Loan and nonloan flows in the Australian interbank network

Abstract

High-value transactions between banks in Australia are settled in the Reserve Bank Information and Transfer System (RITS) administered by the Reserve Bank of Australia. RITS operates on a real-time gross settlement (RTGS) basis and settles payments and transfers sourced from the SWIFT payment delivery system, the Austraclear securities settlement system, and the interbank transactions entered directly into RITS. In this paper, we analyse a dataset received from the Reserve Bank of Australia that includes all interbank transactions settled in RITS on an RTGS basis during five consecutive weekdays from 19 February 2007 inclusive, a week of relatively quiescent market conditions. The source, destination, and value of each transaction are known, which allows us to separate overnight loans from other transactions (nonloans) and reconstruct monetary flows between banks for every day in our sample. We conduct a novel analysis of the flow stability and examine the connection between loan and nonloan flows. Our aim is to understand the underlying causal mechanism connecting loan and nonloan flows. We find that the imbalances in the banks’ exchange settlement funds resulting from the daily flows of nonloan transactions are almost exactly counterbalanced by the flows of overnight loans. The correlation coefficient between loan and nonloan imbalances is about −0.9 on most days. Some flows that persist over two consecutive days can be highly variable, but overall the flows are moderately stable in value. The nonloan network is characterised by a large fraction of persistent flows, whereas only half of the flows persist over any two consecutive days in the loan network. Moreover, we observe an unusual degree of coherence between persistent loan flow values on Tuesday and Wednesday. We probe static topological properties of the Australian interbank network and find them consistent with those observed in other countries.

Introduction

Financial systems are characterised by a complex and dynamic network of relationships between multiple agents. Network analysis offers a powerful way to describe and understand the structure and evolution of these relationships; background information can be found in Refs. [1], [2], [3]. The network structure plays an important role in determining system stability in response to the spread of contagion, such as epidemics in populations or liquidity stress in financial systems. The importance of network studies in assessing stability and systemic risk has been emphasised in Ref. [4] in the context of integrating economic theory and complex systems research. Liquidity stress is of special interest in banking networks. The topology of a banking network is recognised as one of the key factors in system stability against external shocks and systemic risks [5]. In this respect, financial networks resemble ecological networks. Ecological networks demonstrate robustness against shocks by virtue of their continued survival and their network properties are thought to make them more resilient against disturbances [6]. Often they are disassortative in the sense that highly connected nodes tend to have most of their connections with weakly connected nodes (see Ref. [7] for details). Disassortativity and other network properties are often used to judge stability of financial networks.

There has been an explosion in empirical interbank network studies in the last years thanks largely to the introduction of electronic settlement systems. One of the first, reported in Ref. [8], examines the Austrian interbank market, which involves about 900 participating banks. The data are drawn from the Austrian bank balance sheet database (MAUS) and the major loan register (GKE) containing all high-value interbank loans above $€0.36\times {10}^6$; smaller loans are estimated by means of local entropy maximisation. The authors construct a network representation of interbank payments for ten quarterly periods from 1999 to 2003. They find that the network exhibits small-world properties and is characterised by a power-law distribution of degrees. Specifically, the degree distribution is approximated by a power law with the exponent −2.01 for degrees ≳40. This result, albeit with different exponents, holds for the in- and out-degree distributions too (the exponent is −3.1 for out-degrees and −1.7 for in-degrees). A recent study of transactional data from the Austrian real-time interbank settlement system (ARTIS) reported in Ref. [9] demonstrates a strong dependence of network topology on the time-scales of observation, with power-law tails exhibiting steeper slopes when long time-scales are considered.

The network structure of transactions between Japanese banks, logged by the Bank of Japan Financial Network system (BOJ-NET), is analysed in Ref. [10]. The authors consider several monthly intervals of data from June 2001 to December 2002 and construct monthly networks of interbank links corresponding to 21 transactions or more, i.e. one or more transaction per business day on average. Truncating in this way eliminates about 200 out of 546 banks from the network. The resulting monthly networks have a low connectivity of 3% and a scale-free cumulative distribution of degrees with the exponent −1.1.

