Payment Systems ABM Model
How do banks' individual liquidity management strategies, payment queuing decisions, and bilateral netting behavior generate systemic gridlock risk, cascading settlement failures, and the need for central bank intraday liquidity provision in large-value payment systems?
Agent-based models · Model guide
Payment Systems ABM: question, structure, and use cases
How do banks' individual liquidity management strategies, payment queuing decisions, and bilateral netting behavior generate systemic g...
How do banks' individual liquidity management strategies, payment queuing decisions, and bilateral netting behavior generate systemic gridlock risk, cascading settlement failures, and the need for central bank intraday liquidity provision in large-value payment systems?
Background
Large-value payment systems (LVPS) are the backbone of modern financial infrastructure. Systems like Fedwire in the United States, TARGET2 in the euro area, and CHAPS in the United Kingdom settle trillions of dollars daily in real-time or near-real-time gross settlement (RTGS). Each participating bank holds intraday reserves and decides when to submit payments, whether to queue them, and how aggressively to conserve liquidity. These micro-level timing and queuing decisions create system-wide interdependencies: Bank A waits for an incoming payment from Bank B before releasing its own payment to Bank C, and that sequencing propagates through the entire network. Traditional analytical models of payment systems - Kahn-Roberds (1998), Freeman (1996) - treat participation decisions in reduced form and cannot capture the granular timing and network feedback that drive gridlock. Agent-based models of payment systems were developed precisely because the aggregate settlement outcome is an emergent property of individual bank strategies interacting through a shared network.
The foundational work in this space includes Beyeler et al. (2007), who studied congestion and cascade behaviour in Fedwire-calibrated payment flows; Arciero et al. (2009), who constructed a multi-agent simulation of the Italian payment system (BI-COMP and TARGET2-Banca d'Italia) to evaluate queuing algorithms and liquidity-saving mechanisms; and Galbiati and Soramaki (2011), whose contribution is a STYLIZED agent-based payment system - banks that learn, over repeated days, how much liquidity to post at the start of the day given delay costs and liquidity costs. It is worth being exact about that last one, because it is routinely over-claimed: it is not a TARGET2 calibration, it contains no 'hoarding parameter' of the kind this page's toy model uses, and it reports no critical threshold value for such a parameter. Diehl (2013) provided a comprehensive simulation framework for payment system design, covering RTGS, deferred net settlement (DNS), and hybrid systems. Bech and Garratt (2003) is a two-player game-theoretic model of intraday liquidity management - not an empirical Fedwire study - and what it supplies to the ABM literature is the strategic structure of the delay problem, not measurements. Cont, Moussa and Santos (2013) is adjacent rather than an extension: it measures default contagion on the Brazilian interbank CREDIT-exposure network, which is a solvency channel, not an intraday settlement one. Paddrik et al. (2020) is the closer relative, extending payment-network analysis to resilience under multi-system operational disruptions.
Central banks use payment-system simulation operationally. The Bank of England's Payment System Simulator has been used for CHAPS analysis; Eurosystem work has evaluated TARGET2 liquidity-saving mechanisms; Canadian and Federal Reserve researchers have studied large-value payment-system liquidity and operational resilience with transaction-level simulation. These models matter because small rule changes can shift intraday liquidity demand, queue lengths, and settlement-delay risk across systems that move very large daily values. The model class has expanded to cover CLS, CCPs, and cross-border payment corridors.
How the Parts Fit Together
The model assembles three core agent populations operating on a shared settlement infrastructure. Banks are the primary agents: each holds an intraday liquidity position (reserves at the central bank), a queue of outgoing payment obligations with associated deadlines and priorities, and a behavioral strategy governing when to release payments versus hold them in queue. Bank populations range from 15 to 80 in realistic calibrations (matching actual LVPS participant counts) and 5 to 20 in browser-scale demonstrations. The central bank is the second agent: it sets intraday credit terms (collateralized or free), operates the settlement mechanism (RTGS queue, netting algorithm, or hybrid), and intervenes when gridlock threatens systemic stability. The third population is an exogenous payment generator that produces the daily flow of payment obligations between banks, calibrated to match the empirical distribution of payment values, timing, and bilateral concentration observed in actual system data.
Interaction happens through the settlement network. When Bank A submits a payment to Bank B, the system checks whether A has sufficient reserves (including any available intraday credit). If yes, the payment settles immediately: A's reserves decrease, B's reserves increase. If not, the payment enters A's queue. Queued payments create dependencies: B may be waiting for that incoming payment to fund its own outgoing obligations to C, D, and E. The netting or offsetting algorithm periodically scans for cycles of queued payments that can be settled simultaneously without any bank's balance going negative. This cycle-detection mechanism is the key institutional feature that separates modern hybrid RTGS systems from pure gross settlement. The network topology - who pays whom and in what amounts - is drawn from empirical transaction data showing heavy concentration among a few core banks.
State variables update within each simulated business day on a minute-by-minute or payment-by-payment basis. The intra-day timeline proceeds: (1) banks receive their day's payment obligations, (2) at each tick, banks decide which payments to release based on their strategy, (3) submitted payments settle immediately if funded or enter the queue, (4) the netting algorithm runs periodically (every N minutes or when the queue exceeds a threshold), (5) the central bank monitors aggregate queue depth and may inject emergency liquidity, (6) at end-of-day, remaining queued payments are either force-settled using end-of-day borrowing or fail. This granular intra-day resolution is what distinguishes payment system ABMs from models that treat settlement as a single daily event.
