Do liquidity or credit effects explain. the behavior of the LIBOR-OIS spread?

Transcription

1 Do liquidity or credit effects explain the behavior of the LIBOR-OIS spread? Russell Poskitt a a Department of Accounting and Finance, University of Auckland, New Zealand Abstract: This paper decomposes the LIBOR-OIS spread into credit risk and liquidity components using premia of credit default swaps written on LIBOR panel banks as a proxy for the credit risk premia. The decomposition analysis reveals that changes in liquidity rather than changes in credit risk lie behind the major swings in the LIBOR-OIS spread over the course of the global financial crisis. Regression analysis shows that the behavior of the credit risk component is well-explained by a structural model of default. The behavior of the residual liquidity component is also well-explained by two liquidity variables derived from the offshore market for three-month US dollar funding. I find evidence of liquidity variables affecting the credit risk premia but no evidence of credit variables affecting the liquidity premia. (198 words) Key words: LIBOR-OIS spread, credit risk, liquidity premia, financial crisis JEL classification: G01, G10, G15 * Corresponding author. Address: University of Auckland Business School, Private Bag 92019, Auckland 1142, New Zealand. Tel: ; Fax: ; r.poskitt@auckland.ac.nz. 1

2 1. Introduction The financial crisis which began in August 2007 impaired the functioning of US dollar funding markets and disrupted a number of traditional money market spread relationships, of which the LIBOR-OIS spread is one of the more significant. 1 Prior to August 2007, threemonth US dollar LIBOR was typically 5 to 10 basis points above three-month OIS but at the height of the financial crisis this spread reached over 350 basis points before declining to more normal levels during In normal circumstances, arbitrage should ensure that the LIBOR-OIS spread does not reach these extreme levels. For example, when three-month US dollar LIBOR is well above the three-month OIS rate, arbitrageurs would make a three-month term loan in the interbank market, funding the loan by rolling over overnight borrowing in the federal funds market and hedging the interest rate risk by purchasing a three-month OIS contract (Gorton and Metrick, 2009). An excessive LIBOR-OIS spread is a sign that stress in 1 LIBOR (London Interbank Offer Rate) is an indicator of the average interbank borrowing rate in the offshore or eurocurrency market. The US dollar LIBOR fixing takes place daily in London for 15 different maturities ranging from overnight to 12 months. Fixings are a trimmed mean of the estimated borrowing rates submitted by a panel of sixteen banks comprising the largest and most active banks in the US dollar market. The sixteen banks on the US dollar LIBOR panel are listed in Table A1 in the Appendix. The OIS (overnight indexed swap) rate is the fixed rate on a fixed/floating interest rate swap where the floating rate is an overnight interest rate such as the federal funds rate. The OIS rate has emerged as the benchmark risk free rate in the money market. Although Treasury bills are acknowledged as risk free, they entail significant liquidity risk due to the convenience yield they offer investors (Feldhutter and Lando, 2008) OIS are not free of credit risk, but the credit risk is generally regarded as minor because the contracts do not involve the exchange of principal and the residual risk is mitigated by collateral and netting arrangements (Michaud and Upper, 2008). They are also considered to have minimal liquidity risk because the contracts do not involve any initial cash flow. 2 This paper focuses on the three-month spread since three-month LIBOR is the benchmark for many floating rate loans and is used in the settlement of derivative contracts such as short-term interest rates futures, forward rate agreements and interest rate and currency swaps. 2

3 interbank money markets is preventing the arbitrage process from working, either because the lender is unsure of the borrower s creditworthiness or is uncertain whether it will continue to enjoy access to funds in the overnight interbank market for the full term of the loan. [insert Fig. 1 about here] From a theoretical perspective the LIBOR-OIS spread can be represented as the sum of credit risk and liquidity premia (Bank of England, 2007). Accordingly, the fluctuations in the LIBOR-OIS spread can be attributed to variations in credit risk and liquidity premia. This paper decomposes the LIBOR-OIS spread into credit risk and liquidity components using credit default swap (CDS) premia on LIBOR panel banks as a proxy for the credit risk component. The focus of this paper is on the behavior of the three-month US dollar LIBOR-OIS spread between July 2005 and June I am interested in two questions. First, what is the relative importance of credit risk versus liquidity premia and how have these components evolved over the course of the financial crisis. Second, and perhaps more importantly, is the behavior of these two components of the LIBOR-OIS spread consistent with the behavior of proxies for credit risk and liquidity. The decomposition of the spread must be meaningful if there is to be any merit to the subsequent analysis and discussion of the relative importance of credit risk and liquidity premia. Furthermore, it is also important that any divergence between LIBOR and the OIS rate can be explained by fundamental economic forces since a poorly functioning money market will impinge on the cost and availability of credit, with the potential to affect the real economy, and will likely jeopardize the effectiveness of monetary policy (Taylor and Williams, 2008). The unprecedented fluctuations in the LIBOR-OIS spread is a challenge for the existing theoretical literature on the functioning of interbank markets. This literature acknowledges 3

