This project demonstrates a production-ready ML pipeline that predicts gold market movements by analyzing financial news using Large Language Models. It combines fine-tuned Mistral-7B with FinBERT sentiment analysis to generate trading signals across multiple time horizons.

Business Value

• Automated Trading Signals: Predict gold price direction and magnitude from financial news
• Multi-Horizon Forecasting: 6h, 12h, 24h, and 48h prediction windows
• Scalable Infrastructure: AWS SageMaker with GPU acceleration
• Cost-Effective: LoRA fine-tuning reduces training costs by 90%

LoRA Fine-tuning Strategy

Instead of fine-tuning all 7 billion parameters of Mistral, we use Low-Rank Adaptation (LoRA) which adds small trainable matrices to the model's attention layers.

# LoRA Configuration
lora_config = LoraConfig(
    r=16,              # Rank of update matrices
    lora_alpha=32,     # Scaling factor
    target_modules=[   # Which layers to adapt
        "q_proj", "k_proj", "v_proj", "o_proj"
    ],
    lora_dropout=0.05,
    bias="none",
    task_type="CAUSAL_LM"
)

# Training Configuration
training_args = TrainingArguments(
    output_dir="./mistral-gold-lora",
    num_train_epochs=3,
    per_device_train_batch_size=4,
    gradient_accumulation_steps=4,
    learning_rate=2e-4,
    fp16=True,                    # Mixed precision
    logging_steps=50,
    save_strategy="epoch",
    evaluation_strategy="epoch"
)

Data Pipeline Architecture

The system processes financial news through multiple stages, each adding value:

Keywords: ["gold", "XAU", "precious metals", "Fed", "inflation", 
           "dollar", "treasury yields", "safe haven"]

# API Call Example
news = alpaca.get_news(
    symbols=["GLD", "IAU"],
    start=datetime.now() - timedelta(hours=24),
    limit=100
)
filtered = [n for n in news if any(kw in n.headline.lower() 
            for kw in gold_keywords)]

{
  "sentiment": "Positive",
  "confidence": 0.94
}

-- Athena Query for Feature Aggregation
SELECT 
    news_timestamp,
    AVG(sentiment_score) as avg_sentiment,
    xau_close - LAG(xau_close, 1) as price_change,
    dxy_close as dollar_strength,
    treasury_10y as yield_level,
    gld_volume / AVG(gld_volume) OVER (ORDER BY timestamp 
        ROWS BETWEEN 20 PRECEDING AND CURRENT ROW) as volume_ratio
FROM market_data
WHERE timestamp BETWEEN news_timestamp - INTERVAL '2' HOUR 
    AND news_timestamp

Model Output Format

The fine-tuned model outputs structured JSON predictions for 4 time horizons:

{
  "6h": {
    "direction": "Up",
    "magnitude": "Medium",
    "confidence": 0.82
  },
  "12h": {
    "direction": "Up", 
    "magnitude": "High",
    "confidence": 0.76
  },
  "24h": {
    "direction": "Neutral",
    "magnitude": "Low",
    "confidence": 0.65
  },
  "48h": {
    "direction": "Down",
    "magnitude": "Medium",
    "confidence": 0.58
  }
}