Graph Neural Network-Based Cross-Asset Pricing Model with Adaptive Factor Learning and Temporal Attention for High-Dimensional Financial Markets

Abstract

This paper proposes a novel Adaptive Temporal Graph Factor Network (ATGFN) for cross-asset pricing in high-dimensional financial markets, integrating dynamic graph neural networks with temporal attention to jointly learn latent factor structures and evolving inter-asset dependencies. Unlike traditional linear factor models such as Fama–French and Carhart, ATGFN captures nonlinear and time-varying relationships through a graph construction module that encodes asset similarity based on both return co-movements and fundamental attributes. The model incorporates an adaptive factor learning layer that updates latent representations across rolling windows, combined with a temporal attention mechanism that emphasizes regimerelevant signals. We evaluate ATGFN against six benchmark methods: Fama–French 5-Factor Model, Arbitrage Pricing Theory (APT), Principal Component Analysis (PCA)-based factor models, Long Short-Term Memory (LSTM) networks, Temporal Convolutional Networks (TCN), and standard Graph Convolutional Networks (GCN). Empirical results on U.S. equity and multi-asset datasets demonstrate that ATGFN significantly improves out-of-sample pricing accuracy, reduces mean absolute pricing error, and enhances Sharpe ratios in portfolio construction tasks. The model also shows superior robustness during periods of market stress, particularly in capturing contagion effects and structural breaks. Furthermore, interpretability analysis reveals that ATGFN dynamically adjusts factor importance across sectors and time, offering economically meaningful insights into risk premia evolution. This framework contributes to the literature by bridging graph-based representation learning with financial factor modeling, providing a scalable and adaptive approach for modern asset pricing.