🤖 AI

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

This paper introduces TiMi (Trade in Minutes), a rationality-driven multi-agent system that decouples strategic development from deployment to provide efficient, mathematically optimized, and emotion-free quantitative trading across diverse financial markets.

Original authors: Zifan Song, Kaitao Song, Guosheng Hu, Ding Qi, Junyao Gao, Xiaohua Wang, Dongsheng Li, Cairong Zhao

Published 2026-02-10

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Zifan Song, Kaitao Song, Guosheng Hu, Ding Qi, Junyao Gao, Xiaohua Wang, Dongsheng Li, Cairong Zhao

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 you want to run a high-stakes professional kitchen that serves thousands of customers every minute. You have two ways to do it:

The "Human Chef" Method: You hire a world-class chef to stand at the stove. Every time an order comes in, the chef stops, thinks about the ingredients, contemplates the seasoning, and then cooks. This is how most current AI trading "agents" work. They are smart, but they are too slow. By the time the chef finishes thinking, the customer has left, and the food is cold.
The "TiMi" Method (The Paper's Approach): You hire a brilliant Executive Chef to work in a back office. They spend their time studying food trends, writing perfect, foolproof recipes, and testing them in a practice kitchen. Once the recipe is perfect, they hand it to a High-Speed Robot in the kitchen. The robot doesn't "think"—it just follows the recipe with lightning speed and perfect precision.

This paper introduces "TiMi" (Trade in Minutes), a system that works exactly like that second method for the stock and crypto markets.

The Three Big Problems TiMi Solves

The researchers noticed that current AI traders suffer from three "human" flaws:

Emotional Drama: Most AI agents try to "act" like humans (e.g., "I feel bullish about Bitcoin!"). This introduces human-like biases and emotions that lead to bad trades.
Distracted News-Watching: Many AIs spend too much time reading social media or news, which is often "noise" or fake news that can mislead them.
The "Thinking" Delay: If an AI has to "reason" through a complex problem every single time it wants to buy a stock, it will always be too late to catch the best price.

How TiMi Works: The "Brain vs. Brawn" Split

TiMi splits the work into three distinct stages to ensure it is both smart and fast:

1. The Strategy Stage (The Executive Chef)

Instead of just guessing, TiMi uses specialized AI "experts" to build a plan:

The Pattern Spotter: Looks at the big picture (e.g., "Is the whole market going up or down?").
The Customizer: Looks at specific details (e.g., "This specific coin is much more volatile than others; we need a different plan for it").

2. The Optimization Stage (The Taste Test)

Before any real money is spent, TiMi creates a "Trading Bot" (the Robot) and runs it through a simulator. It’s like a chef testing a new sauce in a small pot before serving it to a banquet. If the bot makes a mistake, a "Reflection Agent" looks at the math, figures out exactly why it failed, and rewrites the code to fix it.

3. The Deployment Stage (The High-Speed Robot)

This is the "secret sauce." Once the plan is perfected, the AI doesn't "think" anymore. It converts the strategy into pure, fast computer code. This code runs on a simple, lightning-fast loop. It can react to market changes in minutes, whereas other AIs are still stuck in the "thinking" phase.

Does it actually work?

The researchers tested TiMi on over 200 different trading pairs (stocks and crypto) and found:

It’s a Money Maker: It outperformed traditional math-based methods and even other "smart" AI agents, especially in the wild, unpredictable world of "Altcoins" (smaller cryptocurrencies).
It’s Calm Under Pressure: While other AIs might make huge profits one day and lose everything the next (like a gambler), TiMi was much more stable. It avoids the "heart attack" moments of massive losses.
It’s Incredibly Fast: Because the "thinking" is done beforehand, the actual trading happens with almost zero delay.

The Bottom Line

TiMi moves AI trading away from "AI pretending to be a human trader" and toward "AI acting as a master architect who builds perfect, high-speed machines." It combines the deep wisdom of a strategist with the cold, unshakeable speed of a machine.

