The federal funds network and monetary policy transmission: Evidence from the 2007–2009 financial crisis

The federal funds network and monetary policy transmission: Evidence from the 2007–2009 financial crisis

The Federal Funds Network and Monetary Policy Transmission: Evidence from the 2007-2009 Financial Crisis Journal Pre-proof The Federal Funds Network...

1015KB Sizes 0 Downloads 8 Views

The Federal Funds Network and Monetary Policy Transmission: Evidence from the 2007-2009 Financial Crisis

Journal Pre-proof

The Federal Funds Network and Monetary Policy Transmission: Evidence from the 2007-2009 Financial Crisis Daniel O. Beltran, Valentin Bolotnyy, Elizabeth Klee PII: DOI: Reference:

S0304-3932(19)30219-3 https://doi.org/10.1016/j.jmoneco.2019.12.006 MONEC 3208

To appear in:

Journal of Monetary Economics

Received date: Revised date: Accepted date:

5 February 2015 18 December 2019 18 December 2019

Please cite this article as: Daniel O. Beltran, Valentin Bolotnyy, Elizabeth Klee, The Federal Funds Network and Monetary Policy Transmission: Evidence from the 2007-2009 Financial Crisis, Journal of Monetary Economics (2019), doi: https://doi.org/10.1016/j.jmoneco.2019.12.006

This is a PDF file of an article that has undergone enhancements after acceptance, such as the addition of a cover page and metadata, and formatting for readability, but it is not yet the definitive version of record. This version will undergo additional copyediting, typesetting and review before it is published in its final form, but we are providing this version to give early visibility of the article. Please note that, during the production process, errors may be discovered which could affect the content, and all legal disclaimers that apply to the journal pertain. © 2019 Published by Elsevier B.V.

Highlights • In summer 2007, liquidity shocks led to hoarding and fed funds lending declined • After Lehman Brothers, counterparty credit concerns drove fed funds network dynamics • Early-crisis monetary policy easing surprises dampened hoarding and credit concerns • Less complete fed funds network and abundant reserves muted monetary policy transmission

1

The Federal Funds Network and Monetary Policy Transmission: Evidence from the 2007-2009 Financial Crisis✩ Daniel O. Beltrana,1 , Valentin Bolotnyyb,2 , Elizabeth Kleea,∗ a Federal

Reserve Board of Governors, 20th and C St. NW, Washington, DC 20551 b Hoover Institution, Stanford University

Abstract Using a network approach, we show how the federal funds market was transformed through the financial crisis through the collapse of the ABCP market in 2007, changes in monetary policy implementation, and an increase in counterparty credit risk. For both aggregate and bank-level network metrics, we find that increases in counterparty and liquidity risk are associated with reduced lending activity within the network. We also provide evidence that network peer effects are strong and influence banks’ holdings of reserve balances and rates paid in the federal funds market. Finally, we document how these changes to the network structure dampened the transmission of monetary policy. Keywords: Networks, Fedwire, federal funds, interbank lending JEL Classification: G2, E5

1. Introduction The federal funds market has traditionally been the cornerstone of U.S. monetary policy.3 Prior to the 2007-2009 financial crisis, banks traded scarce, non-interest earning reserve balances in order to meet reserve requirements and offset payment shocks. The Federal Reserve conducted open market operations so that this trading occurred at rates near the Fed’s target policy rate. As such, the federal funds rate—the average rate traded in the federal funds market– influenced banks’ funding costs, and in turn, interest rates on longer-term loans. This chain of events is central to traditional monetary policy transmission. Indeed, changes in the federal funds rate are typically thought to transmit to broader financial markets, thereby influencing financial conditions so that the Federal Reserve can achieve its goals of maximum employment and price stability. The federal funds market was hit by three major shocks during the 2007-2009 financial crisis. The first shock came in August of 2007, when global banks sponsoring asset-backed commercial paper (ABCP) conduits faced a liquidity squeeze, which strained interbank lending markets globally.4 The second shock was the dramatic change in Federal Reserve monetary policy implementation in response to the financial crisis. To provide liquidity and support the economy, the Fed established a number of emergency lending facilities, pumped trillions of dollars in reserve ✩ We appreciate comments from Ricardo Reis (the editor), three anonymous referees, Neil Ericsson, Ruth Judson, Michiel de Pooter, and Serafin Martinez-Jaramillo. We also thank conference participants from: 20th Annual Conference on Computing in Economics and Finance in Oslo, 13th Annual Conference of Financial Stability and Banking of the Banco Central do Brasil, 14th Oxmetrics User Conference in Washington, D.C. ∗ Corresponding author

Email addresses: [email protected] (Daniel O. Beltran), [email protected] (Valentin Bolotnyy), [email protected] (Elizabeth Klee)

1 The views in this paper are the responsibility of the authors and should not be interpreted as the views of the Board of Governors, other members of its staff, or the Federal Reserve System. 2 Bolotnyy acknowledges support from the National Science Foundation and the Paul & Daisy Soros Fellowship for New Americans.

