Hypergraph-Enhanced Kernel Initialization for Convolutional LSTM Networks: Insights from Asset Correlation Forecasting

Lo Pang-Yun Ting, Hua-Cheng Cheng, Yu-Hua Zeng, Kun-Ta Chuang

Abstract: Spatiotemporal data is crucial in applications such as financial markets, where accurately modeling spatial and temporal dependencies is essential. Weight initialization significantly impacts the performance of deep learning models for such tasks. However, current methods for weight initialization often ignore spatiotemporal characteristics, limiting their ability to effectively analyze dynamic systems. In this paper, we propose HIFA, a framework designed to improve weight initialization in ConvLSTM networks, with a specific focus on asset correlation forecasting. HIFA integrates external stock attributes into hypergraphs to model complex spatial dependencies and utilizes past asset returns to capture temporal dependencies among assets. Graph Convolutional Networks are empolyed to generate feature representations, which are then transformed into convolutional kernel weights through an efficient multi-layer perceptron. Experiments on real-world financial datasets demonstrate that HIFA achieves superior performance in asset correlation prediction with minimal extra computational overhead. A case study further validates its robustness across varying asset compositions, highlighting its practicality for diverse scenarios.