PaperlayerView exact revision on arXiv

Source-linked AI summary

A Multimodal Foundation Agent for Financial Trading: Tool-Augmented, Diversified, and Generalist

Wentao Zhang, Lingxuan Zhao, Haochong Xia, Shuo Sun, Jiaze Sun, Molei Qin, Xinyi Li, Yuqing Zhao, Yilei Zhao, Xinyu Cai, Longtao Zheng, Xinrun Wang, Bo An

arXiv:2402.18485v3q-fin.TRcs.AI

TL;DR

LLM trading capabilities remain underexplored, especially for multimodal interpretation and tool use. FinAgent combines multimodal market analysis with dual-level reflection; experiments report over 36% average profit improvement across six financial datasets.

  • Problem

    LLMs’ comprehensive autonomous trading capabilities remain underexplored, particularly for multimodal interpretation and diverse tool use.

    Source: §1 · p. 2

  • Method

    FinAgent combines market intelligence over numerical, textual, and visual data with low- and high-level reflection on price movements and past trading decisions.

    Source: §1 · p. 2, §4 · p. 4

  • Results

    Across 6 financial datasets and 6 metrics, FinAgent significantly outperforms state-of-the-art baselines, with over 36% average improvement on profit.

    Source: §1 · p. 2

  • Takeaways & Limitations

    FinAgent is presented as the first advanced multimodal foundation agent designed for financial trading tasks.

    Source: §1 · p. 2

    On ETH, FinAgent returns slightly less than FinMem because its auxiliary agents target stocks rather than higher-frequency cryptocurrency trading.

    Source: §6 · p. 9

Abstract

from arXiv · show

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.

1 INTRODUCTION

FinAgent is introduced as a multimodal foundation agent for financial trading, designed to address limits in existing systems through multimodal analysis, diversified retrieval, reflection, and expert-informed decisions. The introduction reports improved results across six financial datasets.

1 INTRODUCTION

  • Rule-based systems are rigid, while reinforcement-learning systems [1] need extensive training, lack explainability, generalize poorly, and are noise-sensitive; trading agents also struggle with multimodal intelligence.The introduction identifies five challenges: multimodal processing, imprecise retrieval, rapid adaptation, domain-knowledge integration, and reasoned actions.Source: §1 · pp. 1–2
  • Although BloombergGPT [47] and FinGPT [49] analyze financial data, QA remains distinct from sequential trading, and FinMEM [55] leaves multimodal interpretation and diverse tool use underexplored.The introduction situates LLM agents in the broader shift toward memory, planning, multimodal processing, and tool use.Source: §1 · p. 2
  • FinAgent combines numerical, textual, and visual market intelligence with dual-level reflection and diversified retrieval, while incorporating expert guidance and auxiliary trading strategies into decisions.Its reflection analyzes price movements at a low level and past trading decisions at a high level; retrieval query fields target specific information types.Source: §1 · p. 2
  • Across 6 stock and Crypto datasets, FinAgent outperformed 12 state-of-the-art baselines on 6 financial metrics, improving profit by over 36% on average; one dataset returned 92.27% (84.39% relative improvement).The authors describe FinAgent as the first advanced multimodal foundation agent designed for financial trading tasks.Source: §1 · p. 2

2 RELATED WORK

Prior work has advanced LLM agents and financial prediction, but text-dependent agents and price-focused trading methods leave multimodal market perception and generalization as gaps that FinAgent targets.

2.1 LLM Agents for Decision Making

  • BloombergGPT [47] introduced a finance-domain LLM without public access, while FinGPT [49] proposed open-source finance LLMs incorporating reinforcement learning with human feedback.The paper situates these developments amid the broader emergence of LLMs such as ChatGPT and GPT-4 [24].Source: §2.1 · pp. 2–3
  • LLM agents expanded from language processing to complex reasoning and collaboration; trading examples include FinMEM [55], while tool-augmented models retrieve current information and specialized knowledge.Tool-augmented language models have been evaluated on benchmarks including ScienceQA and TabMWP.Source: §2.1 · p. 3
  • Text-dependent LLM agents and NLP-focused finance models leave trading potential underexplored; FinAgent bridges this gap as a multimodal, tool-augmented foundation agent.Multimodal agents use visual capabilities to perform tasks unavailable to text-only agents.Source: §2.1 · p. 3

