NexQuant/QWEN.md

Predix - QWEN.md Context File

Project Overview

Predix is an autonomous AI-powered quantitative trading agent for EUR/USD forex markets. Built on the RD-Agent framework, it automates the full research and development cycle for trading strategies.

Core Purpose

• Generate trading factors (signals) autonomously using LLMs
• Backtest and validate factors on 1-minute EUR/USD data
• Optimize portfolios using modern portfolio theory
• Target: 1-3% monthly returns with Sharpe > 2.0

Key Technologies

• Python 3.10/3.11 - Primary language
• PyTorch - Deep learning models
• Qlib - Backtesting engine
• LLM (Qwen3.5-35B) - Factor generation via local llama.cpp
• Flask - Web dashboard API
• SQLite - Results database
• Rich/Typer - CLI interface

Architecture

Predix/
├── rdagent/                    # Core agent framework
│   ├── app/
│   │   └── cli.py              # Main CLI entry point (rdagent command)
│   ├── components/
│   │   ├── backtesting/        # Backtest engine, metrics, database
│   │   ├── coder/
│   │   │   └── factor_coder/   # Factor generation & EURUSD-specific modules
│   │   ├── loader.py           # Prompt loader (auto-loads local prompts)
│   │   └── model_loader.py     # Model loader (auto-loads local models)
│   └── scenarios/
│       └── qlib/               # Qlib integration for FX trading
├── prompts/                    # LLM Prompts
│   ├── standard_prompts.yaml   # Standard prompts (in Git)
│   └── local/                  # Your improved prompts (NOT in Git!)
│       ├── factor_discovery_v2.yaml
│       ├── factor_evolution_v2.yaml
│       └── model_coder_v2.yaml
├── models/                     # ML Models
│   ├── standard/               # Standard models (in Git)
│   │   ├── xgboost_factor.py
│   │   └── lightgbm_factor.py
│   └── local/                  # Your improved models (NOT in Git!)
│       ├── transformer_factor.py
│       ├── tcn_factor.py
│       ├── patchtst_factor.py
│       └── cnn_lstm_hybrid.py
├── results/                    # Backtest results (NOT in git)
│   ├── backtests/              # Individual factor backtests (JSON/CSV)
│   ├── db/                     # SQLite database
│   ├── factors/                # Factor analysis
│   ├── runs/                   # Run results & risk reports
│   └── logs/                   # Backtest logs
├── web/                        # Dashboard frontend
│   ├── dashboard_api.py        # Flask API backend
│   └── dashboard.html          # Web UI
├── .env                        # Environment config (API keys, etc.)
├── data_config.yaml            # EURUSD data configuration
└── requirements.txt            # Python dependencies

Open Source vs. Closed Source

🟢 OPEN SOURCE (Public on GitHub):

• rdagent/ - Core framework
• models/standard/ - Base models (XGBoost, LightGBM)
• prompts/standard_prompts.yaml - Base prompts
• web/ - Dashboards
• test/ - Tests

🔒 CLOSED SOURCE (Local Only - NOT on GitHub):

• models/local/ - Your improved models (Transformer, TCN, PatchTST, CNN+LSTM)
• prompts/local/ - Your improved prompts (v2.0 optimized)
• .env - API keys
• results/ - Backtest results
• git_ignore_folder/ - Trading data
• QWEN.md, TODO.md - Internal docs

Protection:

• .gitignore excludes all local/ directories
• Your competitive edge (alpha) stays private
• Framework is open, but your best models/prompts are closed

Building and Running

Installation

# Clone repository
git clone https://github.com/PredixAI/predix
cd predix

# Create conda environment
conda create -n predix python=3.10
conda activate predix

# Install in editable mode
pip install -e .[test,lint]

Configuration

1. Create .env file:

# Local LLM (llama.cpp)
OPENAI_API_KEY=local
OPENAI_API_BASE=http://localhost:8081/v1
CHAT_MODEL=qwen3.5-35b

# Embedding (Ollama)
LITELLM_PROXY_API_KEY=local
LITELLM_PROXY_API_BASE=http://localhost:11434/v1
EMBEDDING_MODEL=nomic-embed-text

