Using Kernel Method to Include Firm Correlation for Stock Price Prediction.

Area of Science:

• Quantitative Finance
• Machine Learning
• Financial Time Series Analysis

Background:

• Stock price prediction is complex due to market volatility and inter-firm relationships.
• Existing models often struggle to capture dynamic spatial and temporal correlations effectively.

Purpose of the Study:

• To propose AGKN (attention-based graph learning kernel network) for end-to-end stock price prediction.
• To integrate correlated firm information for enhanced predictive accuracy.
• To capture both spatial and temporal correlations in financial markets.

Main Methods:

• Constructed a stock-axis attention module using kernel methods for spatial correlations.
• Developed a graph learning module for integrating accurate firm information.
• Applied an ensemble time-axis attention module for temporal correlations.
• Utilized a transformer encoder for aggregating multi-level correlated information.

Main Results:

• AGKN demonstrated superior performance over state-of-the-art baseline methods on Chinese stock market data.
• Achieved up to 4.3% lower prediction error compared to the best competitors.
• Ablation studies confirmed the model's effectiveness in leveraging hidden stock correlations.

Conclusions:

• AGKN provides a novel and effective framework for stock price prediction by incorporating inter-firm correlations.
• The attention-based graph learning approach significantly improves prediction accuracy.
• The model's ability to capture hidden correlations is crucial for enhanced financial forecasting.