Research map / Multi-agent systems and games / LLM agents

Trade in Minutes! Rationality-Driven Agentic System for Quantitative Financial Trading

ICLR 2026 · Multi-agent systems and games › LLM agents · approach: Agent data synthesis and behavior studies

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Presents TiMi, a multi-agent LLM system that develops trading strategies offline (DeepSeek-V3, Qwen2.5-Coder-32B, DeepSeek-R1 as specialized agents), refines them via mathematical reflection, and deploys CPU-only programmatic trading bots for minute-level trading.

Problem

Existing LLM financial trading agents simulate anthropomorphic roles that introduce emotional biases, rely on peripheral news, and need continuous LLM inference during deployment, which limits latency and efficiency.

Approach

Decouple strategy development (LLM agents for semantic analysis, code, and math reasoning) from deployment by compiling strategies into layered Python trading bots that run without LLM inference, optimized in a closed loop where a reflection agent turns risk scenarios into linear constraints over bot parameters and escalates fixes across parameter, function, and strategy layers.

Key results

• Live trading across 213 pairs: TiMi achieves ARR 6.4% (US stock index futures), 8.0% (mainstream coins), and 13.7% (altcoins) with SR 0.74, 0.79, 0.86 respectively, outperforming all listed baselines on ARR.
• Backtest on 2024 altcoin futures: TiMi reaches ARR 23.7% and SR 1.27, with Sortino 0.91 vs 0.67 for the best baseline (PatchTST).
• Ablation on 2024 crypto: removing the feedback reflection agent drops ARR from 20.9% to 1.1%, while removing the strategy adaptation agent raises MDD from 15.3% to 28.4%.

Surprising findings

• Parameter-only optimization yields backtest gains but produces logically inconsistent code that fails in live deployment.
• Semantic-only reflection keeps runtime stable but leaves profitability stagnant (ARR 1.1%).

Limitations

• Optimization stage required, so zero-shot performance on new markets is limited.
• Ethical concerns about market fairness and widening institutional-retail gaps.

Experimental setup

Datasets: 2024 historical market data (US stock index futures, cryptocurrency futures) 2025 live market data (Jan-Apr)

Benchmarks: 2024 historical backtest (US stock index futures, mainstream and altcoin futures) 2025 live trading (Jan-Apr)

Metrics: ARR Sharpe ratio Maximum drawdown Sortino ratio capital utilization

Baselines: MACD Momentum Grid Trading Pairs Trading TSMOM LSTM DDPG PatchTST FinGPT FinMem TradingAgents

Builds on: FinMem TradingAgents FinGPT

Closest papers by meaning

• HiveMind: Contribution-Guided Online Prompt Optimization of LLM Multi-Agent Systems (AAAI 2026 · similarity 0.87)
• FinCon: A Synthesized LLM Multi-Agent System with Conceptual Verbal Reinforcement for Enhanced Financial Decision Making (NeurIPS 2024 · similarity 0.87)
• R&D-Agent-Quant: A Multi-Agent Framework for Data-Centric Factors and Model Joint Optimization (NeurIPS 2025 · similarity 0.86)
• TwinMarket: A Scalable Behavioral and Social Simulation for Financial Markets (NeurIPS 2025 · similarity 0.86)
• MarS: a Financial Market Simulation Engine Powered by Generative Foundation Model (ICLR 2025 · similarity 0.85)
• Win Fast or Lose Slow: Balancing Speed and Accuracy in Latency-Sensitive Decisions of LLMs (NeurIPS 2025 · similarity 0.85)