💰 quantitative finance

Integrating granular data into a multilayer network: an interbank model of the euro area for systemic risk assessment

This paper constructs an empirically grounded, multilayer network of euro area banking groups by integrating heterogeneous supervisory datasets to demonstrate how capturing distinct transmission channels reveals critical structural heterogeneities and systemic risk dynamics that are obscured by traditional aggregated network models.

Original authors: Ilias Aarab, Thomas Gottron, Andrea Colombo, Jörg Reddig, Annalauro Ianiro

Published 2026-02-12

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Ilias Aarab, Thomas Gottron, Andrea Colombo, Jörg Reddig, Annalauro Ianiro

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine the banking system not as a collection of isolated banks, but as a massive, bustling city where every bank is a building. In this city, money doesn't just sit in vaults; it flows constantly between buildings through different types of "roads."

This paper is about building a super-detailed, 3D map of this city to see what happens when a building catches fire (a bank gets into trouble).

Here is the breakdown of their work using simple analogies:

1. The Problem: The "Flat Map" vs. The "3D City"

For a long time, regulators tried to understand the risk of the whole banking system by looking at a flat, 2D map. They would add up all the money owed between banks and draw a single line connecting them.

The Analogy: Imagine trying to understand traffic in a city by only looking at a flat piece of paper. You see that Building A is connected to Building B. But you don't know how they are connected. Is it a narrow footpath? A highway? A subway tunnel? A bridge?

The Flaw: If you flatten everything, you miss the details. A bank might look "safe" on the flat map because it doesn't owe much money in total, but it might be dangerously exposed on a specific, narrow bridge that, if it collapses, brings down the whole block.

2. The Solution: The "Multilayer Network"

The authors (from the European Central Bank) decided to stop using the flat map. Instead, they built a multilayer network. Think of this as a stack of transparent sheets, where each sheet represents a different "road" or type of connection between banks.

They used real, granular data (like a massive digital ledger of every single loan and trade) to build these layers:

Layer 1 (Long-term Loans): The "Highways." These are big, slow-moving loans banks give each other for years.
Layer 2 (Short-term Loans): The "Sprint Tracks." These are quick loans for a few days or weeks. If these dry up, banks panic immediately.
Layer 3 (Securities Cross-holdings): The "Shared Investments." Bank A owns stock in Bank B, and Bank B owns stock in Bank A.
Layer 4 (Repo Market/Short-term Funding): The "Emergency Water Supply." This is where banks borrow cash overnight using assets as collateral. It's the most sensitive layer.
Layer 5 (Common Assets): The "Shared Garden." Even if Bank A and Bank B don't know each other, they might both own the same type of stock. If Bank A sells it all at once, the price crashes, hurting Bank B too. This is called a "fire-sale."

3. The Discovery: "Who is the Real Boss?"

When they looked at their 3D map, they found something surprising: The most important bank changes depending on which "layer" you look at.

The Analogy: Imagine a city where one building is the biggest skyscraper (the "Big Bank"). On the flat map, it looks like the most important building.
The Reality: On the "Long-term Loan" layer, the Big Bank is fine. But on the "Emergency Water Supply" layer, a tiny, medium-sized building is the one holding the main valve. If that small building fails, the water stops flowing to the whole city, even though the Big Bank is still standing.
The Lesson: If you only look at the flat map (aggregated data), you might protect the wrong bank and miss the real weak link.

4. The Simulation: "What If?" Scenarios

The authors ran two types of simulations on their 3D map:

A. The "DebtRank" (The Domino Effect)
They asked: "If Bank X fails, how much of the city's total value disappears?"

Result: They found that in the "Short-term" layers (the sprint tracks and emergency water), the risk of a chain reaction is much higher than in the long-term layers. A small shock here can cause a massive liquidity crunch (a "run on the bank") much faster than a long-term loan default.

