Source-linked AI summary
Systemic risk analysis in reconstructed economic and financial networks
Giulio Cimini, Tiziano Squartini, Diego Garlaschelli, Andrea Gabrielli
TL;DR
The paper tackles systemic-risk estimation when privacy restrictions leave economic and financial networks only partially observed. It reconstructs directed weighted network ensembles from node fitnesses and degrees for a limited node subset, and reports strong agreement with real-network systemic-risk properties. The method’s accuracy depends on its configuration-model and fitness assumptions, as well as the available calibration information.
Problem
Privacy restrictions limit mutual-exposure data, complicating structural and systemic-risk estimation in economic and financial networks.
Method
The method uses node-specific fitnesses and partial in/out-degree information to calibrate a fitness-induced directed configuration-model ensemble and estimate properties from ensemble averages.
Results
The method reconstructs directed weighted networks and shows excellent agreement between synthetic and real networks for percolation, shortest-path, and DebtRank properties.
Takeaways & Limitations
The approach provides a valuable way to estimate systemic-risk-related network properties using minimal information that may be publicly available for privacy-protected systems.
Takeaways & Limitations
Accuracy depends on configuration-model, fitness-model, and partial-degree-information assumptions; the fitness model’s empirical accuracy is especially important.
Abstract
from arXiv · showhide
We address a fundamental problem that is systematically encountered when modeling complex systems: the limitedness of the information available. In the case of economic and financial networks, privacy issues severely limit the information that can be accessed and, as a consequence, the possibility of correctly estimating the resilience of these systems to events such as financial shocks, crises and cascade failures. Here we present an innovative method to reconstruct the structure of such partially-accessible systems, based on the knowledge of intrinsic node-specific properties and of the number of connections of only a limited subset of nodes. This information is used to calibrate an inference procedure based on fundamental concepts derived from statistical physics, which allows to generate ensembles of directed weighted networks intended to represent the real system, so that the real network properties can be estimated with their average values within the ensemble. Here we test the method both on synthetic and empirical networks, focusing on the properties that are commonly used to measure systemic risk. Indeed, the method shows a remarkable robustness with respect to the limitedness of the information available, thus representing a valuable tool for gaining insights on privacy-protected economic and financial systems.
Introduction
The paper addresses reconstructing economic and financial networks when confidentiality limits access to mutual exposures. It develops a directed, weighted reconstruction procedure for estimating systemic-risk properties from partial topology and node-specific information.
- Incomplete network information makes estimating structural properties an unsolved challenge with important applications.
- Financial-network topology affects resilience to institutional default or distress, but confidentiality restricts regulators’ knowledge of mutual exposures.
- Dense reconstruction methods assume fully connected networks, creating a major limitation for realistic financial-network reconstruction.
- The proposed procedure reconstructs link directionality and assigns link weights for systemic-risk estimation.
- The method is evaluated on synthetic networks and the empirical World Trade Web and Electronic Market for Interbank Deposits.
Method
The method calibrates a fitness-induced directed configuration-model ensemble from node fitnesses and partial degree information, then estimates network properties from ensemble averages. Its weight prescription preserves the real network’s strength sequences on average while allowing topology and systemic-risk properties to be evaluated.
- Method: The method assumes a weighted directed network whose matrix element wi→j is the weight from node i to node j.
- Method: Node in-degree and out-degree count incoming and outgoing connections, while undirected degree counts incident connections after symmetrizing directed links.
- Method: Given fitnesses for all nodes and degrees for only a subset I, the procedure estimates a network property compatible with these constraints.
- Method: The directed configuration model represents maximally random networks constrained by ensemble-average in-degrees and out-degrees through node-specific Lagrange multipliers.
- Method: Fitnesses are assumed linearly related to the in-degree- and out-degree-induced multipliers, defining a fitness-induced configuration model.
- Method: The calibration solves for the proportionality parameter z using partial degree sequences, then generates links probabilistically and estimates X(G0) as ⟨X⟩Ω ± σΩX.
- Method: The procedure assigns weights to generated directed links using fitnesses and a normalization W representing the ensemble-average total network weight.
- Method: The weight prescription makes ensemble-average in-strength and out-strength proportional to the corresponding fitnesses and preserves the real network’s strength sequences and total weight on average.
Empirical Dataset
The evaluation uses two empirical economic and financial networks: international trade among countries and interbank loans among banks. The study analyzes 2000 WTW trade volumes and 1999 E-mid transactions, with Figure 1 comparing observed and ensemble-average strengths.
