INTEGRATED FEATURE FUSION APPROACH FOR STOCK MARKET MOVEMENT FORECASTING

Abstract

ThStock market movement forecasting remains a challenging task due to the nonlinear dynamics of financial time series and the influence of heterogeneous factors such as technical indicators, macroeconomic variables, and textual sentiment. Recent studies show that hybrid deep learning architectures—such as CNN-BiLSTM-Attention networks [1], mixed-feature sentiment–price fusion frameworks [2], and multimodal stable-fusion mechanisms [5]—significantly enhance predictive performance compared to single-stream models. Traditional deep models, including LSTM and CNN, capture temporal and local patterns effectively [3], [15], but fail to exploit complementary information across diverse feature sources. Multimodal methods integrating price data, macroeconomic indicators, and textual sentiment have demonstrated improved accuracy and robustness in movement prediction [7], [13], [22], [24].In this work, we propose an Integrated Feature Fusion Approach designed to combine multiview financial features—including historical price series, technical indicators, sector-level signals, and sentiment representations—into a unified predictive architecture. Motivated by advances in hybrid temporal models [4], feature-selection–enhanced prediction [17], mixed-frequency fusion networks [9], and diffusion-based graph–sentiment integration [13], the proposed system employs a multi-branch encoder with attention-driven cross-feature fusion. This enables the model to learn both shared and modality-specific representations to address volatility, noise, and non-stationarity in stock data [11], [12], [20].Experimental evidence reported across recent literature supports the advantages of integrating multiple modalities for financial forecasting [8], [10], [14], [16], [18], [19], showing substantial gains in trend classification, long-term prediction stability, and generalization across markets. Aligning with this direction, our integrated fusion framework aims to deliver a more reliable and adaptive forecasting system that outperforms traditional deep learning models and hybrid baselines [21], [25].