Agent-Driven LLM Quant Research Engine — Autonomous Factor Mining at WorldQuant BRAIN Submission Quality
LLM Agent 自治因子挖矿 → 批量回测 → 多维评分 → 反过拟合验证 → WQ BRAIN 自动提交 | 全程零人工干预
Quick Start · Architecture · API Docs · MCP Guide · Factor Mining · Contributing
QuantGPT is an agent-driven factor research engine — not a backtest library, not a chatbot wrapper. It gives an LLM Agent (Claude, via MCP) a complete toolkit to autonomously discover, evaluate, iterate, and submit alpha factors to WorldQuant BRAIN, with zero human intervention per research cycle.
The core architecture:
LLM Agent (Claude Code / Claude Desktop)
│
├── MCP Tools (8 个) ← Agent 的工具箱
│ ├── run_backtest ← 全市场分组回测
│ ├── score_factor ← 0-100 综合评分
│ ├── diagnose_factor ← 失败模式诊断
│ ├── run_anti_overfit ← 4 项反过拟合检验
│ ├── run_rolling_validation ← Walk-forward 验证
│ ├── validate_expression ← 语法校验
│ ├── list_operators ← 50+ 算子文档
│ └── list_universes ← 股票池和基准
│
├── Evolution Engine ← 因子进化引擎
│ ├── MutationEngine (8 方向突变)
│ ├── CrossoverEngine (高分因子交叉)
│ ├── MetaEvolutionSelector (自适应策略)
│ └── TrajectoryAnalyzer (轨迹分析)
│
├── WQ BRAIN Integration ← WorldQuant 直连
│ ├── Dollar-neutral 模拟
│ ├── IS 检测对齐
│ ├── Fitness 评分
│ └── 一键正式提交
│
└── Knowledge Base ← 跨会话知识积累
├── rules/ (已验证规则)
├── findings/ (经验发现)
└── failures/ (已证伪路径)
传统工具（包括 ChatGPT + 回测库）的模式是：人类想因子 → 工具跑回测 → 人类看结果。Agent 是执行者，人类是决策者。
QuantGPT 的模式是：人类定义目标 → Agent 自治研究 → 人类审阅产出。Agent 是研究者，人类是审阅者。
这不是接口层的区别（自然语言 vs. 代码），而是决策权的区别。Agent 自主决定：探索哪个方向、生成什么表达式、评估哪些指标、何时迭代、何时放弃、何时提交。
| Metric | Value |
|---|---|
| 累计回测任务 | 370+ |
| 单轮迭代（8 候选因子） | ~15 分钟 |
| 表达式算子标准 | WorldQuant BRAIN 对齐 |
| BRAIN 正式提交 | 3 个因子 IS 全部 PASS，已提交（最佳 Fitness 1.26） |
| WQ BRAIN 集成 | 内置 API — 一键模拟 + 自动提交 |
QuantGPT Agent 已产出 3 个正式提交因子，全部通过 WQ BRAIN IS 检测：
| Factor | Expression | WQ Sharpe | WQ Fitness | WQ Returns | IS Tests | Status |
|---|---|---|---|---|---|---|
| Debt-Momentum Composite | -1 * rank(ts_av_diff(close, 10)) + rank(debt / enterprise_value) |
1.77 | 1.26 | 20.18% | ALL PASS | Submitted |
| VWAP Decay Reversal | -1 * rank(ts_decay_linear(close / vwap, 10)) |
1.69 | 1.07 | 18.63% | ALL PASS | Submitted |
| Returns-Volume Momentum | -1 * rank(ts_decay_linear(returns * volume / adv20, 5)) |
1.60 | 1.03 | 24.15% | ALL PASS | Submitted |
3 个因子代表不同的 alpha 来源：Debt-Momentum 结合动量反转与基本面（债务/企业价值），行业中性化；VWAP Decay Reversal 捕捉价格偏离 VWAP 的衰减回归，市场中性化；Returns-Volume Momentum 捕捉收益率与相对成交量的衰减动量，市场中性化。全程 Agent 自治完成。
Debt-Momentum Composite — 已正式提交：Sharpe 1.77, Fitness 1.26, Returns 20.18%, IS 全部 PASS
VWAP Decay Reversal — 已正式提交：Sharpe 1.69, Fitness 1.07, Returns 18.63%, IS 全部 PASS
Returns-Volume Momentum — 已正式提交：Sharpe 1.60, Fitness 1.03, Returns 24.15%, IS 全部 PASS
This is QuantGPT's defining capability.