- World Trade Web: The World Trade Web represents countries as nodes and trade volumes as directed links between them.
- E-mid: The E-mid network represents banks as nodes and directed links as loan amounts granted between banks.
- Data snapshots: The displayed empirical snapshots use WTW trade-volume data from 2000 and aggregated E-mid transaction data from 1999.
- Fitnesses: For WTW and E-mid, fitnesses are real node in-strengths and out-strengths, interpreted as trade import/export volumes or bank liquidity borrowed/lent.
- Figure 1: Figure 1 compares observed node in-strengths and out-strengths with ensemble averages for WTW and E-mid.
Topological Properties
The paper evaluates systemic-risk properties on reconstructed networks, covering undirected core and percolation measures, directed connectivity measures, and DebtRank.
- Undirected properties: The main core is the highest-degree k-core, whose nodes are the network’s most influential potential spreaders of shocks.A k-core is a largest connected subgraph whose nodes each have at least k connections within it.
- Undirected properties: At the reference threshold p*=k̄^-1, giant-component size measures the fraction of nodes in a percolating cluster.This threshold is appropriate for homogeneous graphs in the infinite-volume limit, whereas it tends to zero for scale-free networks.
- Directed properties: Link reciprocity r measures mutual exposure through the ratio of bidirected links to total network connections.It is treated as a systemic-risk indicator because it measures direct mutual exposure between node pairs.
- Directed properties: Average shortest path length λ measures the average number of directed links needed for signals or shocks to propagate between reachable node pairs.The formulation avoids divergence from pairs that cannot reach one another.
- Directed properties: Group DebtRank measures the total economic value potentially affected by distress across all nodes through recursive network impacts.The initial distress can range from 0<Φ<1, with Φ=1 representing default; the final distress amounts enter the DebtRank calculation.
Results
The method is evaluated against modeling and information limitations, using FiCM ensembles to reproduce systemic-risk properties and testing accuracy as the known-node subset grows. Real and synthetic networks show strong agreement, while several properties remain accurate with sparse degree information.
- Method limitations: The method faces errors from CM modeling, the FiCM fitness assumption, and limited degree information; only the latter two are examined quantitatively here.The first error cannot be controlled because estimating the descriptive CM requires the complete degree sequences.
- FiCM validation: Real degrees closely follow FiCM expected degrees in empirical networks, supporting the model’s qualitative description of WTW and E-mid.Partial-information validation can compare expected and observed degrees on the known-node subset.
- Systemic-risk properties: Synthetic and real networks agree strongly on giant-component percolation, shortest-path distributions, and DebtRank, with all real–synthetic curve correlations above 0.99.Figure 3 reports correlations from 0.989 to 0.999 across the six plotted comparisons.
- Limited-information test: The relative error generally falls rapidly as the known-node fraction increases, reaching half its n = 1 value at n/N = 5% and one quarter at n/N = 10%.At n/N = 10%, the error is close to the final error obtained with complete degree information.
- Property-specific accuracy: With degree information for 10% of nodes, SGC, λ, kmain and Smain all have relative errors below 10%, whereas reciprocity remains difficult to reconstruct.DebtRank remains especially accurate for real networks, with relative errors around 0.5% even when information is minimal.
- Conclusion: The method estimates systemic-risk-related network properties well using a relatively small fraction of known node connections, provided all node fitnesses are known.This conclusion covers the reconstruction setting evaluated in the analysis.
Discussion
The method reconstructs directed weighted networks while preserving both strengths and topology, enabling systemic-risk property estimation from limited information. Its effectiveness is supported across synthetic and empirical networks, but depends on fitness information and model accuracy.
- Method: It estimates systemic-risk properties using minimal partial topological information plus intrinsic fitness parameters for every node.The assessed properties include k-core structure, percolation threshold, mean shortest path length, and DebtRank.
- Validation: Validation covers fitness-induced synthetic networks and two empirical systems: the World Trade Web and E-mid.The study also analyzes different temporal snapshots to assess effectiveness and robustness.
- Method: The method reconstructs directed weighted networks by reproducing strengths and tuning topology through degree-based connection probabilities.These probabilities encompass dense and sparse reconstruction as special cases.
- Limitations: The method’s accuracy depends strongly on the fitness model’s ability to describe how links are established across nodes.For WTW and E-mid, the authors report that the fitness model describes link formation well.
- Scope: The approach is presented as applicable beyond privacy-protected economic and financial systems to other directed weighted dependency networks.The discussion names ecological, metabolic, and functional brain networks as additional settings with incomplete topology.