Technical Summary: F2Agent

Problem Statement

Financial trading relies on a complex interplay of heterogeneous information sources, ranging from quantitative historical prices to qualitative textual narratives. While the Efficient Market Hypothesis suggests prices incorporate available information, effectively digesting these multimodal signals remains challenging due to structural disparities across modalities and high noise levels inherent in financial streams.

Existing Large Language Model (LLM)-based trading agents face three primary limitations:

1. Limited Multimodal Modeling: Most methods either focus primarily on price/technical indicators or rely solely on textual sentiment, underutilizing complementary information and overlooking interactions between modalities (e.g., "sell-the-news" scenarios where positive sentiment coincides with bearish price movements).
2. Ineffective Fusion Mechanisms: Current solutions often rely on prompt-level concatenation, treating numerical time-series as textual tokens. This ad-hoc approach fails to resolve the modality gap, allowing the LLM's textual bias to dominate joint representations and drowning out subtle quantitative cues.
3. Inadequate Robustness: Financial streams are stochastic with low signal-to-noise ratios. Existing agents lack explicit mechanisms to disentangle persistent market trends from transient perturbations (e.g., microstructure noise), leading to overreactions to irrelevant cues and unstable decisions in volatile markets.

Methodology

The paper proposes F2Agent, a novel multimodal agentic paradigm driven by the "Financial Fusion of Agentic Intelligence." The system architecture consists of three core components:

1. Hierarchy of Specialized Agents

F2Agent deploys a hierarchy of specialized agents to extract complementary, modality-specific signals from heterogeneous sources, rather than processing all data through a monolithic model:

• Market Analysis Agent: Encodes raw OHLCV (Open, High, Low, Close, Volume) sequences using a causal Transformer encoder to capture long-range temporal dependencies and non-linear price-volume coupling.
• Technical Analysis Agent: Processes expression-based alpha factors (e.g., MACD, RSI) using a causal Transformer to model non-linear interactions among technical indicators.
• Sentiment Analysis Agent: Leverages DeepSeek-R1 with tool-augmented retrieval to query external news APIs, summarize articles, and infer binary sentiment labels and rationales for each news item.
• News Analysis Agent: Utilizes a fine-tuned Qwen-based LLM (Qwen2.5-7B-Instruct) with Chain-of-Thought (CoT) prompting. It retrieves and consolidates news evidence to generate transparent, natural-language reasoning explanations and auxiliary directional signals (UP/DOWN).

2. Modality-Aware Adaptive Fusion Mechanism

To address the limitations of simple concatenation, F2Agent introduces a fusion layer that coordinates specialized agents:

• Alignment: Each agent's output is projected into a shared latent space. For LLM-based agents, last-token pooling is used; for Transformer encoders, the [CLS] representation is utilized.
• Adaptive Attention: The mechanism computes modality-specific queries, keys, and values. It introduces learnable modality vectors ($r_m$) injected into the keys and values to encode instance-independent, modality-specific biases, encouraging diversity among modalities while preserving instance-independent evidence.
• Dynamic Weighting: The system dynamically captures fine-grained inter-modality dependencies, allowing the model to weigh modalities based on their relevance to the current market context.

3. Noise-Robust Consistency Regularization

To enhance stability against market noise and regime shifts, the framework employs a consistency regularization strategy:

• Perturbation: During training, a modality is randomly selected and perturbed (e.g., by adding Gaussian noise).
• Consistency Loss: The model is penalized if the predicted logits for the perturbed input diverge significantly from the original input.
• Attention Suppression: A specific term discourages the model from increasing reliance on the perturbed modality, forcing it to rely on consistent signals across modalities.

System Integration

F2Agent operates as an end-to-end system integrating a news summarizer (to reduce token budgets and noise), the modality processing module, and a backtesting engine. It produces trading decisions, natural language explanations, and performance metrics under practical trading constraints.

Key Contributions

1. Hierarchy of Specialized Agents: Introduction of a multi-agent framework to extract complementary signals from heterogeneous sources (market, technical, news, sentiment) rather than relying on a single unified model.
2. Modality-Aware Adaptive Fusion: A novel fusion mechanism that dynamically captures fine-grained inter-modality dependencies, mitigating the modality gap caused by prompt-level concatenation.
3. Noise-Robust Consistency Regularization: A regularization technique designed to improve robustness against market noise and regime shifts, yielding stable trading signals.
4. Comprehensive Evaluation: Implementation of an end-to-end Multimodal Agentic Trading System and extensive experiments demonstrating consistent outperformance over diverse baselines.

Experimental Results

The authors evaluated F2Agent on six assets (AAPL, AMZN, GOOG, MSFT, TSLA, and BTCUSD) against 16 competitive baselines, including rule-based strategies, ML/DL models, Reinforcement Learning (RL) agents, general LLMs, and specialized financial LLMs.

• Performance Metrics: F2Agent consistently outperformed all baselines across Annualized Return (ARR), Sharpe Ratio (SR), Cumulative Return (CR), and Maximum Drawdown (MDD).
• Key Achievements:
  • Achieved an average rank of 1.00 in ARR across all tested assets.
  • Delivered 120.48% ARR on GOOG and 148.41% ARR on TSLA.
  • Showed an average relative improvement of over 20% in annualized return compared to the best-performing baseline.
  • Demonstrated superior risk-adjusted performance, maintaining stability even on high-volatility assets like BTCUSD where other methods exhibited severe reversals.
• Ablation Studies: Removing the adaptive fusion module or the robustness regularizer significantly degraded performance, confirming the necessity of both components.
• Robustness: F2Agent maintained stable performance as more modalities were added, whereas baseline models (including strong general LLMs) often degraded when handling conflicting cross-modal cues.
• Efficiency: The system demonstrated superior token efficiency compared to frameworks relying on monolithic LLM calls for all data processing.

Significance and Claims

The paper claims that F2Agent represents a significant step forward in AI-driven financial trading by addressing the critical gap in multimodal fusion and robustness.

• Beyond Pattern Extraction: Unlike previous methods that focus on simple pattern extraction or shallow fusion, F2Agent provides a principled mechanism for fusing heterogeneous modalities, enabling the capture of nuanced cross-modal dependencies.
• Robust Decision Making: The system demonstrates that explicitly modeling the interactions between modalities and regularizing against noise leads to more reliable trading signals in volatile markets, avoiding "sentiment traps" that isolated modalities might trigger.
• Scalability and Generalization: The framework is shown to generalize across different asset classes (stocks and cryptocurrencies) and market conditions, suggesting that the agentic paradigm is a viable approach for real-world, information-rich trading environments.

The authors conclude that F2Agent effectively organizes a hierarchy of specialized agents to provide a holistic view of market dynamics, producing coherent trading representations that are robust to noise and shifting regimes. Future work is noted to extend this paradigm to portfolio management for multimodal, risk-aware allocation.