Deep Learning Methods for Stock Prediction

Abstract

Stock market prediction is a challenging task due to the inherent volatility, non-stationarity, and complex interdependencies within financial markets. Although deep learning models have demonstrated strong capabilities in modeling time-series and textual data, their practical accuracy and ability to generalize across varying market conditions remain limited. A key limitation is the difficulty of effectively capturing the dynamic and evolving relationships among stocks, particularly those driven by both market behavior and external textual information. Moreover, the lack of a unified framework capable of integrating heterogeneous modalities such as numerical indicators, candlestick chart images, and semantic insights from financial news significantly impairs the robustness of models when faced with diverse market conditions. The reasons for these challenges include: First, traditional models often fail to capture long-range dependencies and complex topological structures among stocks. Second, there is limited integration of multimodal data sources including technical indicators, price patterns, and financial news. Third, existing approaches lack effective mechanisms to uncover latent inter-stock relationships using textual signals from sources such as news articles, social media, or analyst reports such relationships could be more effectively modeled through motif-based graph structures. Fourth, the interpretability of current deep learning models remains limited, impeding their adoption in high-stakes financial decision-making scenarios. To address these challenges, this thesis proposes three distinct yet complementary modeling frameworks, each designed to target specific aspects of the stock prediction problem through multimodal data fusion and dynamic relational modeling. This thesis aims to advance the field of stock market prediction by proposing a suite of models that integrate temporal, textual, image-based, and graph structural in-formation. In particular, this research addresses the aforementioned challenges through the combined application of deep learning and graph-based modeling. The proposed approaches introduces novel frameworks that embed semantic, structural, and temporal features into unified architectures, enhancing predictive robustness and interpretability. Given the persistent challenges in financial forecasting, there is a clear need for a comprehensive framework capable of leveraging multimodal information, encoding meaningful inter-stock and stock-news relationships, and improving the interpretability and accuracy of predictions in a scalable manner.

The thesis follows three key research questions. The first part of this thesis addresses the challenge of capturing hidden associations between stocks and financial news, which are often overlooked by traditional models. These implicit relationships carry valuable signals that can influence stock price movements. To resolve this, we propose a Motif-based Graph Convolutional Network (MGCN) that constructs motif graphs by linking stocks and news entities based on semantic patterns extracted from financial texts. A Transformer encoder is further employed to refine both price and text features, enabling the model to capture long-range dependencies.. Experimental on the S&P 500 dataset show that this model effectively integrates textual and temporal signals, achieving improved prediction accuracy over standard baselines. The second part of this thesis tackles the difficulty of integrating diverse financial signals such as temporal patterns, visual structures, and contextual sentiment within a unified predictive framework. Most existing models rely on a single modality, limiting their ability to generalize across dynamic market conditions. To address this, we propose a hybrid deep learning framework that fuses three complementary modalities. A Linear Transformer processes historical stock prices and technical indicators to extract temporal dependencies. Candlestick chart images are encoded via a CNN to capture spatial features. Concurrently, a Large Language Model (LLM) generates daily textual analyses, which are embedded using FinBERT to represent semantic sentiment. These multimodal features are combined through a feedforward network to produce final predictions. Empirical results demonstrate that this integrated approach significantly improves forecasting accuracy compared to unimodal baselines. The final part of this thesis addresses the challenge of capturing dynamic inter-stock relationships that are influenced by textual semantics but are often overlooked by conventional models. Most existing graph-based approaches rely on static price correlations, ignoring latent sentiment-driven connections revealed in financial discourse.To overcome this, we introduce an LLM-Augmented Enhanced Graph Transformer frame-work. In this approach, a large language model (LLM) generates concise daily analyses from technical indicators, which are then embedded via FinBERT to reflect semantic context. These embeddings are used to construct a graph where edges represent se-mantic similarity between stocks, enabling a Graph Transformer to model nuanced relational dependencies. Experimental results on the S&P 500 dataset show that this framework significantly outperforms time-series baselines (e.g., LSTM, Transformer, Informer) and prior graph models (e.g., GCN, GAT), delivering superior accuracy and interpretability in stock movement prediction.