Capital markets play a strategic role in the global economy; however, high volatility often poses challenges for investors in making optimal investment decisions. Stock price prediction has become one of the key elements in effective portfolio management, but the complexity of market data demands more advanced analytical approaches. This study aims to explore the potential of Long Short-Term Memory (LSTM) models in predicting stock price movements and integrating these predictions into portfolio optimization strategies based on modern portfolio theory. The research methodology utilizes historical stock data from the S&P 500 and IDX Composite indices with daily frequency. The LSTM model is compared with Recurrent Neural Networks (RNN), Support Vector Machines (SVM), and Artificial Neural Networks (ANN) using metrics such as Mean Squared Error (MSE), Root Mean Squared Error (RMSE), and accuracy. The LSTM model demonstrated superior performance with an MSE of 0.0018, RMSE of 0.042, and prediction accuracy of 91.5%, significantly outperforming RNN (87.3%), SVM (80.2%), and ANN (82.7%). The prediction results were then integrated into portfolio optimization using the Mean-Variance Optimization approach, resulting in an increase in expected returns from 8.5% to 12.3%, a reduction in risk from 6.2% to 5.8%, and an improvement in the Sharpe ratio from 1.37 to 2.12. This study demonstrates that the LSTM model not only excels in stock price prediction accuracy but also contributes significantly to enhancing portfolio management efficiency. This Deep Learning-based approach offers a more adaptive, data-driven investment strategy, supporting more informed decision-making in the dynamic capital market