2.2 AI for Financial Trading

  • Trading research spans temporal RNNs, graph and tree models, tweet-derived latent features, professional behavior, news, earnings calls, ensembles, and reinforcement learning.Examples include GRU, LSTM, graph models [9], a tweet-based variational autoencoder, and ensemble learning.Source: §2.2 · p. 3
  • Price-focused methods often generalize poorly, motivating multimodal intelligence to navigate complex market dynamics.Prior work also draws on economic news [12] and earnings calls to improve prediction.Source: §2.2 · p. 3

3 PROBLEM FORMULATION

The section formulates financial trading as an MDP and extends the RL setup with a multimodal LLM agent whose reasoning, memory, and tools support action selection. FinAgent’s objective is to optimize total discounted returns.

3.1 Financial Trading as MDP

  • The trading agent learns a policy over a full trading period to maximize expected discounted rewards, representing overall profit.The MDP models an investor interacting with the market through discrete investment decisions.Source: §3.1 · p. 3
  • For single-asset trading, states reflect market information, actions are buy, sell, or hold, and rewards account for capital changes and commission fees.State inputs include price information, limited order books, technical indicators, trend predictions, financial news, expert behavior, and overall market status.Source: §3.1 · p. 3

3.2 Problem Formulation

  • FinAgent integrates multimodal LLMs into RL and defines, learns, and applies reasoning processes independently.Its state contains textual, numerical, and visual information.Source: §3.2 · p. 4
  • For a task, FinAgent generates a relevant prompt, obtains a multimodal LLM response, and parses it into an environment-compatible action.The action process uses task-associated memory.Source: §3.2 · p. 4
  • FinAgent comprises market intelligence, memory, low-level reflection, high-level reflection, and decision-making modules, with module-specific prompts and Kline and Trading charts.Memory is associated with the market intelligence and reflection modules.Source: §3.2 · p. 4
  • With memory, augmented tools, and designed modules, FinAgent seeks policies that optimize total discounted returns.The objective uses the policies defined in the formulation’s Eq. (2).Source: §3.2 · p. 4

4 FINAGENT FRAMEWORK

FinAgent organizes trading around market intelligence, memory, low- and high-level reflection, and tool-augmented decision-making. Together, these modules analyze market information and past trading outcomes to inform investment decisions.

4 FINAGENT FRAMEWORK

  • The framework comprises five core modules: market intelligence, memory, low-level reflection, high-level reflection, and tool-augmented decision-making.Memory supports the intelligence and reflection modules, while the decision module integrates their information with tools and trader preferences.Source: §4 · p. 4

4.1 Market Intelligence Module

  • Market intelligence analyzes current and historical news, financial reports, and prices, then uses retrieval-specific queries and diverse retrieval types to find and summarize relevant past information.The retrieval types support searches from multiple perspectives, with top K historical items retrieved separately for each type.Source: §4.1 · pp. 4–5

4.2 Memory Module

  • Memory stores market-intelligence analyses and both reflection types in vector storage, supporting acuity, adaptation to changing markets, and learning from past trading successes and mistakes.Each module’s summaries create query text for storage and retrieval; market intelligence uses vector similarity to match past data.Source: §4.2 · p. 5

4.3 Reflection Module

  • Low-level reflection relates market intelligence, Kline charts, and indicators to price movements, while high-level reflection evaluates past decisions and derives lessons for future trades.Low-level analysis considers short-, medium-, and long-term price changes; high-level reflection identifies decision successes and mistakes and recommends improvements.Source: §4.3 · pp. 5–6

4.4 Tool-Augmented Decision-making Module

  • The decision-making module combines market and reflection insights with professional guidance, trading strategies, financial position, and sentiment analysis to choose buy, sell, or hold and provide reasoning.It uses Chain-of-Thought and in-context learning, and considers strategies including MACD Crossover, KDJ with RSI Filter, and Mean Reversion.Source: §4.4 · p. 6

5 EXPERIMENT SETUP

The experiment section specifies FinAgent’s evaluation questions and protocol: multimodal asset datasets, six financial metrics, diverse trading baselines, and shared benchmark conditions.

