Computing Systemic Risk Measures with Graph Neural Networks

Lukas Gonon, Thilo Meyer-Brandis, Niklas Weber

Abstract

This paper investigates systemic risk measures for stochastic financial networks of explicitly modelled bilateral liabilities. We extend the notion of systemic risk measures from Biagini, Fouque, Fritelli and Meyer-Brandis (2019) to graph structured data. In particular, we focus on an aggregation function that is derived from a market clearing algorithm proposed by Eisenberg and Noe (2001). In this setting, we show the existence of an optimal random allocation that distributes the overall minimal bailout capital and secures the network. We study numerical methods for the approximation of systemic risk and optimal random allocations. We propose to use permutation equivariant architectures of neural networks like graph neural networks (GNNs) and a class that we name (extended) permutation equivariant neural networks ((X)PENNs). We compare their performance to several benchmark allocations. The main feature of GNNs and (X)PENNs is that they are permutation equivariant with respect to the underlying graph data. In numerical experiments we find evidence that these permutation equivariant methods are superior to other approaches.

Introduction

In a modern, highly interconnected financial system, the failure of a single institution can trigger a "domino effect" of defaults. Regulators and "lenders of last resort" face the critical challenge of determining the minimum amount of bailout capital required to secure the system and, crucially, how to distribute that capital optimally before a collapse occurs.

This paper addresses systemic risk by treating financial liabilities as graph-structured data. We extend existing theoretical frameworks to account for the stochastic nature of these networks and propose the use of permutation-equivariant neural network architectures to find optimal capital allocations. By leveraging the inherent structure of financial networks, these models provide a more robust and scalable solution than traditional deep learning approaches.

The Problem of Systemic Risk

Traditional systemic risk measures often struggle with two primary factors: stochasticity and graph complexity.

  1. Stochasticity: The network of bilateral liabilities (who owes what) is often uncertain or changes based on specific market scenarios. A static model cannot capture the range of potential outcomes.
  2. Graph Complexity: Financial networks are essentially directed graphs. Standard Feedforward Neural Networks (FNNs) treat inputs as fixed-order vectors. If the positions of two banks in a data list are swapped, an FNN might produce a completely different output, even though the underlying network is identical. This lack of "permutation equivariance" makes FNNs inefficient and poor at generalizing to new or reordered network structures.

Methodology

We propose an iterative optimization algorithm to determine both the "inner risk" (the optimal distribution of a fixed amount of capital) and the "outer risk" (the minimum total capital needed to meet a safety threshold). The aggregation function used to calculate the system's "shortfall" is based on the Eisenberg-Noe clearing mechanism, which mathematically models how defaults propagate through a network of bilateral liabilities.

Neural Network Architectures

The core technical contribution lies in the architectures used to approximate these allocations:

  • Graph Neural Networks (GNNs): Utilizing SAGEConv layers, these models use "message passing" to allow nodes (banks) to learn about the risk profiles of their neighbors.
  • Permutation-Equivariant Neural Networks (PENNs): This architecture processes graph data by summing signals from all nodes, ensuring the output remains consistent regardless of node ordering.
  • Extended PENNs (XPENNs): A novel extension introduced in this work. While standard PENNs can lose specific node information during global summation, XPENNs add a node-specific component that considers both incoming and outgoing edges for each institution. This mimics a single round of message passing with higher expressivity.

Experimental Setup

The models were evaluated on several synthetic but representative financial network types:

  • Stylized Networks: "Default Cascades" (a linear chain of liabilities) and "Star-Shaped Networks" (multiple banks owing a central hub) were used to test if models could learn known optimal strategies.
  • Erdős-Rényi (ER) Networks: Homogeneous networks where links between banks are formed with a fixed probability.
  • Core-Periphery (CP) Networks: Inhomogeneous networks featuring a small group of highly connected "core" banks and many "small" peripheral banks.

Key Results and Insights

The Importance of Equivariance

Our findings suggest that for financial networks, the identity of a bank is less important than its position within the network topology. Models that inherently respect this (GNNs and XPENNs) require significantly fewer parameters and generalize much better than massive FNNs. Standard FNNs tend to "memorize" the specific order of banks in the training set, leading to performance collapse when presented with the same network in a different list order.

Performance and Efficiency

In tests on unseen data, GNNs and XPENNs reached near-perfect bailout allocations in stylized networks. In Core-Periphery experiments, XPENNs achieved the lowest systemic risk. For a fixed risk threshold, these permutation-equivariant models required significantly less bailout capital to secure the system compared to naive benchmarks like uniform distribution.

Scalability

As the number of institutions ($N$) increased from 10 to 100, the training time for FNNs grew dramatically due to the $N^2$ liability matrix. In contrast, GNN and XPENN training times scaled much more gracefully, demonstrating their viability for larger real-world systems.

Conclusion

This research demonstrates that permutation-equivariant architectures like GNNs and XPENNs are superior for calculating systemic risk in financial networks. By respecting the graph-based nature of liabilities, these models offer better generalization, higher capital efficiency, and superior scalability. Future work may extend this framework to include more complex contagion dynamics, such as fire sales or liquidity hoarding.

Summary updated

Abstract

This paper investigates systemic risk measures for stochastic financial networks of explicitly modelled bilateral liabilities. We extend the notion of systemic risk measures from Biagini, Fouque, Fritelli and Meyer-Brandis (2019) to graph structured data. In particular, we focus on an aggregation function that is derived from a market clearing algorithm proposed by Eisenberg and Noe (2001). In this setting, we show the existence of an optimal random allocation that distributes the overall minimal bailout capital and secures the network. We study numerical methods for the approximation of systemic risk and optimal random allocations. We propose to use permutation equivariant architectures of neural networks like graph neural networks (GNNs) and a class that we name (extended) permutation equivariant neural networks ((X)PENNs). We compare their performance to several benchmark allocations. The main feature of GNNs and (X)PENNs is that they are permutation equivariant with respect to the underlying graph data. In numerical experiments we find evidence that these permutation equivariant methods are superior to other approaches.
BibTeX
@misc{gonon2024computing,
  author = {Lukas Gonon and Thilo Meyer-Brandis and Niklas Weber},
  title = {{Computing Systemic Risk Measures with Graph Neural Networks}},
  year = {2024},
  eprint = {2410.07222},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2410.07222},
}
Figure