The paper presents TiMi (Trade in Minutes), a rationality-driven multi-agent system for quantitative financial trading that addresses limitations of existing LLM-based trading agents, such as emotional biases from anthropomorphic role-playing and high computational costs from continuous inference. TiMi architecturally decouples strategy development (policy and optimization stages) from minute-level deployment, enabling complex reasoning offline while executing lightweight trading bots in real time. The system comprises four specialized agents: a macro analysis agent that identifies market patterns using technical indicators, a strategy adaptation agent that customizes strategies for specific trading pairs, a bot evolution agent that implements trading bots via layered programming (strategy, function, parameter layers), and a feedback reflection agent that performs closed-loop optimization by formulating mathematical problems (e.g., linear programming) from risk scenarios to refine parameters hierarchically.

Empirical evaluations across over 200 trading pairs in U.S. stock index futures and cryptocurrency markets demonstrate TiMi's superior performance. In backtesting on 2024 data, TiMi achieves the highest Sharpe ratios (1.25 for mainstream coins, 1.27 for altcoins) and competitive annual returns (up to 23.7% in altcoins), outperforming quantitative methods (e.g., MACD, TSMOM), ML/RL approaches (e.g., LSTM, DDPG), and LLM-based agents (e.g., FinGPT, TradingAgents). Live trading from January to April 2025 confirms robustness, with TiMi achieving 13.7% ARR and 1.14 Sharpe in altcoins while maintaining low drawdowns. The system also demonstrates scalability (213 trading pairs) and data efficiency (requiring only 4 hours of market data), highlighting its practical applicability for retail and institutional traders seeking automated, rational, and efficient trading solutions.