5 EXPERIMENT SETUP

  • The experiments test overall competitiveness, component contributions, augmented-tool effects, and diversified retrieval while assessing performance with a reduced historical-data training window.The research questions ask whether FinAgent outperforms current agents, how its components contribute, and whether augmented tools and diversified retrieval improve trading performance.Source: §5 · p. 7

5.1 Datasets

  • Six real-world datasets cover five US stocks and Ethereum, combining daily prices, Kline and trading charts, news, and expert guidance.The penultimate year (2022-06-01 ∼2023-06-01) is used for training and the latter half-year (2023-06-01 ∼2024-01-01) for testing; assets are AAPL, AMZN, GOOGL, MSFT, TSLA, and ETHUSD.Source: §5.1 · p. 7

5.2 Evaluation Metrics

  • Following [28], evaluation uses ARR, SR, CR, SOR, MDD, and VOL, covering return, risk-adjusted profit, and risk.The metrics are annual return rate, Sharpe ratio, Calmar ratio, Sortino ratio, maximum drawdown, and volatility.Source: §5.2 · p. 7

5.3 Baselines

  • Baselines comprise B&H, MACD, KDJ&RSI, ZMR, LGBM, LSTM, Transformer [51], SAC [10], PPO [33], DQN [20], FinGPT [49], and FinMem [55].The comparison spans rule-based strategies, ML/DL and reinforcement-learning algorithms, an LLM framework, and an LLM-agent framework.Source: §5.3 · p. 7

5.4 Implementation Details

  • FinAgent requires no GPU for training or inference, but benchmarks used one NVIDIA RTX A6000; all benchmarks shared an RL environment, and FinAgent experiments used diversified retrieval unless noted.Further benchmark and experiment setup details are referred to Appendix ??.Source: §5.4 · p. 8

6 EXPERIMENTAL RESULTS

FinAgent outperformed 9 baselines across 6 financial metrics and was the only method to consistently exceed the broader market in profitability across datasets. Its results also reflect an aggressive-trader risk–return tradeoff, with a weaker result on ETH than FinMem.

Comparison with Baselines (RQ1)

  • FinAgent significantly outperformed 9 baselines across 6 financial metrics, with five-stock ARR% gains of at least 10% and SR gains of at least 19% over the best baseline.The reported comparisons are in Table 4 and Figure 4.Source: §6 · p. 8
  • Across all datasets, FinAgent was the only method to consistently outperform the broader market in profitability.The passage notes that FinMem’s AMZN ARR% of 40% fell below the Buy & Hold strategy’s 42%.Source: §6 · p. 8
  • FinAgent accepts slightly higher risk for substantially greater returns, consistent with its aggressive-trader investor preference.Rule-based methods control risk well but are less effective at capturing returns; the authors characterize FinAgent’s approach as a risk–reward balance.Source: §6 · pp. 8–9
  • FinAgent’s TSLA cumulative returns benefited from a short position anticipating a significant price drop after September 14, 2023.The passage attributes this anticipation to market intelligence and reflection, and describes the short position as hedging potential losses while generating high returns.Source: §6 · p. 9
  • On ETH, FinAgent returned 44%, slightly below FinMem, while a generalized cryptocurrency auxiliary agent could potentially raise returns to 54%.The passage attributes the weaker ETH result to auxiliary agents specialized for stocks rather than higher-frequency cryptocurrency trading; the 54% figure is a potential ablation result.Source: §6 · p. 9

7 ABLATION STUDIES

The ablations find that reflection improves trading metrics, while stock-specific auxiliary tools can hurt cryptocurrency performance; diversified retrieval improves AAPL results and separates retrieved intelligence types.

