💻 computer science

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

The paper introduces QuantHarness, a novel multi-agent LLM framework specifically designed for high-frequency trading that leverages specialized agents to process structured, short-horizon signals, demonstrating superior predictive accuracy and real-time decision-making capabilities compared to existing long-horizon, text-driven financial models.

Original authors: Fei Xiong, Xiang Zhang, Aosong Feng, Siqi Sun, Chenyu You

Published 2026-07-28

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Fei Xiong, Xiang Zhang, Aosong Feng, Siqi Sun, Chenyu You

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine the stock market as a giant, chaotic ocean. For decades, sailors (traders) have tried to predict the waves using two main maps. The first map is a weather report full of news, rumors, and social media chatter—what people say is happening. The second map is the water itself: the actual rise and fall of the waves, recorded in precise numbers. Most modern computer programs trying to predict the market have been obsessed with the weather reports, reading millions of news articles to guess where the water will go next. But here's the catch: by the time a news story hits the internet, the water has often already moved. The waves have already shifted, and the news is just a slow echo.

This is where a new idea comes in: what if we stopped reading the weather reports and just watched the water? This approach, called "technical analysis," assumes that every piece of news, every fear, and every hope is already baked into the price of the asset. If you can read the patterns in the price movements themselves, you might be able to predict the next wave faster than anyone reading the news. Now, imagine giving a super-smart computer brain (a Large Language Model) the job of reading these water patterns. Usually, these brains are great at writing stories or chatting, but they can be slow and easily distracted by too much text. This paper asks a simple, bold question: Can we teach these AI brains to ignore the noise and become lightning-fast experts at reading the raw numbers of the market to trade in real-time?

The paper introduces QuantHarness, a new system designed to answer that question. Think of QuantHarness not as a single trader, but as a tiny, high-speed trading team of four specialized robots, all working together inside a computer. Unlike other AI traders that spend hours reading news headlines, this team only looks at the "candlestick" charts—the raw, colorful bars that show the opening, high, low, and closing prices of an asset over a short period.

The team is made up of four distinct agents, each with a specific job:

The Indicator Agent is the math whiz. It takes the raw price bars and quickly calculates famous "speedometers" like RSI and MACD. These tools tell the team if the market is moving too fast, if it's tired, or if it's about to turn around.
The Pattern Agent is the artist. It looks at the shape of the price bars and tries to spot familiar drawings, like a "V" shape for a quick recovery or a triangle that suggests a big explosion is coming. It turns the squiggly lines into recognizable shapes.
The Trend Agent is the navigator. It draws lines on the chart to see if the market is generally heading up, down, or just stuck in traffic. It checks if the price is bouncing off a floor (support) or hitting a ceiling (resistance).
The Risk Agent is the safety officer. Before anyone makes a move, this agent calculates exactly how much money could be lost and sets a "stop-loss" (a point where you quit to save money) and a "take-profit" (a point where you cash out).

Finally, a Decision Agent acts as the captain. It listens to reports from all four teammates, weighs the evidence, and makes the final call: Buy (Long) or Sell (Short). The system is designed to work in the "high-frequency" zone, meaning it makes decisions in hours or even minutes, not days.

The researchers tested this team on nine different financial markets, including Bitcoin, crude oil, and major stock indices like the S&P 500. They compared QuantHarness against a random guesser, a simple math model, and a standard machine learning model. The results were promising: in these tests, the QuantHarness team consistently made more accurate predictions about which way the market would move in the next few hours. For example, on the Nasdaq futures, it improved prediction accuracy by about 26.5% compared to a random guess. It also managed to limit losses better than the other methods in most cases.

However, the paper is careful not to claim this is a magic money machine. The authors note that the system works best on 1-hour and 4-hour timeframes. When they tried to use it on ultra-short timeframes (like 1 to 15 minutes), the system got confused by the "noise" of the market, and its accuracy dropped. They also admit that the system isn't truly "real-time" yet; because the AI needs to think and call its tools, there is a slight delay that might be too slow for the fastest, split-second trades.

Ultimately, the paper suggests that by combining the pattern-spotting power of AI with the speed and structure of raw price data, we can build trading systems that are not only faster but also easier to understand. Instead of a black box making mysterious decisions, QuantHarness gives a clear, written explanation for every trade, showing exactly which pattern or indicator triggered the move. It's a step toward making high-speed trading less like a gamble and more like a disciplined, transparent science.

Technical Summary: QuantHarness

Problem Statement

Recent advancements in Large Language Models (LLMs) have demonstrated significant capabilities in financial reasoning, particularly within multi-agent frameworks like TradingAgent and FINMEM. However, these existing systems are primarily designed for long-horizon investment tasks, relying heavily on unstructured textual inputs such as news articles, social media sentiment, and earnings reports. This reliance presents two critical limitations for High-Frequency Trading (HFT):

Latency and Lag: Textual signals often lag behind price discovery, as markets incorporate information into prices before it is reflected in news or reports.
Noise and Structure: Unstructured text is noisy and difficult to validate, whereas HFT requires rapid, precision-critical decisions driven by structured, short-horizon signals.

Current LLM-based financial frameworks are ill-suited for the speed and structured nature of HFT, which depends on technical indicators, chart patterns, and trend features encoded directly in Open-High-Low-Close (OHLC) data. There is a lack of frameworks that align LLM reasoning with structured price-based signals for real-time, high-frequency algorithmic trading.

