Artificial Intelligence Models for Predicting Stock Returns Using Fundamental, Technical, and Entropy-Based...

Area of Science:

• Quantitative Finance
• Artificial Intelligence in Finance
• Computational Economics

Background:

• Traditional machine learning (ML) models have limitations in capturing nuanced market sentiment.
• Large Language Models (LLMs) offer advanced semantic understanding, potentially improving financial predictions.
• Predictive portfolio strategies require integrating diverse data sources and analytical methods.

Purpose of the Study:

• To evaluate the synergistic effect of combining LLM-derived semantic intelligence with traditional ML algorithms.
• To develop and test novel predictive portfolio strategies for NASDAQ-100 stocks.
• To determine optimal fusion methods of ML and LLM insights across different predictive frameworks.

Main Methods:

• Utilized three predictive frameworks: fundamental, technical, and entropy-based.
• Integrated ML algorithms with semantic metrics derived from LLMs (e.g., ChatGPT-4o).
• Analyzed NASDAQ-100 stock data for the 2020-2025 period with monthly rebalancing.

Main Results:

• Technical methodology performed best with ML predictions alone, yielding ~1978% cumulative returns.
• Fundamental methodology showed maximum potential when primarily using LLM-derived semantic insights.
• Entropy methodology improved with a balanced mix of ML and LLM signals, demonstrating LLM's contextual value.

Conclusions:

• The optimal blending of ML and LLM for predictive portfolio strategies is methodology-dependent.
• LLMs provide interpretative context for complex market interactions, enhancing predictive power.
• Tailoring semantic-algorithmic fusion to data nature and investment horizon is crucial for effective portfolio management.