Agent 读知识库、设计假设、批量实验、分析结果、积累知识、自我迭代，每个结论经过双模型交叉验证。一个研究循环产出经过验证的、可提交 WQ BRAIN 的因子。
┌─────────────────────────────┐
│ Research Notes & Knowledge │
│ (Rules / Findings / Fails) │
└──────────┬──────────────────┘
│ read
▼
┌──────────┐ ┌──────────────────────────┐ ┌──────────────────┐
│ Phase 0 │───▶│ Phase 1: Factor Design │───▶│ Phase 2: Batch │
│ Context │ │ Hypothesis → Expression │ │ Backtest (10-20 │
│ Loading │ │ 1-3 candidates per idea │ │ concurrent) │
└──────────┘ └──────────────────────────┘ └────────┬─────────┘
│
┌─────────────────────────┘
▼
┌──────────────────────────────────────────┐
│ Phase 3: Four-Step Analysis │
│ │
│ ① Fact Collection (metrics vs baseline) │
│ ② Independent Judgment (Agent) │
│ ③ Cross-Review (DeepSeek Reasoner) │
│ ④ Consensus or Divergence Resolution │
└──────────────────┬───────────────────────┘
│
┌──────────────┴──────────────┐
▼ ▼
┌─────────────────┐ ┌──────────────────┐
│ Phase 4: Update │ │ Phase 5: Stop? │
│ Notes + Knowledge│ │ Converged / │
│ Base │◀────────│ Time / Rounds │
└─────────────────┘ └──────────────────┘
│ │ no
│ └──▶ back to Phase 1
▼
┌──────────────────┐
│ Phase 6: Report │
│ A/B factors + │
│ new knowledge │
└──────────────────┘
|
Dual-LLM Cross-Review 每个结论性判断（采用/不采用/关闭方向）必须经过第二个 LLM 独立评审。把事实数据和第一个模型的推理链一起发给 DeepSeek Reasoner，要求独立评估推理是否合理、是否有遗漏角度。 共识 → 直接输出。分歧 → 呈现双方证据，采用更保守结论。 这解决了单模型因子研究的核心问题：confirmation bias。 |
Persistent Knowledge Base 知识库跨会话积累。第 10 次研究会话可以直接利用前 9 次的所有发现，避免重复实验，遵守已验证规则，绕开已证伪路径。 这不是 chat history——是结构化的研究资产。 |
|
Batch Concurrent Evaluation 单次提交 10-20 个因子表达式，并发回测 + 三波重试。结果按 fitness 降序排列。hs300 fitness < 0.1 时自动跳过 csi500 验证，节省算力。 from scripts.factor_miner import batch_evaluate
results = batch_evaluate(
server, expressions, params,
max_concurrent=10
) |
Research Discipline (Enforced) 不是建议，是硬性规则：
|
上面 Validated Results 中的因子就是这个流程的产出。 多轮迭代，3 个因子正式提交 WQ BRAIN（IS 全部通过）。完整方法论见 Factor Mining Guide。
┌────────────────────────────────────────────────────────────────────┐
│ QuantGPT Research Engine │
├─────────────┬──────────────────────────────┬───────────────────────┤
│ │ Core Engine │ │
│ Agent │ ┌──────────────────────┐ │ Data Layer │
│ Interface │ │ Expression Parser │ │ ┌─────────────────┐ │
│ │ │ 50+ operators │ │ │ baostock (free) │ │
│ MCP Tools │ │ WQ BRAIN compatible │ │ │ akshare (free) │ │
│ REST API │ └──────────┬───────────┘ │ │ PolarDB (opt) │ │
│ Web UI │ ┌──────────▼───────────┐ │ │ Parquet cache │ │
│ (monitor) │ │ Backtest Engine │ │ └─────────────────┘ │
│ │ │ Rank-based grouping │ │ │
│ │ │ WQ BRAIN aligned │ │ AI Layer │
│ │ └──────────┬───────────┘ │ ┌─────────────────┐ │
│ │ ┌──────────▼───────────┐ │ │ DeepSeek LLM │ │
│ │ │ Validation Suite │ │ │ Factor design │ │
│ │ │ Anti-overfit (4x) │ │ │ Cross-review │ │
│ │ │ Walk-forward │ │ │ Mutation engine │ │
│ │ │ WQ BRAIN simulation │ │ └─────────────────┘ │
│ │ └──────────────────────┘ │ │
│ │ │ Storage │
│ │ Evolution Engine │ ┌─────────────────┐ │
│ │ Trajectory → Meta-Evo → │ │ SQLite (default)│ │
│ │ Mutation / Crossover │ │ PostgreSQL (opt)│ │
│ │ │ └─────────────────┘ │
├─────────────┴──────────────────────────────┴───────────────────────┤
│ Agent Orchestrator: Claude Code skill loop / Claude Desktop MCP │
└────────────────────────────────────────────────────────────────────┘
| Layer | Technology |
|---|---|
| Agent | Claude Code (skill loop) / Claude Desktop (MCP) |