3 Federal funds transactions are unsecured loans of reserve balances at Federal Reserve Banks between depository institutions and certain other institutions, including government-sponsored enterprises (GSEs). The vast majority of transactions are spot and the duration is typically overnight. See the online appendix for additional detail. 4 Ashcraft et al. (2011) find evidence of precautionary holdings of reserves during the crisis, especially by banks sponsoring ABCP conduits that faced a liquidity squeeze.

Preprint submitted to Elsevier

December 31, 2019

balances into the banking system, lowered the fed funds target rate to near zero, and began paying interest on required and excess reserves (IOER) to depository institutions—but not to the government-sponsored enterprises (GSEs). The third shock was a massive confidence shock from the failure of Bear Stearns in March 2008, and of Lehman Brothers later that year. Counterparty credit risk intensified during this period, which greatly exacerbated stresses in global financial markets.5 This paper examines how these three shocks impacted the network structure of borrowing and lending activity in the federal funds market, and in turn, the transmission of monetary policy. The analysis is carried out in two stages. In the first stage, we divide the network into components based on borrowing and lending relationships (Section 2), and document how the structure of the network changed during the financial crisis (Section 3). We pay special attention to the role of liquidity shocks, counterparty risk, and excess reserves (Section 4). We further explore how changes in the network structure affect the transmission of monetary policy. The second stage of the analysis uses detailed micro data on bank borrowing in the federal funds market to examine how network and peer effects interact with holdings of reserve balances and rates paid in the federal funds market. We also explore how interactions of these effects with monetary policy surprises influence fed funds market outcomes (Section 5). We find that: • The federal funds network started unraveling in the summer of 2007, when global banks experienced severe funding stress stemming from troubles in the ABCP market. This stress made banks reluctant to lend, and most lenders exited the market.6 • After the collapse of Lehman Brothers, counterparty credit concerns became a dominant factor driving network dynamics. Heightened credit risk dampened overall activity in the market, reduced intermediation through the network, and made the network less complete. In addition, if a bank had less creditworthy peers, it had fewer counterparties, regardless of its own solvency. • As the system became awash in liquidity, banks borrowed from fewer counterparties and held more reserves. These outcomes were amplified by peer effects in liquidity hoarding. • IOER-earning banks pulled back from lending federal funds, leaving the GSEs to represent most of the lending. All else equal, this shift led to larger loans, fewer counterparties and lower rates. • Dispersion in rates fell through the crisis, as only the best credits remained in the market. • Early-crisis monetary policy easing surprises dampened peer hoarding effects and credit concerns, so that borrowers had more lenders, and borrowed at lower rates. • The less complete network and abundant reserves dampened the transmission of monetary policy to other shortterm interest rates. 2. Data and network components This section describes how we construct the daily data, and how we divide the network into components based on borrowing and lending relationships. 5 Measures of counterparty credit risk, such as spreads on credit default swap (CDS) contracts of large global banks surged following both events. Afonso et al. (2011) find that, following Lehman Brothers’ bankruptcy, banks with higher levels of non-performing loans experienced sharper increases in borrowing spreads in the fed funds market and a drop in loan amounts, suggesting that counterparty risk played an important role during the crisis. Williams and Taylor (2009) show that the spread between interest rates on term lending and overnight federal funds lending began to grow in August 2007, which they attribute in part to heightened counterparty risk. 6 These results are consistent with findings in Afonso et al. (2011) and Ashcraft et al. (2011).

