qusong0627/QuantMind

QuantMind（量化大脑）开源版是一款面向个人开发者与投研团队的 AI 原生多市场量化交易平台。深度集成微软 Qlib、RD-Agent 因子演化与 QuantBot全能工作台，提供从 300+ 维因子挖掘、13 种机器学习与深度学习模型工场、Qlib 高性能回测、截面批量推理、7x24 实时舆情情绪分析，到通达信深度联动（板块推送/预警雷达/闪电下单）与实盘模拟交易的完整闭环。支持 A股、港股、美股、期货及区块链。可免费商用，系统采用 Docker Compose 一键私有化部署，数据与模型完全本地化，保障策略隐私，功能零门槛无限制。

QuantMind (量化大脑) – Overview

What it is – QuantMind is an open‑source, AI‑first platform for quantitative trading. It stitches together the Microsoft Qlib quant‑research framework, the RD‑Agent large‑language‑model (LLM) based “auto‑alpha” engine, and a conversational robot called QuantBot that can be reached from chat apps (WeChat, QQ, Feishu, DingTalk, etc.). The whole stack runs in Docker containers and provides a complete end‑to‑end workflow:

Data → Factor mining → Model training → Batch inference → Portfolio back‑test → Simulated trading → Production monitoring

The system targets individual researchers, research teams and professional institutions and supports Chinese‑A‑shares, Hong‑Kong, US stocks, futures and blockchain markets.

Core Components & Features

Component	What it does	Key technologies
QuantBot conversational robot	Natural‑language entry point for the whole platform. Users type a command in a chat app and the bot fetches data, runs factor mining, trains or runs a model, back‑tests, generates a research report and even places simulated orders. Supports 17 IM channels (WeChat, QQ, Feishu, DingTalk, etc.).	QwenPaw robot framework, FastAPI, SSE streaming, skill‑pool system
Market & data layer	Centralised QuantDB with >300 pre‑computed micro‑structure and flow features. Data stored as Parquet and accessed via DuckDB for sub‑second columnar reads.	Parquet, DuckDB, Docker volumes
Factor autonomous evolution	Uses Microsoft RD‑Agent (AutoAlpha 2.0) to turn a natural‑language hypothesis into a factor formula, then runs genetic‑algorithm back‑testing and stores the best factors.	RD‑Agent, LLM‑driven pipeline
13 model factories	Classic tree models (LightGBM, XGBoost, CatBoost) and deep‑learning sequence models (GRU, LSTM, Transformer, TabNet, TCN, NativeTFT). Supports stacking ensembles with a Ridge meta‑learner. Training can run locally (CPU/GPU) or on user‑provided GPU nodes (AutoDL).	PyTorch 2.x, Docker GPU support
Batch inference & signal hub	Daily market‑wide scoring, Top‑N recommendation, dynamic signal fusion, automatic Rank‑IC calculation, SHAP importance and PSI drift alerts.	PyTorch inference, SHAP, custom monitoring
Microsoft Qlib back‑test engine	Event‑driven back‑testing with fine‑grained fee, slippage and limit‑up/down simulation. Provides full return attribution and risk metrics.	Qlib, FastAPI
Simulated trading & risk control	Local T+1 matching engine that reproduces A‑share rules (price limits, halts, commission, taxes, slippage). Tracks positions, orders, daily capital snapshots and pre‑flight checks.	Custom engine, Redis for state
Real‑time news & RSS monitor	24/7 ingestion of financial news, entity extraction, sentiment classification and linkage to stocks.	RSS feeds, NLP pipelines
AI‑IDE	Built‑in code editor with an “AI Copilot” that can lint, suggest, back‑test and publish strategies.	VSCode‑like editor, LLM assistance
Model Hub & Community Square	Browse, import and publish community models with performance metrics (Sharpe, IC, max‑drawdown, etc.).	Metadata store, Docker registry

Quick‑Start / Deployment

One‑click online deployment – a script pulls a pre‑built Docker image, restores a PostgreSQL dump, loads QuantDB data and pretrained models.
```
curl -fsSL https://gitee.com/qusong0627/QuantMind/raw/master/deploy/full-deploy.sh | sudo bash
```
After it finishes you can reach the web UI at http://<host>:3000 and the API at http://<host>:8000.
Data preparation – the full bundle already contains a basic dataset. For production you can either:
- Bind a QuantDB API key in the UI and let the system sync daily increments, or
- Download the full offline dataset (≈56 GB) from ModelScope and unpack it to /opt/quantmind/data/quantdb/.
Update – deploy/update.sh --force refreshes code while keeping the database and model assets.

Typical Use Cases

Remote command via chat – “帮我回测 CSI300 上近一年动量策略” → the bot runs the back‑test and pushes a PDF report back to the chat.
Factor research – describe a hypothesis in plain Chinese, the LLM creates a factor expression, runs genetic evolution, and stores the best factor for later use.
Model training – select a model type (e.g., LightGBM) in the visual training UI, set a rolling‑window WFA config, and let the system schedule GPU training on a self‑hosted node.
Simulation trading – launch the local T+1 engine, monitor daily P&L and position snapshots, and iterate on strategy parameters without risking real capital.

Who Might Benefit?

Quant researchers who want an LLM‑augmented workflow for factor generation and back‑testing.
Small hedge funds / trading desks looking for a turnkey, Docker‑based platform that can be controlled from mobile chat apps.
Students & hobbyists interested in the full pipeline from data ingestion to simulated execution, with ready‑made examples and tutorials.

License & Disclaimer

The code is released under AGPL‑v3. All AI‑generated signals are for research only; the repository includes a clear disclaimer that users bear all investment risk and must handle any real‑world broker integration themselves.

For more details, see the README sections on architecture, fast deployment, and the extensive screenshot‑rich product preview.