Multi-attributeSD: An Adaptive Self-Knowledge Distillation Method for Financial Time Series Prediction

Intelligent financial forecasting models play a crucial role in warnings of market crashes, market recovery, portfolio optimization, and risk control. Among deep learning methods, graph neural networks have achieved remarkable results in time series prediction due to their ability to model complex relationship between nodes, but they still have limitation in capturing dynamic associations within attribute features, which restricts their ability to fully understand financial data. On the other hand, most deep learning methods are difficult to quantify the relationship between different financial data attributes (e.g., closing price, highest price, and lowest price), which prevents them from comprehensively analysing the attribute features, which in turn affects the prediction accuracy. To address this gap, this study proposes a new self-knowledge distillation model for multi-attribute financial time series data. This approach constructs an adaptive graph to represent different attribute features, performs self-distillation across mixed attributes, and integrates these techniques into a teacher network. Furthermore, this study introduces a novel multi-attribute feature attention fusion module to capture and integrate key features from multiple attributes, thereby enhancing the model’s predictive effectiveness and robustness. This research validates the superiority of the proposed method using time series data from three different financial markets: futures, cryptocurrency, and stock, demonstrating superior prediction performance. Given the cross-market and cross-asset correlations present in financial data, we have decided to open-source this dataset for time series prediction, data imputation, and correlation studies.