3

2.1. Data We use proprietary transaction-level data from the Fedwire funds transfer service. Similar data and matching algorithm were first used by Furfine (1999) and since then by Ashcraft and Duffie (2007), Afonso et al. (2011), and Kuo et al. (2013), among others.7 We capture monetary policy surprises using the change in the 6-month ahead futures contract on 3-month Eurodollar deposits during a 20-minute window surrounding each FOMC announcement on FOMC meeting days, as in Kuttner (2001) and Gertler and Karadi (2015), for example. We then combine these data with information on daily market rates, including the effective federal funds rate and Treasury general collateral finance (GCF) repurchase agreement (repo) rates. We control for liquidity needs using the (daily) change in the log amount of ABCP outstanding, the level of excess reserve balances, and the log aggregate payments volume in the Fedwire system (excluding federal funds transactions identified by our algorithm), which proxy for payment shocks.8 We measure counterparty risk using adjusted spreads on CDS of large US banks. Specifically, we first take the average of the spread on the 5-year CDS contract for selected large domestic banks that had liquid CDS contracts for the entire sample period.9 We then isolate the counterparty credit risk portion of the spread.10 Specifically, we extract the component of the CDS spread that is orthogonal to ABCP outstanding by regressing the CDS spread on ABCP outstanding and retaining the residuals.11 We estimate our models on daily data from January 3, 2005 to August 31, 2012. The length of this sample period provides roughly the same amount of time before the advent of interest on reserves as after.12 We use the Bai and Perron (2003) algorithm for simultaneous estimation of multiple breakpoints to test for series breaks in the logarithm of fed funds dollar volume, measured at the daily frequency.13 This test reveals a break on June 29, 2007, just before the rating agencies Moody’s and S&P downgraded hundreds of U.S. subprime residential mortgage backed securities (Moody’s Investor Service, 2007; Standard & Poor’s, 2007). Another break occurs on December 5, 2008, a period of heightened investor concerns about the health of financial institutions. Using these breakpoints, we create an “Early” crisis dummy (7/2/2007-12/5/2008) and a “Late” crisis dummy (12/8/2008-8/31/2012), and interact these dummies with a time trend and with explanatory variables of interest. For our daily panel analysis, we use disaggregated versions of the data discussed above. For the network variables, we include both the number of lenders (nodes) of a borrower and the daily amount borrowed from these lenders. As indicators of the monetary policy implementation regime, we use the level of reserve balances held by borrowers.14 Finally, we use weighted average rates paid by borrowers in the federal funds market. We also use quarterly bank Call Report data and daily bank reserves-holding data. We control for bank size using the log of total assets for the relevant quarter in the Call Report.15 In addition, to capture the effect of counterparty credit risk, we incorporate proxies for borrower risk characteristics. As in Laeven and Levine (2009), we use bank 7 Further

details about our matching algorithm and its performance can be found in the online appendix. on ABCP outstanding are from the Federal Reserve, Commercial Paper Data, available at https://www.federalreserve.gov/ releases/cp/. 9 We chose the 5-year tenor because it is the most liquid tenor (CDS price data are missing over many consecutive weeks for shorter-dated CDS contracts). Nearly half of the trading volume in single-name CDS occurs in the 5-year tenor (Chen et al., 2011). For robustness, we also use a spread averaged over a broader sample of banks (including those whose CDS contracts are less liquid), and the 1-year CDS contract; our results are broadly unchanged. 10 During the crisis, counterparty credit risk and liquidity risk became closely intertwined, as uncertainty surrounding the quality of assets on banks’ balance sheets (e.g., their holdings of asset-backed securities backed by U.S. subprime mortgages) made banks more reluctant to lend to each other. The resulting instability in interbank funding markets, in turn, made it more likely that a bank could lose access to funding and consequently default on its financial obligations. This increase in the probability of default would again make banks more hesitant to lend, reinforcing the cycle. Consistent with this cycle, Acharya and Merrouche (2012) find that after August 2007, large U.K. settlement banks with high counterparty risk saw the largest increase in demand for liquidity, especially on days with high payment activity. 11 This regression has an adjusted R2 of 0.42. The coefficient on log ABCP outstanding is negative with a t-statistic of -37.2, indicating that reductions in ABCP were associated with increases in CDS spreads. 12 Interest on reserves was implemented on October 9, 2008; the before- and after-IOER samples are each a little over 3 1/2 years long. 13 The test is based on a regression of log fed funds dollar volume on a constant, trend, average of 5 year CDS spreads, aggregate payments volume in the Fedwire system, and the spread between the effective and target federal funds rate. These breakpoints are fairly robust to the regression specification. 14 The distributions of the daily volume of funds and the level of reserve balances exhibit considerable skew; consequently, we use the natural logarithm of these variables in the analysis. 15 Results are little changed if the previous quarter Call Report variables are used. 8 Data

4

ROA +K/A

q q z-scores, defined as z = σ(ROA) . ROAq is the bank’s return on assets in quarter q, K/Aq is the bank’s capital-to-asset ratio in quarter q, and σ(ROA) is the standard deviation of the bank’s annual return on assets from 1999 to 2012. The z-score is a measure of the inverse probability of solvency that captures this pattern; a higher z-score suggests a greater likelihood of solvency. Because the z-score for our population is highly skewed, we use the natural logarithm of this measure in our calculations.16 Finally, we also include indicators on jurisdictions and institution type. Specifically, we include indicator variables for whether the borrower is a domestic bank, and also include the share of a borrower’s transactions where the lender is a GSE.

