[Paper Review] FinMem: A Performance-Enhanced LLM Trading Agent with Layered Memory and Character Design

Yangyang Yu, Haohang Li+7 more|arXiv (Cornell University)|Nov 23, 2023

Stock Market Forecasting MethodsDecision Sciences7 citations

TL;DR

FinMem introduces a modular LLM-based trading agent with profiling, layered memory, and adaptive character design to enhance stock trading performance, enabling memory-driven decision making and rapid adaptation to market cues.

ABSTRACT

Recent advancements in Large Language Models (LLMs) have exhibited notable efficacy in question-answering (QA) tasks across diverse domains. Their prowess in integrating extensive web knowledge has fueled interest in developing LLM-based autonomous agents. While LLMs are efficient in decoding human instructions and deriving solutions by holistically processing historical inputs, transitioning to purpose-driven agents requires a supplementary rational architecture to process multi-source information, establish reasoning chains, and prioritize critical tasks. Addressing this, we introduce extsc{FinMem}, a novel LLM-based agent framework devised for financial decision-making. It encompasses three core modules: Profiling, to customize the agent's characteristics; Memory, with layered message processing, to aid the agent in assimilating hierarchical financial data; and Decision-making, to convert insights gained from memories into investment decisions. Notably, extsc{FinMem}'s memory module aligns closely with the cognitive structure of human traders, offering robust interpretability and real-time tuning. Its adjustable cognitive span allows for the retention of critical information beyond human perceptual limits, thereby enhancing trading outcomes. This framework enables the agent to self-evolve its professional knowledge, react agilely to new investment cues, and continuously refine trading decisions in the volatile financial environment. We first compare extsc{FinMem} with various algorithmic agents on a scalable real-world financial dataset, underscoring its leading trading performance in stocks. We then fine-tuned the agent's perceptual span and character setting to achieve a significantly enhanced trading performance. Collectively, extsc{FinMem} presents a cutting-edge LLM agent framework for automated trading, boosting cumulative investment returns.

Motivation & Objective

• Motivate the need for autonomous agents that can process multi-source financial data with varying timeliness.
• Propose FinMem’s three-module architecture (Profiling, Memory, Decision-making) to address memory and interpretability limitations in DRL and prior LLM agents.
• Demonstrate how layered memory and adaptive character settings improve trading decisions and robustness in volatile markets.
• Show that FinMem achieves superior trading performance on real-world datasets with limited training data and general-purpose LLMs.

Proposed method

• Introduce FinMem’s Profiling module to customize agent characteristics and risk inclination.
• Develop a Memory module with working memory and layered long-term memory for time-sensitive information processing.
• Define a decision-making module that translates memory-derived insights into trading actions (Buy/Sell/Hold).
• Employ a novel memory scoring function combining recency, relevancy, and importance across shallow, intermediate, and deep layers.
• Use top-K memory events from each layer to inform immediate and extended reflections guiding trading decisions.
• Ground memory updates and retention with a layered decay mechanism and an access-counter to prioritize pivotal events.

Figure 1: The prompt template for FinMem ’s profiling module. It includes two key elements of its character setting: professional background knowledge and three distinct investment risk inclinations. In the self-adaptive risk inclination option, the omitted texts align with the detailed descriptions

Experimental results

Research questions

• RQ1 RQ1: Does FinMem outperform contemporary algorithmic trading agents on real-world data?
• RQ2 RQ2: Are there tasks that challenge other trading algorithms but are manageable by FinMem?
• RQ3 RQ3: Which LLM backbone yields the best performance for FinMem?
• RQ4 RQ4: Do different trader risk inclinations meaningfully differentiate FinMem’s trading performance?
• RQ5 RQ5: Can FinMem effectively filter and prioritize information to inform trading decisions?

Key findings

• FinMem demonstrates leading trading performance versus representative algorithmic agents on a scalable real-world dataset.
• FinMem’s memory and profiling modules enable better stratification and leveraging of market insights, improving decision quality.
• Adjusting backbone LLMs, working memory capacity, and character settings significantly affects trading performance.
• The dynamic risk inclination feature allows FinMem to adapt to market conditions and mitigate downturns.
• Ablation studies show FinMem can learn from past trading experiences and evolve its knowledge base through market interaction.
• The framework achieves robust trading performance with limited training data (days to months) and remains promising with larger datasets and specialized financial LLMs.

Figure 2: Memory module structure of FinMem with a detailed view of components, operations, and workflow. The cognitive architectures of FinMem ’s memory module have two core components – Working Memory and Layered Long-term Memory.