B. The "Agent-Based Model" (The Panic Party)
They created a computer simulation where banks act like real people.

The Scenario: Bank A fails.
The Reaction: Bank B sees this and gets scared. Instead of lending money, Bank B starts hoarding cash (liquidity hoarding). Bank C, who needed that cash, can't pay its bills and fails. Bank D, who owned the same stocks as Bank A, sees the stock price crash and loses money.
The Result: They found that a tiny number of banks (about 5% of the system) could trigger massive cascades of failure, while many huge banks did very little. It proved that size doesn't equal safety; position matters more.

5. Why This Matters

This paper is a wake-up call for regulators. It says:

"Stop looking at a blurry, flat photo of the banking system. You need a high-definition, 3D model that shows exactly how banks are connected through loans, trades, and shared investments."

By understanding these different "layers," regulators can:

Spot the real weak links before they break.
Design better safety nets (like requiring banks to hold more cash if they are heavily involved in the "emergency water" layer).
Prevent the next crisis by knowing that a shock in one specific market (like short-term funding) can spread much faster than we thought.

In short: The banking system is a complex web of different relationships. To keep it safe, we need to understand the web, not just the knots.

1. Problem Statement

The paper addresses a critical gap between theoretical models of financial contagion and empirical reality.

Theoretical vs. Empirical Disconnect: While micro-structural models emphasize that shock propagation occurs through multiple channels (counterparty exposure, short-term funding/rollover risk, securities cross-holdings, and common-asset fire-sales), empirical studies often rely on stylized or simulated networks or focus on a single exposure dimension.
Data Fragmentation: Real-world bilateral exposures are fragmented across heterogeneous supervisory and statistical datasets (e.g., AnaCredit, EMIR, SFTDS, FINREP). These datasets differ in frequency, definitions, reporting structures, and entity identifiers, making it difficult to reconcile them into a coherent, consistent network representation.
Limitations of Aggregation: Collapsing these diverse channels into a single "flattened" network can mask economically relevant structures, misidentify systemically important institutions, and obscure channel-specific transmission mechanisms.

2. Methodology

The authors construct an empirically grounded multilayer network for significant banking groups in the Euro Area (EA) using a rigorous data integration and modeling process.

A. Data Integration

The study integrates multiple granular datasets to form layer-consistent exposure matrices on a common node set (banking groups):

Node Identification: Uses RIAD (Register of Institutions and Affiliates Data) and the ROSSI list (Significant Institutions) to define banking groups and their consolidation boundaries (including subsidiaries and financial holdings).
Layer Construction:
1. Long-term Credit: Derived from AnaCredit (loans >€25k, maturity ≥3 months).
2. Short-term Credit: Derived from AnaCredit (loans <3 months).
3. Cross-Securities: Derived from CSDB (issuance) and SHSG (holdings), representing equity and debt cross-holdings.
4. Short-term Funding: Derived from SFTDS (Securities Financing Transactions), covering repo and buy-sell-back transactions.
5. External Securities (Fire-Sale): A projected layer derived from SHSG. Instead of a bipartite network (banks vs. securities), it is projected onto the bank node set. Edge weights represent the market value of overlapping portfolios (common assets), calculated as $w_{ij} = \min(\text{holdings}_i, \text{holdings}_j) \times \text{price}$ .
6. Flattened Network: An artificial aggregate layer summing all other layers to serve as a benchmark.
Enrichment: Nodes are enriched with balance-sheet data from FINREP (Total Assets) and COREP (Tier 1 Capital) to support model calibration.

B. Analytical Frameworks

Two distinct systemic risk frameworks are applied to the constructed network:

Propagation-Based Measure (DebtRank):
- Adapts the DebtRank algorithm to calculate the systemic impact of a node's distress.
- Credit Risk Calibration: Uses capital buffers (Tier 1) to measure counterparty loss propagation in long-term credit and cross-securities layers.
- Liquidity Risk Calibration: Uses liquidity buffers (Cash - $\beta \times$ Deposits) to measure rollover/funding risk propagation in short-term credit and funding layers.
- Metric: $DR_i = \sum_j [h_j(T) - h_j(1)] v_j$ , measuring the additional distress propagated excluding the initial shock.
Micro-Structural Agent-Based Framework:
- Simulates a shock cascade across layers:
  - Phase 1: Long-term market shock (counterparty default) leads to capital losses.
  - Phase 2: If capital ratios drop below a threshold, banks hoard liquidity in the short-term market, causing funding stress.
  - Phase 3: Fire-sales in external securities markets depress asset prices, affecting overlapping portfolios.
- Metrics: Number of additional defaults and cumulative capital base of failed institutions.

