QuantAgent: Price-Driven Multi-Agent LLMs for High-Frequency Trading

In one sentence

Summary

Existing LLM-based financial agents primarily rely on textual inputs like news or sentiment, which lag price discovery and are ill-suited for the low-latency, precision-critical demands of high-frequency trading (HFT). QuantAgent addresses this gap by operating exclusively on structured, short-horizon price signals derived from OHLC bars. The framework decomposes the trading process into four specialized agents: IndicatorAgent, which computes technical metrics like RSI and MACD; PatternAgent, which uses multimodal reasoning to identify chart formations; TrendAgent, which fits support and resistance channels to determine directional bias; and RiskAgent, which integrates these signals to define stop-loss and take-profit boundaries. This modular design allows for traceable, language-native rationales for each trade decision, bridging the interpretability of LLMs with the speed and structure required for algorithmic trading. The system is evaluated on a multi-asset benchmark comprising nine financial instruments, including cryptocurrencies (Bitcoin), equity indices (S&P 500, Nasdaq), commodities (Crude Oil), and volatility indices (VIX). Experiments utilize 1-hour and 4-hour OHLC data, with 5,000 historical bars per asset, testing the model against random selection, linear regression, and XGBoost baselines. The evaluation focuses on directional accuracy over the next three candlesticks and risk-constrained rate-of-return metrics (Rcc, Rmax, Rmin) that simulate realistic execution with fixed stop-loss thresholds. The agents operate in a zero-shot setting without supervised fine-tuning, relying on structured prompts and tool-use capabilities to process market data and generate executable trade orders. QuantAgent consistently outperforms all baselines across most assets, achieving up to 80% directional accuracy in rolling-window validations and significant improvements in risk-adjusted returns. Notably, it shows pronounced gains in equity markets like SPX and NQ, demonstrating robust generalization across different market regimes. However, the system faces limitations in ultra-short timeframes (1-15 minutes), where noise degrades prediction quality, and inference latency remains a barrier to true real-time deployment. The findings suggest that coupling structured price signals with multi-agent LLM reasoning offers a viable path for interpretable, high-frequency decision systems, though further optimization is needed for micro-second execution environments.

Research problem

Current LLM-based financial agents are designed for long-horizon, text-driven tasks and are unsuitable for high-frequency trading (HFT) due to latency and reliance on lagging textual signals. This paper addresses the need for an LLM framework that operates directly on structured, short-horizon price data to enable rapid, risk-aware, and interpretable trading decisions.

Core contributions

• Introduction of QuantAgent, the first multi-agent LLM framework explicitly designed for high-frequency algorithmic trading.
• Decomposition of trading into four specialized agents (Indicator, Pattern, Trend, Risk) for modular and interpretable analysis.
• Development of a multimodal reasoning pipeline that converts OHLC data into technical indicators, chart patterns, and trend channels.
• Creation of a comprehensive multi-asset benchmark for evaluating LLM-based HFT systems across cryptocurrencies, equities, and commodities.
• Demonstration of superior directional accuracy and risk-adjusted returns compared to traditional quantitative baselines.

Method

• IndicatorAgent computes RSI, MACD, ROC, Stochastic, and Williams %R from OHLC data.
• PatternAgent generates candlestick charts and identifies formations like double bottoms or triangles.
• TrendAgent fits OLS regression lines to highs and lows to determine support/resistance and trend slope.
• RiskAgent sets fixed stop-loss (0.05%) and LLM-predicted take-profit ratios.
• DecisionAgent aggregates signals from upstream agents to issue LONG/SHORT orders.
• LangGraph orchestrates the multi-agent workflow with structured prompts.
• Zero-shot inference using GPT-4o for reasoning and text generation.

Results & findings

• QuantAgent outperforms baselines in 7 out of 8 assets for risk-constrained return (Rcc).
• Achieves up to 80% directional accuracy in rolling-window case studies on SPX.
• Significant accuracy gains in equity markets: +26.5% over baseline in NQ, +34.6% in SPX.
• Consistently highest Rmax in 6 out of 8 markets, indicating effective upside capture.
• Robust downside protection with Rmin ranking among the least negative values.
• Performance degrades on ultra-short candles (1-15 min) due to market noise.
• Inference latency exceeds the window for exploitable 1-minute opportunities.

Data & experiments

• Benchmark includes 9 assets: Bitcoin, Crude Oil, DJI, ES, NQ, QQQ, SPX, VIX, and DAX.
• Data sources: 1-hour and 4-hour OHLC bars from TradingView and YFinance APIs.
• 5,000 historical bars per asset, sampled into 100 evaluation segments of 100 consecutive bars.
• Baselines: Random selection, Linear Regression, and XGBoost trained on technical indicators.
• Metrics: Mean Directional Accuracy (over 3 candlesticks) and risk-constrained Rate of Return (Rcc, Rmax, Rmin).
• Zero-shot evaluation without supervised fine-tuning or external sentiment data.

Limitations & caveats

• Predictive accuracy drops significantly on ultra-short candles (1-15 minutes).
• Inference latency is too high for true real-time HFT execution.
• Relies on historical OHLC data, ignoring macroeconomic or news events.
• Performance may vary in extreme market regimes not covered in the benchmark.
• Zero-shot setting may limit performance compared to fine-tuned models.
• Risk parameters are fixed or heuristically set, not dynamically optimized.

Curator notes

Discovery status: trading_agent_focus_promoted. Source links: FinMem: A Performance-Enhanced Large Language Model Trading Agent with Layered Memory and Character Design