4 that the existence of an interbank market allows banks to access large volumes of short-term funds to manage liquidity shocks, thus avoiding the need to liquidate long-term assets (Bhattacharya and Gale, 1987). However one of the notable features of the global financial crisis has been the extreme difficulty banks have faced rolling over short-term funding in wholesale markets (Brunnermeier, 2009; Shin, 2009). It has become evident that one of the risks that banks face in tapping the interbank market is that suppliers of funds may withdraw funding based on noisy signals of bank solvency (Huang and Ratnovski, 2008). Several recent papers highlight the critical role played by asymmetric information in the functioning of the interbank market. In these models the credit losses suffered by a bank are private information and the lack of transparency over where credit losses reside gives rise to an adverse selection problem. Heider et al. (2008) develop a model that explains the phenomena of very high unsecured rates in interbank markets, liquidity hoarding by banks and the ineffectiveness of central banks liquidity injections in restoring interbank activity. In their model the type of interbank regime that arises depends on the level and distribution of credit risk. When the level and dispersion of credit risk is low, the adverse selection problem is minor, the interest rate penalty safe borrowers pay for the presence of riskier borrowers is low and the interbank market functions smoothly with full participation. If the level and dispersion of credit is significant, however, the penalty built into the interbank rate rises and safe borrowers drop out of the market due to higher adverse selection costs. When the level and dispersion of credit risk is high, the prohibitive adverse selection costs cause the complete breakdown of the interbank market: interest rates are not high enough to compensate lenders for lending to even riskier borrowers and banks with surplus funds abandon the market, leading to hoarding of liquidity. In addition, some riskier borrowers find interest rates too high and prefer to borrow elsewhere. 4

5 Baglioni (2009) presents a model that allows a more central role for liquidity shocks. Although there is no aggregate shortage of liquidity in this model, individual banks do not know whether their liquidity shock is transitory or permanent. A bank with excess liquidity can lend in the interbank market on either a short-term or a long-term basis. If the bank makes a long-term loan and the liquidity shock is permanent then this bank will be forced to borrow in the interbank market or sell illiquid assets (and possibly incur liquidation losses). The liquidity risk a bank faces will be more severe during a period of financial turmoil when liquidity shocks have greater volatility. The model also allows participants in the interbank market to be hit by a negative credit shock in the future. If this shock is large enough some banks may be pushed into insolvency. However the distribution of credit losses is private information giving rise to an adverse selection problem. In the Bagliono (2009) model the interplay of liquidity and credit risks can lead to gridlock in the interbank market during a period of financial turmoil when the volatility of liquidity shocks is high and adverse selection problems are severe. Banks short of liquidity will want to avoid being forced to sell illiquid assets at a heavy discount to their face value while banks long on liquidity will demand a premium on interbank term loans. This situation is likely to lead to the emergence of a spread between term and overnight interbank rates and to the cessation of lending in the term segment of the interbank market despite their being no aggregate shortage of liquidity. The unprecedented fluctuations in the LIBOR-OIS spread since August 2007 has prompted a number of researchers to examine the respective roles of credit risk and liquidity premia in the determination of the LIBOR-OIS spread. In one of the earliest studies, Michaud and Upper (2008) examined the LIBOR-OIS spread in a number of currencies including the US dollar. They find that the LIBOR-OIS spreads tracked measures of credit risk such as the premia of CDS written on LIBOR panel banks although they acknowledge that the looseness 5

6 in this relationship reflects the impact of liquidity factors. The authors also noted that of the ten extraordinary liquidity management operations conducted by central banks during their sample period, LIBOR-OIS spreads declined in seven cases while CDS premia on LIBOR panel banks declined in only five cases, offering this as evidence of the important role played by liquidity factors. 3 More relevant to this paper are the studies that have focused on the impact of the Federal Reserve s Term Auction Facility (TAF) on the US dollar LIBOR-OIS spread. McAndrews et al. (2008) regress the daily change in the three-month LIBOR-OIS spread on the lagged level of the spread, the daily change in the JP Morgan Banking Sector CDS Index and separate TAF announcement date and operations date dummy variables. The authors report negative and significant estimates for both types of TAF dummies but the level of significance is stronger for the announcement date dummy variable. The authors find similar results when they use the CDS data to decompose the LIBOR-OIS spread into credit risk and non-credit risk components using CDS premia as a proxy for the credit risk component and regress the daily change in the non-credit risk component of the LIBOR-OIS spread against the TAF dummies. Frank and Hesse (2009) also decompose the US dollar LIBOR-OIS spread into credit risk and non-credit risk components using CDS premia on LIBOR panel banks. The authors show that the daily changes in the non-credit risk component were overwhelmingly positive prior to the December 2007 monetary policy initiatives of the Federal Reserve in mid-december Using e-mid data from the overnight money market, Michaud and Upper (2008) find less evidence that the behavior of the euro LIBOR-OIS spread can be explained by changes in market liquidity, proxied by the number of trades, trading volume, bid/ask spreads and the price impacts of trades. However the authors warn that their results should be treated with caution because the overnight market appears to have been much less affected by the financial turmoil than the market for term deposits. 6

