Source-linked AI summary
QuantHarness: Price-Driven Multi-Agent LLMs for High-Frequency Trading
Fei Xiong, Xiang Zhang, Aosong Feng, Siqi Sun, Chenyu You
TL;DR
Short-horizon HFT needs rapid decisions from structured price signals, while most LLM financial frameworks use textual inputs. QuantHarness coordinates four specialized agents over price-derived signals, and the paper reports stronger results than baselines across 1-hour and 4-hour bars.
Problem
Most LLM-based financial frameworks rely on textual inputs that typically lag price discovery and are noisy, while the authors identify no prior LLM-based HFT framework operating directly on OHLC data.
Source: §1 · pp. 1–2Method
Results
Across 1-hour and 4-hour bars, QuantHarness consistently outperforms baselines on directional accuracy and return-based metrics, with rolling-window validation reaching up to 80% directional accuracy.
Source: §1 · p. 2Takeaways & Limitations
The authors conclude that price-grounded agent collaboration is a viable foundation for future real-time, data-efficient financial systems without external sentiment or supervision.
Source: §6 · p. 11Prediction accuracy degrades on approximately 1–15-minute candles, and inference latency can exceed the exploitable window for a 1-minute opportunity.
Source: §7 · p. 11
Abstract
from arXiv · showhide
Recent advances in Large Language Models (LLMs) have shown remarkable capabilities in financial reasoning and market understanding. Multi-agent LLM frameworks such as TradingAgent and FINMEM augment these models to long-horizon investment tasks by leveraging fundamental and sentiment-based inputs for strategic decision-making. However, these approaches are ill-suited for the high-speed, precision-critical demands of High-Frequency Trading (HFT). HFT typically requires rapid, risk-aware decisions driven by structured, short-horizon signals, such as technical indicators, chart patterns, and trend features. These signals stand in sharp contrast to the long-horizon, text-driven reasoning that characterizes most existing LLM-based systems in finance. To bridge this gap, we introduce QuantHarness, the first multi-agent LLM framework explicitly designed for high-frequency algorithmic trading. The system decomposes trading into four specialized agents--Indicator, Pattern, Trend, and Risk--each equipped with domain-specific tools and structured reasoning capabilities to capture distinct aspects of market dynamics over short temporal windows. Extensive experiments across nine financial instruments, including Bitcoin and Nasdaq futures, demonstrate that QuantHarness consistently outperforms baseline methods, achieving higher predictive accuracy at both 1-hour and 4-hour trading intervals across multiple evaluation metrics. Our findings suggest that coupling structured trading signals with LLM-based reasoning provides a viable path for traceable, real-time decision systems in high-frequency financial markets.
1 Introduction
The introduction motivates price-based signals as a better fit for short-horizon trading than lagging, noisy textual inputs. It presents QuantHarness as a four-agent HFT framework and reports stronger benchmark performance at 1-hour and 4-hour intervals.
1 Introduction
- Existing LLM financial frameworks rely mainly on textual inputs, which typically lag price discovery and are noisy, unstructured, and difficult to validate.The introduction argues that short-horizon dynamics are already encoded in OHLC bars.Source: §1 · p. 1
- The paper identifies a gap in LLM-based HFT: to the authors’ knowledge, no prior framework operates directly on OHLC data.It proposes aligning LLM reasoning with structured price-based signals.Source: §1 · p. 2
- QuantHarness decomposes high-frequency trading into IndicatorAgent, PatternAgent, TrendAgent, and RiskAgent, producing traceable language-native rationales alongside execution.The agents capture complementary technical-analysis dimensions, from indicators and chart formations to directional bias and risk–reward integration.Source: §1 · p. 2, Figure 1 · p. 2
- Across a multi-asset benchmark, QuantHarness outperforms baselines on directional accuracy and return-based metrics at 1-hour and 4-hour bars, reaching up to 80% directional accuracy in rolling-window validation.The reported gains are particularly pronounced in equity markets.Source: §1 · p. 2
2 Related Works
Prior work spans LLM-based financial agents and conventional indicator- and pattern-based trading systems. The section frames their textual-input, resolution, and noisy-market constraints as motivation for combining structured price signals with LLM reasoning.
