The List / Agents and Multi-Agent Systems / paper
Trade in Minutes! Rationality-Driven Agentic System for Quantitative Financial Trading
Abstract
Recent advancements in large language models (LLMs) and agentic systems have shown exceptional decision-making capabilities, revealing significant potential for autonomic finance. Current financial trading agents predominantly simulate anthropomorphic roles that inadvertently introduce emotional biases and rely on peripheral information, while being constrained by the necessity for continuous inference during deployment. In this paper, we pioneer the harmonization of strategic depth in agents with the mechanical rationality essential for quantitative trading. Consequently, we present TiMi (Trade in Minutes), a rationality-driven multi-agent system that architecturally decouples strategy development from minute-level deployment. TiMi leverages specialized LLM capabilities of semantic analysis, code programming, and mathematical reasoning within a comprehensive policy-optimization-deployment chain. Specifically, we propose a two-tier analytical paradigm from macro patterns to micro customization, layered programming design for trading bot implementation, and closed-loop optimization driven by mathematical reflection. Extensive evaluations across 200+ trading pairs in stock and cryptocurrency markets empirically validate the efficacy of TiMi in stable profitability, action efficiency, and risk control under volatile market dynamics.
In one sentence
Summary
The paper addresses the limitations of existing LLM-based financial agents, which often suffer from emotional biases, reliance on noisy peripheral data, and high inference latency due to continuous multi-agent reasoning. The authors propose TiMi, a system that harmonizes strategic depth with mechanical rationality by architecturally decoupling the offline policy and optimization stages from the online deployment stage. This design allows for complex, time-consuming reasoning to occur offline, while the live deployment relies on lightweight, pre-compiled trading bots, ensuring minute-level execution efficiency. TiMi employs a multi-agent architecture comprising a macro analysis agent for pattern recognition, a strategy adaptation agent for pair-specific customization, a bot evolution agent for code generation, and a feedback reflection agent for mathematical optimization. The system utilizes specialized LLMs: DeepSeek-V3 for semantic analysis, Qwen2.5-Coder for programming, and DeepSeek-R1 for mathematical reasoning. Strategies are transformed into layered trading bots (strategy, function, parameter levels) and optimized via closed-loop feedback where risk scenarios are formulated as linear programming problems to solve for optimal parameters. Experiments cover 200+ pairs in U.S. stock index futures and cryptocurrency markets, comparing against quantitative, ML/RL, and other LLM-agent baselines. Empirical results demonstrate that TiMi achieves competitive Annual Rate of Return (ARR) and superior Sharpe ratios, particularly in volatile altcoin markets, while maintaining low action latency comparable to traditional quantitative methods. The decoupling mechanism significantly reduces computational costs during deployment. Limitations include the reliance on specific LLM capabilities for the offline stages and the potential for overfitting during the optimization phase if risk scenarios are not diverse enough. The system is most relevant for quantitative trading infrastructure rather than discretionary investment advice.
Research problem
Current LLM-powered financial trading agents often simulate anthropomorphic roles, introducing emotional biases and relying on noisy, peripheral information. Furthermore, they suffer from high inference latency due to continuous multi-agent reasoning during deployment, which is detrimental to high-frequency or minute-level trading. The paper seeks to address these issues by creating a system that combines the strategic depth of LLMs with the mechanical efficiency and rationality required for quantitative trading.
Core contributions
- Introduction of TiMi, a rationality-driven multi-agent system that decouples strategy development from real-time deployment.
- A two-tier analytical paradigm that moves from macro market patterns to micro pair-specific customization.
- A layered programming design for trading bot implementation, separating strategy, function, and parameter layers.
- A closed-loop optimization system driven by mathematical reflection, using linear programming to refine parameters based on feedback.
- Empirical validation across 200+ trading pairs in stock and crypto markets, demonstrating superior risk-adjusted returns and low latency.
Method
- Multi-agent architecture with specialized LLMs: DeepSeek-V3 (semantic), Qwen2.5-Coder (code), DeepSeek-R1 (math).
- Decoupled pipeline: Offline policy/optimization stages followed by online deployment.
- Macro analysis agent identifies general strategies from technical indicators.
- Strategy adaptation agent customizes strategies for specific trading pairs.
- Bot evolution agent generates code for trading bots based on layered programming laws.
- Feedback reflection agent formulates linear programming problems to optimize parameters based on simulation feedback.
- Minute-level grid trading strategy with dynamic order placement and position management.
- Risk control via mathematical constraints on position size and order boundaries.
Results & findings
- TiMi achieves an ARR of 8.9% and SR of 0.84 in U.S. stock index futures backtesting.
- In live trading, TiMi achieved ARR of 6.4%, 8.0%, and 13.7% for stock, mainstream coin, and altcoin markets respectively.
- TiMi demonstrated a Sharpe Ratio of 1.27 in altcoin backtesting, outperforming most baselines.
- Action latency for TiMi was 137ms, significantly faster than LLM-agent baselines like FinGPT (18,645ms) and FinMem (25,071ms).
- TiMi matched the market coverage (213 pairs) of quantitative methods while surpassing other LLM agents.
- The feedback reflection agent significantly improved performance, with optimized bots achieving >20% return compared to near break-even for unoptimized bots.
- Strategy adaptation agent improved cross-pair stability, reducing the standard deviation of returns.
Data & experiments
- Backtesting on 2024 historical data for U.S. stock index futures and cryptocurrency markets.
- Live trading experiments from January to April 2025 across U.S. stock index futures, mainstream coin futures, and altcoin futures.
- Evaluation across 200+ trading pairs, comparing TiMi against quantitative methods (MACD, Momentum, Grid), ML/RL methods (LSTM, DQN, DDPG), and other LLM agents (FinGPT, FinMem).
- Metrics include Annual Rate of Return (ARR), Sharpe Ratio (SR), Maximum Drawdown (MDD), action latency, and capital utilization rate.
- Ablation studies on agent components and strategy adaptation effectiveness.
Limitations & caveats
- The system relies on the capabilities of specific LLMs for the offline stages, which may vary.
- The optimization phase may overfit to historical or simulated risk scenarios if not diverse enough.
- The system's performance in live trading may be affected by external factors like network latency and exchange API changes.
- The paper does not discuss the potential for adversarial attacks on the LLM components.
- The decoupling mechanism requires significant offline computational resources for strategy development and optimization.
Curator notes
Discovery status: round3_promoted_seed_for_round4. Source links: FinCon: A Synthesized LLM Multi-Agent System with Conceptual Verbal Reinforcement for Enhanced Financial Decision Making; FinGPT: Democratizing Internet-scale Data for Financial Large Language Models; When AI Meets Finance (StockAgent): Large Language Model-based Stock Trading in Simulated Real-world Environments