Abstract
Type of the article: Research Article AbstractSentiment analysis of financial text data plays a crucial role in investment decision-making, yet existing approaches often rely on single-model sentiment scores that may suffer from biases or hallucinations. This study aims to enhance portfolio optimization by integrating sentiment signals from multiple Large Language Models (LLMs) into the Black-Litterman framework. The proposed method aggregates sentiment scores from three finance-domain fine-tuned LLMs using a Long Short-Term Memory network, which captures non-linear relationships and temporal dependencies to produce a robust Meta-LLM sentiment score. This score is then incorporated into the Black-Litterman model as investor views to derive optimal portfolio weights. The methodology is tested on a portfolio of S&P 500 stocks. The results show that the proposed approach significantly improves portfolio performance, achieving an annualized return of 31.22%, compared to 24.57% for the market capital-weighted portfolio. Additionally, the model attains a Sharpe Ratio of 3.02, an Omega Ratio of 2.48, and a Jensen’s Alpha of 1.95%, outperforming both the benchmark portfolios and portfolios based on single-LLM sentiment. The findings demonstrate that aggregating sentiment from multiple LLMs enhances risk-adjusted returns while mitigating model-specific limitations. Future research could explore the integration of LLMs with different architectures to further refine sentiment-aware portfolio strategies.