7.1 Effectiveness of Each Component (RQ2)

  • Adding low-level reflection to M and ML raises ARR% 45% to 101% across TSLA and ETHUSD and cuts risk 14% to 44%; high-level reflection raises ARR% and SR while reducing risk.High-level reflection entails a slight 7% rise in MDD% for TSLA; adding tools slightly improves stock profitability but lowers ETH performance by over 20%.Source: §7.1 · p. 9, Table 5 · p. 9

7.2 Effectiveness of Augmented Tools (RQ3)

  • On ETHUSD, tools-only T yields 16% ARR% versus B&H's 29% ARR%, showing stock-specific rule-based agents can harm cryptocurrency performance.The results suggest selecting auxiliary agents to match the market rather than adding them indiscriminately.Source: §7.2 · p. 9

7.3 Effectiveness of Diversified Retrieval (RQ4)

  • Diversified retrieval improves AAPL ARR and SR, while t-SNE embeddings distinguish retrieved market-intelligence types.The comparison is FinAgent with versus without diversified retrieval, and the embeddings are extracted from LLM representations of retrieved intelligence on the validation set.Source: §7.3 · p. 9, Figure 5 · p. 9

8 CONCLUSION AND FUTURE WORK · A DETAILS OF NOTATIONS

FinAgent combines multimodal market understanding, auxiliary-tool analysis, and diverse retrieval to support trading decisions. The paper proposes extending it to portfolio management, while its notation appendix lists key symbols and high-level-reflection terms.

8 CONCLUSION AND FUTURE WORK

  • FinAgent is a financial trading agent powered by an LLM, designed for high reasoning ability and generalizability.Source: §8 · p. 9
  • It integrates textual and visual data to understand market dynamics and historical trading behaviors.Source: §8 · p. 9
  • FinAgent independently uses auxiliary tools to analyze detailed market data across different time scales.Source: §8 · p. 9
  • Its multi-perspective, diverse retrieval approach relates current market conditions to past patterns and trends, integrating market information into final decisions.Source: §8 · p. 9
  • Future work will apply FinAgent to portfolio management, using observed market intelligence to rank stocks and select them.Source: §8 · p. 9

A DETAILS OF NOTATIONS

  • The notation appendix presents the paper’s main notations in Table 6 and identifies HLR_t as high-level reflection results for day t, with QHLR_t as its query and PHLR_t as its retrieved top-k results.Source: A DETAILS OF NOTATIONS · p. 12, Table 6 · p. 12

B DETAILS OF DATASETS AND PROCESSING · C DETAILS OF COMPARISON WITH BASELINES

The evaluation uses six real-world financial datasets and combines numerical, textual, and visual market information with expert guidance. FinAgent is compared with nine baselines on six financial metrics, with the paper reporting significant outperformance, especially in profitability.

B DETAILS OF DATASETS AND PROCESSING

  • The evaluation spans six real-world datasets—five from US stock markets and one cryptocurrency—and names five stocks, a foreign-exchange pair, and Ethereum in its asset selection.The named stocks are AAPL, AMZN, GOOGL, MSFT, and TSLA.Source: B DETAILS OF DATASETS AND PROCESSING · pp. 12–13
  • Daily open, high, low, close, and adjusted-close prices are paired with news coverage from financial sources, including Bloomberg Technology, Seeking Alpha, and CNBC Television.The price and news data were acquired through Financial Modeling Prep.Source: B DETAILS OF DATASETS AND PROCESSING · pp. 12–13
  • FinAgent receives historical Kline charts and trading line charts as visual input, plotted with pyecharts.Source: B DETAILS OF DATASETS AND PROCESSING · p. 13
  • Seeking Alpha professional analysis supplies expert guidance as auxiliary information, adding perspectives from professional analysts and individual investors.Source: B DETAILS OF DATASETS AND PROCESSING · p. 13
  • For the dataset split, data from the penultimate year, 2022-06-01 to 2023-06-01, is used for training, while latter-half-year data, 2023-06-017 to 2024-01-01, is allocated for testing.Source: B DETAILS OF DATASETS AND PROCESSING · p. 13

C DETAILS OF COMPARISON WITH BASELINES

  • FinAgent significantly outperforms nine baselines across six financial metrics, with especially notable profitability improvements, as reported in Table 7 and Figure 4.The supplied passages do not provide metric values or dataset-level winners.Source: C DETAILS OF COMPARISON WITH BASELINES · p. 13, Table 7 · p. 13

D DETAILS OF BENCHMARK METHODS … G.1 Case Study I

This section describes FinAgent’s benchmark implementation and prompt-based workflow for analyzing multimodal market inputs, reflecting on past trades, and structuring trading decisions. It also presents performance analysis and case studies, including a reasoned AAPL BUY decision.