3. Key Contributions

Measurement Innovation: Successfully harmonizes heterogeneous, granular supervisory data (AnaCredit, EMIR, SFTDS, etc.) into a single, semantically consistent multilayer network with a unified node set.
Structural Evidence: Documents pronounced cross-layer heterogeneity. Connectivity and centrality distributions vary significantly across layers (e.g., credit layers show fat-tailed, scale-free distributions, while cross-securities layers are flatter).
Methodological Shift: Demonstrates that aggregating layers into a single network masks channel-specific structures and can lead to the misidentification of systemically important institutions.
Empirical Application: Extends standard systemic risk models (DebtRank and Agent-Based Models) to operate directly on observed, real-world exposures rather than simulated data.

4. Key Results

Topological Heterogeneity:
- Degree Distribution: Credit and funding layers exhibit right-skewed, fat-tailed distributions (core-periphery structure), whereas the cross-securities layer is more uniform. The external securities layer shows a left-skewed fat-tailed structure.
- Centrality: A flattened network fails to capture the specific topology of individual layers. Institutions central in one layer (e.g., repo market) may be peripheral in another (e.g., long-term credit).
Systemic Risk Findings (DebtRank):
- Channel Specificity: The identity of the "most systemic" bank changes depending on the transmission channel.
- Liquidity Sensitivity: Short-term funding (repo) markets exhibit significantly higher contagion potential than short-term credit markets. Systemic impact in the repo layer is an order of magnitude larger under base scenarios.
- Non-Linearity: Tighter liquidity conditions (higher $\beta$ ) amplify contagion non-linearly. Increasing the liquidity buffer scaler from 0.1 to 0.2 causes a much larger jump in DebtRank than increasing from 0.05 to 0.1.
- Aggregation Effects: Aggregating layers generally overstates systemic impact compared to a linear superposition of layer-specific impacts, though the difference is modest for most nodes.
Agent-Based Simulation Results:
- Cascades: A small subset of institutions triggers large default cascades (up to 43 additional defaults).
- Size vs. Position: Systemic impact is not solely determined by bank size; smaller institutions with specific network positions can trigger larger cascades than larger, well-connected banks.
- Liquidity Buffer Impact: Increasing $\beta$ from 0.05 to 0.2 leads to an extreme increase in defaults (from ~8 to ~75 additional defaults for top nodes) and a 500% increase in cumulative defaulted capital, highlighting the critical role of liquidity buffers.

5. Significance and Policy Implications

Layer-Aware Stress Testing: The paper argues that stress tests must be "layer-aware." Assessing resilience based on a single aggregated network is insufficient because institutions play different roles in different markets.
Targeted Regulation: The findings suggest that macro- and micro-prudential instruments should be designed to address specific transmission channels (e.g., liquidity requirements for repo markets vs. capital buffers for long-term credit).
Data-Driven Supervision: The methodology provides a blueprint for regulators to utilize existing granular data collections (AnaCredit, SFTDS, etc.) to build realistic, empirically calibrated models for systemic risk assessment, moving beyond reliance on stylized assumptions.
Fire-Sale Dynamics: The explicit modeling of overlapping portfolios provides a mechanism to quantify how asset liquidation in one bank can destabilize others through common-asset price depreciation.

In conclusion, the paper establishes that granular data integration into a multilayer framework is essential for accurately identifying systemic risks, as it reveals structural heterogeneities and channel-specific vulnerabilities that are invisible in traditional, aggregated network models.