More than half a million overnight loans from the Italian electronic broker market for interbank deposits (e-MID), covering the period from 1999 to 2002, are analysed in Ref. [11]. There are about 140 banks in the network, connected by about 200 links. The degree distribution is found to exhibit fat tails with power-law exponent 2.3 (2.7 for in-degrees and 2.15 for out-degrees), the network is disassortative, with smaller banks staying on its periphery. In a related paper [12], the authors make use of the same dataset to uncover liquidity management strategies of the participating banks, given the reserve requirement of 2% on the 23rd of each month imposed by the central bank. Signed trading volumes are used as a proxy for the liquidity strategies and their correlations are analysed. Two distinct communities supporting the dichotomy in strategy are identified by plotting the correlation matrix as a graph. The two communities are mainly composed of large and small banks respectively. On average, small banks serve as lenders and large banks as borrowers, but the strategies reversed in July 2001, when target interest rates in the Euro area stopped rising and started to decrease. The authors also note that some mostly small banks tend to maintain their reserves through the maintenance period. The evolution of the network structure over the monthly maintenance period is examined in Ref. [13].

A study of the topology of the Fedwire network, a real-time gross settlement (RTGS) system operated by the Federal Reserve System in the USA, is reported in Ref. [14]. The study covers 62 days in the first quarter of 2004, during which time Fedwire comprised more than 7500 participants and settled 3.45×10⁵ payments daily with total value $1.3 trillion. It reveals that Fedwire is a small-world network with low connectivity (0.3%), moderate reciprocity (22%), and a densely connected sub-network of 25 banks responsible for the majority of payments. Both in- and out-degree distributions follow a power law for degrees ≳10 (exponent 2.15 for in-degrees and 2.11 for out-degrees). The network is disassortative, with the correlation of out-degrees equal to −0.31. The topology of overnight loans in the federal funds market in the USA is examined in Ref. [15], using a large dataset spanning 2415 days from 1999 to 2006. It is revealed that the overnight loans form a small-world network, which is sparse (connectivity 0.7%), disassortative (assortativity ranging from −0.06 to −0.28), and has low reciprocity of 6%. The reciprocity changes slowly with time and appears to follow the target interest rate over the period of several years. A power law is the best fit for the in-degree distribution, but the fit is only good for a limited range of degrees. A negative binomial distribution, which requires two parameters rather than one for a power law, fits the out-degree distribution best.

A comprehensive survey of studies of interbank networks is given in Ref. [16]. The number of interbank markets being analysed continues to increase. For example, a study of the interbank exposures in Brazil for the period from 2004 to 2006 was reported in Ref. [17]. A topological analysis of money market flows logged in the Danish large-value payment system (Kronos) in 2006 was reported in Ref. [18], where customer-driven transactions are compared with the bank-driven ones. Empirical network studies have been used to guide the development of a network model of the interbank market based on the interbank credit lending relationships [19].

Establishing basic topological features of interbank networks is essential for understanding these complex systems. Fundamentally, however, interbank money markets are flow networks, in which links between the nodes correspond to monetary flows. The dynamics of such flows has not been examined in depth in previous studies, which mostly viewed interbank networks as static or slowly varying. But the underlying flows are highly dynamic and complex. Moreover, monetary flows are inhomogeneous; loan flows are fundamentally different from the flows of other payments. Payments by the banks’ customers and the banks themselves cause imbalances in the exchange settlement accounts of the banks. For some banks, the incoming flows exceed the outgoing flows on any given day; for other banks, the reverse is true. Banks with excess reserves lend them in the overnight money market to banks with depleted reserves. This creates interesting dynamics: payment flows cause imbalances, which in turn drive compensating flows of loans. Understanding this dynamic relationship is needed for advancing our ability to model interbank markets effectively.

In this paper, our objective is to define empirically the dynamics of interbank monetary flows. Unlike most studies cited above, we aim to uncover the fundamental causal relationship between the flows of overnight loans and other payments. We choose to specialise in the Australian interbank market, where we have privileged access to a high-quality dataset provided by the Reserve Bank of Australia (RBA). Our dataset consists of transactions settled in the period from 19 to 23 February 2007 in the Australian interbank market. We separate overnight loans and other payments (which we call nonloans) using a standard matching procedure. The loan and nonloan transactions settled on a given day form the flow networks, which are the main target of our statistical analysis. We compare the topology and variation of the loan and nonloan networks and reveal the causal mechanism that ties them together. We investigate the dynamical stability of the system by testing how individual flows vary from day to day. Basic network properties such as the degree distribution and assortativity are examined as well.