Agent-Based LLMs for Financial Decision-Making
- FINCON uses a manager–analyst hierarchy trained through verbal reinforcement learning, while TradingAgents models institutional workflows but prioritizes interpretability over HFT’s low-latency, price-driven logic.RD-Agent(Q) shifts toward structured, data-centric signals and factor–model co-optimization, but remains limited to daily-resolution strategies and slower research-feedback cycles.Source: §2 · pp. 2–3
Quantitative Trading Based on Indicators and Patterns
- Technical indicators, momentum strategies, and curated patterns support interpretable intraday trading, but prior systems face overfitting, researcher-bias, and volatile or noisy-market challenges (Lo et al., 2000; Chen and Chen, 2016).The section motivates fusing structured price signals with LLM reasoning for more adaptive and interpretable systems.Source: §2 · p. 3
3 QuantHarness
QuantHarness combines price-derived technical analysis with prompt-structured LLM reasoning in specialized agents. Its modules transform OHLC data into indicators, chart-pattern interpretations, trend structure, and risk-aware trade decisions.
3 QuantHarness
- QuantHarness uses collaborative, specialized agents and prompt-structured LLM reasoning, operating solely on price-derived signals rather than external news or social-media sentiment.The framework is presented as bridging traditional HFT and multi-agent LLM systems; its four agents cover numerical indicators, geometric patterns, directional momentum, and integrated decision-making.Source: §3 · p. 3
3.1 IndicatorAgent
- IndicatorAgent maps OHLC sequences to structured quantitative signals, using RSI, MACD, RoC, STOCH, and WILLR to summarize momentum, turning points, and potential reversals.It interprets the indicators in context rather than applying fixed rules.Source: §3.1 · p. 4
3.2 PatternAgent
- PatternAgent detects visual features in candlestick charts and compares market structure with a language-described pattern library, assessing matches in their market context.Its multimodal perspective addresses cases where numerical indicators become unclear during stalls or regime changes.Source: §3.2 · pp. 4–5
3.3 TrendAgent
- TrendAgent fits resistance and support lines to recent highs and lows, then uses their slopes and geometric relationships to assess direction, strength, breakout zones, and trend stability.The agent also distinguishes patterns consistent with the prevailing trend from potential reversals or non-directional congestion.Source: §3.3 · p. 5
3.4 RiskAgent and DecisionAgent
- RiskAgent sets a fixed stop-loss and an LLM-predicted take-profit ratio; DecisionAgent aggregates upstream signals and outputs LONG or SHORT decisions with justification and risk–reward ratio.DecisionAgent forecasts the next three candlesticks without a hold option and proceeds when most upstream signals align and receive confirmation.Source: Figure 2 · p. 4, §3.4 · pp. 5–6, Figure 3 · p. 5
4 Experiments
Experiments test autonomous short-term trading against random, linear-regression, and XGBoost baselines on sampled OHLC benchmarks. Evaluation combines directional scoring with return measures under a three-candlestick risk-control setup.
