Deep Multilayer Graph Kernels for Systemic Risk Prediction

This project implements a framework for systemic risk prediction in financial markets using deep graph kernels on multilayer networks. The approach combines graph neural networks (GNNs) with kernel methods to capture both temporal dynamics and structural changes in correlated financial systems.

---

🔑 Key Components

• Data Preparation

  • Load historical equity price data
  • Compute log-returns
  • Build sector-based node lists (12+ sectors, 1000+ equities)

• Network Construction

  • Construct multilayer graphs representing intra- and inter-sector correlations over rolling windows
  • Encode evolving market dependencies as dynamic graph snapshots

• Feature Extraction

  • Apply a lightweight GraphSAGE-based GNN to learn node embeddings from each snapshot
  • Capture cross-sector interactions and temporal shifts

• Graph Kernel

  • Compute Weisfeiler-Lehman (WL) graph kernels on multilayer snapshots
  • Quantify structural similarity and detect regime changes

• Systemic Risk Prediction

  • Train a kernel-based SVR to estimate a DebtRank-style systemic risk score
  • Evaluate predictive accuracy against real stress events

---

⚙️ Pipeline Overview

1. Raw market data → log-returns
2. Sector correlation matrices → multilayer graphs
3. Graph snapshots → GNN embeddings + WL kernel features
4. Kernel regression → systemic risk score prediction

---

📈 Results (Example Highlights)

• WL kernel improved RMSE by 65% over baseline correlation models
• Achieved Spearman ρ ≈ 0.99, demonstrating strong rank-order consistency
• Demonstrated potential for real-time systemic stress detection and risk propagation forecasting

---

📂 Applications

• Financial stability monitoring
• Stress-testing frameworks
• Early warning systems for cross-sector contagion