FactorDGL: factor-driven dynamic graph learning for stock price forecasting

Predicting stock price movements is highly challenging due to the noisy, nonlinear, and multi-factor nature of financial markets. Most traditional and deep learning models mainly rely on the observed price sequence, making it difficult for them to capture complex latent temporal patterns associated with short-term fluctuations. These models also struggle to represent higher-order temporal dependencies and often fail to organize different sources of market variation in a structured manner. To address these limitations, we propose FactorDGL: factor-driven dynamic graph learning for stock price forecasting (FactorDGL), a framework designed to progressively learn and refine hierarchical latent representations underlying price movements. FactorDGL uses a temporal hypergraph to represent complex relationships among multiple historical time points and incorporates a cascading residual mechanism to progressively extract latent representations from residual signals that remain after accounting for known influences. In addition, we formulate stock forecasting as a joint price-risk representation learning problem by integrating auxiliary volatility and Sharpe-ratio prediction tasks, which helps the model better adapt to dynamic market regimes and structural shifts. Experiments on MasterCard, Visa, and DJIA datasets indicate that FactorDGL consistently outperforms existing baselines under the evaluated market settings.