Evaluation Setup
- QuantHarness generates short-term forecasts autonomously, using recent candlestick data and basic context without prior demonstrations or supervised fine-tuning.Its structured trade report includes a LONG or SHORT prediction, a brief rationale, and an estimated risk–reward ratio; the directional decision and ratio are used for quantitative evaluation.Source: §4 · p. 6
Benchmark Construction and Evaluation Protocol
- The benchmark samples 100 segments per asset from 5,000 historical bars at 1-hour and 4-hour resolutions, withholding the final three candles to prevent test-time leakage.QuantHarness is compared with random, linear-regression, and XGBoost approaches; Table 2 presents these methods across trading symbols, while Table 1 covers Sharpe, Sortino, CumReturn, and MaxDrawdown.Source: §4 · p. 6, Table 1 · p. 8, Table 2 · p. 9
Evaluation Metrics
- Directional accuracy compares each LONG or SHORT forecast with the next three candlesticks, scoring up to three correct candle movements per test.The average score is computed over samples and aligns with Mean Directional Accuracy.Source: §4 · p. 6
Case Study of Agent Reasoning
- The PatternAgent case identifies a descending triangle from lower highs and flat support near 78, summarizing its bearish bias and triangular convergence.Dashed edges and EMA overlays are visual aids; the classification is derived solely from bar geometry.Source: Figure 4 · p. 7, §4 · p. 7
Evaluation Metrics
- Risk-constrained returns exit at the first of the next three candles to hit a stop-loss or reward threshold, using a fixed 0.05% stop-loss.Rmax and Rmin bound the most favorable and adverse price movements over that interval under the issued decision.Source: §4 · p. 7
5 Results
Across the reported results, QuantHarness leads the baselines on key accuracy and return measures, with especially strong results in equity markets. The paper also illustrates short-horizon SPX forecasts and PatternAgent’s interpretation of a descending triangle.
5.1 Main Results
- QuantHarness accuracy outperforms all three baselines across eight markets, with a 26.5% increase over the random baseline on NQ; it also leads Rcc in 7 of 8 assets and Rmax in 6 of 8.Rmin ranks among the least negative values across most assets, and the authors report generalization across asset classes at the 4-hour timeframe.Source: §5.1 · pp. 9–10, Figure 5 · p. 10
5.2 Case Study on Continuous Short-Term Prediction
- On a randomly selected 100-bar SPX segment, QuantHarness achieves 80% directional accuracy across 10 overlapping windows offset by five bars.The system issued correct sell signals during resistance stalls and negative momentum shifts, and buy signals on momentum flips, support bounces, and recovery patterns; two errors involved incomplete patterns and premature bullish calls.Source: §5.2 · pp. 10–11, Figure 6 · p. 10
5.3 Case Studies of Agent Reasoning
- PatternAgent’s descending-triangle case separates lower highs from flat support and reports a bearish breakdown bias with triangular convergence.The example illustrates how it decomposes raw price action into interpretable features.Source: §5.3 · p. 11
6 Conclusion
QuantHarness presents price-grounded agent decomposition as a viable basis for accurate, transparent, risk-aware HFT decisions across diverse markets.
6 Conclusion
- Specialized price-grounded agents support accurate, transparent, and risk-aware decisions for high-frequency trading.The conclusion frames the approach as viable across diverse markets and as a potential foundation for future financial systems operating without external sentiment or supervision.Source: §6 · p. 11
- Cross-agent validation and specialization promote robustness while the multi-agent structure enhances interpretability.The paper describes these as benefits of dividing trading among agents with complementary roles.Source: §6 · p. 11
- The paper proposes structured agent collaboration grounded purely in price data as a scalable foundation for future real-time, data-efficient financial systems.This proposed scope excludes external sentiment and supervision.Source: §6 · p. 11
7 Limitations and Future Work
QuantHarness is constrained by reduced accuracy on ultra-short candles and inference latency that can outlast the window for exploiting a one-minute opportunity.
7 Limitations and Future Work
- The zero-shot LLM ensemble loses predictive accuracy on 1–15-minute candles relative to 30-minute–4-hour bars.The authors attribute this degradation to noise and rapid regime shifts that make transient spikes difficult to separate from tradable signals.Source: §7 · p. 11
- At ultra-short horizons, noisy price series and rapid regime shifts challenge separation of transient spikes from tradable signals.These are the conditions the authors identify alongside the observed accuracy degradation.Source: §7 · p. 11
- Each inference cycle includes an LLM call and several bound tools, so latency can exceed the exploitable window of a one-minute opportunity.The authors suggest streamlining orchestration, caching intermediate features, or moving critical logic to lighter edge models as possible directions.Source: §7 · p. 11
A Why Technical Analysis Alone Can Suffice for Trading
The paper motivates price-only technical analysis by treating prices as records of market reactions and by relying on recurring, non-random patterns in price movement.
