BoolQuant

English | 中文

LLM-Driven Boolean Rule Evolution for Stock Selection

100% Local. Zero API calls. Your data, your models, your alpha. No leakage, ever.

Example Evolved Indicator

Evolved Rule:

(((high + low) / 2) - close) > ((((high + low) / 2) - close).rolling_std(10))

Metric	Train (1899 days)	Test (664 days)
Sharpe	1.44	1.16
Total Return	600.8%	145.0%
Max Drawdown	-74.5%	-44.8%
Win Rate	52.4%	50.6%

---

Quick Start

1. Install Dependencies

pip install -r requirements.txt

2. Set Up Local LLM

Install and run a local LLM server using LM Studio or Ollama.

Recommended GGUF models (4GB+ VRAM):

• Thinker (reasoning/analysis):
  Qwen3-4B-Instruct-2507-Gemini-3-Pro-Preview-Distill-GGUF or Falcon-H1R-7B-GGUF

• Summarizer (final output):
  Qwen3-4B-Instruct-2507-GGUF

⚠️ Always use different models for thinker and summarizer—this separation yields cleaner reasoning and more concise final responses. Never reuse the same model for both roles.

3. Configure Settings

Edit config.json to set key parameters:

Parameter	Description
generations	Number of evolution cycles
attempts	Rule mutations per generation
amount_rank_n	Stock pool size (e.g., 5-20)
return_type	o2c, o2o, c2c, or c2o
transaction_fee	Default: 0.0007
train_test_ratio	Train/test data split

4. Run the App

streamlit run app.py

UI Usage

1. Configure parameters in the left sidebar:

   • Stock Selection: Data source and prefix filters
   • LLM Configuration: Choose local LLM service and models
   • Evolution Settings: Generations, attempts, return type
   • Data Filter: Optional amount ranker for stock pool size
   • Date Range: Select backtest period
   • Train/Test Split: Set ratio for validation

2. Run Evolution tab: Click "Run Evolution" to start the genetic algorithm

3. Results tab: View top performing rules, plots, and performance metrics

4. Stock Selector tab: Test boolean conditions on specific dates

---

How It Works

BoolQuant evolves human-readable boolean rules to select stocks based on LLM generated indicators from basic ohlcv:

((close - low) / (high - low) * volume) > ((close - low) / (high - low) * volume).shift(1)

Evolution Process

1. LLM generates rule mutations based on backtest feedback
2. Rules are evaluated on historical train data
3. Top performers are selected and mutated
4. Process repeats for N generations
5. Best rules are tested on held-out test data

Focused Stock Pool Selection

Unlike traditional portfolio optimization methods that may output hundreds of candidates, BoolQuant's Stock Pool Ranker is designed to produce a small, focused list (typically 5-20 stocks) for research analysis.

This compact approach enables:

• In-depth review of each candidate's characteristics
• Manageable scope for technical or fundamental analysis
• Clear interpretability of what the rules are selecting

The system ranks stocks by trading volume and applies evolved boolean rules to filter the pool. Always follow the provided selection hint (e.g., "select from the first 5 stocks whose open price is between 0.97× and 0.99× preclose") to ensure:

• No lookahead bias: Uses only information available at decision time
• Reproducible methodology: Deterministic, rules-based selection
• Research discipline: Prevents arbitrary cherry-picking

This focused output is designed for research and educational exploration, not as trading recommendations.

---

Requirements

• Python 3.10+
• Local LLM server (Ollama or LM Studio)

---

Data Quality

⚠️ yfinance data quality: Data may be missing, inaccurate, or retroactively corrected without notice—use at your own discretion.

⚠️ Data source consistency: Same rules may behave differently across data sources—stick to one source for consistent results.

Disclaimer

BoolQuant is a personal research project for educational purposes only.

• Not intended for production trading or investment decisions
• No warranties or guarantees of any kind
• Past backtest performance does not predict future results
• Use at your own risk for research purposes only

---

License

Personal Open Research License. See LICENSE file for details.