F²Agent trading AI beats 16 baselines, 148% returns on TSLA
New multimodal agent fuses text, charts, and news to deliver 120% returns on GOOG
Deep Dive
F²Agent is a novel multimodal AI trading framework that uses a hierarchy of specialized agents to extract modality-specific signals, then applies a modality-aware adaptive fusion mechanism and noise-robust consistency regularization to capture fine-grained cross-modal dependencies and resist market noise. Tested on six stocks and cryptocurrency assets, it outperformed 16 competitive baselines across multiple trading metrics, delivering over 20% relative improvement in annualized return on average—including 120.48% on GOOG and 148.41% on TSLA.
Key Points
- F²Agent uses a hierarchy of specialized agents to extract modality-specific trading signals from text, charts, and market data
- A modality-aware adaptive fusion mechanism plus noise-robust consistency regularization improves cross-modal dependency capture
- Outperforms 16 baselines with >20% average annualized return boost; delivers 120.48% on GOOG and 148.41% on TSLA
Why It Matters
Multimodal AI trading agents that handle market noise could reshape algorithmic trading profitability across assets