📈 Language Models for Rational, Evidence-Driven Trading

This project explores the use of Large Language Models (LLMs) and machine learning to build an intelligent, explainable trading agent.
By combining textual signals (from financial news, political speeches, and social media) with numerical market indicators, we aim to forecast XAU-USD trends with actionable insights and human-readable justifications.

---

✨ Key Features

• 📊 Multimodal Feature Extraction

  • Numerical: OHLCV data + 15+ technical indicators (SMA, EMA, MACD, RSI, Bollinger Bands, OBV, etc.).
  • Textual: 330K+ financial news articles (2020–2025), clustered into interpretable event categories.

• 🧠 Hybrid Modeling

  • XGBoost for structured + textual features.
  • ReAct LLM Agent for reasoning-based decision-making using few-shot prompting and chain-of-thought strategies.

• 🗣️ Explainable Decisions

  • SHAP values for feature attribution.
  • LLM-generated natural language explanations aligned with SHAP and textual inputs.

• 💬 Multi-role LLM Agents

  • Roles: filtering, clustering, predicting, explaining.

• 📈 Profitability + Explainability Evaluation

  • Metrics: Accuracy, Sharpe Ratio, Cumulative Profit, Per-Trade Profit.
  • Explanation quality: coverage, fidelity, stability, sentiment alignment.

---

📂 Dataset

• Numerical Data:

  • Daily OHLCV for gold (XAU-USD), oil, and equities.
  • Technical indicators engineered with leakage-safe lagging.

• Textual Data:

  • 331,689 financial articles (Financial Post, Yahoo Finance, political speeches).
  • Clustered via:
    • Transformer embeddings + k-means (scalable).
    • LLM-based semantic clustering (interpretable).

---

⚙️ Methodology

1. Profit-Optimal Labeling

   • BUY / SELL / HOLD labels derived from utility-maximizing actions with transaction costs.

2. Hybrid Prediction

   • Path A: Accuracy-based XGBoost.
   • Path B: Utility-weighted (profit-aware) XGBoost.
   • Convex ensemble + HOLD-threshold for risk control.

3. Explainability

   • TreeSHAP for local/global attribution.
   • LLM explanations referencing SHAP features + relevant news.

4. ReAct Agent

   • LLM agent integrates daily technical indicators + news summaries.
   • Multi-step reasoning before outputting decision + rationale.

---

🧪 Experiments

• Models compared:

  • XGBoost + LLM explanations.
  • ReAct LLM agent (with/without CoT and few-shot).

• Evaluation metrics:

  • Accuracy
  • Sharpe Ratio
  • Cumulative Profit
  • Per-Trade Profit
  • Explanation quality

---

📌 Discussion

• XGBoost outperforms standalone LLMs in profitability and robustness.
• Hybrid modeling (XGBoost + LLM) achieves the best trade-off between accuracy and interpretability.
• Textual features provide unique early signals (e.g., “rate hike,” “inflation fears”) not captured by indicators.
• Explanations are concise, news-aware, and semantically aligned, though ReAct agents sometimes provide more faithful reasoning.

---

🚀 Conclusion

This project demonstrates that multimodal, explainable AI can advance algorithmic trading by balancing accuracy, interpretability, and profitability.

Our hybrid framework—combining XGBoost feature modeling with LLM-based reasoning and explanations—achieves superior financial performance while maintaining human-aligned justifications.

---