Technical Summary: TiMi (Trade in Minutes)

Title: TiMi: Rationality-Driven Agentic System for Quantitative Financial Trading
Venue: Published as a conference paper at ICLR 2026

1. Problem Statement

Current Large Language Model (LLM)-based financial trading agents face three critical bottlenecks that hinder their practical application in quantitative finance:

Anthropomorphic Bias: Existing agents often simulate human roles (e.g., news analysts or emotional traders), which inadvertently introduces subjective biases and emotional noise into decision-making.
Data Reliability: Many agents rely on unstructured, peripheral information (social media, news) that is often plagued by temporal lags and misleading signals.
Deployment Latency: The high computational cost of continuous multi-agent reasoning (inference) during live trading leads to significant execution delays (slippage) and missed opportunities in high-volatility environments.

2. Methodology

The authors propose TiMi (Trade in Minutes), a system that shifts the paradigm from "simulating humans" to "implementing mechanical rationality." The core innovation is the decoupling of strategy development from real-time execution through a three-stage architecture:

A. Three-Stage Architecture:

Policy Stage (Offline): Specialized agents perform complex reasoning to develop prototype trading bots.
Optimization Stage (Offline): Bots are tested in simulated environments. A "Feedback Reflection Agent" uses mathematical reasoning to identify failures and refine the bots.
Deployment Stage (Live): Highly optimized, programmatic trading bots are deployed. These bots run on lightweight code (Python) rather than continuous LLM inference, enabling minute-level execution with minimal latency.

B. Multi-Agent Design & Specialized Capabilities:
TiMi utilizes four specialized agents, each leveraging a specific LLM strength:

Macro Analysis Agent ( $A_{ma}$ ): Uses semantic analysis ( $\phi$ ) to identify broad market patterns from technical indicators.
Strategy Adaptation Agent ( $A_{sa}$ ): Uses semantic analysis ( $\phi$ ) and mathematical reasoning ( $\gamma$ ) to customize macro strategies into pair-specific rules (e.g., adjusting for a specific coin's volatility).
Bot Evolution Agent ( $A_{be}$ ): Uses code programming ( $\psi$ ) to transform strategies into modular, three-layered bots (Strategy, Function, and Parameter layers).
Feedback Reflection Agent ( $A_{fr}$ ): Uses mathematical reasoning ( $\gamma$ ) to convert trading feedback into linear programming problems, solving for optimal parameters to maximize performance while respecting risk constraints.

C. Layered Programming & Hierarchical Optimization:
The system follows strict "programming laws" (functional cohesion, unidirectional dependency, and parameter externalization). Optimization is hierarchical: the system first attempts to fine-tune numerical parameters; if unsuccessful, it modifies functions; and finally, it alters the core strategy logic.

3. Key Contributions

Strategic Decoupling: Pioneered a mechanism that separates time-consuming cognitive reasoning from time-sensitive execution.
Two-Tier Analytical Paradigm: Introduced a workflow that moves from macro-market patterns to micro-asset customization.
Mathematical Reflection Loop: Developed a closed-loop system where LLMs don't just "chat" about mistakes but formulate and solve mathematical optimization problems to refine trading logic.
Layered Bot Implementation: Provided a modular framework for programmatic trading that allows for systematic, automated evolution of trading code.

4. Experimental Results

The system was evaluated across 200+ trading pairs in U.S. stocks and cryptocurrency markets.

Profitability & Risk: TiMi outperformed traditional quantitative methods (MACD, Grid), ML/RL methods (LSTM, DQN), and existing LLM agents (FinGPT, TradingAgents). Notably, it achieved a high Sharpe Ratio (1.27) and Annual Rate of Return (ARR) in the highly volatile Altcoin market.
Efficiency: Due to the decoupling mechanism, TiMi's action latency is comparable to traditional quantitative methods, making it 180× faster than continuous-inference agent approaches.
Capital Management: TiMi demonstrated superior capital utilization and a higher profit/loss ratio per unit of invested capital (1.53) compared to baselines.
Robustness: Ablation studies confirmed that the "Full System" (including both adaptation and mathematical reflection) is essential for stability; without these, the system suffers from logical inconsistencies or high drawdowns.

5. Significance

TiMi represents a significant step toward autonomic finance. By treating LLMs as "architects" and "optimizers" rather than "traders," the paper provides a blueprint for building agentic systems that possess high-level strategic intelligence without sacrificing the mechanical speed and rigor required for professional-grade quantitative trading.