Meta-Fusion Trading System

Advanced automated trading system leveraging deep learning ensemble for Korean stock market (KOSPI/KOSDAQ). Combines LSTM sequence learning, multi-model gradient boosting, meta-fusion network, and adversarial training for robust performance.

Performance Results

Backtest Performance (2020-2023, Samsung Electronics)

Metric	Value
Total Return	57.98%
Annualized Return	12.18%
Sharpe Ratio	0.76
Maximum Drawdown	-33.33%
Win Rate	100%
Total Trades	10

Portfolio value evolution from 10M KRW initial capital, compared with buy-and-hold benchmark.

Strategy returns vs benchmark over the full backtest period.

Peak-to-trough decline showing maximum drawdown of -33.33%.

---

Architecture Overview

AI Ensemble System (80%)                    Traditional (20%)
┌──────────────────────────────────┐       ┌───────────────┐
│  LSTM (30%)                      │       │  Momentum     │
│  XGBoost (25%)                   │       │  Indicators   │
│  LightGBM (15%)                  │       │               │
│  CatBoost (10%)                  │       └───────┬───────┘
│                                  │               │
│  ┌────────────────────┐          │               │
│  │ Meta-Fusion        │◄─────────┼───────────────┤
│  │ Network            │          │               │
│  └────────────────────┘          │               ▼
│  + Market Context                │        ┌─────────────┐
└──────────────────────────────────┘        │   Signal    │
                                            │ + Position  │
                                            └─────────────┘

Core Components:

1. LSTM Sequence Learner (30%) - Temporal pattern recognition in 20-day windows
2. Multi-Model Gradient Boosting (50%) - XGBoost, LightGBM, CatBoost ensemble
3. Meta-Fusion Network - Neural network for optimal model combination
4. Adversarial Training - Robustness enhancement
5. Momentum Support (20%) - Traditional technical indicators

---

Quick Start

Installation

git clone <repository-url>
cd ai-automated-trade
python -m venv venv
source venv/bin/activate
pip install -r requirements.txt

Training

# Train all models (LSTM + XGBoost + LightGBM + CatBoost + Meta-Fusion)
python train.py --config config.yaml --device cpu

# With GPU (if available)
python train.py --device cuda

Training Time: 2-3 hours on GPU, 5-6 hours on CPU

Testing

# Run integration test
python test.py

# Run backtest
python backtest.py

---

Project Structure

ai-automated-trade/
├── train.py                       # Training pipeline
├── test.py                        # Integration test
├── backtest.py                    # Backtesting script
├── config.yaml                    # Configuration
├── requirements.txt               # Dependencies
│
├── src/
│   ├── data/
│   │   ├── pipeline.py            # Data loading and preprocessing
│   │   └── sequences.py           # LSTM sequence preparation
│   ├── features/
│   │   └── indicators.py          # Technical indicators + market context
│   ├── models/
│   │   ├── lstm_predictor.py      # LSTM sequence learner
│   │   ├── feature_learner.py     # Neural feature extraction
│   │   ├── gb_ensemble.py         # Multi-model gradient boosting
│   │   └── meta_fusion.py         # Meta-fusion network
│   └── utils/
│       └── config.py              # Configuration utilities
│
├── models/                        # Trained model storage
├── results/                       # Backtest results and charts
└── data/                          # Market data cache

---

System Features

Neural Sequence Learning

• LSTM Architecture: 2-layer LSTM with 128 hidden units
• Sequence Length: 20-day temporal patterns
• Feature Learning: Auto-extraction from raw OHLCV data
• Parameters: 220,859 trainable parameters

Multi-Model Ensemble

• XGBoost: 200 estimators, all features (25% weight)
• LightGBM: 300 estimators, top 70% features (15% weight)
• CatBoost: 250 iterations, random 80% features (10% weight)
• Feature Diversity: Each model sees different feature subsets

Intelligent Fusion

• Meta-Fusion Network: 13 → 32 → 16 → 3 architecture
• Inputs: 4 model predictions + 5 market context features
• Learning: Optimal combination learned from historical data
• Adaptability: Adjusts to changing market conditions

Robustness Enhancement

• Adversarial Training: Hard example mining and noise injection
• Early Stopping: Prevents overfitting during training
• Validation: Comprehensive out-of-sample testing

---

Configuration

Edit config.yaml to customize:

models:
  lstm:
    sequence_length: 20
    hidden_size: 128
    num_layers: 2
    dropout: 0.3
    learning_rate: 0.001
    num_epochs: 100
  
  xgboost:
    n_estimators: 200
    max_depth: 6
    learning_rate: 0.1
  
  lightgbm:
    n_estimators: 300
    feature_fraction: 0.7
  
  catboost:
    iterations: 250
    depth: 4
  
  meta_fusion:
    hidden_dims: [32, 16]
    dropout: 0.2
  
  weights:
    lstm: 0.30
    xgboost: 0.25
    lightgbm: 0.15
    catboost: 0.10
    momentum: 0.20

data:
  ticker: "005930.KS"  # Samsung Electronics
  start_date: "2015-01-01"
  end_date: "2023-12-31"

---

Dependencies

Deep Learning:

• torch>=2.0.0 (LSTM, Meta-Fusion)
• xgboost>=2.0.0
• lightgbm>=4.0.0
• catboost>=1.2.0

Data & Utilities:

• pandas, numpy
• pandas-ta (Technical indicators)
• yfinance (Market data)
• scikit-learn (Preprocessing)

See requirements.txt for complete list.

---

Testing

# Run all tests
pytest tests/ -v

# Integration test
python test.py

# Full backtest
python backtest.py

---

Technical Specifications

LSTM Architecture

Input: (batch, 20, 18)  # 20 days, 18 features
  ↓
LSTM Layer 1: 128 units, dropout 0.3
  ↓
LSTM Layer 2: 128 units, dropout 0.3
  ↓
FC Layers: 128 → 64 → 3 (Buy/Hold/Sell)

Meta-Fusion Network

Input: 13 features
  - 4 models × 2 (signal, confidence) = 8
  - 5 market context features = 5
  ↓
Dense: 13 → 32, ReLU, Dropout(0.2)
  ↓
Dense: 32 → 16, ReLU
  ↓
Output: 16 → 3 (Buy/Hold/Sell probabilities)

References

Deep Learning:

• Hochreiter & Schmidhuber (1997). "Long Short-Term Memory"
• Goodfellow et al. (2014). "Explaining and Harnessing Adversarial Examples"

Gradient Boosting:

• Chen & Guestrin (2016). "XGBoost: A Scalable Tree Boosting System"
• Ke et al. (2017). "LightGBM: A Highly Efficient Gradient Boosting Decision Tree"
• Prokhorenkova et al. (2018). "CatBoost: Unbiased Boosting with Categorical Features"

Financial Machine Learning:

• López de Prado (2018). "Advances in Financial Machine Learning"

---

Disclaimer

This software is for educational and research purposes only. Trading involves significant financial risk. Past performance does not guarantee future results. Never risk capital you cannot afford to lose.