D DETAILS OF BENCHMARK METHODS

E.1 Main Entry

  • The main entry point initializes the dataset and constructs the environment for training and inference.Source: §E.1 · p. 14

E.2 Run Step

  • The run step obtains latest intelligence, retrieves and stores market context, performs low- and high-level reflection, and then makes a trading decision.Source: §E.2 · pp. 14–15

F DETAILS OF PROMPT DESIGN

  • Prompt templates are modular, use XML for flexible parsing, and are converted through HTML parsing into the JSON message format required by the GPT-4 API.The templates separate latest and past market intelligence, low-level reflection, high-level reflection, and decision-making.Source: F DETAILS OF PROMPT DESIGN · p. 15

F.1 Market Intelligence

  • Market-intelligence templates distinguish current from past news and financial reports, include asset prices, and prompt analysis of visual market data and trading curves.Source: §F.1 · pp. 17–18
  • The market-intelligence guidance ignores unrelated news and classifies relevant information by effect duration and sentiment, balancing favorable and unfavorable implications.When duration is unclear, it is treated as long-term.Source: §F.1 · p. 19
  • Market-intelligence analysis filters unrelated items, extracts key insights with effect-duration and sentiment labels, and combines related findings into a reasoned, ID-linked investment summary.The overall summary states an explicit POSITIVE, NEGATIVE, or NEUTRAL sentiment and is limited to 300 tokens.Source: §F.1 · p. 20
  • Retrieval queries group asset-related intelligence by effect duration and use concise keywords drawn from the original items.The prompt specifies separate SHORT-TERM, MEDIUM-TERM, and LONG-TERM queries, each capped at 100 tokens.Source: §F.1 · p. 20

F.2 Low-level Reflection

  • Low-level reflection templates combine latest and past intelligence with real-time prices, news, reports, visual market data, and auxiliary strategy explanations to support BUY, HOLD, or SELL decisions.Visual inputs include Kline charts with technical indicators, historical trading curves, and cumulative return curves.Source: §F.2 · p. 22
  • Low-level reflection compares past and expected short-, medium-, and long-term price movements, attending to trajectory and trend shifts.It directs attention to the effect of latest market intelligence and to Kline-chart price changes.Source: §F.2 · pp. 23–24
  • Low-level reflection treats momentum and asset-specific price patterns as lessons to incorporate into further analysis when applicable.Momentum is described as persistence in the current price direction, used to anticipate short-term movements from historical trends.Source: §F.2 · p. 23
  • The low-level prompt requests step-by-step reasoning for each horizon and concise key-sentence queries to retrieve prior price-movement reasoning.Reasoning is output in separate short-, medium-, and long-term fields; each query condenses the key information for retrieval.Source: §F.2 · p. 24

F.3 High-level Reflection

  • High-level reflection combines past and latest market intelligence, three-horizon price analysis, trading decisions, and cumulative-return charts to review trading outcomes.The prompts also provide expert guidance and the asset’s prior decisions and reasoning as context.Source: §F.3 · pp. 25–26
  • Past trading decisions’ correct and incorrect lessons are intended to guide subsequent decisions toward maximizing profit.The review uses prior decision reasoning and visualized trading outcomes, including cumulative returns, as context.Source: §F.3 · p. 26
  • High-level reflection evaluates each trade, explains successes or mistakes using market and technical factors, proposes corrective actions, distills lessons, and creates concise queries for retrieving past reflections.The requested output includes reasoning, improvement, and query fields.Source: §F.3 · p. 27

F.4 Decision-making

  • The decision-making template combines past and latest market intelligence, multi-horizon price analysis, low-level and high-level reflections, visual charts, technical indicators, and professional guidance to support BUY, HOLD, or SELL decisions.It asks the agent to identify relevant price patterns and incorporate them into further analysis and reflections.Source: §F.4 · pp. 28–29