2.2. Network components The federal funds network describes a collection of nodes (banks), and the links (federal funds transactions) between them. As in Bech and Atalay (2010), we model daily federal funds flows as a directed network linking the lender to the borrower using network topology measures described in Soramaki et al. (2006). We group nodes into components based on how they interact with other nodes on any given day. Banks may switch groups from day to day, depending on their borrowing and lending relationships with other banks. The top panel of Figure 1 illustrates the components of our network. The giant strongly connected component (GSCC) is defined as the largest set of nodes that can reach each other via directed paths (Jackson (2008)). For example, if bank A lends to bank B, and bank B lends to bank C, then one can trace a directed path from bank A to bank C. Hereafter, we refer to the GSCC group as “the core” of the network. Banks in the core are engaged in both borrowing and lending, amongst each other and with banks in the giant in-component (GIN) and giant out-component (GOUT). Banks (and other nonbank entities with Federal Reserve accounts, such as the GSEs) in the GIN group have directed paths to the core, but not from it. That is, they lend federal funds to the core, but they do not borrow from the core. The opposite is true for banks in the GOUT group: they borrow from the core, but do not lend to it. Tendrils are in the “suburbs” of the network because they have no directed paths to the core—they typically either borrow from GIN, lend to GOUT, or both. The GWCC (giant weakly connected component) is the union of the core, GIN, GOUT, and tendrils.17 Nodes can also be grouped into communities. A community is a set of nodes that have many edges within it, and few edges outside of it. We identify such communities using the spin-glass algorithm, which optimizes an energy function.18 3. A network view of the fed funds market during the 2007-2009 financial crisis As shown by the blue line in the top panel of Figure 2, federal funds volume remained robust even through Lehman Brothers’ bankruptcy in September 2008, consistent with Afonso et al. (2011). However, this finding masks important changes in the network. Following the collapse of Bear Stearns in March 2008, we find that the number of links in the federal funds network fell notably (purple line in bottom panel). Subsequently, the number of links plummeted after Lehman’s bankruptcy, as fewer banks transacted with each other. As many of the smaller banks exited the market, the number of nodes declined (red line in top panel), while the average transaction size surged (not shown).19 Most lenders exit the market The top panel of Figure 3 shows the evolution of the size (number of nodes) of the various components of the federal funds network. The size of the network was fairly stable between 2000 and 2007, with GIN being the largest 16 Market-based measures such as the CDS spread of an institution may be a more direct way to elicit investors’ beliefs about the solvency of an institution. A limit of the CDS measure is that the universe of banks that had CDS traded over our sample period is small relative to the universe of banks identified as participating in overnight unsecured funding markets. 17 The few nodes which are outside the GWCC, but which connect to each other via undirected paths, are the disconnected components (DCs). We ignore them because they represent a negligible fraction of the overall federal funds transaction volume. 18 We implement this method using the cluster-spinglass function provided by the igraph package in R Reichardt and Bornholdt (2006), with the number of spins (maximum number of communities to search for) set to 15, and γ = 1 (γ governs the balance between the importance of present and non-present edges in a community). 19 Transcripts of FOMC meetings at the time reveal that Federal Reserve officials were well aware of the deteriorating conditions in interbank markets, despite steady federal funds volume. (Board of Governors of the Federal Reserve System, 2008).