7 and overwhelmingly negative in the days that followed these initiatives. The authors also employ a bivariate VAR model the dependent variables are the daily changes in the LIBOR-OIS spreads for the euro and US dollar to capture the impact of the monetary policy initiatives of both the Federal Reserve and European Central Bank (ECB). The authors report that the US dollar LIBOR-OIS spread shows evidence of significant narrowing following the announcement of the Federal Reserves s TAF and the ECB s long-term refinancing operations and following cuts in the federal funds rate and ECB s discount rate. However the authors acknowledge that the economic magnitudes are not large and not particularly effective in restoring LIBOR-OIS spreads to their pre-crisis levels. Taylor and Williams (2009) estimate an OLS regression model of the US dollar LIBOR-OIS spread in levels form. 4 They regress the three-month LIBOR-OIS spread on three proxies for credit risk the median premia on five-year CDS contracts written on LIBOR panel banks, the LIBOR-TIBOR spread and the LIBOR-REPO spread and alternative TAF auction date dummy variables. 5 The estimates of the credit risk proxies are found to have the expected sign and are usually significantly different from zero. However the sums of the estimates of the TAF auction date dummies are not negative and statistically significant. These broad results hold when separate tests are conducted using a single TAF dummy variable and/or when regressions are corrected for first-order serial correlation. The authors also test other 4 McAndrews et al. (2008) observe that estimating the OLS regression model in levels form will only yield negative estimates on the TAF dummy if the TAF has a permanent effect on the LIBOR-OIS spread; the more likely temporary effect of the TAF can only be captured by a OLS regression model estimated in first differences. McAndrews et al. (2008) also note that the presence of a unit root in the LIBOR-OIS spread suggests that the OLS regression model should be estimated in first differences. 5 The authors use five dummy variables: one dummy variable is set to one on the day of the TAF auction and the other four are set to one on each of the four days following the auction. The authors claim this specification is appropriate when the effects of the TAF auction on spreads diminish over time. 7

8 specifications of the TAF variable. When they set the TAF dummy to zero before the TAF was first announced on 12 December 2007 and one thereafter, they find a negative and significant effect of the TAF on the LIBOR-OIS, but the economic magnitude, at 8 basis points, is small. Not surprisingly, when the authors also follow the procedure employed by McAndrews et al. (2008) and include separate TAF announcement date and TAF operation date dummy variables in their regression model, they obtain similar results to those reported in the earlier study. Their regression results using other money market spreads as the dependent variable find less evidence of significant negative estimates on the TAF dummies. The authors conclude that, overall, they find more evidence that counterparty credit risk rather than liquidity premia drives money market spreads during the global financial crisis. The studies cited above suffer from a number of limitations. First, the sample periods employed have not extended beyond April 2008 and thus have arguably missed the most interesting phases of the global financial crisis, the period of intense turmoil that followed the collapse of Lehman Brothers in September 2008 and the return to more benign conditions in interbank markets from the middle of 2009 to early The sample period in the current paper extends to the end of June 2010 and provides a much richer data set. Second, not all prior research decomposes the LIBOR-OIS spread into credit risk and noncredit risk (i.e. liquidity) components and seeks to identify the driving forces behind the behavior of each component. The current study starts with this decomposition which is fundamental to understanding the interplay between credit risk and liquidity factors. Third, the studies cited above have employed dummy variables to capture the liquidity impacts of the emergency monetary policy measures. This is not necessarily the most appropriate way to understand the evolution of liquidity premia in the interbank market. The current paper introduces three new measures of liquidity conditions. Two of these measures are developed from a time series of intraday quote data from the offshore market for three- 8

9 month US dollar funding and the third comes from the domestic US commercial paper market. Fourth, and lastly, the results of prior research appear to be influenced by whether the regression models are estimated in levels or first differences. I seek to circumvent this problem by using both specifications. I decompose the LIBOR-OIS spread into credit risk and liquidity components using CDS premia on LIBOR panel banks as a proxy for the credit risk components, treating the residual as the liquidity component. The decomposition analysis shows that the liquidity component invariably exceeded the credit risk component and that variations in the liquidity component were largely responsible for the dramatic swings in the LIBOR-OIS spread. For example, on 10 th October 2008, at the height of the turmoil following the collapse of Lehman brothers, the LIBOR-OIS spread reached basis points, of which 23.0 basis points was attributed to credit risk and the remaining basis points was attributed to liquidity factors. Over the following months, the LIBOR-OIS spread contract while the credit risk component increased! These events demonstrated clearly that changes in the LIBOR-OIS spread were primarily a liquidity phenomenon. Following the literature on credit spreads and CDS premia, I then estimate a structural model of the credit risk components using data in both levels and first differences form. When using daily data in levels form, I find strong evidence that variations in the credit risk component can be explained by variations in standard proxies for credit risk, namely bank stock prices, stock price volatility and the risk free rate. Results are weaker but still supportive of the structural model when data in weekly first differences is employed. Turning to the liquidity component of the LIBOR-OIS spread, I introduce two new measures of liquidity in the offshore market for three-month US dollar funding. I find strong evidence 9

