A Spatiotemporal Deep Learning Framework Based on GCNN and BiLSTM for Stock Price Forecasting
Year: 2026
Type: article
Abstract: Stock price forecasting remains a challenging problem due to the nonlinear, dynamic, and highly interconnected nature of financial markets. Traditional models and conventional deep learning approaches often fail to adequately capture complex interstock relationships and temporal dependencies simultaneously. To address these limitations, we propose a hybrid spatiotemporal forecasting model that combines a graph convolutional neural network (GCNN), a one‐dimensional convolutional neural network (1D‐CNN), and a bidirectional long short‐term memory (BiLSTM) network for stock price forecasting. In our model, stocks are represented as nodes in a graph, and their relationships are encoded through a correlation‐based adjacency matrix, enabling effective learning of spatial dependencies among multiple stock indices. The 1D‐CNN extracts local temporal features from graph‐enhanced representations, whereas the BiLSTM captures long‐term bidirectional temporal dependencies. The experiments were conducted on real‐world data from 10 major stock indices obtained from Yahoo Finance. A systematic evaluation was conducted to analyze the impact of different graph construction strategies, correlation thresholds, and input timestep lengths. The results demonstrate that a moderately sparse correlation‐based graph and an input timestep of 10 days provide the most stable and accurate forecasting performance. Furthermore, comprehensive comparisons with well‐known forecasting methods demonstrate that the proposed model outperforms several baseline approaches.
Source: Journal of Engineering
DOI: 10.1155/je/9973211
Language: English
Field: Decision Sciences
Domain: Social Sciences
Open Access status: gold
Estimated APC paid: $1,400
OpenAlex ID: W7172016362