Spatiotemporal Graph Neural Network Modeling of Financial Systemic Risk Transmission and Multi-Market Linkages for Early Warning

Abstract

This repository contains the data, source code, configuration files, trained-model implementation, and supporting materials required to reproduce the experiments presented in the associated manuscript on systemic-risk forecasting. The repository provides the processed dataset comprising 2,516 observations and 16 integrated financial and macroeconomic variables. The data are organized into market-based and macroeconomic feature groups and are used to construct the input sequences for systemic-risk prediction. The repository includes the complete Python implementation of the proposed artificial intelligence and machine learning framework, including data preprocessing, chronological dataset splitting, feature standardization, sliding-window sequence construction, graph-based representation learning, graph attention processing, spatiotemporal modeling, temporal feature extraction, cross-market attention, systemic-risk aggregation, and multi-horizon prediction. The implementation supports prediction at 1-, 5-, 10-, and 20-step forecasting horizons and includes the training and evaluation procedures used for the experimental analysis. Source-specific experiments and component-level ablation experiments are also provided where applicable. The package includes configuration files specifying the principal experimental parameters, training scripts, model definitions, evaluation metrics, and scripts for reproducing the experimental workflow. The evaluation framework includes ROC-AUC, PR-AUC, precision, recall, and F1-score.