Hybrid Ensemble Model Approaches for Stock Price Forecasting Using LSTM, Random Forest, ARIMA, and Linear Regression as Meta-Learner

Predicting stock market price remains a complex and challenging task due to the influence of various financial, economic, and behavioral factors. To address these challenges, numerous machine learning and deep learning techniques such as ARIMA, Random Forest, and Long Short-Term Memory (LSTM) have been proposed. However, each model exhibits inherent strengths and limitations in capturing specific data characteristics. This study proposes a hybrid ensemble framework that integrates ARIMA for linear pattern, Random Forest, for non-linear dependencies, and LSTM for temporal dynamics. A meta-learning strategy based on Linear Regression is implemented to perform weighted ensembling of individual model outputs. The experiment results across five major stocks demonstrate that the proposed approach significantly outperforms standalone models. The hybrid ensemble model achieved low MAPE scores of 0.5590 for AAPL, 0.7126 for XOM, and 5.944 for TSLA. Additionally, the MAE results across three stocks were 0.8155 to 33.82. The RMSE results were from 1.027 to 40.86. And lastly, the hybrid ensemble model achieved high R2 scores across the three stocks with range from 0.9304 to 0.9601. These findings highlight the effectiveness and robustness of the hybrid ensemble approach in stock price forecasting.