# Paths
QLIB_DATA_DIR=~/.qlib/qlib_data/eurusd_1min_data

2. Start LLM server (llama.cpp):

~/llama.cpp/build/bin/llama-server \
  --model ~/models/qwen3.5/Qwen3.5-35B-A3B-Q3_K_M.gguf \
  --n-gpu-layers 36 \
  --ctx-size 80000 \
  --port 8081

Running the Trading Loop

# Start trading loop (24/7)
./start_loop.sh

# Or single run
rdagent fin_quant

# With dashboard
rdagent fin_quant --with-dashboard

# With CLI dashboard
rdagent fin_quant --cli-dashboard

Running the Dashboard

# Web dashboard (runs with fin_quant --with-dashboard)
# Access at: http://localhost:5000/dashboard.html

# Or standalone
python web/dashboard_api.py

Testing

Integration Test Suite (ALL Features)

Comprehensive test system that validates ALL 13 implemented features:

# Run ALL integration tests (60 tests, ~7.5 seconds)
pytest test/integration/test_all_features.py -v

# Run with coverage report
pytest test/integration/test_all_features.py --cov=rdagent.components.backtesting -v

# Run via test runner script
./scripts/run_all_tests.sh

# Test specific features only
pytest test/integration/test_all_features.py -k "backtest or database" -v

# Skip slow tests
pytest test/integration/test_all_features.py -m "not slow" -v

Tested Features (60 Tests, ALL MUST PASS):

#	Feature	Tests	Status
1	Factor Evolution	5	✅ LLM generates trading factors autonomously
2	Model Evolution	5	✅ ML models auto-improved
3	Quant Loop (fin_quant)	4	✅ Main 24/7 trading loop
4	Backtesting Engine	5	✅ IC, Sharpe, Drawdown, Win Rate
5	Results Database	5	✅ SQLite with queries
6	Risk Management	6	✅ Correlation, Portfolio Optimization
7	CLI Dashboard	4	✅ Rich live-progress display
8	Web Dashboard	4	✅ Flask API + HTML
9	Health Check	4	✅ Environment validation
10	Streamlit UI	3	✅ Alternative dashboard
11	LLM Integration	5	✅ llama.cpp (Qwen3.5-35B)
12	Embedding	3	✅ Ollama (nomic-embed-text)
13	Security Scanning	5	✅ Bandit pre-commit hook

⚠️ MANDATORY: These tests run BEFORE every commit and MUST pass!

Unit Tests

# Run all unit tests
pytest test/

# Run with coverage
pytest --cov=rdagent --cov-report=html

# Test backtesting module
python rdagent/components/backtesting/backtest_engine.py
python rdagent/components/backtesting/results_db.py
python rdagent/components/backtesting/risk_management.py

Code Quality

# Linting
ruff check rdagent/

# Type checking
mypy rdagent/

# Format
black rdagent/

# Pre-commit (install first)
pre-commit install
pre-commit run --all-files

Development Conventions

Language Policy

ALL code comments and documentation MUST be in English.

❌ Wrong (German):

# Inspiriert von: TradingAgents
# Berechnet den Sharpe Ratio
# Achtung: Division durch Null möglich!
# Hinweis: Diese Funktion ist experimentell

✅ Correct (English):

# Inspired by: TradingAgents
# Calculates the Sharpe ratio
# Warning: Division by zero possible!
# Note: This function is experimental

Rationale:

• International collaboration
• Better searchability
• Professional codebase
• Consistent with commit messages (also English-only)

Enforcement:

• All new code must have English comments
• Existing German comments should be translated when modified
• PRs with German comments will be rejected

Code Style

• Line length: 120 characters (configured in pyproject.toml)
• Type hints: Required for all public functions
• Docstrings: Google style for public APIs
• Imports: Sorted automatically with isort

Testing Practices

• Unit tests in test/ directory
• Test files named test_*.py
• Use pytest fixtures for common setup
• Mock external APIs (LLM, yfinance)
• Minimum 80% coverage target

Commit Conventions

git commit --author="TPTBusiness <tpt.requests@pm.me>" -m "type: description"

# Types:
# - feat: New feature
# - fix: Bug fix
# - docs: Documentation
# - style: Formatting
# - refactor: Code restructuring
# - test: Tests
# - chore: Maintenance

Module Structure

"""
Module Name - Brief description

Longer description if needed.
"""

import numpy as np
import pandas as pd
from typing import Dict, List, Optional
from datetime import datetime

class ClassName:
    """Class docstring."""
    
    def __init__(self, param: type) -> None:
        """Initialize."""
        pass
    
    def method(self, param: type) -> ReturnType:
        """
        Method docstring.
        
