Document Type : Original Article
Authors
1 MSc student in Financial management, Shandiz Institute of Higher Education, Mashhad, Iran.
2 Assistant professor, Department of accounting, Shandiz Institute of Higher Education, Mashhad, Iran.
Abstract
The primary purpose of this study is to enhance stock price forecasting accuracy by jointly modeling numerical market data, semantic information from financial news, investor sentiment, and structural relationships among stocks. Addressing the limitations of traditional price-based and standalone deep learning models, this research focuses on uncovering hidden nonlinear and relational dynamics in an emerging stock market. This study proposes an integrated framework that combines a Stock Knowledge Graph (SKG), sentiment analysis, and a hybrid deep learning architecture. Structural and semantic relationships among stocks, investors, and financial entities are modeled using Node2Vec and TransE embeddings. These features are fused with technical indicators, financial news sentiment, and user-generated sentiment, and processed through a CNN–LSTM–Attention model. The framework is empirically evaluated using 4,319 trading days of data from the Iranian stock market, with multiple benchmark models for comparison. Empirical results demonstrate that the proposed model significantly outperforms baseline CNN, LSTM, and Transformer models across multiple evaluation metrics. The hybrid CNN–LSTM–Attention architecture achieves the lowest RMSE (0.754) and the highest explanatory power (R² = 0.91). Incorporating sentiment and semantic features reduces prediction error by more than 28% compared with models relying solely on technical indicators. Knowledge graph embeddings effectively capture latent inter-stock and semantic relationships, further improving forecasting performance. The findings confirm that integrating structural, semantic, and behavioral information within a unified deep learning framework substantially enhances stock price prediction accuracy.
Keywords
)