Systemic Risk Radar: A Multi-Layer Graph Framework for Early Market Crash Warning

TL;DR

Systemic Risk Radar (SRR) models financial markets as evolving multi-layer graphs to detect regime transitions before price moves. By combining a correlation-focused layer with potential sector/factor and sentiment layers and applying temporal graph neural networks, SRR aims to produce early-warning signals for systemic fragility with interpretability. Early results across the Dot-com, Global Financial Crisis, and COVID shocks show graph topology carries meaningful predictive information beyond traditional feature-based models, while highlighting the need for richer multi-layer temporal architectures. The framework emphasizes modularity, reproducibility, and multi-scale risk perspectives to inform regulators and risk managers about macro-level dynamics and contagion pathways.

Abstract

Financial crises emerge when structural vulnerabilities accumulate across sectors, markets, and investor behavior. Predicting these systemic transitions is challenging because they arise from evolving interactions between market participants, not isolated price movements alone. We present Systemic Risk Radar (SRR), a framework that models financial markets as multi-layer graphs to detect early signs of systemic fragility and crash-regime transitions. We evaluate SRR across three major crises: the Dot-com crash, the Global Financial Crisis, and the COVID-19 shock. Our experiments compare snapshot GNNs, a simplified temporal GNN prototype, and standard baselines (logistic regression and Random Forest). Results show that structural network information provides useful early-warning signals compared to feature-based models alone. This correlation-based instantiation of SRR demonstrates that graph-derived features capture meaningful changes in market structure during stress events. The findings motivate extending SRR with additional graph layers (sector/factor exposure, sentiment) and more expressive temporal architectures (LSTM/GRU or Transformer encoders) to better handle diverse crisis types.

Figures (12)

• Figure 1: Illustrative multi-layer graph for SRR. The correlation layer captures return co-movement, the sector/factor layer captures structural exposure, and the sentiment layer captures behaviorally driven co-movement. Multiple edge types enable the model to disentangle different forms of market connectivity.
• Figure 2: Temporal GNN view of SRR. Each graph snapshot $G_{t-k},\ldots,G_t$ is encoded into a latent vector $z_{t-k},\ldots,z_t$ by a graph encoder $f_\theta$. A temporal encoder aggregates these embeddings and produces a crash-regime probability $\hat{y}_t$.
• Figure 3: High-level SRR pipeline. Raw market and macro data are transformed into features, used to construct multi-layer graphs, encoded with a GNN, and fed into a temporal encoder that produces a systemic risk score or warning.
• Figure 4: System architecture and performance summary. Top: Data is transformed into features, multi-layer graphs, and evaluated using baseline ML, snapshot GNN, and Temporal GNN models. Bottom: The temporal GNN prototype shows improved AUROC on the Dot-com period relative to the snapshot GNN, while performance in other crises reflects the limitations of correlation-only temporal modeling.
• Figure 5: Aggregate comparison of logistic regression, Random Forest, and a snapshot GNN baseline across crisis periods. Bars summarize AUROC, precision, recall, and accuracy for each model.