Abstract: This paper presents a rigorous comparative analysis of ensemble tree-based models and deep sequential networks for financial time-series prediction, focusing on Apple Inc. stock data. The methodology employs a sophisticated feature engineering pipeline incorporating technical indicators (Relative Strength Index, Moving Average Convergence Divergence, Bollinger Bands), volatility metrics, and lagged returns. We integrate an anomaly detection component using Isolation Forest for both data cleaning and market surveillance. The study targets two primary objectives: regression (forecasting daily Close Price) and classification (predicting 5-day directional movement, H=5. Our results show that the CNN-LSTM regression model achieves an R2 score of 0.6695, demonstrating a strong statistical fit for continuous value prediction21. However, the Ensemble Classification approach, specifically the Stacked Ensemble, offers a superior and more actionable directional signal, achieving 80.54% accuracy after optimization via threshold tuning on a validation set. This is supplemented by a parallel GARCH(1,1) volatility analysis, which provides a robust framework for forecasting risk. The analysis confirms the critical role of Isolation Forest in identifying and mitigating the impact of outliers. The discussion highlights the crucial trade-off between the high interpretability and efficiency of tree-based models and the potential temporal dependency capture of deep learning architecture. Practical deployment recommendations favour tree-based models for high-volume, real-time trading signals, reserving the resource-intensive sequential models for strategic, offline risk analysis.

Keywords: Ensemble Methods, Stock Price Prediction, Anomaly Detection, Isolation Forest, CNN-LSTM, GARCH

Downloads: | DOI: 10.17148/IJIREEICE.2025.131115