Applying Deep Learning Models in Forecasting Real Estate Stock Prices: Empirical Evidence from the Vietnamese Stock Market

Abstract

This study analyzes the applicability of deep learning models in forecasting stock prices of 15 real estate companies listed on the Vietnamese stock market during the period 2015–2025. The traditional ARIMA model is compared with three deep learning architectures: LSTM, GRU, and CNN-LSTM. The models are evaluated using four metrics: RMSE, MAE, MAPE, and Directional Accuracy (DA). The experimental results show that GRU outperforms all others, achieving the lowest RMSE and MAE across most stocks, while ARIMA records the highest errors. Although deep learning models significantly improve absolute forecasting accuracy, DA remains at only around 45–46%, reflecting limitations in predicting market directions. The findings indicate that GRU is the most suitable model for short-term stock forecasting in Vietnam and suggest that future research should integrate macroeconomic and unstructured data to enhance directional prediction performance.