Sph-net: Co-Attention Hybrid Model Achieves Accurate Stock Price Prediction on Eight Diverse Datasets

Predicting stock price fluctuations remains a significant challenge in financial analysis, complicated by market volatility and unpredictable patterns. Yiyang Wu, Hanyu Ma, and Muxin Ge, alongside colleagues, address this problem with SPH-Net, a new deep learning framework designed to improve the accuracy of stock price forecasting. The team, including researchers from Nanyang Technological University and the University of Nebraska-Lincoln, developed a model that uses a co-attention mechanism to analyse market data, capturing both broad trends and subtle details. Through rigorous testing on eight different stock datasets, utilising key market indicators, SPH-Net consistently outperforms existing prediction models, offering a potentially valuable tool for investors and financial analysts seeking to navigate complex and rapidly changing market conditions.

Due to the inherent volatility, non-stationarity, and nonlinear characteristics of market data, scientists introduce SPH-Net, an innovative deep learning framework designed to enhance the accuracy of time series forecasting in financial markets. The proposed architecture employs a novel co-attention mechanism that initially processes temporal patterns through a Vision Transformer, followed by refined feature extraction via an attention mechanism, thereby capturing both global and local dependencies in market data. To rigorously evaluate the model’s performance, the team conducts comprehensive experiments on eight diverse stock datasets, including AMD and Ebay.

Time Series Prediction with Vision Transformers

Scientists developed SPH-Net, a novel deep learning framework for stock price prediction, by integrating a Vision Transformer with Transformer encoder-decoder mechanisms. The study pioneers a method of transforming temporal financial data into image-like patches, enabling the Vision Transformer to effectively extract patterns from time series data. Researchers standardized eight diverse stock datasets, AMD, Ebay, Facebook, FirstService Corp, Tesla, Google, Mondi ADR, and Matador Resources, using six fundamental market indicators: Open, High, Low, Close, Adjusted Close, and Volume, creating a comprehensive feature set for analysis. The core of the methodology involves processing the standardized data into patch representations, allowing the Vision Transformer to analyze the time series as if it were an image, capturing spatial relationships within the temporal data.

Subsequently, the system employs a co-attention mechanism, simultaneously modeling temporal dependencies and interactions between different market indicators, refining feature extraction and capturing both global and local dependencies within the financial time series. Experiments rigorously evaluated SPH-Net’s performance against existing stock prediction models across all evaluation metrics, demonstrating consistent outperformance across the eight datasets. This comprehensive evaluation confirms the model’s ability to capture complex temporal patterns while maintaining robustness against market noise, ultimately improving prediction accuracy in volatile financial conditions. The innovative approach provides valuable decision-support capabilities for investors and financial analysts seeking more informed investment strategies.

SPH-Net Predicts Stock Prices Accurately

Scientists developed SPH-Net, a novel deep learning framework designed to improve the accuracy of stock price prediction. The work introduces a hybrid architecture that combines a Vision Transformer with Transformer encoder-decoder mechanisms, enabling comprehensive analysis of financial time series data. This innovative approach utilizes co-attention mechanisms to simultaneously model temporal dependencies and cross-feature interactions within the data, capturing complex patterns crucial for accurate forecasting. The team preprocessed data from eight diverse stock datasets, AMD, Ebay, Facebook, FirstService Corp, Tesla, Google, Mondi ADR, and Matador Resources, using six fundamental market indicators: Open, High, Low, Close, Adjusted Close, and Volume.

This standardized dataset served as a comprehensive input for the SPH-Net model, allowing for rigorous evaluation across varied market conditions. Experiments demonstrate that SPH-Net consistently outperforms existing stock prediction models, showcasing its ability to effectively capture intricate temporal patterns while maintaining robustness against market noise. Results reveal that the model’s hybrid architecture successfully integrates the strengths of both Vision Transformers and Transformer components. By transforming temporal financial data into image-like patch representations, SPH-Net facilitates enhanced pattern extraction through the Vision Transformer, leading to improved predictive performance. The team validated the framework in both regression and classification scenarios, demonstrating its versatility and adaptability to different forecasting tasks. This breakthrough delivers valuable decision-support capabilities for investors and financial analysts, potentially enabling more informed investment strategies and risk assessment in volatile market conditions.

SPH-Net Achieves Superior Stock Price Prediction

SPH-Net, a novel hybrid neural network, represents a significant advancement in stock price prediction, effectively capturing complex temporal dependencies within financial time-series data. The team developed this model by integrating a Vision Transformer with a Transformer encoder-decoder architecture, transforming conventional stock data into a structured, image-like format to improve feature extraction and trend recognition. Comprehensive experiments conducted on eight diverse stock datasets consistently demonstrate SPH-Net’s superior performance compared to existing state-of-the-art models, achieving improved R2 scores, reduced mean squared error, and enhanced classification metrics including accuracy, precision, and recall. Ablation studies confirm the effectiveness of each component within SPH-Net’s architecture, validating the co-attention mechanism and hybrid structure.

While the model demonstrates robust results, the researchers acknowledge that Google’s stock data presents a particular challenge due to its discrete and irregular distribution. Future research directions include incorporating additional data modalities such as financial news and sentiment analysis, investigating the model’s applicability to high-frequency trading data, and developing more scalable and interpretable versions of SPH-Net for real-time deployment and enhanced decision-making transparency. These advancements promise to further refine the model’s capabilities and broaden its potential applications beyond stock market analysis.