Integrating Large Language Models and CNN-LSTM for Enhanced Portfolio Optimization

Portfolio optimization is a core financial problem that involves balancing risk and return. Although Modern Portfolio Theory (MPT) gives a fundamental framework, its reliance on Gaussian distributions of returns often fails in real-world practice. Recent deep learning developments present data-driven solutions; such approaches are typically marred by a lack of interpretability and theoretical foundation. In this paper, we present a hybrid deep learning approach (CNN-LSTM) augmented with large language models (LLMs) for enhanced prediction and explainability. Our approach employs a convolutional LSTM network to forecast asset prices, followed by mean-variance optimization. For additional performance improvement, we integrate LLM-based sentiment analysis of financial news, enabling real-time portfolio weight adjustments. On a range of ETFs (VTI, AGG, DBC, VXX), experiments demonstrate our model achieves a 1.623 Sharpe ratio outperforming traditional approaches (MVO, MAD, CVaR) by $$\varvec{104-300\%}$$ 104 - 300 % and the deep learning state-of-the-art baselines. The incorporation of LLM sentiment signals improves the Sharpe ratio by $$\varvec{28.18 \%}$$ 28.18 % , and sensitivity tests guarantee robustness across regimes. By combining deep learning with interpretable LLM knowledge, our model bridges the gap between data-driven performance and theoretical interpretability, delivering actionable value to institutional and retail investors.