10 from daily levels data that the liquidity premia can be explained by tightness in interbank markets proxied by the average bid/ask spread on three-month US dollars and by competition for interbank US dollar deposits proxied by the number of active dealers. Results are similar when data in weekly first differences is employed. I also find that the liquidity component of the LIBOR-OIS spread is inversely related to the liquidity of the domestic US commercial paper market for financial institutions. Lastly I investigate the relationship between credit risk and liquidity factors. I find some evidence that credit risk premia reflect liquidity factors, consistent with the argument that banks facing funding difficulties are in greater risk of default. However I find no evidence that liquidity premia reflect credit risk factors. I make several contributions to the literature. First, prior research has not settled the issue as to the respective roles played by credit risk and liquidity factors in the dislocation of the LIBOR-OIS spread following the onset of the global financial crisis. Our decomposition of the LIBOR-OIS spread into credit risk and liquidity components shows that the liquidity component underwent the more dramatic changes over the sample period, suggesting that the importance attached to fluctuations in liquidity as a fundamental driving force of the LIBOR- OIS spread cannot be overstated. Second, the principal results presented in this paper are consistent with evidence from other markets disrupted by the financial crisis. For example, the finding that measures of liquidity in the offshore US dollar market can explain the behavior of the liquidity component of the LIBOR-OIS spread is consistent with recent research which highlights the role played by the shortage of US dollar funding in the dislocation of the FX swap market (Baba and Packer, 2009). Thus this paper adds to our understanding of the critical role played by market illiquidity in the disruption of traditional money market relationships during the global financial crisis. 10

11 Third, the evidence I present of the linkage between various liquidity measures and the liquidity component of the LIBOR-OIS spread is an improvement on prior research which has been limited to using monetary policy dummy variables and other money market spreads to help identify liquidity effects in US dollar funding markets (e.g., McAndrews et al., 2008; Frank and Hesse, 2009; Taylor and Williams, 2009). The remainder of the paper is organized as follows. Section 2 describes the methodology employed in this paper, namely the decomposition of the LIBOR-OIS spread into credit risk and liquidity risk components. Section 3 presents the results of the decomposition analysis and subsequent regression analysis of the credit risk and liquidity risk components of the LIBOR-OIS spread. Section 3 also addresses the issue of the interrelationship between credit risk and liquidity components. Section 4 concludes. 2. Methodology 2.1 Decomposition of LIBOR-OIS spread In the absence of any frictions, the spread between an interbank rate and a risk free rate reflects two risks: the risk of the borrower defaulting on repayment (default or credit risk) and the ease with which funding can be raised (liquidity risk). 6 Thus the LIBOR-OIS spread on day t is written as the sum of credit risk and liquidity premia in the offshore market: LIBOR OIS CRD LIQ (1) t t t 6 In reality the assumption that the spread can be decomposed into separate credit risk and liquidity risk premia is problematic since the two risks can be related. For example, a bank facing difficulty in raising funding is also in greater risk of default and the inability to raise funds will likely be factored into CDS premia. Similarly, the uncertainty over the creditworthiness of banks could lead some banks to withdraw from the interbank market, raising liquidity premia. 11

12 where LIBOR-OIS t is the LIBOR-OIS spread on day t, CRD t represents the credit risk premium associated with the LIBOR panel on day t and LIQ t represents the liquidity premium faced by the LIBOR panel in the offshore market on day t. I use CDS premia to decompose the LIBOR-OIS spread over the risk free rate into credit risk and liquidity risk components (Bank of England, 2007). I require the premia of three-month CDS contracts written on LIBOR panel banks to accomplish this decomposition. In the absence of market data on premia on three-month CDS contracts, I take one-quarter of the premia of a one-year CDS contract as a crude approximation. 7 I use the median of these estimated premia for three-month CDS contracts written on LIBOR panel banks as a proxy for CRD, thus enabling the identification of the liquidity component of the spread, LIQ Modeling the credit risk component I start with the framework of a structural model of default to help identify variables that can explain changes in CRD. 9 This approach has been used to examine the theoretical 7 DataStream does not provide premia on CDS contracts with a maturity of less than one year. Although CDS premia are unlikely to be linear in their term for example, premia on five-year CDS are less than five times the premia on one-year CDS I use this linear approximation since any bias is likely to be small for shorter-term CDS contracts and alternative rules of thumb are just as likely to lead to bias. 8 This approach implicitly assumes that changes in CDS premia reflect changes in default expectations and/or recovery rates and not liquidity premia in CDS markets. Blanco et al. (2005) note that the CDS market is relatively young and demand/supply imbalances can often cause sharp movements in CDS premia unrelated to changes in default probabilities or recovery rates. 9 A structural model of default describes the relationship between the credit spread on a firm s debt and firm value (Merton, 1974; Longstaff and Schwartz, 1995). Changes in credit spreads can be likened to changes in the premium of the put option held by shareholders. Thus factors that raise the value of the put option will widen credit spreads. Examples include a fall in the value of the firm, an increase in volatility and a reduction in the risk free rate. 12

