The List / Agents and Multi-Agent Systems / paper
MASS: Multi-Agent Simulation Scaling for Portfolio Construction
Abstract
The application of LLM-based agents in financial investment has shown significant promise, yet existing approaches often require intermediate steps like predicting individual stock movements or rely on predefined, static workflows. These limitations restrict their adaptability and effectiveness in constructing optimal portfolios. In this paper, we introduce the Multi-Agent Scaling Simulation (MASS), a novel framework that leverages multi-agent simulation for direct, end-to-end portfolio construction. At its core, MASS employs a backward optimization process to dynamically learn the optimal distribution of heterogeneous agents, enabling the system to adapt to evolving market regimes. A key finding enabled by our framework is the exploration of the scaling effect for portfolio construction: we demonstrate that as the number of agents increases exponentially (up to 512), the aggregated decisions yield progressively higher excess returns. Extensive experiments on a challenging, self-collected dataset from the 2023 Chinese A-share market show that MASS consistently outperforms seven state-of-the-art baselines. Further backtesting, stability analyses and the experiment on data leakage concerns validate its enhanced profitability and robustness. We have open-sourced our code, dataset, and training snapshots at https://github.com/gta0804/MASS/ to foster further research.
In one sentence
Summary
The paper addresses the limitations of existing LLM-based financial agents that rely on static workflows or intermediate stock prediction steps, which hinder adaptability in dynamic markets. It proposes Multi-Agent Scaling Simulation (MASS), a framework that shifts focus to direct, end-to-end portfolio construction by simulating a market of heterogeneous investor agents. The core innovation is a backward optimization process that dynamically learns the optimal distribution of agent types to maximize portfolio returns, allowing the system to adapt to evolving market regimes without predefined procedural constraints. The method employs a daily cycle of forward propagation and backward optimization. In forward propagation, heterogeneous agents with distinct styles and partial market views generate investment signals based on macroeconomic data and stock features. These signals are aggregated using the market disagreement hypothesis, which rewards consensus and penalizes disagreement. Backward optimization uses simulated annealing to adjust the agent type distribution based on historical performance over a look-back window. Experiments were conducted on a self-collected dataset from the 2023 Chinese A-share market, covering SSE 50, CSI 300, and ChiNext 100 indices, using Qwen-2.5-72B-Instruct as the underlying LLM. Results show MASS significantly outperforms seven state-of-the-art baselines across all metrics, including RIC and ICIR, with backtesting confirming higher cumulative returns and lower drawdowns. A key finding is the scaling effect: increasing the number of agents exponentially up to 512 yields progressively higher excess returns. The framework also demonstrates robustness to data leakage concerns when tested on 2025 data and adapts swiftly to market regime shifts. Limitations include the computational cost of API calls and the reliance on a specific market context, though the authors note potential for broader application.
Research problem
Existing LLM-based financial agents often rely on intermediate steps like individual stock prediction or static, predefined workflows, which limits their adaptability and effectiveness in constructing optimal portfolios. The paper addresses the need for a dynamic, end-to-end portfolio construction method that can adapt to evolving market regimes and leverage the collective intelligence of multiple agents without being constrained by fixed procedural structures.
Core contributions
- Introduction of MASS, a novel framework for end-to-end portfolio construction using multi-agent simulation and backward optimization.
- First demonstration of a scaling effect in multi-agent simulation for portfolio construction, showing performance improves with agent count up to 512.
- Development of a backward optimization mechanism that dynamically learns the optimal distribution of heterogeneous agent types.
- Release of a comprehensive, challenging dataset from the 2023 Chinese A-share market, including firm-level and macroeconomic data.
- Extensive empirical validation showing MASS outperforms seven state-of-the-art baselines in profitability, stability, and robustness.
Method
- Multi-agent simulation with heterogeneous investor types (16 types) and individual agents (32 per type).
- Forward propagation: Agents generate daily investment strategies based on macroeconomic data and their specific style.
- Partial market view: Each agent observes a random subset of stocks to manage context length and simulate realistic constraints.
- Score aggregation: Uses market disagreement hypothesis, combining weighted consensus and disagreement metrics.
- Backward optimization: Simulated annealing optimizes the agent type distribution to maximize correlation with historical returns.
- Daily cycle: Continuous loop of forward signal generation and backward distribution adjustment.
- LLM usage: Qwen-2.5-72B-Instruct for strategy formulation and stock selection by agents.
Results & findings
- MASS outperforms all seven baselines across all three stock pools (SSE 50, CSI 300, ChiNext 100) in RIC and ICIR metrics.
- On CSI 300, MASS achieved a RIC of 6.50 and RICIR of 43.49, significantly higher than the next best baseline (TradingAgents: 3.01 RIC, 10.14 RICIR).
- Backtesting on CSI 300 showed MASS generated substantially higher cumulative excess returns with lower drawdowns compared to the index and baselines.
- Scaling experiment showed a clear, approximately linear growth in RankIC as the number of agents increased up to 512.
- Data leakage test on Q1 2025 data confirmed MASS maintained significant effectiveness (e.g., RIC 3.91 on CSI 300), proving methodological robustness.
- Ablation studies confirmed that backward optimization and the market disagreement hypothesis are critical for performance.
- MASS demonstrated rapid adaptation to market regime shifts in 2023, maintaining stable excess returns during volatility.
Data & experiments
- Dataset: Self-collected 2023 Chinese A-share market data covering SSE 50, CSI 300, and ChiNext 100 indices, including news, financial reports, price-volume, and macroeconomic indicators.
- Baselines: Comparison against proxy indicators, LightGBM, DTML, MASTER, SEP, FinCON, and TradingAgents.
- Metrics: Rank Information Coefficient (RIC), Information Coefficient (IC), and their respective Information Ratios (RICIR, ICIR).
- Backtesting: Weekly rebalanced portfolio simulation with transaction costs to evaluate cumulative returns and drawdowns.
- Data Leakage Test: Evaluation on unseen data from Q1 2025 to verify generalization beyond memorization.
- Scaling Experiment: Analysis of performance as the number of agents increases exponentially from 64 to 512.
Limitations & caveats
- Experiments are limited to the Chinese A-share market, which may affect generalizability to other markets.
- Relies on Qwen-2.5-72B-Instruct, so performance is tied to the capabilities of this specific LLM.
- Computational cost and API fees, while optimized, may still be a barrier for high-frequency or large-scale deployment.
- The backward optimization process requires a look-back window, which may lag in rapidly changing markets.
- The dataset, while comprehensive, is from a specific volatile period (2023), and long-term stability beyond 2025 Q1 is not fully validated.
- Agent types are predefined, and the framework's ability to discover new, emergent agent types is not explored.
Curator notes
Discovery status: round3_promoted. Source links: Fin-R1: A Large Language Model for Financial Reasoning through Reinforcement Learning; FinCon: A Synthesized LLM Multi-Agent System with Conceptual Verbal Reinforcement for Enhanced Financial Decision Making; MarS: a Financial Market Simulation Engine Powered by Generative Foundation Model; When AI Meets Finance (StockAgent): Large Language Model-based Stock Trading in Simulated Real-world Environments