Ddg: Domain-Knowledge Distilled Dual-Gating for Event-Driven Stock Prediction
Year: 2026
Type: conference-paper
Abstract: External event-driven stock prediction is difficult due to low signal-to-noise ratios, sparse event-asset relations, and lagged responses. Existing deep models usually fuse text and price series in black-box architectures, making them prone to spurious correlations and unable to explicitly model which assets are affected and when the impact appears. We propose DDG, a domain-informed framework that converts empirical financial evidence into differentiable inductive biases. DDG contains two components: CRD learns an event-asset relevance gate using structural priors derived from LLM-based extraction of industry chains and economic linkages under strict timestamp constraints, while TDK learns an$H=5$response distribution using temporal priors summarized from event-study abnormal return patterns. Both priors are injected only during training through soft regularization, introducing no additional dependencies at inference. Experiments on ACL18, CMIN-US, and CMIN-CN show consistent gains in ACC and MCC. Ablations further verify the complementary roles of the two branches and the effectiveness of training-phase soft prior injection, while improving interpretability consistent with financial intuition.
Language: English
Field: Decision Sciences
Domain: Social Sciences
Open Access status: hybrid
OpenAlex ID: W7202357253