RDP 2021-01: The Role of Collateral in Borrowing 2. Institutional Background and Data

This section provides background on the markets that we analyse and describes our data. Section 2.1 details characteristics of Australian interbank markets. Section 2.2 summarises the financial environment in Australia during the period under analysis. Section 2.3 describes our data sources and Section 2.4 describes the variables in our regressions.

2.1 Interbank Markets in Australia

In Australia the repo and unsecured interbank markets are over-the-counter and settled bilaterally (more detail on the infrastructure is in Appendix A). Banks use both markets to manage short-term liquidity needs, and the repo market also acts as a source of funding for securities holdings (Hing, Kelly and Olivan 2016; Becker and Rickards 2017).

In the unsecured market, the overnight interest rate is the Reserve Bank of Australia's (RBA's) monetary policy implementation target. The rate has historically displayed very little cross-sectional variance and has not been indicative of counterparty risk. The RBA keeps the rate at target by controlling the aggregate supply of reserve balances via RBA repo auctions each morning, and the lack of cross-sectional variance is due to strong commitment by the RBA combined with relatively low interbank default risk (although these dynamics have changed since COVID-19, see Kent (2020)). Statistics on market-wide rates outside of our sample reveal some instances of loans not at the target (Debelle 2007), although these appear to be driven by fund surpluses, with deviations mainly occurring below the target. Reactions to counterparty risk tend to occur through lending quantities (e.g. Brassil and Nodari 2018), consistent with OTC unsecured interbank markets in Europe (Abassi et al 2014) and the United States (Afonso et al 2011). Afonso et al (2011) write about the period around the Lehman Brothers default:

… interest rates in the fed funds markets did not become increasingly sensitive to bank performance metrics in a consistent manner ... Rather, the results suggest that lenders seem to be more likely to manage their risk exposure by the amount they lend to a particular bank or even whether they lend to the bank at all (p 1127).

The total value of RBA repos is large relative to the size of Australian interbank markets, typically representing more than half of the total repo value outstanding. However, there is some segmentation between RBA repos and interbank repos. RBA repos tend to have maturities ranging between one and several weeks, whereas the interbank market is focused on maturities up to one week. Also, RBA repos can be against private securities – and in our sample they mostly are – whereas the interbank repo market is heavily concentrated in AGS and securities issued by the Australian state governments (referred to as ‘semis’). AGS play a similar role in Australia to Treasuries in the United States, as the safest and most liquid assets, all rated AAA (or equivalent) in our sample, while semis are still considered very high quality, all rated AAA or AA+ (or equivalent) in our sample. AGS and semis are both included in the regulatory definition of high-quality liquid assets.

The unsecured and repo markets have strong overlap in participation – in our sample all of the banks participating directly in the repo market are also active in the unsecured market. Roughly three-quarters of the banks in the unsecured market and half of the banks in the repo market have foreign parents, which are mostly large international banks from the United States, the United Kingdom, Europe and Asia.[2] The Australian banks include all of the major Australian banks and some smaller Australian financial institutions.

2.2 The Australian Financial System Leading into the Crisis

The Australian financial system is not conspicuous among advanced economies. In 2007, Australia's bank assets were 114 per cent of GDP, relative to an average of 131 per cent across high-income OECD countries (Davis 2011). Australian banks were (and are) relatively concentrated in residential real estate loans, which in 2009 comprised 59 per cent of all loans, compared to, for example, 15 per cent in the United Kingdom and 38 per cent in the United States. Nonetheless, each of the major Australian banks had positive profits throughout the crisis, in part due to low investment in subprime mortgages relative to banks in other advanced economies. This helped Australia to avoid a severe downturn, with one quarter of negative GDP growth at end 2008.

Notwithstanding this, in the second half of 2008, funding conditions in Australia tightened substantially. The spread on Australian banks' short-term paper reached historical highs (Figure 1). In September banks' total bond issuance dropped to under a third of typical monthly issuance, and by November it had declined to almost zero. On 28 November, the Australian Government implemented a guarantee on banks' large deposits and wholesale debt (following similar guarantees in other countries), and a surge in bond issuance followed.

The funding tightness is also evident from the rapid rise of Exchange Settlement Account (ESA) balances to historical highs around when Lehman Brothers defaulted on 15 September (Figure 2). The expansion of ESA balances was driven by RBA repos against lower-quality securities, as banks preferred to hold onto their safest securities for other purposes, which the RBA deliberately facilitated through its auction allocations. On 8 October, immediately after our sample, the RBA expanded its eligible repo collateral to include a wider range of residential mortgage-backed securities and asset-backed commercial paper, and offered repos of extended maturities. On the same day, the RBA lowered its policy rate by 1 percentage point, the first change greater than half a percentage point in magnitude since 1992. This rapid easing occurred a day before the unscheduled and coordinated decisions by six other central banks to each cut policy rates by half a percentage point, including the Bank of England, the European Central Bank and the Federal Reserve.