A Why Technical Analysis Alone Can Suffice for Trading
- Technical analysis assumes relevant information is reflected in market prices, which record buying and selling responses to new developments.On this premise, price movement can indicate collective reactions without explicitly processing the underlying information.Source: A Why Technical Analysis Alone Can… · p. 16
- Technical analysis assumes prices are not entirely random and that rising or falling trends can persist before reversing.QuantHarness is described as capturing market patterns under this price-movement assumption.Source: A Why Technical Analysis Alone Can… · p. 16
- Repeated price patterns are attributed to stable behavioral tendencies, allowing predictions from recognizable structures without identifying each movement’s cause.The passage connects this repetition to technical systems’ use of recurring peaks, dips, and reversals.Source: A Why Technical Analysis Alone Can… · p. 16
- Price changes can precede official disclosures, so technical analysis can respond to observed market behavior without access to its underlying cause (Cao et al., 2025).The passage presents this as a way to respond before delayed or incomplete public disclosures.Source: A Why Technical Analysis Alone Can… · p. 16
- The paper characterizes price-only forecasting as a practical, self-contained approach based on information being incorporated into prices and human reactions tending to follow consistent patterns.Under these assumptions, the agent forecasts price direction without external inputs.Source: A Why Technical Analysis Alone Can… · p. 16
B Why LLMs Are Well-Suited for Price-Based Technical Analysis
This section argues that price-based technical analysis aligns with LLM capabilities for structured reasoning, chart perception, tool use, and agent coordination. It describes specialized agents that turn OHLC data into indicator, pattern, trend, and decision outputs.
B Why LLMs Are Well-Suited for Price-Based Technical Analysis
- Technical analysis combines heterogeneous symbolic signals into short-horizon judgments, for which LLMs can provide language-native rationales and machine-checkable action schemas.The passage describes indicator combinations such as MACD, RSI, and resistance interactions as inputs to decisions that are human-auditable and executable.Source: B Why LLMs Are Well-Suited for… · p. 17
C.1 Indicator Agent
- IndicatorAgent is prompted to interpret recent OHLC data through tool-assisted analysis under latency constraints.The prompt dynamically instantiates the data and timeframe so outputs can adapt across market conditions.Source: §C.1 · p. 18
C.2 Pattern Agent
- PatternAgent transforms OHLC input into a visual chart and evaluates it against canonical formations, explaining matches through structure, direction, and symmetry.Its textual pattern library includes shapes such as U- and W-forms and triangles.Source: §C.2 · p. 19
C.3 Trend Agent
- TrendAgent analyzes tool-generated charts and price interactions with support and resistance lines to predict an upward, downward, or sideways short-term trend.Its structured output includes the prediction, reasoning, and signals.Source: §C.3 · p. 20
C.4 Decision Agent
- The decision prompt calls for aligned, confirmed signals, favors strong momentum, and resolves conflicts using the stronger, more recent confirmation.For consolidation or unclear setups, it also directs the agent to defer to the dominant trendline slope and suggest a risk–reward ratio between 1.2 and 1.8.Source: §C.4 · p. 21
- DecisionAgent integrates indicator, pattern, and trend reports into a directional decision, prioritizing consistent evidence and providing a structured justification and estimated risk–reward ratio.The prompt directs it to avoid speculative outputs and weigh the reports together.Source: §C.4 · p. 21
C.5 Pattern Tool Sample Output / C.6 Trend Tool Sample Output
- A PatternAgent sample shows contracting lower highs and higher lows in NQ intraday candles, while a TrendAgent sample depicts a symmetrical-triangle wedge between converging support and resistance.The examples describe these structures as potential precursors to a breakout once a boundary is breached.Source: Figure 7 · p. 22, Figure 8 · p. 23
QuantHarness:
- The interface supports asset, timeframe, and analysis-window selection, live input, and local execution; by default it uses the most recent 40–50 price bars.Its final directional decision is accompanied by a forecast horizon, risk–reward ratio, and textual justification grounded in pattern recognition.Source: Figure 9 · p. 24, Figure 10 · p. 25
E.1 Benchmark Construction
The benchmark is designed to compare trading systems reproducibly across asset classes and trading environments. It uses standardized 1-hour and 4-hour OHLC data and fixed-length evaluation segments drawn from liquid, diverse instruments.
