Stock Strategy Lab: A Multi-Agent Quantitative Strategy Research Platform for A-Shares

Stock Strategy Lab is a multi-agent quantitative strategy research platform designed specifically for the A-share market. Unlike traditional fixed-strategy tools, it enables end-to-end automated strategy exploration and optimization—from single-asset timing to asset allocation and portfolio integration—through collaboration among three layers of agents (signal layer, budget layer, and portfolio layer). It supports human-machine collaboration, autonomous iteration, and the precipitation of standardized research results.

Most traditional quantitative tools provide fixed strategy templates, while Stock Strategy Lab builds an agent-based research system that allows long-term interaction, repeated operation, and continuous result precipitation. It aims to help quantitative researchers systematically explore the strategy space and quickly validate ideas. Positioned as a quantitative research assistant rather than an automatic money-making machine, it explicitly does not constitute investment advice.

The platform adopts a three-layer agent architecture consisting of signal layer, budget layer, and portfolio layer:

1. Signal layer: Solves single-asset timing problems, uses over 30 types of signal components to explore Alpha switching capabilities through closed-loop processes such as data acquisition, strategy writing, and backtest evaluation;
2. Budget layer: Handles fund allocation for asset pools, generates candidate strategies by combining asset pool profiles and user preferences;
3. Portfolio layer: Integrates the capabilities of the first two layers and generates final portfolio strategies through complementary correction mechanisms.

The platform has five core capabilities:

1. Autonomous iteration loop: Agents at each layer cyclically optimize strategies until meeting the standards;
2. Human-machine collaborative interaction: Users can intervene to adjust preferences, risks, etc., and session saving and restoration are supported;
3. Memory management: Records research progress and historical ideas to avoid repeated exploration;
4. Multi-method parameter search: Supports grid/random/genetic algorithms, etc., and uses rolling validation to test robustness;
5. Standardized outputs: Each task outputs independent directories, reports, etc., facilitating comparison and precipitation.

The system is built on the DeepAgents and LangChain frameworks, encapsulating independent agent modules such as review and data. It adopts an extensible strategy framework, supporting users to customarily introduce methods like machine learning and Chan Theory to enrich the strategy exploration space.

In terms of data, it supports local data sources such as MiniQMT and xtquant (for Windows QMT environment); in terms of models, it is compatible with OpenAI-compatible APIs like DeepSeek and Moonshot, and each agent can independently configure model parameters; it uses hierarchical configuration management, supporting global defaults and agent-level overrides.

The typical usage flow includes: initializing the environment → configuring API keys and data sources → creating research tasks → interacting and iterating with agents → viewing backtest results and reports → saving session state. It emphasizes that agents are responsible for strategy exploration and code generation, while users control the direction and make decisions.

The project's value lies in providing an automated strategy research pipeline for quantitative researchers, and the three-layer architecture offers reference ideas for multi-step decision-making problems. At the same time, it emphasizes risk awareness: backtest results do not represent future returns, do not constitute investment advice, and should be treated rationally.