5

component. The GIN component began to shrink in the summer of 2007, while the GOUT component began to expand. This increase in the ratio of borrowers to lenders suggests increased demand for short-term funding at a time when global banks began to experience severe funding stress from their inability to roll over maturing ABCP liabilities. The GIN component shrank rapidly following the bankruptcy of Lehman Brothers in September 2008. This trend accelerated throughout the fall of 2008, after reserve balances began to rise and the Federal Reserve started to pay interest on reserves. By the end of 2008, half of the institutions that had participated in mid-2007 had left the network, and the GIN component was just a quarter of its pre-crisis size. The remaining GIN lenders participated less frequently in the market. The green line in the bottom panel shows one measure of connectedness, the degree of completeness, defined as the number of links over the number of possible links.20 Interestingly, the federal funds network became more complete starting in late 2008. Even though the number of links fell sharply, the decline in the number of participants (nodes) in the network lowered the number of possible links. Core hollows out The bottom panel of Figure 3 shows the quarterly averages of the shares of daily dollar volume between the key components in the network. Between mid-2007 and the end of 2009, the core of the network diminished in importance, as lenders in the GIN component began to lend directly to borrowers in the GOUT component or in the tendrils, bypassing the core. Also, lending between banks in the core declined considerably. The bottom-left panel of Figure 1 shows the federal funds network on September 15, 2008, the day Lehman Brothers filed for bankruptcy. Despite the turbulence in financial markets that day, the network structure looked fairly typical. By December 2008, however, there were considerably fewer nodes, particularly in the GIN and in the core, as well as fewer links (right panel). During this transition period in late 2008, the structure of the federal funds network changed dramatically from day to day.21 Communities shrink The peak of the crisis took its toll on trading relationships. Throughout our sample period, the number of communities identified by the spin-glass algorithm is stable at around 10. However, the size of the largest community shrinks from about 140 in early 2007, to just 40 in 2009, mostly due to a sharp decline in the fall of 2008. Meanwhile, on average, borrowers and lenders are identified as being in the same community in 50 percent of the observations; in the other 50 percent, borrowers and lenders are in different communities. We use the heterogeneity in community size and membership to identify peer effects in our panel regressions. 4. Network activity, monetary policy transmission, and the crisis: Aggregate analysis This section explores the drivers of activity in the fed funds network, and the interaction of the network with monetary policy transmission. 4.1. How did the network respond to the crisis? We estimate the following regression: yt =α + ρyt−1 + β0 Xt + γ0 (Xt ∗ Earlyt ) + η0 (Xt ∗ Latet ) + δ1 Earlyt + δ2 Latet + δ3 T rendt + δ5 (T rendt ∗ Earlyt ) + δ6 (T rendt ∗ Latet ) + ξ0 Z t + t ,

(1)

where yt is one of our network measures of interest. Xt are the key counterparty risk and liquidity controls that are interacted with the “Early” and “Late” dummies. Z t comprises other explanatory variables that are not interacted with the “Early” and “Late” dummies. They include the spread between the effective federal funds rate and the target federal funds rate (lagged one day), the spread between the effective fed funds rate and the overnight repo rate (lagged one day), S&P500 log trading volume, monetary policy surprises, and controls for known calendar effects in the m a directed network, the degree of completeness is given by α = n(n−1) , where m is the number of links and n is the number of nodes. visual demonstration is available in the animated version of the network plot in the online appendix.