        Parameters
        ----------
        param : type
            Description
        
        Returns
        -------
        ReturnType
            Description
        """
        pass

Backtesting Module Usage

from rdagent.components.backtesting import (
    FactorBacktester,
    ResultsDatabase,
    PortfolioOptimizer,
    AdvancedRiskManager
)

# Run backtest
backtester = FactorBacktester()
metrics = backtester.run_backtest(
    factor_values=factor_series,
    forward_returns=forward_returns,
    factor_name="MyFactor"
)

# Save to database
db = ResultsDatabase()
db.add_backtest("MyFactor", metrics)

# Query top factors
top = db.get_top_factors('sharpe_ratio', limit=20)

# Portfolio optimization
optimizer = PortfolioOptimizer()
weights = optimizer.mean_variance(expected_returns, cov_matrix)

# Risk management
risk_manager = AdvancedRiskManager()
report = risk_manager.generate_risk_report(returns, weights)

Key Metrics

Metric	Target	Minimum
IC (Information Coefficient)	> 0.05	> 0.02
Sharpe Ratio	> 2.0	> 1.0
Max Drawdown	< 15%	< 25%
Win Rate	> 55%	> 45%
Annualized Return	> 10%	> 5%

Important Files

• rdagent/app/cli.py - Main CLI entry point
• rdagent/components/backtesting/ - Backtest engine
• rdagent/components/coder/factor_coder/ - Factor generation
• results/README.md - Results documentation
• data_config.yaml - EURUSD configuration
• web/dashboard_api.py - Dashboard API
• requirements.txt - Dependencies

External Dependencies

• llama.cpp - Local LLM inference (Qwen3.5-35B)
• Ollama - Embedding models
• Qlib - Backtesting engine
• yfinance - Live market data

Common Issues

1. LLM Connection Errors: Ensure llama.cpp server is running on port 8081
2. Embedding Errors: Check Ollama is running with nomic-embed-text loaded
3. Database Lock: Close all connections before running multiple processes
4. Memory Issues: Reduce batch size or context length for LLM

Project Status

• ✅ Factor Generation (110+ factors created)
• ✅ Backtesting Engine (IC, Sharpe, Drawdown)
• ✅ Results Database (SQLite with queries)
• ✅ Risk Management (Correlation, Portfolio Optimization)
• ✅ Dashboards (Web + CLI)
• ⏳ Live Trading (Paper trading pending)

Next Steps

1. Backtest all 110 factors
2. Select top 20 by IC/Sharpe
3. Portfolio optimization
4. 4 weeks paper trading
5. Live trading with small capital

---

Git Commit Guidelines

Language Policy

ALL commit messages MUST be in English.

❌ Wrong (German):

git commit -m "feat: Neue Funktion hinzugefügt"
git commit -m "fix: Fehler behoben"
git commit -m "chore: QWEN.md zu .gitignore hinzugefügt"

✅ Correct (English):

git commit -m "feat: Add new feature"
git commit -m "fix: Fix bug"
git commit -m "chore: Add QWEN.md to .gitignore"

Pre-Commit Checklist

BEFORE every commit, you MUST:

1. Run git status and verify:

   • Only intended files are staged
   • No generated files (.qwen/, results/, *.db, etc.)
   • No sensitive data (.env, API keys, etc.)