13 determinants of corporate bond spreads (Collin Dufresne et al., 2001) and CDS premia (Ericcson et al., 2009). The results in the latter study indicate that the minimal set of determinants include financial leverage, firm-specific volatility and the risk-free rate. The authors find that the explanatory power of these determinants is approximately 55% 60% for levels data and 20% 25% for data in first difference form. The first structural variable is STK, the value of an equally-weighted portfolio of stocks of the LIBOR panel. The value of this portfolio should capture changes in the credit quality and the perceived financial stability of its panel members. As stronger relative stock price performance should be reflected in lower perceived credit risk, I expect to observe a negative relationship between STK and CRD. 10 The next structural model variable is stock price volatility, IVOL, measured by the implied volatility of put options written on the stocks of LIBOR panel banks. I use implied volatility rather than historical volatility because prior research finds that implied volatility is more successful than historical volatility in explaining variations in CDS premia (Cao et al., 2010). I construct IVOL from implied volatility data on put options written on the stocks of LIBOR panel banks where the individual options satisfy two criteria: the term to expiration is between 61 and 120 days and the ratio of the exercise price-to-stock price is between 0.85 and Increases in implied volatility reflect a higher volatility of firm value and if the volatility of the firm s assets increases, there is a greater probability of default (Merton, 1974). Thus I expect to observe a positive relationship between IVOL and CRD. The third structural model variable is RATE, the yield on five-year Treasury bonds. Longstaff and Schwartz (1995, p 808) argue that an increase in the interest rate increases the drift rate 10 Although leverage is a key variable in the structural model it is difficult to measure this variable accurately over short sampling intervals such as a day or a week. Instead I proxy changes in a bank s financial health with the bank s stock price (Blanco et al., 2005). 13

14 of the risk-neutral process for firm value, which in turn makes the risk-neutral probability of default lower. Thus I expect to observe a negative relationship between RATE and CRD. An additional variable employed in the empirical literature is SLOPE, the slope of the yield curve (Collin Dufresne et al., 2001). SLOPE is designed to capture the term structure of interest rates and is constructed by subtracting the yield on one-year Treasury bonds from the yield on ten-year Treasury bonds. An upward-sloping yield curve implies higher short-term interest rates in the future. Following the argument of Longstaff and Schwartz (1995), an increase in interest rates should lower the probability of default. Hence I expect to observe a negative relationship between SLOPE and CRD. 11 To summarize, the full structural regression model is as follows: CRD t 0 1STKt 2IVOLt 3RATEt 4 SLOPE u (2) t t where CRD t is the default component of the LIBOR-OIS spread on day t, STK t is the value of an equally-weighted portfolio of stocks of the LIBOR panel on day t, IVOL t is the median implied volatility of near-the-money put options on the stocks of LIBOR panel banks on day t, RATE t is the risk free rate on day t proxied by the yield on five-year Treasury bonds, SLOPE t is the slope of the yield curve measured by subtracting the yield on ten-year Treasury bonds less the yield on one-year Treasury bonds on day t and u t is a random error term. 11 I choose not to include credit ratings of panel banks in the regression model because changes in credit ratings tend to lag market information. More immediate market measures of creditworthiness such as share price and CDS premia are likely to be more relevant measures of changes in the credit quality of banks. The often poor timeliness of credit ratings changes can be attributed to two features of the through-the-cycle methodology employed by ratings agencies (Altman and Rijken, 2006). First, ratings agencies disregard short-term fluctuations in default risk in an attempt to avoid excessive rating reversals. Second, ratings agencies typically employ a conservative migration policy whereby ratings are only partially adjusted towards the level suggested by the permanent risk component. 14

15 Daily data on the stock prices of the LIBOR panel banks, Treasury bond yields and the premia of one-year CDS contracts written on LIBOR panel banks are all obtained from DataStream. Stock price data is only available for 13 banks as Norinchukin Bank, Rabobank and West Landesbank are not publicly-listed. CDS data is not available for Royal Bank of Canada. Daily data on the implied volatility of put options written on the stocks of LIBOR panel banks are obtained from Thomson Reuters Tick History database of Securities Industry Research Centre Asia-Pacific (SIRCA) Modeling the liquidity component It was suggested earlier that liquidity was the principal part of the non-default component of the LIBOR-OIS spread. Testing this proposition requires suitable measures for liquidity, which in the context of the interbank money market, is broadly defined as the ability of a bank to raise uncollateralized funding promptly and without incurring much in the way of transaction costs. Liquidity measures can be divided in two broad categories: trade-based measures such as trading volume, the number of trades and the turnover ratio; and order-based measures such as the bid-ask spread and price impact costs (Aitken and Comerton-Forde, 2003). While this dichotomy is useful when considering liquidity measures in exchange-traded markets, it is less useful in over-the-counter (OTC) interbank markets which are typically less transparent than exchange-traded markets. For example, it is usually very difficult to obtain data on trading volumes and transaction prices; often the only data that is available are the indicative 12 Options data is not available for the three non-listed panel banks, Norinchukin Bank, Rabobank and West Landesbank. In addition, only a small number of implied volatility observations are available on options on Mitsubishi UFJ, the parent company of Bank of Tokyo-Mitsubishi UFJ. I calculate the mean implied volatility each day for each of the remaining 12 bank from data on the implied volatility of individual option series. I then calculate the median of these 12 bank-specific daily averages. 15