| Backend | Python 3.10+, FastAPI, uvicorn, SQLAlchemy 2.0 async |
| Database | SQLite (default, zero-config) / PostgreSQL (optional) |
| AI/LLM | DeepSeek (factor generation + cross-review) |
| Market Data | baostock + akshare (free) → Parquet cache |
| Frontend | React 18 + TypeScript + Tailwind CSS 4 (monitoring dashboard) |
| MCP | FastMCP (stdio / SSE / streamable-http) |
| Report | QuantStats HTML |
自研的表达式解析器（expression_parser.py, 870+ lines）是整个系统的核心：
- 50+ 算子：截面（
rank,zscore）、时序（ts_corr,decay_linear）、非线性（sign_power）、条件（where,trade_when）、技术指标（rsi,macd,atr） - 双模式：
mode="wq"仅允许 WQ BRAIN 兼容算子（提交前校验），mode="local"开放全部算子 - 语义正确的截面/时序分离：
rank()按trade_date分组（截面），ts_mean()按stock_code分组（时序），自动处理分组逻辑 - 安全约束：递归深度限制、窗口上限、表达式长度限制，防止恶意输入
| Layer | Module | Method |
|---|---|---|
| Statistical Tests | anti_overfit.py |
IC 稳定性 + 子样本压力测试（牛/熊/震荡）+ 安慰剂检验 + 半衰期估计 |
| Walk-Forward | rolling_validator.py |
滚动 train/valid/test 窗口，评估样本外 IC 衰减 |
| WQ Simulation | wq_simulate.py |
Dollar-neutral 多空模拟，对齐 BRAIN 的 Sharpe/Turnover/Fitness 计算 |
| WQ BRAIN API | wq_brain_client.py |
直连 BRAIN 平台 — 真实模拟 + IS 检测 + 一键正式提交 |
受 QuantaAlpha 启发的三阶段自动搜索：
TrajectoryAnalyzer → MetaEvolutionSelector → Strategy Execution
(质量指标评估) (EXPLOIT/EXPLORE/ (MutationEngine ×8 方向
RECOMBINE/SIMPLIFY) / CrossoverEngine)
8 种定向突变：时间窗口变异、算子替换、复杂度调整、截面变换叠加等。5 维评分驱动迭代方向。
| Mode | Role | Use Case |
|---|---|---|
| MCP (primary) | Agent toolkit | Claude Code / Claude Desktop 通过 MCP 调用所有研究工具，驱动自治研究循环 |
| REST API | Programmatic access | 批量回测、外部系统集成、CI/CD 因子验证 |
| Web UI | Monitoring dashboard | 任务监控、报告查看、因子库管理 |
MCP Tools (8 个)
| Tool | Description |
|---|---|
list_operators |
全部算子文档 |
list_universes |
股票池和基准 |
validate_expression |
语法校验 |
run_backtest |
完整回测 |
score_factor |
评分（0–100, A/B/C/D） |
diagnose_factor |
失败模式诊断 + 改进建议 |
run_anti_overfit |
4 项反过拟合检验 |
run_rolling_validation |
Walk-forward 验证 |
| Capability | JoinQuant | Backtrader | ChatGPT + Backtest | QuantGPT |
|---|---|---|---|---|
| Research mode | Human writes code | Human writes code | Human prompts, tool executes | Agent autonomously researches |
| Factor discovery | Manual | Manual | One-shot LLM | Multi-round evolution + knowledge base |
| Anti-bias | Researcher judgment | None | None | Dual-LLM mandatory cross-review |
| Knowledge accumulation | Personal notes | None | Lost between sessions | Structured KB across sessions |
| WQ BRAIN integration | -- | -- | -- | Operator-aligned + direct submission |
| Anti-overfit | -- | -- | -- | 4 statistical tests + walk-forward |
| MCP / AI Agent | -- | -- | -- | 8 tools, skill-loop orchestration |
| Live trading | Yes | Limited | -- | -- |
| Intraday data | Yes | Yes | -- | Daily only |
git clone https://github.com/Miasyster/QuantGPT.git && cd QuantGPT
make setup # creates venv, installs deps, generates .env
make run # starts server at http://localhost:8003Add MCP configuration to Claude Code or Claude Desktop:
{
"mcpServers": {
"quantgpt": {
"command": "python",
"args": ["-m", "quantgpt"]
}
}
}Then let the Agent work: "在沪深300上挖掘高 fitness 的因子，目标 WQ BRAIN 可提交"
# Direct expression backtest via API
curl -X POST http://localhost:8003/api/v1/auto_backtest \
-H "Content-Type: application/json" \
-H "Authorization: Bearer <token>" \
-d '{"expression": "rank(close / ts_mean(close, 20))", "universe": "hs300"}'Zero config by default: SQLite database, baostock + akshare free data. See full Quick Start guide for details.