Methodology: QuantHarness

To address this gap, the authors introduce QuantHarness, the first multi-agent LLM framework explicitly designed for high-frequency algorithmic trading. Unlike prior works that incorporate external sentiment, QuantHarness operates solely on price-derived market signals (OHLC data), adhering to the Efficient Market Hypothesis which posits that prices aggregate all available information.

The system is built on LangGraph and decomposes the trading process into four specialized, collaborative agents:

IndicatorAgent:
- Function: Transforms raw OHLC sequences into structured quantitative signals.
- Mechanism: It computes five standard technical indicators: Relative Strength Index (RSI), Moving Average Convergence Divergence (MACD), Rate of Change (ROC), Stochastic Oscillator (STOCH), and Williams %R.
- Role: Provides a noise-resistant summary of market momentum and sentiment, abstracting low-level price data into high-level features for downstream reasoning.
PatternAgent:
- Function: Performs visual and structural multi-modal reasoning to detect chart formations.
- Mechanism: It utilizes LLM-bound tools to generate simplified candlestick charts from raw data. The agent then compares these visualizations against a library of canonical patterns (e.g., Double Bottoms, Head and Shoulders, Triangles) described in natural language.
- Role: Identifies geometric structures such as peaks, troughs, and consolidations, bridging raw chart data with high-level reasoning.
TrendAgent:
- Function: Analyzes directional momentum and trend dynamics.
- Mechanism: It fits Ordinary Least Squares (OLS) regression lines to recent local highs (resistance) and lows (support) to generate a trend channel. It calculates the average slope ( $\kappa_t$ ) to classify trends as Uptrend, Downtrend, or Sideways.
- Role: Determines the structural context of the market, assessing whether detected patterns align with the prevailing trend or signal a reversal.
RiskAgent & DecisionAgent:
- RiskAgent: Integrates outputs from the previous agents into a unified risk-reward framework. It sets a fixed stop-loss ( $\rho = 0.0005$ ) and computes a dynamic take-profit level based on an LLM-predicted risk-reward ratio ( $r \in [1.2, 1.8]$ ).
- DecisionAgent: Acts as the execution layer. It aggregates signals from the Indicator, Pattern, and Trend agents to forecast the market direction for the next three candlesticks. It outputs a structured decision (LONG or SHORT), a concise justification, and a risk-reward ratio. The system prohibits "HOLD" decisions, forcing a directional choice.

Key Contributions

First HFT-Oriented LLM Framework: QuantHarness is the first multi-agent system designed specifically for high-frequency trading, operating directly on OHLC data without reliance on lagging textual inputs.
Specialized Multi-Agent Architecture: The decomposition of trading into Indicator, Pattern, Trend, and Risk agents allows for modular, interpretable, and cross-validated decision-making, mimicking institutional trading desk workflows.
Structured Reasoning over Unstructured Text: The paper demonstrates that aligning LLM reasoning with structured price signals (technical indicators and chart geometry) is a viable path for traceable, real-time decision systems in HFT.
Interpretability: Unlike traditional "black box" algorithmic strategies, QuantHarness generates natural-language rationales for every trade, providing transparency into the decision-making process.

Experimental Results

The authors evaluated QuantHarness on a benchmark comprising nine financial instruments (including Bitcoin, Nasdaq futures, S&P 500, and Crude Oil) across 1-hour and 4-hour time resolutions. The system was tested without prior demonstrations or supervised fine-tuning, relying on zero-shot reasoning.

Performance Metrics: QuantHarness was compared against Random Baselines, Linear Regression (LR), XGBoost, and the TradingAgent framework.
Directional Accuracy: QuantHarness consistently outperformed baselines. For instance, on the Nasdaq (NQ) 4-hour interval, it achieved a 26.5% improvement in directional accuracy over the random baseline. In the 1-hour timeframe, it showed pronounced gains in assets like SPX, QQQ, and BTC.
Risk-Adjusted Returns: The system achieved the highest Rate of Return (RoR) metrics ( $R_{cc}$ , $R_{max}$ , $R_{min}$ ) in the majority of tested assets, demonstrating an ability to capture profitable trends while effectively limiting downside risk.
Generalization: Rolling-window validation showed robust generalization, achieving up to 80% directional accuracy in continuous short-term prediction tasks on the S&P 500 (SPX).

Significance and Claims

The paper claims that QuantHarness validates the hypothesis that coupling structured trading signals with LLM-based reasoning creates a superior framework for high-frequency markets compared to text-driven approaches.

Viability of Price-Only Analysis: The results suggest that technical analysis based purely on price data is sufficient for high-frequency decision-making, as prices inherently reflect all relevant market information.
Traceability: The framework offers a degree of traceability and interpretability often missing in traditional algorithmic strategies, as every decision is accompanied by a language-native rationale derived from specific indicators and patterns.
Scalability: The multi-agent structure provides a scalable foundation for future real-time, data-efficient financial systems that do not require external sentiment analysis or human supervision.

Limitations

The authors modestly acknowledge current constraints:

Micro-Horizon Accuracy: Predictive precision degrades on ultra-short timeframes (1–15 minutes) due to market noise and rapid regime shifts.
Latency: The current architecture is not truly real-time for sub-minute trading, as the inference cycle involving multiple LLM calls and tool bindings introduces latency that may exceed the exploitable window for 1-minute opportunities. Future work suggests streamlining tool orchestration or moving logic to edge models to address this.