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

🕒 Published (v1): 2026-01-01 · Source: ICLR · Venue: ICLR 2026 · link

Why this paper was selected

Rationality-driven agentic trading system with structured critique-refine reasoning loop

Ask a follow-up

Open an assistant pre-loaded with this paper's context.

TL;DR¶

TiMi (Trade in Minutes) is a rationality-driven multi-agent system for quantitative financial trading that architecturally decouples offline LLM-driven strategy development from lightweight, latency-free live deployment via programmatic trading bots. It composes three specialized LLMs for semantic analysis, code programming, and mathematical reasoning across a policy→optimization→deployment pipeline. Evaluated on 200+ trading pairs in U.S. stock and cryptocurrency markets, TiMi achieves competitive returns with superior risk control relative to quantitative, ML/RL, and LLM-agent baselines.

Problem¶

Existing LLM trading agents simulate anthropomorphic roles (sentiment analysts, debaters, news readers) that introduce emotional biases and depend on unstructured peripheral signals (social media, project reports) prone to noise and temporal lag. Critically, they require continuous multi-agent inference during deployment, inducing execution slippage and latency costs intolerable in high-frequency, high-volatility markets. Classical quantitative strategies lack adaptability to non-linear dynamics and black-swan events.

Method¶

TiMi decomposes trading into three stages:

Policy stage (offline): A macro analysis agent $\mathcal{A}_{ma}$ uses semantic analysis $\phi$ and code programming $\psi$ to extract technical-indicator-based patterns from market $\mathcal{M}$ across time windows $\mathcal{W}$, producing a general strategy set $\mathcal{S}$. A strategy adaptation agent $\mathcal{A}_{sa}$ then applies $\phi$ and mathematical reasoning $\gamma$ to customize $\mathcal{S}$ into pair-specific strategies $\mathcal{S}_\mathcal{P}$ with initialized parameters $\Theta_\mathcal{P}$.

Optimization stage (offline simulation): A bot evolution agent $\mathcal{A}_{be}$ (Code LLM: Qwen2.5-Coder-32B) compiles strategies into layered Python trading bots $\mathcal{B}$ governed by three programming laws (functional cohesion, unidirectional dependency, parameter externalization). A feedback reflection agent $\mathcal{A}_{fr}$ (DeepSeek-R1) receives action feedback $\mathcal{F}$ and formulates constrained optimization problems: $$\Theta^* = \arg\max_{\Theta \in C(\Theta)} \sum_i \omega_i J_i(\Theta, \mathcal{F}) \quad \text{s.t.} \quad C(\Theta) = \{\Theta \in \mathbb{R}^n \mid A(R)\Theta \preceq b(R)\}$$ where risk scenarios $R = \gamma(\mathcal{F})$ define the constraint matrix. Hierarchical refinement escalates from parameter → function → strategy layer as needed, producing advanced bots $\mathcal{B}^*$.

Deployment stage (live, CPU-only): Pre-compiled bots execute at minute-level granularity without any LLM inference. The efficiency ratio $\eta = \frac{c_\text{agent} \times n}{c_\text{policy} + c_\text{optim} + c_\text{bot} \times n} \to \frac{c_\text{agent}}{c_\text{bot}}$ as $n \to \infty$; since $c_\text{bot} \ll c_\text{agent}$, gains scale with trading frequency.

Backbone assignment: DeepSeek-V3 for $\phi$ (semantic analysis), Qwen2.5-Coder-32B-Instruct for $\psi$ (code programming), DeepSeek-R1 for $\gamma$ (mathematical reasoning).

Key Contributions¶

TiMi system: first rationality-driven multi-agent architecture for quantitative trading that replaces anthropomorphic simulation with mechanical reasoning grounded in technical indicators and mathematical optimization.
Analysis-deployment decoupling: offline LLM reasoning generates compiled bots; live execution requires no LLM inference, enabling minute-level latency.
Two-tier analytical paradigm: macro-level strategy generation ($\mathcal{A}_{ma}$) followed by pair-specific customization ($\mathcal{A}_{sa}$), capturing both market-wide patterns and pair heterogeneity.
Layered programming design: three-law governance (cohesion, unidirectional dependency, parameter externalization) enabling systematic bot evolution across optimization cycles.
Closed-loop mathematical reflection: feedback is formalized as linear programming with risk-scenario-derived constraints, enabling Pareto-efficient parameter solutions.

Results¶

Backtesting on 2024 historical data (Table 1); metrics: Annual Rate of Return (ARR%), Sharpe Ratio (SR), Maximum Drawdown (MDD%).

U.S. Stock Index Futures: TiMi ARR=8.9%, SR=0.84, MDD=10.5% — best SR and lowest MDD among all methods; TSMOM achieves higher ARR (10.5%) but with MDD=19.7%.
Mainstream Coin Futures: TiMi ARR=16.5%, SR=1.25, MDD=12.1% — highest SR; TradingAgents achieves higher ARR (17.9%) but MDD=20.3% vs TiMi's 12.1%.
Altcoin Futures: TiMi ARR=23.7%, SR=1.27, MDD=26.0% — best ARR and SR by large margin; next-best LLM agent (TradingAgents) achieves ARR=9.7%, SR=0.72; best ML method (PatchTST) achieves ARR=17.5%, SR=0.98 with MDD=23.5%.
TiMi outperforms all LLM-agent baselines (FinGPT, FinMem, TradingAgents) across all three market segments on risk-adjusted metrics.

Limitations¶

Backtesting relies on 2024 historical data; live trading results are reported qualitatively (bot evolution cycles C1–C4) without quantitative comparison against baselines.
No ablation isolating contributions of individual agents ($\mathcal{A}_{ma}$, $\mathcal{A}_{sa}$, $\mathcal{A}_{fr}$) or the mathematical reflection module specifically.
Backbone LLMs (DeepSeek-V3, Qwen2.5-Coder, DeepSeek-R1) are used off-the-shelf without domain fine-tuning; generalization to regime shifts beyond 2024 data is untested.
Efficiency analysis ($\eta$) is theoretical; no empirical latency benchmarks are reported.
The two-tier strategy paradigm assumes technical-indicator-identifiable periodicity, which may break down under structural market shifts.

Relevance to Agentic AI / LLM Agents¶

TiMi demonstrates a concrete instantiation of capability-specialized multi-agent composition: rather than a single general LLM or homogeneous agent pool, it assigns distinct LLM variants to distinct cognitive primitives ($\phi/\psi/\gamma$), a design principle applicable broadly to agentic pipelines requiring heterogeneous reasoning types. The analysis-execution decoupling is a significant architectural contribution — it resolves the tension between LLM reasoning depth and real-time deployment latency by separating the planning horizon from the execution horizon, a pattern relevant to any agentic system requiring low-latency action in high-frequency environments. The closed-loop mathematical reflection (formalizing environment feedback as constrained optimization solvable by a reasoning LLM) offers a template for grounded, verifiable agent self-improvement beyond prompt-based reflection. Finally, the layered bot evolution with hierarchical feedback (parameter → function → strategy) demonstrates how agentic systems can implement structured, progressive self-refinement in real-world deployment.