Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

The systemic impact of edges in financial networks

Michel Alexandre, Thiago Christiano Silva, Francisco A. Rodrigues

TL;DR

This paper investigates how marginal changes in financial network interconnectedness, viewed at the edge level, affect systemic risk (SR). It uses the differential DebtRank ($S_\zeta$) framework to quantify edge impact $I_{e,\zeta}$ in a two-network Brazilian dataset, applying machine learning (TPOT AutoML) and SHAP for interpretability to predict edge criticality and the sign of edge impact. Key findings include substantial heterogeneity in edge effects, a heavier-tailed distribution of edge impacts well-described by a log-normal, and a rising fraction of null/negative impacts as the initial shock grows. The study identifies borrower PageRank and lender centrality as primary predictors of edge criticality, and reveals a tipping-point behavior in the sign of edge impact around a small lender PageRank, with important macroprudential implications for edge-level interventions and policy design.

Abstract

In this paper, we assess how the stability of financial networks is affected by interconnectedness considering its tiniest variation: the edge. We compute the impact of edges as the percentage difference in the systemic risk (SR) of the whole network caused by the inclusion of that edge. We apply this framework to a thorough Brazilian dataset to compute the impact of bank-firm edges. After observing that (i) edges are heterogeneous regarding their impact on the SR, and (ii) the fraction of edges whose impact on the SR is non-positive increases with the level of the initial shock, we use machine learning techniques to try to predict two variables: the criticality of the edges (defining as critical an edge whose impact on SR is significantly greater than that of the others) and the sign of the edge impact. The level of accuracy obtained in these prediction exercises was very high. These results have important implications for the development macroprudential policies aimed at financial stability. Our framework allows to identify, based on features related to the origin and destination nodes of the edge (i.e., the lending bank and the borrowing firm), whether an additional loan will have a significant and positive impact on the SR.

The systemic impact of edges in financial networks

TL;DR

This paper investigates how marginal changes in financial network interconnectedness, viewed at the edge level, affect systemic risk (SR). It uses the differential DebtRank () framework to quantify edge impact in a two-network Brazilian dataset, applying machine learning (TPOT AutoML) and SHAP for interpretability to predict edge criticality and the sign of edge impact. Key findings include substantial heterogeneity in edge effects, a heavier-tailed distribution of edge impacts well-described by a log-normal, and a rising fraction of null/negative impacts as the initial shock grows. The study identifies borrower PageRank and lender centrality as primary predictors of edge criticality, and reveals a tipping-point behavior in the sign of edge impact around a small lender PageRank, with important macroprudential implications for edge-level interventions and policy design.

Abstract

In this paper, we assess how the stability of financial networks is affected by interconnectedness considering its tiniest variation: the edge. We compute the impact of edges as the percentage difference in the systemic risk (SR) of the whole network caused by the inclusion of that edge. We apply this framework to a thorough Brazilian dataset to compute the impact of bank-firm edges. After observing that (i) edges are heterogeneous regarding their impact on the SR, and (ii) the fraction of edges whose impact on the SR is non-positive increases with the level of the initial shock, we use machine learning techniques to try to predict two variables: the criticality of the edges (defining as critical an edge whose impact on SR is significantly greater than that of the others) and the sign of the edge impact. The level of accuracy obtained in these prediction exercises was very high. These results have important implications for the development macroprudential policies aimed at financial stability. Our framework allows to identify, based on features related to the origin and destination nodes of the edge (i.e., the lending bank and the borrowing firm), whether an additional loan will have a significant and positive impact on the SR.

Paper Structure

This paper contains 13 sections, 7 equations, 11 figures, 6 tables.

Table of Contents

  1. Introduction
  2. The dataset
  3. Methodology
  4. Systemic risk
  5. Classification of critical edges
  6. Machine learning techniques
  7. Results
  8. General results
  9. Predicting edges' critical status
  10. Predicting the sign of the edges' impact
  11. Concluding remarks
  12. Potential explanatory variables
  13. SHAP partial dependence plots

Figures (11)

  • Figure 1: A stylized representation of a financial network
  • Figure 2: Example of SR computation through the DDR approach
  • Figure 3: Overview of the TPOT pipeline search. Source: olson2016evaluation
  • Figure 4: Fraction of edges with null or negative impact on the systemic risk of the financial network
  • Figure 5: Distribution of edges' impact (in log scale)
  • ...and 6 more figures