Figure 1: Interbank Stress Measures – 2008
US and Australia
Figure 1: Interbank Stress Measures – 2008

Sources: Federal Reserve Bank of St. Louis; RBA

Figure 2: ESA Balances and New Borrowing from RBA – 2008
Figure 2: ESA Balances and New Borrowing from RBA – 2008

Source: RBA

2.3 Data Sources

Our transaction-level repo data, and data on participants' collateral holdings, come from two proprietary datasets that contain all overnight holdings and intraday changes in holdings (i.e. transactions) of most categories of debt securities held in Australia's debt securities infrastructure, Austraclear.[3] To identify the securities transactions that comprise repos, we use the algorithm from Garvin (2018), which detects pairs or groups of transactions that involve the same counterparties, the same type and face value of securities, and money quantities consistent with a feasible repo rate. We calibrate the algorithm to detect repos open for eight days or less with implied interest rates between 3 and 7.25 per cent, which the data indicate are realistic bounds.[4] The analysis in Garvin (2018) indicates that the algorithm detects repos with high accuracy.

The unsecured lending data come from a proprietary transaction-level dataset that contains all payments through the Reserve Bank Information and Transfer System (RITS). We use the algorithm in Brassil, Hughson and McManus (2016) to identify which payments are interbank loans. Brassil et al (2016) enhance the Furfine (1999) algorithm, which detects pairs of transactions that comprise interbank loans, to detect loans that involve more than two transactions. This picks up many otherwise-missed loans. Both our repo and unsecured datasets therefore capture loans whose principal is increased or decreased (or both) before it is fully repaid. Due to the transaction-level nature of the data, we (and other users of related algorithms) cannot explicitly distinguish between overnight loans that are subsequently rolled over without transacting, and loans that upon initiation are agreed to be for multiple days (i.e. short-maturity term loans). Our analysis groups these loan types together and, within the short-term category, puts little emphasis on maturities, instead focusing on quantity of loans outstanding at any point in time.

Data relating to entities' balance sheets are from S&P Global Market Intelligence (formerly from SNL Financial) or, if unavailable or problematic, from public financial reports. In Section 4.3 we also analyse banks' borrowing from the RBA. These data come from the securities transactions dataset, and are inferred from banks' transactions with the RBA's open market operations account.

2.4 Sample and Variables

We focus on loans outstanding during a four-week period from Monday 8 September to Friday 3 October 2008. By starting the sample shortly before the Lehman Brothers default, we can construct an ex ante (i.e. exogenous) measure of collateral holdings that remains relevant to the stressed period we examine. Our sample ends on Friday 3 October just prior to major RBA interventions that targeted interbank markets (see Section 2.2). The effect of these interventions on individual banks was likely related to the bank characteristics we use as explanatory variables, so we end the sample just before this period. For placebo regressions we also analyse a four-week sample of corresponding data covering Monday 11 September 2006 to Friday 6 October 2006.

Banks are grouped at the parent company level, in some instances aggregating across multiple Austraclear accounts. Some accounts banks use on behalf of clients – indicated by the account name – which we treat as entities separate from the bank holding the account, with missing observations for counterparty characteristics (which appropriately removes them from some regressions).[5] We remove one non-major Australian bank owing to in-sample corporate activity and a relatively inactive financial company with unavailable balance sheet data.

The key dependent variable is a balanced panel of outstanding loan balances at the lender-borrower-day-market level (LoanValueslbdm). These include loans of 8 days or less that were open at any point during the 20 business days from Monday 8 September to Friday 3 October 2008. To construct LoanValueslbdm, we sum all loans from lender l to borrower b that occurred in market m and were open at the end of day d. Usually m {unsecured, repo}; in Section 4.4 repo is separated into two markets by collateral type. If lender l never lent to borrower b in the sample or sub-sample analysed, the lb counterparty pair is excluded; if lender l lent to borrower b but not on day d in market m then LoanValueslbdm is zero. We measure outstanding loans in millions of AUD, add one, then take the natural logarithm; the raw distribution of loan sizes across entities is heavily skewed. The pre-log values of LoanValueslbdm are summarised in Panel A of Table 1. To separately analyse the extensive margin, we construct a Participationlbdm indicator variable equal to one if and only if LoanValueslbdm is greater than zero.

The key explanatory variables are collateral holdings ( Collaterali ), counterparty risk ( CPRiski ) and a Lehman default indicator variable ( LBd ).LBd is equal to 1 for all observations from Monday 15 September, the first business day after Lehman Brothers declared bankruptcy. Collateral holdings and counterparty risk are at the bank level, with banks indexed by i, which is replaced by l or b when included in regressions. In some instances collateral holdings are split into subcomponents (i.e. AGSi and Semisi). Collateral (or, where used, its subcomponents) and counterparty risk are standardised to have zero mean and unit variance. Panel B of Table 1 reports the means and standard deviations before standardisation.

