when considering “herd behavior,” is the severe lack of real-world user data. This data scarcity makes it difficult to effectively train LLMs to mimic human investment patterns through a process called Supervised Fine-Tuning (SFT).
Traditional SFT methods rely heavily on vast amounts of authentic user data. Collecting such data in the financial sector is not only expensive but also raises significant privacy concerns, as investors are often reluctant to share their sensitive financial decisions. This fundamental limitation has hindered the full potential of LLMs in behavioral finance.
To address this challenge, researchers have introduced a new framework called InvestAlign. This innovative approach tackles data scarcity by generating high-quality SFT datasets. Instead of relying on scarce real-user data from complex scenarios, InvestAlign leverages theoretical solutions derived from simpler, yet similar, optimal investment problems. The core idea is that if a simpler version of an investment problem has a known theoretical solution, this solution can be used to create a large, high-quality dataset for training LLMs.
The research demonstrates that training LLMs with data generated by InvestAlign leads to faster parameter convergence compared to using real-user data. This suggests a superior learning efficiency for the models. To put this into practice, the team developed InvestAgent, an LLM agent that has been fine-tuned using the InvestAlign framework. InvestAgent has shown remarkable success, achieving significantly closer alignment with real-user data than models that haven’t undergone this specialized fine-tuning, both in straightforward and intricate investment problems.
The study specifically examined an optimal investment scenario involving two agents, considering factors like “absolute herd behavior” (mimicking entire portfolios) and “relative herd behavior” (mimicking changing rates of decisions), as well as “unilateral” and “mutual influence” between agents. While complex theoretical solutions exist for these scenarios, InvestAlign simplifies the training by using a problem with absolute herd behavior under unilateral influence as the “simple problem” for data generation. This simple problem, despite its mathematical similarities, differs in how it measures herd behavior, making it an ideal candidate for generating training data.
The findings indicate that pre-SFT LLMs often misalign with real-user data in investment decisions. However, InvestAgent, after being fine-tuned with the theoretically-derived data, significantly bridges this gap. For instance, in the simple problem (P3), InvestAlign helped reduce the overall Mean Squared Error (MSE) by 45.59% to 61.26% across different LLMs. Even for more complex problems (P1 and P2), the reduction in MSE ranged from 44.52% to 60.92%, showcasing InvestAlign’s ability to generalize its learning to more intricate scenarios.
Furthermore, the research validated that InvestAgents better reflect established economic principles related to herd behavior. For example, as the influence of herd behavior increases, agents’ investment decisions progressively converge, and in mutual influence scenarios, the total terminal funds of agents tend to decrease. These observations align with existing economic hypotheses.
Also Read:
- Unlocking Financial Intelligence in AI: Introducing FEVO
- RLVER: Cultivating Empathetic AI Agents Through Verifiable Emotion Rewards
The code for InvestAlign is publicly available, fostering further research and development in this promising area. This framework offers a potential solution to the persistent data scarcity issue in behavioral finance, paving the way for more accurate and human-aligned LLMs in investment decision-making. You can find more details about this research paper here: InvestAlign Research Paper.
