A CASCADED DEEP LEARNING AND KERNEL-BASED APPROACH FOR ENHANCED STOCK PRICE PREDICTION: INTEGRATING LSTM FEATURE EXTRACTION WITH OPTIMIZED SUPPORT VECTOR REGRESSION

Abstract

Accurately predicting stock prices remains a formidable challenge due to the inherently volatile and nonlinear characteristics of financial markets. This research introduces an innovative cascaded approach, leveraging the deep learning capabilities of Long Short-Term Memory (LSTM) networks for feature extraction in combination with an optimized kernel-based Support Vector Regression (SVR) model for prediction. The LSTM component is adept at capturing temporal dependencies and complex patterns within financial time series, thereby gen- erating informative feature representations that enhance down- stream learning. These features are then provided as input to a finely-tuned SVR, which is empowered by kernel optimization to identify nonlinear relationships in the processed data. This synergy between sequential deep learning and robust kernel regression fosters improved predictive accuracy over standalone methods. Comprehensive experimentation utilizing real-world stock price datasets demonstrates marked improvements in mean squared error, directional accuracy, and adaptiveness to abrupt market shifts. The model’s dual-stage design not only mitigates overfitting but also addresses the limitations of purely statistical and deep learning approaches by blending their strengths. Results underscore the potential of this hybridized framework in algorithmic trading and decision-making systems, offering scalable solutions for practitioners facing the intricacies of high- frequency and multivariate time series forecasting in financial domains.