Optional: DeepSeek API (for factor generation & cross-review)
# Edit .env, add your DeepSeek API key (~$0.001 per query)
DEEPSEEK_API_KEY=sk-your-key-hereOptional: PostgreSQL (for production)
pip install "quantgpt[postgresql]"
# Edit .env:
DATABASE_URL=postgresql+asyncpg://quantgpt:password@localhost:5432/quantgpt
alembic upgrade headExpression Examples
# Debt-momentum composite — BRAIN submitted, Fitness 1.26, Sharpe 1.77
-1 * rank(ts_av_diff(close, 10)) + rank(debt / enterprise_value)
# VWAP decay reversal — BRAIN submitted, Fitness 1.07, Sharpe 1.69
-1 * rank(ts_decay_linear(close / vwap, 10))
# Returns-volume momentum — BRAIN submitted, Fitness 1.03, Sharpe 1.60
-1 * rank(ts_decay_linear(returns * volume / adv20, 5))
# 20-day momentum
rank(close / ts_mean(close, 20))
# Low volatility
rank(-1 * ts_std(close/ts_shift(close,1)-1, 20))
# Decay-weighted correlation
decay_linear(rank(ts_corr(vwap, volume, 10)), 5)quantgpt/
├── quantgpt/ # Backend (Python)
│ ├── expression_parser.py # Factor expression parser (50+ ops, WQ compatible)
│ ├── backtest.py # Rank-based group backtest engine
│ ├── market_data.py # baostock/akshare → Parquet cache
│ ├── api_server.py # FastAPI REST API + SSE
│ ├── mcp_server.py # FastMCP server (8 tools — Agent's toolkit)
│ ├── iteration.py # 3-phase evolutionary iteration
│ ├── mutation_engine.py # 8 directed mutation strategies
│ ├── crossover_engine.py # High-score factor crossover
│ ├── meta_evolution.py # Adaptive strategy selector
│ ├── trajectory_analyzer.py # Trajectory quality metrics
│ ├── anti_overfit.py # 4 statistical anti-overfit tests
│ ├── rolling_validator.py # Walk-forward validation
│ ├── wq_simulate.py # WQ BRAIN dollar-neutral simulator
│ ├── wq_brain_client.py # WQ BRAIN API integration
│ ├── neutralize.py # Industry & cap neutralization
│ ├── paper_engine.py # Paper trading engine
│ ├── daily_summary.py # LLM-powered daily market report
│ └── routes/ # API route modules
├── frontend/ # React 18 + TypeScript + Tailwind CSS 4
│ └── src/components/ # Monitoring dashboard
├── scripts/
│ └── factor_miner.py # Batch factor evaluation toolkit
├── tests/ # 74 tests (parser + backtest + WQ simulate)
├── example_factor/ # BRAIN validation screenshots
└── docs/ # Architecture, API, MCP, Mining guides
- Daily frequency only — no intraday backtesting
- A-share market only — China mainland equities
- Agent quality depends on LLM — better models produce better factors
MIT — Copyright (c) 2026 Miasyster
This repository is the original source of the QuantGPT factor research engine. Derivative works should retain the copyright notice and comply with the MIT License terms. See NOTICE for details.
Past factor performance does not guarantee future returns. This project does not constitute investment advice.