Hybrid Machine Learning Architecture for Stock Market Price Prediction: Integrating Statistical Time-Series Models with Deep Learning and Ensemble Methods
Summary
No summary yet. VIRAT can read this paper and write a plain-language summary of its aim, methods, key findings and limitations.
Summarise with VIRAT Free to try with your Science Catalogs account.Abstract
Stock prediction is hard. Prices are noisy, nonstationary, and nonlinear. We built a hybrid system that combines statistical models (ARIMA, GARCH), deep learning (LSTM, GRU), and Random Forests via Ridge regression meta-learning. The meta-learner uses 5-fold time-series cross-validation to adaptively weight models. Testing across 20 stocks from Technology, Finance, Healthcare, Consumer, and Industrial sectors, we achieved 87.74 percent average RMSE improvement over individual models. Directional accuracy ranged from 42.45 percent to 85.87 percent. Boeing (BA) showed 95.43 percent RMSE improvement with 85.87 percent directional accuracy, U.S. Bancorp (USB) hit 94.31 percent RMSE improvement. Random Forest dominated the learned weights (60-92 percent), while ARIMA and deep learning added complementary signals. Walk-forward validation with 252-day rolling windows ensured that we tested on truly unseen data, not on retrofitted history.
Fields of Study & Subjects
Stock Market Forecasting Methods
Forecasting Techniques and Applications
Financial Markets and Investment Strategies
Publication details
- Journal / source
- International Journal of Emerging Science and Engineering
- Volume
- 14
- Issue
- 4
- Pages
- 13–22
- Publication date
- 28 March 2026
- Language
- EN