20 For 21 A

6

federal funds market: maintenance period day, FOMC announcement day, day before and after a holiday, the first and last day of each month, quarter-end, and year-end. We also include a constant, trend, and the lagged dependent variable to reduce autocorrelation in the residuals. The dependent variables for the aggregate measures of network activity are: log fed funds dollar volume, the number of links in the network, the degree of completeness, and the ratio of the number of borrowers to the number of lenders. The dependent variables for the network components are: log dollar volume from GIN to the core, average number of lenders that banks in the core borrow from, average number of lenders that banks in GOUT borrow from, and the size of the tendrils component. Table 1 shows regression results for aggregate measures of the federal funds network and its components. Four results stand out. Strains in the ABCP market spill over to the fed funds network Overall federal funds market volume is unaffected by ABCP issuance (column 1). However, the counterparty structure, as measured by the number of links, does shift (column 2). In particular, the negative coefficient on “Log change ABCP” indicates that before the crisis, increases in ABCP outstanding were associated with less participation in the federal funds market. Perhaps this reflected some substitution between the two markets for the subset of counterparties that had access to both. Similarly, in the pre-crisis period increases in ABCP outstanding were associated with declines in log dollar volume from GIN to the core (column 5), again suggesting that ABCP markets provided an alternate source of short-term funding for financial institutions. However, once troubles in the ABCP market hit, those strains influenced the federal funds market. That is, the positive coefficient on the interaction of ABCP and lending to the core with the “Early” and “Late” (column 2) indicate an overall pullback in the number of counterparties and in lending to the core on days when large global banks were unable to rollover their maturing ABCP liabilities. The magnitude of the coefficient implies that, during the early and late periods, a one-standard deviation drop in the amount of ABCP outstanding translates into a 5% decline in the number of links relative to the average number of links during this period. Heightened counterparty risk unravels the network late in the crisis Turning to the role of counterparty credit risk, the coefficient on the interaction term “Log CDS (residual) * Late” indicates that heightened counterparty risk damaged trading relationships and caused the network to unravel late in the crisis. In particular, after December 2008, increased counterparty risk reduced overall volume, number of links, degree of completeness, lending to the core, and the number of counterparties lending to the core and to GOUT components (columns 1-3, 5-7). Moreover, heightened counterparty risk is associated with borrowers scrambling for funds (column 4) and increased activity in the suburbs of the network, as institutions bypassed the core (column 8). For the number of links, a one-standard deviation increase in the the CDS spread during the late crisis period translates into a 5% drop in the number of links. Excess reserves dampen activity Pre-crisis, excess reserves were positively associated with the number of links in the network (column 2), and with the average in-degree of institutions both in the core and in the GOUT components (columns 6 and 7). During the crisis, however, the interaction terms with ‘Early’ and ‘Late’ indicate that higher levels of excess reserves were associated with banks borrowing from fewer counterparties (column 2), and with more borrowers than lenders, consistent with reserves hoarding (column 4). Furthermore, excess reserves no longer matter for determining the average number of lenders that banks in the core (and GOUT) borrow from (columns 6 and 7). The system became increasingly flush with liquidity and banks began hoarding reserves. This is consistent with the expansion of the Fed’s balance sheet shifting the market for reserves to a region where demand for reserves is saturated (and the demand curve is horizontal), as described for example, in Reis (2016). Easing monetary policy surprises ameliorate counterparty credit concerns Broadly speaking, monetary policy surprises have little overall effect on the volume in and structure of the federal funds network. However, one important exception is the number of links (column 2). During both the early and late stages of the crisis, an easing surprise is associated with a higher number of links in the network. Against a backdrop of declining ABCP and heightened counterparty risk, an easing surprise tended to bring more trading relationships back into the network. 7

4.2. How did changes in the network affect monetary policy transmission? We explore how monetary policy surprises (the change in the 6-month ahead futures contract on 3-month Eurodollar deposits surrounding FOMC announcements) are transmitted to the effective fed funds rate, 1-month Libor rate, and the rate on repurchase agreements by estimating the following regressions: ∆rt =α +

4 X i=1

(ρi ∆rt−i ) + β0 MPS t + β1 MPS t−1 + β2 MPS t−1 ∗ Latet

+ β0 (MPS t−1 ∗ Xt ) + η0 (MPS t−1 ∗ Xt ∗ Latet ) + ξ0 Xt + ω0 Z t + t ,

(2)

where ∆rt is the daily change in the rate (or spread) we are trying to explain, and MPS is the monetary policy surprise. To test whether the strength of the network can amplify or dampen this transmission, we interact the monetary policy surprises with the level of excess reserve balances, and with the degree of completeness in the network (both included in Xt ). We further interact these variables with the late period dummy. Z t includes variables that are not interacted with the monetary policy surprise: the change in excess reserve balances, and dummies for calendar effects (FOMC meeting day, maintenance period day, end-of-month, day after holiday, twenty-fifth day of the month, and year-end). The regression estimates showing the effects of tightening surprises on fed funds rate changes are shown in the first column of Table 2. Monetary policy tightening surprises have no significant effect on the change in the fed funds rate contemporaneously (row 1).22 However, a tightening surprise does significantly raise the fed funds rate with a one day lag (row 2). Transmission of monetary policy surprises weakened during the crisis, especially when the network was less complete, and when reserves were more abundant. The interaction with the late period dummy indicates much weaker transmission during the crisis. Prior to December 2008, a more complete network would dampen the transmission of monetary policy to the effective fed funds rate. However, this coefficient flips to being positive in the late period, indicating that during the crisis transmission was stronger when the network was more complete. Similarly, in the late period, transmission was weaker when reserves were more abundant. In column 2, we obtain very similar results for the transmission of monetary policy surprises to changes in the spread between the 1-month Libor and the 1-month OIS (Overnight Interest Swap rate). Column 3 examines the transmission to the repo rate, expressed as the spread over the target fed funds rate. The coefficient on monetary policy surprises is positive, but not statistically significant. However, as in Kroeger and Sarkar (2016), we also test for transmission from daily changes in the effective to target fed funds rate spread to the repo spread. The coefficient on the effective to target spread is positive, but not significant. In the late period, we again find evidence of weaker transmission, indicated by the negative and statistically signficant interaction term. Also, a higher degree of completeness seems to amplify transmission in the late period, whereas more abundant reserves seems to dampen this transmission. Together with the results from Table 1, which showed that increases in counterparty risk were associated with declines in the degree of completeness, these findings suggest that higher counterparty risk can dampen the transmission of monetary policy by altering the structure of the network. 5. Network dynamics, monetary policy transmission, and the crisis: Counterparty analysis We now use panel data on participants in the federal funds market to investigate the aggregate dynamics more closely. We first characterize how network and counterparty characteristics affect the market, and how the effect changed through the financial crisis. Next, we gauge how network peer effects influence reserve balance holdings and federal funds rates. Finally, we look at the interaction of both counterparty characteristics and peer effects with monetary policy shocks, to gauge how monetary policy transmission through the network changed over the course of the crisis. 22 While this may seem unusual, the effective federal funds rate was measured as the weighted average of all brokered trades in a day. Given that monetary policy announcements were made late in the day, the effect of the surprise is diluted by the existence of trades before the announcement.

