Abstract
The stock market is well-known for volatility and unpredictability, which can pose significant risks to investment strategies. This manuscript compares LSTM, GRU, CNN, and LSTM-CNN models for stock price prediction using standard performance metrics. A structured methodology is applied, including detailed data preprocessing, hyperparameter tuning, and splitting the data into training and test sets. The results show that each model has its own strengths and limitations in learning short- and long-term patterns, as well as differences in computational efficiency. Model performance is evaluated using RMSE, MAE, and R-squared metrics to support portfolio optimization strategies. All models are trained and tested under consistent data inputs, preprocessing steps, and experimental settings. This method helps maintain reliable and comparable results across different models. Practical applications include dynamic asset allocation, risk-based position sizing, sector-specific model deployment, and integration of LSTM forecasts into institutional portfolio strategies. The analysis examines model performance using datasets from six different companies. The LSTM model outperformed the other models due to its strong ability to learn long-term patterns in volatile stock price data. In comparison, CNN and GRU models showed weaker results as they were less effective at retaining long-range temporal information under changing market conditions. These findings provide practical guidance for investors in selecting appropriate deep learning models for real-world financial forecasting.
Keywords: Convolutional Neural Network, Gated Recurrent Unit, Long Short-Term Memory, Metrics, Multi-Stock Prediction Portfolio Optimization.