2. Check .gitignore is working:

   git status
   # Verify .qwen/, results/, *.db are NOT shown
   

3. Review staged changes:

   git diff --staged
   # Review what will be committed
   

4. Run tests (if applicable):

   pytest test/backtesting/ -v
   # Ensure all tests pass
   

Commit Message Format

Use Conventional Commits:

<type>: <description in English>

[optional body]

Types:

• feat: - New feature
• fix: - Bug fix
• test: - Tests
• docs: - Documentation
• chore: - Maintenance
• style: - Formatting
• refactor: - Code restructuring

Examples:

feat: Add backtesting tests with 98% coverage
fix: Remove .qwen/ from Git tracking
test: Add unit tests for ResultsDatabase
docs: Update QWEN.md with commit guidelines
chore: Add pytest to requirements.txt

Protected Files (NEVER commit)

These files/directories MUST NEVER be committed:

.qwen/              # AI agent files (generated)
results/            # Backtest results (sensitive data)
*.db                # SQLite databases
.env                # Environment variables (API keys!)
git_ignore_folder/  # Generated data
*.log               # Log files

If you accidentally commit any of these:

# Remove from last commit (keeps files locally)
git reset HEAD~1

# Or remove from tracking
git rm -r --cached .qwen/
git commit -m "chore: Remove .qwen/ from tracking"

Fixing Past Commits

To fix the last 3-5 commits:

# For last 5 commits
git rebase -i HEAD~5

# In the editor, change 'pick' to 'reword' for commits to rename
# Save and close
# Write new English message for each commit

To fix older commits (advanced):

# Find the commit hash
git log --oneline

# Start rebase from that commit
git rebase -i <commit-hash>^

# Follow same process as above

Current German commits to fix (as of April 2026):

73140b68 test: Backtesting Tests mit 98.77% Coverage
     → test: Add backtesting tests with 98.77% coverage

5148d17d chore: QWEN.md zu .gitignore hinzugefügt
     → chore: Add QWEN.md to .gitignore

df93e162 feat: Intelligent Embedding Chunking statt Kürzung
     → feat: Intelligent embedding chunking instead of truncation

01aa183a fix: CLI Dashboard in separatem Terminal-Fenster
     → fix: CLI dashboard in separate terminal window

df356978 feat: predix.py Wrapper für Dashboard-Support
     → feat: predix.py wrapper for dashboard support

89d01f5d feat: Beautiful CLI Dashboard + korrigierter Start-Befehl
     → feat: Beautiful CLI dashboard + corrected start command

48e4f44e feat: Auto-Start Dashboard für fin_quant
     → feat: Auto-start dashboard for fin_quant

59122a19 feat: Dashboard + Live-Daten Integration (Phase 4)
     → feat: Dashboard + live data integration (Phase 4)

a0f414ed feat: EURUSD Trading-Verbesserungen (Phase 2 & 3)
     → feat: EURUSD trading improvements (Phase 2 & 3)

e8b962b5 feat: EURUSD Trading-Verbesserungen implementiert (Phase 1)
     → feat: Implement EURUSD trading improvements (Phase 1)

⚠️ Warning: Rewriting history changes commit hashes. If you've already pushed:

# After rebasing locally
git push --force-with-lease origin master

# Tell team members to re-clone:
git clone <repo-url>

Push Policy

BEFORE pushing:

1. Verify commit messages are in English
2. Verify no protected files are included
3. Run tests one final time

git status
git log -3 --oneline  # Verify last 3 commits
pytest test/backtesting/ -v  # Quick test
git push origin master

Enforcement

• All PRs will be rejected if commit messages are not in English
• Protected files in commits will be rejected
• Tests must pass before merging

Remember: Consistent English commit messages ensure:

• International collaboration
• Better searchability
• Professional project history

---

Implementation Guide: Prompts & Models

Using the Prompt Loader

Auto-Load Prompts (Local First):

from rdagent.components.loader import load_prompt

# Load factor discovery prompt
# Automatically loads from prompts/local/ if exists!
prompt = load_prompt("factor_discovery")

# Load specific section
system_prompt = load_prompt("factor_discovery", section="system")
user_prompt = load_prompt("factor_discovery", section="user")

# Force local only (raise error if not found)
prompt = load_prompt("factor_discovery", local_only=True)

# List available prompts
from rdagent.components.loader import list_available_prompts
available = list_available_prompts()
print(f"Standard: {available['standard']}")
print(f"Local: {available['local']}")

Priority:

1. prompts/local/factor_discovery_v2.yaml (loaded first if exists)
2. prompts/local/factor_discovery.yaml
3. prompts/standard_prompts.yaml (fallback)

