Multi-order Market Modeling with Wavelet Hypergraph Attentions and Temporal Generative Filters

Stock markets exhibit multi-order dynamics (non-pairwise correlations) and internal dynamics (stock-specific behaviors), which traditional deep learning models fail to fully capture. Existing hypergraph-based approaches rely on inefficient Fourier convolutions and overlook stock-specific patterns, leading to information loss. We propose Multi-order Market Modeling with Wavelet Hypergraph Attentions and Temporal Generative Filters, a novel framework that enhances stock prediction by integrating waveletbased hypergraph attentions for localized feature extraction and temporal generative filters for learning individual stock behaviors. Experiments on six years of US market data show superior profitability and stability over state-of-the-art methods.