16 bid and ask quotes of dealers. In light of these constraints, I develop two liquidity measures based on indicative quote data from the offshore market for three-month US dollar funding: the daily average bid-ask spread and the number of dealers active in the market. A time series of indicative quotes from the OTC market for three-month US dollars is obtained from the Thomson Reuters Tick History database of SIRCA. After filtering this data to remove one-sided quotes (i.e. either the bid or ask quote is missing) and quotes posted outside the hours 6.00AM to 4.00PM GMT on weekdays, I am left with a final sample of 694,305 quotes posted by 147 dealers. This sample of quotes is used to construct daily time series of two variables, BAS and DLRS. 13 BAS is defined as the equally-weighted bid/ask spread across all dealers with a market share of at least 0.5% and quoting on at least 25% of the days across the sample period. A total of 19 dealers meet these two criteria. 14 DLRS is defined as the number of dealers posting at least two quotes per day on the Reuters network. I expect that BAS will widen when liquidity in the interbank market worsens; thus I expect to observe a positive relationship between LIQ and BAS. However the relationship between LIQ and DLRS is not so straightforward. Normally, one would expect that an illiquid market is associated with the presence of fewer dealers and thus for there to be an inverse relationship between LIQ and DLRS (Huang and Masulis, 1999). However if banks are finding it difficult 13 The microstructure literature decomposes dealer bid/ask spreads into three components: order processing costs, inventory holding costs and adverse selection costs (Stoll, 1989). To the extent that the adverse selection cost incorporate the costs of transacting with poor credit risks the bid/ask spread will incorporate an element of credit risk. 14 The alternative of using all posted quotes to measure the daily bid/ask spread suffers from the problem that different dealers adopt different spread conventions and thus changes in the bid/ask spread over time will be influenced by changes in the composition of the quoting dealers rather than by fundamental changes in market liquidity. 16

17 to raise funds in other US dollar markets I could expect more dealers to turn to the interbank market to raise funds and there to be a positive relationship between LIQ and DLRS. An additional liquidity variable is FINCPO, the amount of outstanding commercial paper issued by financial institutions in the US commercial paper market. This weekly time series is available from the Federal Reserve Board. I use this variable to measure the availability of funding from alternative markets for US dollar funding. It is expected that increased availability of funds from these alternative markets will reduce liquidity pressures in the interbank market and hence narrow the liquidity premium in the LIBOR-OIS spread. Thus I expect to observe a negative relationship between LIQ and FINCPO. The full liquidity regression model is as follows: LIQ t 0 1BASt 2DLRSt 3 FINCPO e (3) t t where LIQ t is the non-default or liquidity component of the LIBOR-OIS spread on day t, BAS t is the equally-weighted spread across all active dealers on day t, DLRS t is the number of dealers active on day t, and FINCPO t is the amount of outstanding commercial paper issued by financial institutions in the US commercial paper market on day t and e t is a random error term. 3. Empirical results 3.1 Spread decomposition The decomposition of the LIBOR-OIS spread into credit risk and liquidity risk components is illustrated in Fig. 2. Descriptive statistics on the spread and its decomposition are presented in Table 1. Panel A of Table 1 reports a mean (median) LIBOR-OIS spread of 41.9 (11.5) basis points over the full sample period. The mean LIQ (30.6 basis points) was significantly higher than the mean CRD (11.3 basis points), although the difference between the medians was smaller. However both the t- and Wilcoxon test statistics show that the null hypothesis of 17

18 equal CRD and LIQ is strongly rejected. The descriptive statistics also show that LIQ varied enormously over the sample period, ranging between a maximum of and basis points. On the other hand, CRD moved within a much narrower range, with a maximum of 60.2 basis points and a minimum of 0.4 basis points. [insert Fig. 2 about here] I find it useful to partition the five-year sample period into four sub-periods. Period I starts on 1 July 2005 and ends on 8 August 2007 and is designated the pre-crisis period. Period II commences on 9 August 2007 and ends on 15 September 2008, the day Lehman Brothers filed for bankruptcy. The conventional wisdom is that a structual shift occurred in money markets after 9 August 2007 as the first signs of the turmoil to follow became apparent (e.g., Gyntelberg and Wooldridge, 2008; Brunnermeier, 2009). On this day, BNP Paribas suspended three of its investment funds that held mortgage backed securities and AIG warned that defaults were spreading beyond the sub-prime sector. Period III begins on 16 September 2008 and ends on 17 October This period of approximately one-month was a period of intense turmoil in interbank markets sparked by the collapse of Lehman Brothers. The failure of Lehman Brothers is regarded as a seminal event because it shattered the belief that large banks were too big to fail. The resulting turmoil prompted co-ordinated intervention by the governments and central banks of the leading industrial nations to restore investor confidence in the banking system and forestall the complete implosion of wholesale financial markets. 15 The end of this period of intense 15 For example, the UK government introduced plans to recapitalize the banking system and guarantee bank debt funding while the Bank of England injected liquidity into the financial system (Treasury, 2008). The US introduced similar measures to recapitalize banks and guarantee bank debt funding. The Federal Reserve also 18