G.1 Case Study I

  • Examples show BUY on positive Apple AR/VR sentiment and bullish prices, SELL on negative medium- to long-term sentiment, and SELL amid mixed signals, immediate risks, and limited cash despite bullish long-term technicals.The third example prioritizes liquidity and risk mitigation over speculative gains.Source: §G.1 · p. 29
  • The case-study guidance presents MACD crossovers for momentum, KDJ with an RSI filter for ranging markets, and mean reversion for stable markets, while noting limits in choppy, trending, or volatile conditions.The guidance also asks the agent to combine strategy signals with their explanations when deciding.Source: §G.1 · p. 30
  • For final decisions, the guidance downweights neutral sentiment, emphasizes immediate price impacts, and uses professional advice informed by historical prices when market intelligence is mixed.It also directs BUY or SELL responses to bullish or bearish trends and missed trading opportunities, respectively.Source: §G.1 · p. 30
  • Action must be exactly BUY, HOLD, or SELL; decisions should synthesize conflicting evidence and obey cash and position checks.The prompt says not to BUY when cash is below the current Adj Close Price or to SELL without an existing position.Source: §G.1 · p. 30
  • In the AAPL case, FinAgent chooses BUY, weighing positive product-driven sentiment and rising prices against negative news, mixed guidance, HOLD strategies, and prior missed gains.The rationale also cites sufficient cash and no existing position as support for opening a position.Source: §G.1 · p. 31, Figure 8 · p. 31
  • FinAgent’s approach emphasizes balanced analysis, lessons from prior trades, risk management, adaptability, and cautious handling of market volatility.The passage characterizes these practices as supporting potential returns while minimizing exposure to adverse market swings.Source: §G.1 · p. 31

G.2 Case Study: II … H.1.3 Z-score Mean Reversion.

FinAgent’s case studies illustrate context-dependent stock and crypto decisions, while the benchmark analysis describes rule-based strategies whose limitations include delayed or misleading signals and poor fit to persistent trends.

G.2 Case Study: II

  • Alphabet examples show FinAgent buying when long-term prospects and price momentum are bullish and cash permits, but selling to secure profits and preserve liquidity.The examples draw on cloud and AI prospects, investor sentiment, price trends, and reflections on past decisions.Source: §G.2 · p. 33

G.3 Case Study III

  • ETHUSD examples show FinAgent selling amid bearish sentiment, downtrends, and insufficient cash, but buying when MA5, Bollinger Bands, long-term sentiment, and liquidity support an uptrend.The sell decision prioritizes capital protection; the buy decision aims to capitalize on anticipated gains.Source: §G.3 · p. 34
  • FinAgent’s crypto approach is designed to balance risk and reward, but the paper identifies volatility as a challenge and further model refinement as an area for improvement.The discussion calls for gradual refinement to better align strategies with unpredictable crypto markets.Source: §G.3 · pp. 34–35

H DETAILS OF BENCHMARK ANALYSIS / H.1 Analysis of Rule-based Strategies for Trading

  • Four rule-based strategies span trend following and reversal detection, use the same market state and trading environment as RL benchmarks, and receive OPTUNA hyperparameter tuning.The strategies are MACD, stochastic oscillators with Bollinger Bands, KDJ with RSI Filter, and Z-score Mean Reversion.Source: §H.1 · p. 35

H.1.1 Moving Average Convergence Divergence (MACD) Crossover Strategy.

  • MACD can react too slowly to market changes, generate misleading signals in volatile conditions, and incur substantial transaction costs through high turnover.The passage attributes delayed reactions to its reliance on historical moving averages and notes standard thresholds may not adapt to volatility.Source: §H.1.1 · p. 35

H.1.2 KDJ&RSI Strategy.

  • KDJ&RSI lacks generalization across market environments because parameter sensitivity can produce widely varying results across stocks, including unprofitable instruments.The passage notes that hyperparameters trained on historical data still fail to yield profitable results on some instruments.Source: §H.1.2 · pp. 35–36

H.1.3 Z-score Mean Reversion.

  • Z-score mean reversion can sell before a continuing bullish trend and may sustain losses in trending markets that do not revert to historical means.The AAPL example illustrates missing a bullish trend after the price reached the reverting point; the passage also cites TSLA decisions as evidence of limited versatility.Source: §H.1.3 · pp. 36–37, Figure 17 · p. 37, Figure 18 · p. 37

H.2 Analysis of Using FinAgent as a router for technical indicators-driven strategies for trading. · H.3 Analysis of RL for Trading

As a strategy router, FinAgent delivers robust, positive trading results but has limitations tied to strategy disagreement, weak underlying strategies, and prompt or API randomness. The RL analysis benchmarks PPO, DQN, and SAC under uniform training conditions with Optuna hyperparameter optimization.

