Optimizing CNN-LSTM Models for Stock Price Prediction in a Multi-Sector Holding Company

Abstract

Accurate stock price forecasting is difficult because financial time-series data usually demonstrate nonlinear relationships, irregular fluctuations, and interdependent temporal patterns. This research investigates the predictive performance of three neural network models based on deep learning: Convolutional Neural Network (CNN), Long Short-Term Memory (LSTM), and a hybrid CNN-LSTM architecture for forecasting stock prices of a multi-sector holding company. The dataset used in this study contains daily historical price observations collected from 2015 to 2025, where sequential samples are generated using a sliding window approach. To obtain appropriate model settings, hyperparameter optimization is carried out using a grid search procedure. Model performance is assessed using Mean Squared Error (MSE), Root Mean Squared Error (RMSE), Mean Absolute Error (MAE), and Mean Absolute Percentage Error (MAPE). Experiments are first performed using an 80:20 training-testing split and followed by a robustness evaluation using a 70:30 data split. Under the primary evaluation scheme, the experimental results indicate that the LSTM model yields the lowest prediction error, reflected by an RMSE value of 77.86, MAE of 58.23, and MAPE of 1.28%. Meanwhile, the hybrid CNN-LSTM model demonstrates more stable performance across different data proportions, achieving an RMSE of 75.71 and MAPE of 1.23% during the robustness test. The results indicate that LSTM is effective in capturing sequential dependencies inherent in financial time-series data, integrating convolutional feature extraction with sequential learning can improve prediction stability under varying training conditions. The results provide empirical insights into the selection of deep learning architectures for stock price prediction in the context of multi-sector holding companies.