Banking System ABM Model

The 2008 financial crisis demonstrated that banking system risk is not the sum of individual bank risks. A bank that looks solvent in isolation can fail when its counterparties fail, and the network of interbank exposures determines whether a single default stays contained or cascades through the system. Traditional bank regulation - Basel I and II capital ratios, stress tests based on individual bank balance sheets - was built on the assumption that if each bank is sound, the system is sound. That assumption collapsed spectacularly. Banking system ABMs exist to model what single-bank analysis cannot: the emergent network topology of interbank lending, the nonlinear propagation of losses through that network, and the feedback loops between bank behavior, market liquidity, and asset prices that amplify small shocks into systemic crises.

The intellectual foundation comes from several distinct threads. Iori et al. (2006) built one of the first ABMs of the interbank market, showing that the network structure of overnight lending generates liquidity dynamics invisible to representative-bank models. Gai and Kapadia (2010) formalized contagion in random financial networks, demonstrating a 'robust-yet-fragile' property: highly connected networks absorb small shocks well but propagate large shocks catastrophically. Battiston et al. (2012) introduced DebtRank, a network centrality measure that quantifies how much of the system's value is destroyed when a given bank fails - the formal metric behind 'too big to fail.' Georg (2013) added a central bank with emergency lending facilities, showing how lender-of-last-resort policy interacts with network topology. Klimek et al. (2015) built a systemic risk ABM calibrated to the Austrian banking system. Poledna et al. (2017) extended the framework to multi-layer networks where banks are connected through multiple channels (interbank loans, derivatives, securities holdings, payment systems) simultaneously.

Central banks and financial stability authorities now use banking ABMs operationally. The IMF's Agent-Based Book of Analytics (ABBA) framework (Bookstaber et al. 2018) includes an interbank contagion module as one of its three core components. The European Central Bank's research directorate has published multiple banking network ABMs for macroprudential stress testing. The Bank of England uses network models to assess system-wide stress propagation. The Austrian central bank (OeNB) runs multi-layer network ABMs calibrated to granular bilateral exposure data. The shift from single-bank stress testing to system-wide network analysis is one of the major post-crisis regulatory innovations, and ABMs are the primary computational tool for this analysis.

The model family continues to evolve along several active frontiers. Climate risk modeling adds physical and transition risk shocks that propagate through bank portfolios with correlated exposures to carbon-intensive sectors. Cyber risk introduces operational contagion distinct from financial contagion. Cross-border banking networks extend the framework to international capital flows and regulatory arbitrage. Central bank digital currencies (CBDCs) create new deposit migration dynamics that interact with interbank funding. The common thread is that each extension adds a new layer to the network and a new channel for systemic propagation.

The banking system ABM assembles five agent populations interacting through a layered network structure. Commercial banks are the primary agents: each carries a full balance sheet with assets (loans to firms and households, interbank loans, securities holdings, central bank reserves), liabilities (deposits, interbank borrowing, bonds, equity), and behavioral rules for lending, borrowing, portfolio allocation, and default. The bank population ranges from 20 to 500 agents in research implementations and 10 to 30 in browser-scale toy models. Firms form the second population: each borrows from one or more banks, generates revenue, and may default on its loans, imposing credit losses on its lending banks. Households form the third population, primarily as depositors whose withdrawal behavior creates funding risk. The central bank is the fourth agent: it sets the policy rate, operates standing lending facilities, and may intervene as lender of last resort during stress episodes. A regulatory authority is the fifth agent: it sets and enforces capital adequacy requirements (Basel-style risk-weighted capital ratios), conducts stress tests, and may impose systemic risk buffers on designated institutions.

Interaction occurs through four distinct network layers, following Poledna et al. (2017). The interbank lending network connects banks through unsecured overnight and term loans - this is the primary contagion channel for credit losses. The derivatives network connects banks through bilateral OTC contracts whose mark-to-market values create exposure. The securities network connects banks through common asset holdings - fire sales by one bank depress the market value of assets held by others, creating indirect contagion even without direct bilateral exposure. The payment system network connects banks through intraday settlement obligations that create operational liquidity risk. Each layer has its own topology (scale-free, core-periphery, or random), and shocks propagate differently through each layer. The multi-layer structure means that two banks with no direct interbank loans can still be connected through common securities holdings or derivatives exposure.

State variables update each period in a fixed sequence: macro shocks arrive and update firm revenues, firms make default decisions, banks mark loan portfolios to current values, interbank market opens and banks borrow or lend to meet reserve requirements, banks rebalance portfolios (selling securities if capital is tight), fire-sale price impacts are computed, the central bank processes liquidity requests, the regulator checks capital ratios and imposes corrective actions on undercapitalized banks, bank defaults are resolved (either through orderly wind-down or fire sale of assets), contagion losses propagate through the network, and aggregate statistics are recorded. The ordering is critical: interbank lending precedes portfolio rebalancing because banks first try to borrow before they sell assets, and fire-sale prices are computed after all selling decisions are aggregated.

The European Central Bank uses banking network ABMs to conduct system-wide stress tests that go beyond the single-bank EBA methodology. The standard EBA stress test applies a macro scenario to each bank independently and checks whether capital remains above the minimum. The ECB's network extension adds a second round: losses from the first round propagate through the interbank network and common asset holdings, generating amplification that the first-round test misses. Halaj and Kok (2013) at the ECB showed that second-round network effects can amplify first-round losses by 20-40% for concentrated banking systems. Poledna et al. (2017) demonstrated that multi-layer network effects (combining interbank, derivatives, and securities channels) amplify losses by an additional 30% beyond single-layer estimates. These findings directly inform the calibration of systemic risk buffers imposed on designated systemically important institutions (G-SIBs and D-SIBs).

The IMF's ABBA framework (Bookstaber et al. 2018) uses a banking ABM module to evaluate the interaction between bank behavior, market liquidity, and fire-sale dynamics during stress episodes. The ABBA banking module was applied to the 2016 Article IV consultation for the United States, modeling how a sudden spike in corporate bond spreads would propagate through dealer banks, money market funds, and the repo market. The model identified specific amplification pathways - notably the feedback loop between margin calls, asset sales, and further price declines - that static stress tests could not capture. The Austrian central bank (OeNB) has calibrated multi-layer banking ABMs to the full bilateral exposure data of the Austrian banking system, producing DebtRank scores and contagion simulations used in the Financial Stability Report.

The model breaks down in several identifiable settings. Banking systems dominated by a single large bank (extreme concentration) reduce to a single-agent problem where network dynamics are irrelevant. Systems where interbank lending is minimal (e.g., banks primarily funded by retail deposits with little wholesale funding) have weak direct contagion channels, though indirect channels through common asset holdings may still be significant. Cross-border banking networks require detailed bilateral exposure data that is often unavailable or subject to reporting lags. For these cases, reduced-form network models (Eisenberg-Noe clearing) or single-bank DSGE models with a banking sector may be more practical. The model also struggles with sudden institutional changes - the introduction of a deposit insurance scheme, for instance, fundamentally changes depositor behavior and requires re-specification rather than re-parameterization.