Modeling the continually shifting dependencies between traded equities has historically posed significant challenges for automated financial forecasting algorithms. To address these inter-asset relational dynamics, researchers developed the Spatio-Temporal Attention Network with Dynamic Graphs, termed STAN-DG, which explicitly incorporates geographic data alongside market signals.
The first component is a multi-source feature extraction module that combines company fundamental signals and technical indicators through spatio-temporal encoding and an integrated patching mechanism.
The second stage of the framework relies on a dynamic temporal graph learning module to construct evolving stock relationship networks over time. This module uses gated recurrent units alongside multi-head attention mechanisms to track changing inter-stock connections.
To refine the quality of these evolving relational representations, the researchers implemented an adaptive graph enhancement module as the third component. This module couples graph convolution operations with an adaptive sparsification technique designed to strengthen informative connections between companies while suppressing noisy or uninformative links.
In the final operational stage, a cross-modal fusion module integrates the enhanced dynamic graph representations directly with the extracted multi-source features to generate target stock price forecasts.
Comparative evaluations conducted on the CSI 300, CSI 500, CSI 1000, and full datasets demonstrated that STAN-DG achieved the lowest mean absolute error across all four benchmark datasets. Furthermore, the model recorded the lowest root mean squared error on three of the four tested evaluation sets. In benchmark comparisons against Finformer, identified as the closest static-dynamic relational baseline, the proposed framework lowered mean absolute error by approximately 5.1 percent to 21.4 percent across the four datasets.
Ablation studies confirmed that spatial attention, graph enhancement, and cross-modal fusion each contributed substantially to the overall forecasting performance of the system.
Subject of Research: Stock price forecasting with spatio-temporal dynamic graphs and geospatial semantics
Article Title: STAN-DG: Spatio-temporal attention network based on dynamic graphs for stock price prediction with geospatial semantics
Article References: Chen, L., Duolikun, N., Wu, S., Xu, J., Yan, C., Yang, G., & Zhang, D. (2026). STAN-DG: Spatio-temporal attention network based on dynamic graphs for stock price prediction with geospatial semantics. Applied Intelligence, 56(17), Article 564. https://doi.org/10.1007/s10489-026-07611-5
Image Credits: AI Generated
DOI: 10.1007/s10489-026-07611-5
Keywords: stock price prediction, spatio-temporal attention network, dynamic graphs, geospatial semantics, graph convolution, gated recurrent unit, multi-head attention, Finformer, CSI 300, CSI 500, CSI 1000, adaptive sparsification, cross-modal fusion, applied intelligence, financial forecasting, relational graphs
Cite Scienmag News
Denise Maddox. (November 20, 2026). Dynamic Graph Network Incorporating Geospatial Semantics Improves Stock Forecasting Accuracy. Scienmag. https://scienmag.com/dynamic-graph-network-incorporating-geospatial-semantics-improves-stock-forecasting-accuracy/
Denise Maddox. "Dynamic Graph Network Incorporating Geospatial Semantics Improves Stock Forecasting Accuracy." Scienmag, 20 November 2026, https://scienmag.com/dynamic-graph-network-incorporating-geospatial-semantics-improves-stock-forecasting-accuracy/. Accessed 20 November 2026.
Denise Maddox. "Dynamic Graph Network Incorporating Geospatial Semantics Improves Stock Forecasting Accuracy." Scienmag. November 20, 2026. https://scienmag.com/dynamic-graph-network-incorporating-geospatial-semantics-improves-stock-forecasting-accuracy/