H.2 Analysis of Using FinAgent as a router for technical indicators-driven strategies for trading.

  • The router experiment used FinAgent’s decision module alone, with Optuna-fine-tuned strategies and their documented performance data provided to the LLM.Market intelligence, memory, and reflection were excluded, so no FinAgent training phase was needed.Source: §H.2 · p. 37
  • FinAgent’s router consistently delivers robust, positive trading results, though it does not always select the optimal strategy and exhibits less extreme behavior than single-strategy trading.The authors suggest the router can adapt to varying market conditions, while noting that other FinAgent modules address its limitations.Source: §H.2 · p. 37
  • The router can make conflicting sequential decisions because strategies with different assumptions may guide buys and sells in the same short period.For the period in Fig. 19, KDJ&RSI and SO&BB mostly influenced buys, while ZMR guided sells.Source: Figure 19 · p. 37, §H.2 · p. 38
  • Router performance is constrained when all underlying strategies are weak, and without reflection its decision-making capability is limited.The example period in Fig. 20 had no profitable strategy, as referenced alongside Fig. 11, Fig. 15, and Fig. 17.Source: Figure 20 · p. 37, §H.2 · p. 38
  • Trading decisions varied with prompt changes and across trials, while OpenAI API randomness could not be controlled, limiting decision-module robustness.Source: §H.2 · p. 38

H.3 Analysis of RL for Trading

  • The RL analysis benchmarks PPO, DQN, and SAC after training them on the same dataset under uniform conditions and optimizing hyperparameters with OPTUNA.The authors describe these algorithms as having potential but also limitations in stock prediction accuracy and practicality due to large data requirements.Source: §H.3 · p. 38

H.3.1 Analysis of PPO for trading. · H.3.2 Analysis of DQN Algorithm for Stock Trading.

The analyses identify PPO’s prolonged inactivity and poor trade timing, and DQN’s weak risk–reward balance and delayed reactions to volatility, as limitations in stock and crypto trading. They also propose model refinements to improve adaptation and decision-making.

H.3.1 Analysis of PPO for trading.

  • PPO can remain idle while cumulative returns fall below zero, missing potential trading opportunities in AAPL.The analysis attributes this behavior to cautious learning that prioritizes risk avoidance.Source: §H.3.1 · p. 38, Figure 21 · p. 38
  • PPO also mistimed a GOOGL trade on September 27, 2023, missing a low-value buying opportunity and buying at a relatively high price.The passage links this decision to suboptimal income and difficulty predicting short-term price movements.Source: §H.3.1 · pp. 38–39, Figure 22 · p. 38
  • For PPO, proposed refinements include parameter and training optimization, sentiment and macroeconomic features, and regular monitoring to improve adaptation and trade timing.The passage also recommends combining machine learning with domain expertise.Source: §H.3.1 · p. 39

H.3.2 Analysis of DQN Algorithm for Stock Trading.

  • Despite relatively positive long-term results, DQN needs richer market indicators, real-time and diverse data, risk management, and regular updates to adapt to changing conditions.The proposed refinements aim to improve its sensitivity to market dynamics and decision quality.Source: §H.3.2 · p. 39
  • DQN misses bullish gains and adjusts slowly in bearish markets, including a zero position during the October 26–November 23, 2023 bullish phase and missed AMZN gains from October 23–December 2023.The examples illustrate how conservative positioning can forgo gains, while slow adjustment in declining markets can prolong losses.Source: §H.3.2 · p. 39, Figure 23 · p. 39, Figure 24 · p. 39
  • On ETHUSD, DQN lagged rapid price movements during extreme volatility from July 18 to August 15, 2023.The analysis describes delayed reactions as a source of suboptimal decision-making.Source: §H.3.2 · p. 39

H.3.3 Analysis of SAC Algorithm for Stock Trading.

SAC’s entropy regularization supports exploration, but its stock-trading decisions can be too cautious or poorly timed, limiting its ability to capture profitable market movements. The analysis concludes that SAC needs careful tuning and enhancement for dynamic, volatile markets.

H.3.3 Analysis of SAC Algorithm for Stock Trading.

