[論文レビュー] A Multimodal Foundation Agent for Financial Trading: Tool-Augmented, Diversified, and Generalist

Wentao Zhang, Lingxuan Zhao+11 more|arXiv (Cornell University)|Feb 28, 2024

Multi-Agent Systems and Negotiation被引用数 5

ひとこと要約

FinAgent は、金融取引のためのツール拡張を備えた多模態ファウンデーションエージェントで、6つのデータセットにわたり9つのベースラインを上回り、平均利益を36％以上改善し、1つのデータセットで92.27％のリターンを達成します。

ABSTRACT

Financial trading is a crucial component of the markets, informed by a multimodal information landscape encompassing news, prices, and Kline charts, and encompasses diverse tasks such as quantitative trading and high-frequency trading with various assets. While advanced AI techniques like deep learning and reinforcement learning are extensively utilized in finance, their application in financial trading tasks often faces challenges due to inadequate handling of multimodal data and limited generalizability across various tasks. To address these challenges, we present FinAgent, a multimodal foundational agent with tool augmentation for financial trading. FinAgent's market intelligence module processes a diverse range of data-numerical, textual, and visual-to accurately analyze the financial market. Its unique dual-level reflection module not only enables rapid adaptation to market dynamics but also incorporates a diversified memory retrieval system, enhancing the agent's ability to learn from historical data and improve decision-making processes. The agent's emphasis on reasoning for actions fosters trust in its financial decisions. Moreover, FinAgent integrates established trading strategies and expert insights, ensuring that its trading approaches are both data-driven and rooted in sound financial principles. With comprehensive experiments on 6 financial datasets, including stocks and Crypto, FinAgent significantly outperforms 9 state-of-the-art baselines in terms of 6 financial metrics with over 36% average improvement on profit. Specifically, a 92.27% return (a 84.39% relative improvement) is achieved on one dataset. Notably, FinAgent is the first advanced multimodal foundation agent designed for financial trading tasks.

研究の動機と目的

• Address multimodal data processing and generalization gaps in financial trading AI.
• Develop a multimodal foundation agent with memory, reflection, and tool augmentation.
• Integrate domain knowledge and explainability to improve trust in decisions.
• Evaluate on diverse datasets (stocks and crypto) to demonstrate generalization.

提案手法

• Introduce a market intelligence module processing numerical, textual, and visual data.
• Incorporate a dual-level reflection module (low-level and high-level) for rapid adaptation and learning from history.
• Employ a diversified memory retrieval system with separate memory for market intelligence, low-level, and high-level reflections.
• Use a tool-augmented decision-making module that integrates expert guidance and classic trading strategies.
• Formulate FinAgent within an MDP framework and integrate LLMs into the RL pipeline with specialized prompts (phi, mem, tool).
• Apply Chain-of-Thought reasoning and provide justifications for trading decisions.]
• research_questions:[

Figure 1 . The overall architecture of FinAgent. The ordinal numbers in the figure represent the order of execution, where augmented tools are implemented with the decision-making module.

実験結果

リサーチクエスチョン

• RQ1 RQ1: Does FinAgent outperform current state-of-the-art trading agents across diverse tasks?
• RQ2 RQ2: How does each FinAgent component contribute to overall performance?
• RQ3 RQ3: Does augmented-tool integration improve trading performance?
• RQ4 RQ4: How effective is the diversified retrieval mechanism in FinAgent?

主な発見

Model	AAPL ARR%	AAPL SR	AAPL MDD%	AMZN ARR%	AMZN SR	AMZN MDD%	GOOGL ARR%	GOOGL SR	GOOGL MDD%	MSFT ARR%	MSFT SR	MSFT MDD%	TSLA ARR%	TSLA SR	TSLA MDD%	ETHUSD ARR%	ETHUSD SR	ETHUSD MDD%
B&H	13.00	0.60	14.78	42.33	1.08	17.38	22.47	0.71	12.97	22.49	0.84	12.92	37.40	0.72	32.65	29.26	0.87	23.21
MACD	11.86	0.72	10.38	14.27	0.71	7.84	-18.00	-0.89	20.07	15.23	0.77	8.34	-4.90	-0.02	14.15	10.24	0.47	24.32
KDJ&RSI	2.17	0.17	11.88	19.38	0.65	17.27	24.39	2.13	2.03	18.84	1.06	7.78	2.14	0.17	24.73	8.87	0.51	16.95
ZMR	-3.91	-0.22	8.88	18.73	0.84	7.89	32.51	1.45	5.38	9.86	0.71	6.22	-7.28	-0.09	19.90	29.35	1.23	13.11
RL-based	7.92	0.40	14.88	27.43	1.17	5.27	34.40	1.39	7.15	30.44	1.18	10.56	15.07	0.44	28.12	29.81	1.18	9.53
SAC	24.84	1.12	11.98	38.33	1.07	13.84	23.80	0.75	13.07	22.02	0.82	12.92	42.22	0.87	26.19	17.84	0.76	10.06
PPO	13.26	0.61	14.78	21.17	0.70	13.84	38.29	1.30	8.45	11.32	0.48	17.51	33.64	0.78	28.35	34.75	1.31	11.12
FinGPT	-5.46	-0.17	16.23	42.93	1.10	18.94	12.28	0.44	13.00	25.10	0.97	9.84	38.43	0.75	31.47	21.57	0.68	25.56
FinMem	23.78	1.11	10.39	40.07	1.03	18.53	31.27	1.11	8.97	40.58	1.50	7.48	50.04	0.92	25.77	44.71	1.27	13.59
FinAgent	31.89	1.43	10.40	65.10	1.61	13.20	56.15	1.78	8.45	44.74	1.79	5.57	92.27	2.01	12.14	43.08	1.18	12.71
Improve	28.38	27.68	-	51.64	37.61	-	46.64	-	-	10.25	19.33	-	84.39	118.48	-	14.20	-	-

• FinAgent significantly outperforms 9 state-of-the-art baselines on 6 financial datasets across 6 metrics with over 36% average profit improvement.
• On one dataset, FinAgent achieves a 92.27% return (84.39% relative improvement).
• FinAgent demonstrates strong generalization across stocks and crypto (AAPL, AMZN, GOOGL, MSFT, TSLA, ETHUSD).
• The framework is the first advanced multimodal foundation agent designed for financial trading tasks.
• The evaluation uses diversified retrieval and tool augmentation to enhance decision quality and explainability.

Figure 2 . Case studies of FinAgent. We only display the partial prompt for brevity. See Appendix G for the full prompt structure.