E.1 Benchmark Construction
- The benchmark supports systematic generalization tests across asset classes and trading environments, controlled model comparisons, and reproducibility.It uses 1-hour and 4-hour OHLC data, with 100 sampled segments per asset and the final three candlesticks withheld during prediction.Source: §E.1 · p. 28, §E.2 · p. 28
E.2 Data Collection and Asset Selection
- The asset set spans cryptocurrency, crude oil, equity index futures, ETFs or spot indices, and volatility indices, with both stable and volatile instruments.Examples include BTC/USD, CL, ES, NQ, QQQ, SPX, DJI, and VIX; the selection emphasizes widely traded, liquid assets.Source: §E.2 · p. 28
- Each asset contributes 5,000 historical bars; fixed bar counts mean lower-frequency instruments can cover as much as ten years of history.Each of 100 evaluation segments contains 100 consecutive candlesticks, with the final three reserved to validate predictions.Source: §E.2 · p. 28
E.3 Benchmark Asset Properties
The benchmark asset tables document market type, collection dates, and sampled bar counts. The asset descriptions span varied markets, while the standardized resolution and segment format support reproducible comparisons across stable and volatile instruments.
E.3 Benchmark Asset Properties
- Tables 3 and 4 describe benchmark assets by name, market type, data-window start and end dates, and total sampled bars.Table 3 covers 4-hour assets and Table 4 covers 1-hour assets; the captions do not provide the individual table values.Source: Table 3 · p. 28, Table 4 · p. 29
E.4 Benchmark Assets Detail
- The asset descriptions include Bitcoin, crude oil, ES and NQ futures, QQQ, SPX, DJI, VIX, and the DAX Volatility Index.They identify each instrument’s market context, including Bitcoin’s continuous trading, ES’s link to the S&P 500, and DAX volatility as a measure of expected short-term German equity volatility.Source: §E.4 · p. 29
E.5 Conclusion
- The benchmark’s standardized data resolution and segment format enable fair, reproducible comparisons while retaining variety across real-world markets.Including stable, trend-following assets alongside more volatile instruments allows performance stress testing.Source: §E.5 · p. 29
F Case Studies
The case studies show agents extracting geometric and indicator-based signals to characterize market direction and risk. The indicator explanation describes MACD’s EMA-based construction, and the paper states that LLMs assisted with manuscript writing rather than the research itself.
F Case Studies
- A declining sequence of swing highs above clustered lows near 78 forms a Descending Triangle, with the narrowing gap marking range compression.The agent’s summaries identify lower highs, relatively flat support, and a bearish bias with heightened breakdown probability after repeated support tests.Source: F Case Studies · pp. 29–30
- In the DJI trend case, higher highs and lows, an upward resistance line, and flat untested support combine into a bullish assessment near resistance.The agent’s summaries flag a pullback after resistance contact and frame breakout continuation or a tactical move toward support as possibilities.Source: F Case Studies · p. 30, Figure 13 · p. 30
- The indicator case combines positive momentum signals with overbought oscillators, yielding a “bullish but extended” reading and caution about possible pullbacks or consolidation.MACD and ROC indicate upward momentum, while Stochastic and Williams %R indicate overbought conditions.Source: F Case Studies · p. 31
G Indicator Explanation
- MACD compares fast and slow exponential moving averages, then smooths the momentum sequence into a signal line whose crossover indicates upward or bearish pressure.The paper describes a 9-period EMA signal line and states that a crossover can capture mid-term trend shifts while filtering high-frequency price noise.Source: G Indicator Explanation · p. 32
H The Use of Large Language Models (LLMs)
- The authors report using GPT-4o to refine manuscript wording and state that they retained responsibility for the research questions, methods, and findings.Source: H The Use of Large Language Models… · p. 32