Table 1: Summary Statistics
Panel A: Loans outstanding at lender-borrower-day-market level, in AUD millions, pre-logs
  Count 25% Median Mean 75% 99% Max Std dev
Repo All 8,100 0 0 11 0 206 2,591 79
  Non-zero 1,124 10 29 79 70 928 2,591 198
Unsecured All 8,100 0 0 19 0 400 1,480 78
  Non-zero 997 40 100 154 210 779 1,480 169
Panel B: Bank-level explanatory variables, before mean/variance standardisation
  Min Mean Max Std dev   Min Mean Max Std dev
CPRiski 0 1.15 3.49 0.97 Assets (AUDtr) 0.02 1.23 4.17 1.11
Collaterali 0 0.42 1.70 0.56 Liabilities/assets 0.82 0.94 0.98 0.03
AGSi 0 0.12 0.81 0.22 Foreign (indicator)   0.72    
Semisi 0 0.36 1.69 0.50  
Panel C: Number of unique lenders (borrowers) to each borrower (lender) across sample
  Lenders to each borrower   Borrowers to each lender
Min Mean Max Std dev Min Mean Max Std dev
Repo 0 4.67 17 5.54   0 3.82 14 5.11
Unsecured 0 8.31 22 6.26   0 6.80 18 5.33
Notes: Panel A reports statistics on the dependent variable LoanValueslbdm before log transformations. Lender-borrower pairs are included if they appear in any of the regressions and if one of the pair has observed counterparty characteristics (e.g. is not an unidentifiable client). Loans outstanding is zero if that lender has no open loans to that borrower on that day in that market. ‘All’ summarises all observations and ‘Non-zero’ summarises only the positive observations.
Panel B reports statistics on bank-level variables. AGSi and Semisi correspond to the AGS and semis components of Collaterali , also in log AUD billions plus one. ‘Assets’ and ‘Liabilities/assets’ are from the bank's (parent company's) balance sheet at the same date as NPLi is measured. Foreign is an indicator variable for whether the bank (or its parent company) is foreign. The zero minimum CPRiski corresponds to a financial services company that reported zero provisions for loan losses in its 2007 annual report. See the Section 2.4 text for more details.
Panel C reports how many different lender or borrower counterparties each bank had across the full sample, separated by market. All entities that appear in the sample are included.

More specifically, Collaterali measures the total face value of AGS and semis that were held at both 1 September 2008 and 8 September 2008. Including only securities that were held at both dates minimises the influence of high-frequency changes in holdings prior to the sample start.[6] We measure this in billions of AUD, add one, then take the natural logarithm. Because collateral holdings is an ex ante measure, it is best interpreted as a soft constraint on repo borrowing. Banks could have acquired more collateral between the first week of September and the weeks in our sample, but it would have entailed transaction costs that were likely intensified by the tight liquidity conditions.

CPRiski measures the ratio of non-performing loans over total loans in percentage points (NPL), as at end 2007 or the closest available reporting date to end 2007. This is a common measure of counterparty risk in interbank markets, used by, for example, Cocco, Gomes and Martins (2009) and Afonso et al (2011). NPL has several advantages as a measure of counterparty risk in this setting:

it measures banks' assets at the lower tail of the quality distribution, of which subprime mortgages are likely to be a component; it depends directly on balance sheet outcomes, without determination by market speculation or credit rating agencies, which can provide biased forward-looking information before a crisis (Bolton, Freixas and Shapiro 2012); and it is available for all banks in our sample. Section 5 presents further evidence that NPL is a suitable measure of counterparty risk. For a small number of entities, non-performing loans are not reported, in which case we replace non-performing loans with the closest available measure to debt assets that are 90 days or more past due.

The sample includes 30 borrowers and 31 lenders with observations of Collaterali and CPRiski (i.e. banks) and 42 lenders and 48 borrowers in total. In each market, there are typically about 20 active on any day, and around 50 to 60 active counterparty pairs. Panel C of Table 1 provides summary statistics for how many counterparties each bank dealt with in each market.


In Section 5 we assess the relevance of domicile to our research question. [2]

Discount securities issued by non-government entities – that is, private securities without coupon payments – are the only debt securities not present in the datasets. Prudential data indicate these could comprise up to a quarter of the interbank repo market, but the RBA's liaison with financial institutions suggests they are much less. The transaction dataset would also miss transactions that take place within a single Austraclear account (and therefore require no movement of securities). Garvin (2018) discusses the prevalence of such transactions, which include any repo market segment transacted through foreign infrastructure. [3]

Specifically, this is the widest interest rate range at which repos are detected with the following characteristics: the implied interest rate is the same when rounded to four and five decimal places; there are at least two other detected repos at the same rate (rounded to four decimal places); and the repo is not between two ‘client’ designated accounts. [4]

Accounts are assumed to be client accounts if the words ‘nominee’, ‘custodian’ or ‘client’, or any abbreviation of these, appear in the account name. Where multiple client accounts are owned by a single organisation, those client accounts are grouped into a single ‘client’ entity. [5]

Specifically, for each International Securities Identification Number (ISIN) – the lowest level at which securities are identifiable – we take the minimum of the face value held at 1 September and at 8 September, and then sum these minima across all AGS and semis ISINs held by that bank (or for AGSi or Semisi , only the bank's AGS or semis holdings). [6]