---

Using the Model Loader

Auto-Load Models (Local First):

from rdagent.components.model_loader import load_model

# Load XGBoost model
# Automatically loads from models/local/ if exists!
model_factory = load_model("xgboost_factor")

# Create model instance
model = model_factory(max_depth=8, learning_rate=0.03)

# Train
model.fit(X_train, y_train, epochs=50, batch_size=64)

# Predict
predictions = model.predict(X_test)

# Save/Load
model.save("models/my_model.pth")
model.load("models/my_model.pth")

Available Models:

Model	Location	Use Case
xgboost_factor	models/standard/	Tabular data, fast training
lightgbm_factor	models/standard/	Large datasets, faster than XGBoost
transformer_factor	models/local/	Time-series, long-range dependencies
tcn_factor	models/local/	Multi-scale patterns
patchtst_factor	models/local/	SOTA for time-series forecasting
cnn_lstm_hybrid	models/local/	Complex pattern recognition

Priority:

1. models/local/{name}_v2.py (loaded first if exists)
2. models/local/{name}.py
3. models/standard/{name}.py (fallback)

---

Creating Your Improved Prompts

Step 1: Create Local Prompt

mkdir -p prompts/local
nano prompts/local/factor_discovery_v3.yaml

Step 2: Add Your Improvements

# prompts/local/factor_discovery_v3.yaml

factor_discovery:
  system: |-
    YOUR IMPROVED SYSTEM PROMPT HERE
    
    Add your proprietary insights:
    - Specific EURUSD patterns you've discovered
    - Your unique factor formulas
    - Custom session filters
    - Proprietary risk management rules
    
  user: |-
    YOUR IMPROVED USER PROMPT HERE

Step 3: Test

from rdagent.components.loader import load_prompt

# Auto-loads your v3!
prompt = load_prompt("factor_discovery")

---

Creating Your Improved Models

Step 1: Create Local Model

mkdir -p models/local
nano models/local/my_optimized_model.py

Step 2: Implement Model

# models/local/my_optimized_model.py
"""
My Optimized Model v1.0
Better than standard with custom improvements.
"""

import torch
import torch.nn as nn

class MyOptimizedModel(nn.Module):
    def __init__(self, **params):
        super().__init__()
        # Your custom architecture
        pass
    
    def forward(self, x):
        # Your custom forward pass
        pass

def create_my_optimized_model(**params):
    """Factory function."""
    return MyOptimizedModel(**params)

Step 3: Test

from rdagent.components.model_loader import load_model

# Auto-loads your optimized model!
model_factory = load_model("my_optimized_model")
model = model_factory()

---

Backup Your Private Assets

Backup Prompts & Models to Private Repo:

# Create private repo on GitHub: predix-private-assets

# Clone private repo
cd ~/Dev
git clone git@github.com:TPTBusiness/predix-private-assets.git

# Copy local assets
cp -r ~/Predix/prompts/local/* ~/predix-private-assets/prompts/
cp -r ~/Predix/models/local/* ~/predix-private-assets/models/

# Commit to private repo
cd ~/predix-private-assets
git add .
git commit -m "Backup: prompts v2, models (Transformer, TCN, PatchTST, CNN+LSTM)"
git push

Auto-Sync Script:

# ~/Predix/sync_private.sh
#!/bin/bash
echo "Syncing private assets..."
rsync -av prompts/local/ ~/predix-private-assets/prompts/
rsync -av models/local/ ~/predix-private-assets/models/
cd ~/predix-private-assets && git add . && git commit -m "Auto-sync $(date)" && git push
echo "Done!"

---

Security Best Practices

What to Keep Private:

✅ Your proprietary model architectures ✅ Optimized prompt templates ✅ Best-performing factors ✅ Evolution weights ✅ Trade secrets & alpha-generating logic

What NOT to Commit:

❌ Anything in prompts/local/ ❌ Anything in models/local/ ❌ .env (API keys) ❌ results/ (backtest performance) ❌ git_ignore_folder/ (trading data)

Verify Before Committing:

# Check what will be committed
git status
git diff --staged

# Verify .gitignore is working
git status
# Should NOT show prompts/local/, models/local/, .env, results/