Applications
Payment-system simulators are used to compare rule changes before they reach production. For CHAPS-style systems, a simulator can compare tiered participation with broader direct participation by holding the payment matrix fixed and changing who settles directly. The output is not a single policy number; it is a set of liquidity and delay tradeoffs: peak intraday liquidity, queue depth, deadline misses, and the concentration of settlement pressure. For TARGET2-style hybrid systems, Arciero et al. use simulation to evaluate liquidity-saving mechanisms by comparing settlement performance under alternative queue-resolution rules. The classroom point is the design margin: a liquidity-saving algorithm can reduce required liquidity, but the result depends on payment timing, participant behavior, and the exact queue rule.
Fedwire-style simulations study operational resilience by forcing a major participant outage and tracing the payments that become delayed because counterparties expected incoming funds from the disrupted participant. The ABM tracks queue depth, delay duration, affected counterparties, and the liquidity injection needed to keep settlement moving. Multi-system work extends the same logic to cases where CLS, Fedwire, or related settlement infrastructures interact. The credible lesson is the propagation mechanism and the dependence on the payment network; exact dollar impacts are scenario outputs, not portable constants.
Payment system ABMs break down when the relevant dynamics are not about intraday timing and queuing. Cross-border payment corridor design, where the issue is exchange rate risk and correspondent banking charges over multi-day settlement, is poorly served by an intraday-resolution model. Retail payment systems (ACH, card networks) operate on batch cycles where individual timing decisions are irrelevant - a throughput queuing model is more appropriate. Markets where a single CCP centralizes all settlement (e.g., equity clearing through DTCC) reduce the bilateral network structure that drives gridlock in the ABM. For those settings, a queuing-theory model or a network flow optimization is more suitable than an agent-based approach.
Components
A participant bank with state vector (reserve balance, payment queue, intraday credit line, queuing strategy parameters, bilateral exposure limits). Decides when to release queued payments and how aggressively to hoard liquidity.
Bank i's real-time reserve position at the central bank at tick t. Increases with incoming payments, decreases with outgoing settlements. The binding constraint on payment capacity.
Ordered list of Bank i's pending outgoing payments at tick t. Each entry has a value, recipient, deadline, and priority class. Queue depth is the key state variable for gridlock detection.
Bank i's strategic aggressiveness in conserving reserves. Higher lambda means the bank releases payments later, waiting for incoming flows. Drives the coordination failure at the heart of gridlock.
System-level binary variable equal to 1 when total queued payments exceed a threshold fraction of total daily value and no further settlements occur without intervention. Emergence of gridlock is the model's primary target phenomenon.
Netting or offsetting algorithm that periodically scans queued payments for bilateral or multilateral cycles that can settle simultaneously. Reduces gross liquidity needs by exploiting payment offsets.
Maximum intraday overdraft the central bank extends to bank i, typically collateralized. Expands effective settlement capacity beyond own reserves but creates central bank exposure.
Number of payments that fail to settle by end-of-day. The tail risk metric that central banks minimize through system design and intervention rules.
Assumptions
Banks actively manage intraday liquidity by timing payment submissions to maximize incoming flows before releasing outgoing obligations, rather than submitting all payments immediately.
If violated: If banks submit all payments immediately at the start of day, gridlock disappears (sufficient aggregate liquidity exists in most systems) but intraday liquidity usage peaks far above what central banks actually observe. The strategic timing is what creates the coordination problem.
The bilateral payment network is concentrated: a small number of core banks account for a large share of total value, and payment flows are not uniformly distributed.
If violated: A uniform random network underestimates gridlock risk because it misses the hub-and-spoke structure where a single core bank's delay cascades through many counterparties simultaneously.
Settlement dynamics unfold within a single business day on a minute-by-minute or payment-by-payment basis. Overnight or multi-day dynamics are outside scope.
If violated: Aggregating to daily resolution collapses the timing and queuing dynamics that are the entire point of the model. Intraday is non-negotiable for payment system analysis.
The set of payments each bank must make on a given day is determined exogenously (drawn from calibrated distributions). Banks choose when to submit, not whether to submit.
If violated: Endogenous payment generation (banks choosing transaction volumes based on system conditions) would create a second-order strategic interaction that most payment system ABMs do not model. Extensions exist but add substantial complexity.
The central bank provides collateralized intraday credit and may intervene to break gridlock. It does not face a binding budget constraint intraday.
If violated: Without central bank backstop, any queuing equilibrium is fragile: a single bank's delay can cascade into system-wide failure. The model studies when and how much intervention is needed, not whether it should exist.
Banks cannot borrow reserves from each other during the settlement day. Liquidity comes from own reserves, incoming payments, or central bank credit.
If violated: An intraday interbank market would provide an alternative liquidity source that reduces gridlock risk but introduces counterparty credit considerations. Bech and Garratt (2003) is sometimes cited for this channel and should not be: in their game the alternative to paying early is central bank intraday credit - collateral costs and overdraft fees - not borrowing reserves from another bank. An intraday money market is a genuinely separate extension and remains outside the baseline.
Continue reading
Concepts, data, and nearby models
Open the concept, data series, policy setting, or neighboring model that anchors this page.