19 turmoil is dated to the week ending 17 October The final sub-period, Period IV, begins on 20 October 2008 and extends to the end of the sample period and is referred to as the postcrisis period. [insert Table 1 about here] The data in Panel B of Table 1 show a mean (median) LIBOR-OIS spread of 7.9 (8.0) basis points during the pre-crisis period. CRD was virtually insignificant during this period, with a mean (median) of 0.8 (0.7) basis points. The low level of CRD is consistent with the general underpricing of credit risks during this period. The mean (median) level of LIQ was 7.2 (7.2) basis points. The test statistics reported in the last column confirm that the null hypothesis of equal CRD and LIQ is strongly rejected. The data in Panel C of Table 1 show that the LIBOR-OIS spread was significantly higher during period II, with a mean (median) level of 69.2 (70.8) basis points. The mean (median) level of CRD rose to 10.1 (11.0) basis points while the mean (median) level of LIQ rose to 59.1 (59.1) basis points. The test statistics reported in the last column confirm that the null hypothesis of equal CRD and LIQ is strongly rejected. The data in Panel C suggest that although both CRD and LIQ rose during this period the rise in the LIBOR-OIS spread was principally a liquidity phenomenon. The turmoil associated with period III led to a sharp rise in the LIBOR-OIS spread, with the mean (median) spread rising sharply to (258.6) basis points during this period. The data in Panel D of Table 1 show that although the mean (median) CRD approximately doubled to 22.0 (21.6) basis points, LIQ rose nearly four-fold, with a mean (median) of (233.3) basis points. Again, the test statistics reported in the last column confirm that the null began to purchase commercial paper from high quality issuers and established swap lines with foreign central banks to ease pressure in global short-term dollar markets (Paulson, 2008). 19

20 hypothesis of equal CRD and LIQ is strongly rejected. The data in panel D confirm that the rise in the LIBOR-OIS spread during this period was principally a liquidity phenomenon. The data in Panel E of Table 1 show that period IV was accompanied by a dramatic decline in the LIBOR-OIS spread, with the mean (median) spread during this period of 55.6 (30.9) basis points. CRD increased slightly during this period, recording a mean (median) of 24.7 (24.4) basis points. 16 LIQ abated sharply, with a mean (median) of 30.3 (3.5) basis points. The test statistics reported in the last column confirm that the null hypothesis of equal CRD and LIQ is strongly rejected. Although the mean and median for CRD and LIQ data give conflicting signals, the sharp decline in the LIBOR-OIS spread at a time when CRD was rising suggests that the contraction in the spread was driven by the sharp decline in LIQ. The discussion above, supported by casual inspection of Fig. 2, suggests that the variations in liquidity premia are the key to the dramatic swings in the LIBOR-OIS over the course of the global financial crisis. This conclusion is, of course, conditional on an accurate decomposition of the LIBOR-OIS spread into credit risk and liquidity premia components. One potential problem is the decision to use the median CDS premia, rather than the mean CDS premia, of the LIBOR panel as the proxy for the credit risk component of the LIBOR- OIS spread. This decision was motivated by the need to minimize the potential impact of a disproportionately large CDS premia of a single panel bank on the decomposition analysis The rise in both the LIBOR-OIS spread and CRD in the second quarter of 2010 has been attributed to concerns over counterparty credit risk, in particular, the heightened concerns of market participants over the exposure of European banks to the sovereign debt of a number of fiscally-challenged European countries (Gyntelberg et al., 2010). 17 For example, on 10 th March 2009, the estimated three-month CDS premia for Citigroup was basis points. This contributed to a situation where the mean three-month CDS premia of 70.3 basis points exceeded the median three-month CDS premia of 58.5 basis points. 20

21 For a given LIBOR-OIS spread, an artificially high CRD will reduce LIQ, and in extreme cases, lead to a negative LIQ. In fact, using the median CDS premia still gives rise to 145 cases (out of a total of 1,262 sample observations) where LIQ is negative. All but one of these negative LIQ observations occur in period IV. The maximum negative LIQ is basis points and the mean negative LIQ is -3.3 basis points. In contrast, when CRD is proxied by the mean CDS premia of the LIBOR panel, the incidence of negative LIQ increases to 189 observations, again all but one of these occurring in period IV. The maximum negative LIQ is basis points and the mean negative LIQ is -4.4 basis points. The lower incidence and size of negative LIQ when using the median CDS premia as the proxy for CRD suggests that the median CDS premia is a better proxy for CRD than the mean CDS premia. While the incidence of negative LIQ is still troubling, the relatively small size of the negative LIQ suggests that the decomposition analysis has not been adversely impacted. The following sections present a more rigorous check on the accuracy of our decomposition. This involves testing if the variations in the CRD and LIQ components can be explained by variations in fundamental proxies for credit risk and liquidity. 3.2 The credit risk component Descriptive statistics and correlations Table 2 reports summary data on the structural model variables over the sample period. CRD, proxied by the median premia on CDS written on panel banks, has a mean (median) of 11 (7) basis points and ranges between 0 and 60 basis points. STK has a mean (median) of (760.60) versus a level of 1,000 on 8 August The large differences between the maximum and minimum levels of 1, and respectively illustrate the dramatic changes that occurred in the fortunes of many LIBOR panel banks over the sample period. 21

22 IVOL has a mean (median) of 42.89% (37.29%) and ranges between a maximum of % and a minimum of 14.66%. The large difference between the maximum and minimum IVOL also illustrate the large swings that occurred in investor sentiment towards LIBOR panel banks over the sample period. Large changes also occurred in the level and slope of the yield curve during the sample period. RATE, proxied by the five-year Treasury bond yield, ranges between 5.23% and 1.26% while SLOPE ranges between 3.53% and -0.48%. [insert Table 2 about here] Table 3 reports the correlations between the structural model variables over the sample period. CRD is highly positively correlated with IVOL and SLOPE and highly negatively correlated with STK and RATE. These correlations are also consistent with those reported by Ericcson et al. (2009) in their study of the determinants of CDS premia. The sign of the correlations between CRD, STK and IVOL are as predicted by the structural model. The high correlations between the explanatory variables of the structural model suggest that multicollinearity could be an issue if most or all of these variables are included as explanatory variables in the same regression model. [insert Table 3 about here] Regression model estimates I begin the regression analysis by estimating simple versions of the structural model represented by equation (1) using daily data in levels form. Column (1) of Table 4 reports estimation results for the basic regression model. The estimation results show that STK has the expected negative sign and the estimates are significant at the level. IVOL has the 22

