Spaces:
Sleeping
Sleeping
| title: QuantAgent OS - Autonomous Quant Trading & Self-Improving Alpha Discovery | |
| emoji: 📈 | |
| colorFrom: indigo | |
| colorTo: blue | |
| sdk: gradio | |
| sdk_version: 4.42.0 | |
| app_file: app.py | |
| pinned: false | |
| license: mit | |
| # 🏛️ QuantAgent OS: Autonomous Quantitative Trading System | |
| ### *Self-Improving Multi-Agent Quantitative Architecture based on arXiv:2402.03755* | |
|  | |
|  | |
|  | |
|  | |
| QuantAgent OS is an institutional-grade, self-improving quantitative investment system powered by agentic AI. It implements the two-layer loop architecture pioneered in the research paper: | |
| > **QuantAgent: Seeking Holy Grail in Trading by Self-Improving Large Language Model** | |
| > *Saizhuo Wang, Hang Yuan, Lionel M. Ni, Jian Guo (HKUST & IDEA Research)* | |
| > [arXiv:2402.03755](https://arxiv.org/abs/2402.03755) | |
| --- | |
| ## 🏗️ System Architecture | |
| QuantAgent OS solves the core bottleneck of domain-specific financial AI: **how to autonomously discover, test, and accumulate quantitative trading signals (alphas) without manual human data curation.** | |
| ``` | |
| +---------------------------------------+ | |
| | Outer Feedback Loop | | |
| | (Slow / Real-World Backtest Oracle) | | |
| +---------------------------------------+ | |
| | ^ | |
| Submit Alpha | | Ground-truth Metrics | |
| for Backtest v | & Qualitative Review | |
| +--------------------+ | | |
| | Real Environment |-----------+ | |
| | (Backtesting Eng) | | |
| +--------------------+ | |
| | | |
| Updates KB v | |
| +---------------------------------------------------------------------------------+ | |
| | Inner Reasoning Loop (Fast Simulation) | | |
| | | | |
| | +------------------------------------------------------------------+ | | |
| | | Shared Context Buffer (ctx) | | | |
| | +------------------------------------------------------------------+ | | |
| | | ^ | ^ | | |
| | 1. Query | 2. Writer Hint | 3. Generated | 4. Judge | 5. Review | | |
| | v | Alpha Code v Hint | & Score | | |
| | +---------------+ +---------------+ +---------------+ | | | |
| | | KnowledgeBase | | Writer LLM | | Judge LLM |---+ | | |
| | | (KB) | +---------------+ +---------------+ | | |
| | +---------------+ | | |
| +---------------------------------------------------------------------------------+ | |
| ``` | |
| ### 1. The Inner Reasoning Loop (Fast Heuristic Simulation) | |
| - **Shared Context Buffer**: Holds task descriptions, retrieved knowledge hints, generated factor code, and mentor feedback. | |
| - **Quant Alpha Writer**: Translates qualitative trading ideas into vectorized, lookahead-free Python factor code inheriting from the standardized `Factor` base class. | |
| - **Quant Mentor Judge**: Evaluates candidate signals against 6 quantitative criteria: | |
| 1. *Vectorization efficiency* (no single-asset Python loops) | |
| 2. *Lookahead bias prevention* (no `shift(-1)` or future leakage) | |
| 3. *Continuous nuance* (smooth tanh / z-score scaling into `[-1.0, 1.0]`) | |
| 4. *Numerical stability* (`.clip()`, `.fillna()`, zero-division protection) | |
| 5. *Code validity* | |
| 6. *Economic intuition alignment* | |
| - **Iterative Dialogue**: The Writer and Judge iterate for up to $T$ rounds until the candidate meets approval score $\beta$. | |
| ### 2. The Outer Feedback Loop (Real Backtesting Oracle) | |
| - Submits the approved alpha code to the vectorized backtesting engine against historical market cross-sections. | |
| - Calculates empirical metrics: | |
| - **Information Coefficient (IC)**: Cross-sectional Pearson correlation between signal and forward returns. | |
| - **IC Information Ratio (IC_IR)**: $\text{Mean}(IC) / \text{Std}(IC)$. | |
| - **Long-Short Sharpe Ratio, Max Drawdown, and Turnover Rate**. | |
| - **Autonomous Knowledge Base (KB)**: Retains both successful and failed alphas (with failure diagnosis) to enable *pessimistic value updates*. | |
| ### 3. Downstream XGBoost Alpha Ensemble (Paper Figure 5) | |
| - Automatically pools all valid alphas from the KB as feature columns. | |
| - Trains an XGBoost cross-validated non-linear regression model to predict cross-sectional asset returns, demonstrating how prediction Mean Squared Error (MSE) monotonically decreases as the KB accumulates more diverse alphas. | |
| --- | |
| ## 🌐 Free Public Data Sources Integrated | |
| | Source | Description | Authentication | | |
| | :--- | :--- | :--- | | |
| | **Yahoo Finance** | Live & historical OHLCV, volume, dividends, valuation metrics, beta, market cap | 100% Free (Built-in browser session) | | |
| | **SEC EDGAR API** | Official regulatory 10-K and 10-Q XBRL financial facts: Revenue, Net Income, Total Assets, Liabilities, Debt/Equity | 100% Free (Official SEC endpoint) | | |
| | **FRED / Market Proxies** | 10Y Treasury, 3M T-bill, Yield Curve Slope, VIX Volatility Index, Crude Oil, USD Index | 100% Free (FRED key optional) | | |
| | **Yahoo RSS & Finnhub** | Live news headlines, company announcements, financial sentiment polarity scoring | 100% Free (Finnhub key optional) | | |
| --- | |
| ## 🚀 How to Run Locally | |
| ```bash | |
| # 1. Clone the repository | |
| git clone https://github.com/your-username/quantagent-os.git | |
| cd quantagent-os | |
| # 2. Install dependencies | |
| pip install -r requirements.txt | |
| # 3. Launch the application | |
| python app.py | |
| ``` | |
| Open your browser at `http://localhost:7860` to access the interactive dashboard. | |
| --- | |
| ## ☁️ Deploying to Hugging Face Spaces | |
| You can deploy this Space in one of two ways: | |
| ### Method 1: Directly from the Web UI (Tab 6) | |
| 1. Open the application locally or on your server. | |
| 2. Navigate to **Tab 6: Hugging Face Deployment & Settings**. | |
| 3. Enter your Hugging Face API Token (with `write` permission) and target Space ID (`username/quantagent-os`). | |
| 4. Click **Deploy to Hugging Face Spaces**. | |
| ### Method 2: Via Git | |
| ```bash | |
| git remote add space https://huggingface.co/spaces/your-username/quantagent-os | |
| git push space main | |
| ``` | |
| --- | |
| ## 📐 Mathematical Formulation & Regret Guarantees | |
| The framework formulates alpha mining as an Information-State Markov Decision Process $(S, A, \mathcal{T}, r, \gamma)$: | |
| - **Inner Loop**: Under implicit Bayesian inference by the LLM (Assumption 4.1), Bayesian regret is sublinear in inner iterations $T$. | |
| - **Outer Loop**: Grounded in Pessimistic Value Iteration (PEVI, Jin et al. 2021), epistemic uncertainty between simulated KB and real-world environment decays with outer iterations $K$. | |
| - **Overall Regret**: Total Bayesian regret $R(TK)$ decomposes into four bounded terms and is proven **sublinear in $KT$** (Theorem 4.6). | |
| --- | |
| ## ⚖️ License | |
| Released under the MIT License. Educational and quantitative research use only. | |