8

5.1. How do counterparty characteristics affect the federal funds market? To assess the effect of counterparty characteristics on the federal funds market, we evaluate a baseline panel fixed-effects regression from the perspective of a federal funds market borrower, at a daily frequency: yit = β0 Bit + γGS Eit + µi + τt + it .

(3)

yit represents one of four outcomes: number of lenders to a borrower (number of nodes), (log) amount borrowed, (log) reserve balances held, and dollar-weighted average rate paid (expressed as a deviation from the target rate). Bit is a vector of time-varying characteristics of the borrower, and GS Eit is the borrower’s daily share of loans from a GSE. We also include borrower fixed effects, µi , as well as time fixed effects, τt .23 Summary statistics for the panel-level analysis are presented in table 3. Where appropriate, we interact parameters with date indicators to accommodate the structural breaks identified above. The standard errors are adjusted for covariances among it using a conventional panel-regression clustering approach to control for correlation across time and entities. Table 4 displays results from estimating equation 3. Three observations stand out. GSEs extend larger loans at lower rates, especially with the advent of IOER. Regardless of the time period, GSEs extended larger loans than other lenders (column 1, row 1). In addition, rates slipped below the target through the early and late stages of the crisis (column 4). This result confirms other observations, by Kim et al. (2018) and Bech and Klee (2011), that GSEs are willing to lend at lower rates than other institutions because they do not earn IOER. In contrast to volumes and rates, the proportion of loans from GSEs appears to have little correlation with either the number of lending nodes a borrower has (column 2), or with the level of reserve balances held by a borrower (column 3). Larger, domestic banks borrow more, have more counterparties, and command lower rates, particularly early in the crisis. Pre-crisis, larger institutions borrowed more, had more counterparties, held higher levels of reserve balances, and paid lower rates (columns 1 through 4, rows 2 and 3.) Our coefficient estimates imply that a bank in our sample at the top quartile by assets (about $88 billion in assets) paid rates that were roughly 4 basis points lower than banks at the lower quartile (about $8 billion). These advantages were even more pronounced in the early stages of the crisis, as counterparty credit concerns began to mount: a bank in the top asset quartile paid 8 basis points less than the lower quartile bank. Possible reasons for these advantages include that larger institutions may have had more bargaining power, been less risky, or viewed as less likely to default. Larger banks also held proportionally more reserve balances later in the crisis, suggesting elevated liquidity demands from more complicated institutions (column 3). Borrowing banks were hurt by peer credit concerns. Using the communities identified in section 3, we define the borrower’s peers as the other (bank) lenders in the community and examine contagion effects from counterparty credit concerns. Specifically, we use the average z-score for the lenders in a borrower’s community and test for the significance on outcomes. All else equal, credit concerns for peers led banks to pay higher rates in normal times (column 4, row 5), and to have fewer counterparties in the crisis (column 2). Specifically, if the average z-score in the community fell by one standard deviation, then the number of available nodes declined by about 10 percent, and rates paid rose a little. Interestingly, as measured by bank z-scores, there is little effect of a bank’s own solvency on the number of counterparties or rates paid (columns 1 and 4, row 4). 5.2. How do peer effects impact monetary policy outcomes? Because the federal funds market exhibits network characteristics, it may be subject to peer effects, where a borrower’s community affects its own choices. However, estimating the impact of the peer group on an individual is complicated by the individual’s membership in the peer group. This “reflection problem” has three elements (Manski, 1993): (1) the endogenous peer effect, or the effect of peers on an individual’s outcome; (2) the exogenous contextual 23 The time fixed effects reflect calendar effects and year-quarter controls. Pre-crisis, maintenance period, and payment flow dates generated predictable changes in federal funds market conditions.