  • SAC’s entropy regularization helps it explore unknown states and learn strategies through interaction with the environment to maximize expected cumulative reward.Despite these strengths, the analysis notes that SAC still has difficulty handling stock-market intricacies and recommends careful tuning and enhancement for dynamic, volatile markets.Source: §H.3.3 · pp. 39–40
  • In ETHUSD, SAC made only two buys around Nov 15, 2023, during a bullish phase, missing significant market movements with its cautious long-term observations.This conservatism may avoid potential losses but can also prevent the model from fully capitalizing on profitable trends.Source: §H.3.3 · pp. 39–40, Figure 26 · p. 40
  • For MSFT and AMZN, SAC struggled to time entries and exits, selling AMZN on Oct 19, 2023, and MSFT on Nov 10, 2023, while profitable trends continued.The analysis says late responses to uptrends and premature exits from profitable positions reduce potential returns.Source: §H.3.3 · p. 40, Figure 27 · p. 40, Figure 28 · p. 40

H.4 Analysis of FinGPT for Trading

FinGPT’s trading analysis identifies five recurring limitations: missed opportunities, incorrect price-movement predictions, actions that do not follow its reasoning, and generic responses after consecutive failures. These cases raise concerns about its reliability as investment guidance.

Limitation 1: Failure to capitalize on peak price opportunities

  • After buying MSFT at $330.46, FinGPT held through a $358 peak and sold at $336.37, missing the peak-price opportunity.The passage attributes the hold to an expected further increase based on positive developments, market conditions, and anticipated strong earnings.Source: §H.4 · pp. 40–41, Figure 31 · p. 42

Limitation 2: Wrong prediction for a sharp price drop

  • FinGPT recommended buying TSLA at $245.34 on October 8, 2023, assuming an upward trajectory; the passage reports TSLA at $215.49 on August 18, 2023, indicating an inaccurate forecast.Source: §H.4 · p. 41

Limitation 3: Wrong prediction for a stock price surge

  • FinGPT recommended selling AAPL while forecasting a 1-2% decrease, but the stock rose for three consecutive days.The passage says the forecast cited positive developments, including an expected $4 trillion valuation by 2025 and the Vision Pro launch.Source: §H.4 · p. 41, Figure 29 · p. 42

Limitation 4: Mismatch between the action and reasoning

  • FinGPT correctly predicted GOOGL’s price changes from June 6 to July 6, 2023, but kept holding despite forecasting a near-term decline.The example illustrates a mismatch between its prediction and trading action, potentially missing opportunities to act on the forecast.Source: §H.4 · p. 41, Figure 30 · p. 42

Limitation 5: Meaningless prediction after consecutive failures

  • After three consecutive wrong AMZN predictions, FinGPT gave a generic response rather than meaningful analysis or a coherent adjustment.The passage describes its explanation as acknowledging prediction difficulty and offering vague positive factors and potential concerns.Source: Table 8 · p. 41, §H.4 · p. 42

H.5 Analysis of FinMem for Trading

FinMem uses layered memory to retrieve historical news and financial reports, then reflects on their connections to future price movements. Its decision analysis reveals errors tied to incomplete trend reasoning, limited state awareness, and unstable decisions.

Limitation 1: Incorrect decision-making due to partially positive market news

  • A 4 August 2023 AAPL buy ignored the broader downward trend and shifted returns from 6% to -3%.FinMem based its rationale on two positive news items despite mostly negative news and a continuous price decline.Source: §H.5 · pp. 42–43, Figure 33 · p. 43

Limitation 2: The cash is insufficient for more shares but provides a buying decision result

  • Two buying operations despite insufficient remaining cash show FinMem can choose actions inconsistent with its current state.In the authors’ setting, buying means going all in and selling means going all out; FinAgent’s state description iframe module is designed to avoid this issue.Source: §H.5 · p. 43, Figure 34 · p. 43

Limitation 3: Unstable decision-making

  • FinMem makes unstable buy-then-sell and sell-then-buy decisions, risking substantial losses in highly volatile markets.The authors attribute this instability to mixed positive and negative market information and FinMem’s inability to perceive its past decisions.Source: §H.5 · p. 43, Figure 35 · p. 43, Figure 36 · p. 43

Related papers

Loading 2402.18485v3…