23 expected positive sign and is also significant at the level. RATE has the expected negative sign and is significant at the level. 18 The estimation results for STK, IVOL and RATE are consistent with the intuition underlying the structural model: the default components of the spread over the risk free rate is increasing in equity volatility and decreasing in the stock price and interest rate. These estimation results are also broadly consistent with the results of prior research into the determinants of CDS premia (Blanco et al., 2005; Ericcson et al., 2009; Cao et al., 2010). The explanatory power of the structural model is higher than reported by Ericcson et al. (2009) at 85% versus 55% 60%. Column (2) of Table 4 reports the results when SLOPE is added to the regression. I include SLOPE in the model to control for the potential effect of the slope of the yield curve on the default component of the LIBOR-OIS spread. The signs and significance levels of the estimates of the existing explanatory variables are similar to those reported in column (1). More importantly, SLOPE has a negative sign that is significant at the 0.05 level. The estimation results for SLOPE are consistent with those reported in prior research (Cao et al., 2010). 19 [insert Table 4 about here] 18 I also estimate equation (1) using three alternative measures of implied volatility: (i) implied volatility calculated using data from the five UK-based banks on the LIBOR panel (which account for 73.2% of the daily implied volatility estimates on individual options); (ii) implied volatility calculated from all put option on LIBOR panel banks regardless of their term to expiration and moneyness; and (iii) the VIX index which represents a weighted average of the implied volatility of near-the-money options on the S&P500 index. Estimation results are broadly the same. 19 However Ericcson et al. (2009) do not find that SLOPE is significant when included in the structural model. 23

24 One econometric issue that has been ignored to date is that of potential multicollinearity between the explanatory variables of the structural model. This is likely to be more of a concern when the data is measured in levels form. In view of the high correlations between the structural model explanatory variables evident in Table 3 I conduct variance inflation factor (VIF) analysis. In turn each explanatory variable is regressed on all other explanatory variables and the variance inflation factor computed. A common rule of thumb is that multicollinearity could be influencing OLS estimates if the VIF exceeds 10 (Neter et al., 1989). STK, SLOPE and RATE have VIFs of 15.93, and respectively, whilst IVOL has a VIF of This suggests that STK, SLOPE and RATE should be excluded from the regression model but this is not feasible for STK and RATE since both variables are key elements of the structural model and are not easily replaced. The second econometric issue is that of the stationarity of the data. I conduct an Augmented Dickey-Fuller test for a unit root in CRD. The test statistics show that the null hypothesis of a unit root cannot be rejected. When I test for a unit root in the first difference of CRD the null hypothesis of a unit root is rejected. These results suggest that the regression model should be estimated in first differences. The structural model is re-estimated using first differences of weekly average data. The results reported in columns (3) and (4) of Table 4 are weaker than reported for levels data. The estimates of IVOL and RATE have the same signs as reported in columns (1) and (2) and are still significantly different from zero. The estimates of STK are still negative but are no longer significantly different from zero. The estimate of SLOPE is negative and significant. The explanatory power of the regression model is also much reduced when measured in first difference form. This is not surprising given the amount of noise in the data. However it should also be borne in mind that the explanatory power of the structural model estimated in first differences is marginally above the level of 20% 25% reported by Ericcson et al. 24

25 (2009). Overall, the estimation results using first differences are broadly supportive of the predictions of the structural model. The estimation results also lend credibility to the results of the decomposition analysis. 3.3 The liquidity component Descriptive statistics and correlations Table 5 reports summary data on LIQ, BAS and DLRS over the sample period. LIQ has a mean (median) of 31 (8) basis points and ranges between a maximum of 342 and a minimum of -13 basis points. BAS has a mean (median) of 0.149% (0.104%) and ranges between 0.717% and 0.0%. DLRS has a mean (median) of 7.8 (7.0) and ranges between 23 and 0. The wide swings in LIQ, BAS and DLRS show that large changes in liquidity occurred over the course of the global financial crisis. [insert Table 5 about here] Table 6 reports the correlations between LIQ, BAS and DLRS over the sample period. LIQ is positively correlated with BAS and DLRS. BAS and DLRS are negatively correlated, consistent with prior research in OTC markets (Huang and Masulis, 1999). All correlations are significant at the level. The high positive correlation between LIQ and BAS is expected. The positive correlation between LIQ and DLRS suggests that funding pressures in wholesale financial markets are associated with more dealers posting quotes in the interbank market. The relatively low (but significant) correlation between BAS and DLRS suggests that both these proxies for liquidity can be included as explanatory variables in the same regression model without any concerns about potential multicollinearity. [insert Table 6 about here] 25