GLD-Trader

An algorithmic trading model for GLD ETF using hybrid ARIMA-LSTM architecture with multi-source data integration (macro data, stock prices, and news data with FinBERT sentiment analysis)

Model Architecture

The project implements an ensemble approach combining ARIMA and LSTM models:

1. ARIMA Model: Captures linear trends and seasonality using technical indicators and macroeconomic features
2. LSTM Model: Models non-linear patterns in ARIMA residuals for improved prediction accuracy
3. Ensemble Prediction: Final predictions combine ARIMA forecasts with LSTM-predicted residual corrections

Data Sources

Market Data

• GLD ETF: Daily OHLCV data with technical indicators
• Technical Indicators: EMA (30, 60, 200), RSI (14), MACD, Bollinger Bands, Momentum (10)

Macroeconomic Data

• DTWEXBGS: US Dollar Trade Weighted Exchange Rate
• CPIAUCSL: Consumer Price Index (inflation)
• FEDFUNDS: Federal Funds Rate
• GDP: Gross Domestic Product
• OIL_PRICE: Crude Oil Prices

Sentiment Analysis

• FinBERT: Financial news sentiment analysis using pre-trained FinBERT model
• News Sources: Daily financial news related to gold and market conditions

Model Performance

Project Structure

GLD-Trader/
├── data/
│   ├── ingestion/           # Data collection scripts
│   ├── combined_data.csv    # Final processed dataset
│   ├── finbert_gold_news.csv # Sentiment-analyzed news
│   └── macro_data.csv       # Macroeconomic indicators
├── models/
│   └── training.ipynb       # Model development and training
├── images/                  # Performance visualizations
└── logs/                    # Data ingestion logs

Data Features

The model uses 20+ features including:

• Technical indicators (EMAs, RSI, MACD, Bollinger Bands)
• Macroeconomic indicators (interest rates, inflation, USD strength, oil prices)
• Market volume and momentum
• News sentiment scores (FinBERT-based)

Implementation Details

ARIMA Component

• Order: (1, 1, 0) with exogenous variables
• Uses all technical and macroeconomic features
• Captures linear market trends and fundamental relationships

LSTM Component

• 16 hidden units with 0.5 dropout
• 14-day sliding window for residual prediction
• SGD optimizer with momentum and learning rate scheduling
• Trained for 1000 epochs with early stopping

Training Strategy

• 60% ARIMA training, 24% LSTM training, 16% final testing
• StandardScaler normalization for LSTM features
• Residual-based ensemble approach for improved accuracy

In Progress

Currently adding trading logic, integrating with Alpaca, and sourcing better news data.

Usage

1. Run data ingestion scripts in data/ingestion/
2. Execute model training in models/training.ipynb
3. View performance metrics and visualizations
4. Apply trained model for GLD price predictions

Dependencies

See requirements.txt for complete package list including:

• pandas, numpy, scikit-learn
• statsmodels (ARIMA)
• torch (LSTM)
• transformers (FinBERT)
• matplotlib, seaborn (visualization)