Beyond Text: Chart-Vision Fusion for LLM Trading Agents in High-Frequency Environments

TL;DR: What if LLM traders could "see" market charts like human quants? This project trains vision-language models on live price charts and news to test if visual pattern recognition boosts predictive accuracy in LiveTradeBench.

Research Question: Can multimodal LLMs that synthesize textual news and visual chart patterns outperform text-only agents in capturing short-term market movements?

Hypothesis: Integrating technical analysis (chart patterns) with news narratives via VLMs will improve prediction accuracy for high-frequency trades by 10–20% relative to text-only baselines.

Experiment Plan: - Setup: Extend LiveTradeBench with real-time chart feeds (candlestick, volume) and fine-tune VLMs (e.g., GPT-4V) on MME-RealWorld-style financial visual QA tasks.

• Data: Pair 1-minute resolution charts with news from FinSearch’s API corpus; label key patterns (e.g., "head-and-shoulders").
• Metrics: Track prediction accuracy for intraday price movements vs. text-only agents.
• Expected Outcome: Multimodal agents will excel in high-volatility windows where visual cues (e.g., breakouts) precede news.

References: ['Zhang, Y., et al. (2024). MME-RealWorld: Multimodal LLM Benchmarking in High-Resolution Real-World Scenarios. ICLR.', 'Li, J., et al. (2024). FinSearch: Real-Time Financial Information Searching with LLMs. arXiv.']