9

effect, or the effect of individual characteristics on the outcome; and (3) the correlated effect, or the effect of the common environment on both the individual and the peers. During the financial crisis, it is possible that all three of these factors are evident. We use two-stage least squares methods developed in Bramoulle et al. (2009) and Giorgi et al. (2010) to control for endogeneity and to identify peer effects.24 Concretely, suppose we want to know how average rates paid in a community affect an individual borrower’s rate. To control for the reflection bias, we instrument the average rate charged by lenders in the borrower’s community with the average rate paid by other borrowers in the lenders’ communities. This technique will yield an unbiased estimate of the peer effect if community size varies across nodes, and borrower and lender communities do not completely overlap. As discussed above, our communities satisfy these requirements. Furthermore, because we have a sufficiently long time series and a large number of observations, we can take additional precautions, such as using lags of our endogenous variables as controls and using one or two lags as instruments.25, 26 Using the Manski (1993) linear-in-means model, we evaluate yi,t = µyC−i,t−1 + β0 BC−i,t−1 + ζzC−i,t + δ0 Bi,t−1 + γGS Eit + αi + τt + it where Ci,t is borrower i’s community on date t, P P j∈Ci,t y j,t j∈Ci,t B j,t yC−i,t = , BC−i,t = , nCi,t − 1 nCi,t − 1

zC−i,t =

P

j∈Ci,t zk, j,t

nCi,t − 1

(4)

(5)

and j , i. We view µ as capturing the endogenous peer effect, β as the endogenous correlated effect, and ζ as the exogenous contextual effect. They are the estimated coefficients on the “leave-out” means of the characteristics of a borrower’s community; that is, means for members of the community other than the borrower. Similarly, to form instruments for borrower peer effects and correlated effects, we again use “leave-out” means—but this time, for the communities of the lender. Against this backdrop, our first stage specification is: yC−i,t−1 = η + θ P

1 k (nCk

− 1)

X k

(nCk − 1)yC−k,t−1 + µi,t−1

(6)

where yC−k,t−1 is the “leave-out” average outcome for lender k’s community. We use instruments for both the peer effects and the contextual effects, but do not use instruments for the plausibly exogenous contextual effects. We also use lags of both the borrower and lender leave-out means to further control for any endogeneity problems. F-statistics reject the null of weak instruments. To allow coefficients to vary across monetary policy implementation regimes, we provide separate estimates for each of the sample periods identified in section 4. Table 5 displays results from estimating equation 4. Three observations stand out. Peer effects boosted liquidity hoarding. In the early stages of the crisis, peer effects strongly influenced holdings of reserve balances—if a bank’s community members held high levels of reserve balances, the borrowing bank did as well (column 2, row 1). But this also meant that lenders were less willing to supply funds: higher levels of reserve balances were associated with lower levels of lending from the network (column 2, row 3). Still, banks that held high levels of reserve balances continued to borrow, even if it was from fewer counterparties (column 2, rows 7 and 8). Banks paid higher rates to borrow more and from more counterparties. Amid liquidity hoarding, equilibrium rates in the federal funds market rose when banks borrowed more (column 5, row 7). Banks paid on average an extra 6 basis points to increase funds borrowing by one standard deviation, an 24 Silva

(2017) uses a similar strategy. authors including Angrist (2014) are skeptical regarding the appropriateness of the Bramoulle et al. (2009) instruments. 26 Jackson (2008) and others suggest that using the lag of the network effects should help ameliorate endogeneity problems. Other research, including Leary and Roberts (2014), recommend using the contemporaneous factors as a way to control for potential reactions to previous outcomes. The results that follow are robust to either assumption; both qualitatively and quantitatively, parameter estimates are little changed. 25 Some

10

economically meaningful amount, during the early stages of the crisis. The network effect of borrowing from more counterparties also caused rates to rise (column 5, row 4). Interestingly, in the early stages of the crisis, individually, if a bank was able to obtain a range of counterparties, equilibrium rates were lower (column 5, row 8). However, the network effect outweighed the individual effect, boosting rates overall. Fed funds rates converged, and only the best borrowers remained in the market. Over time, rates paid in the fed funds market became more and more a function of the rates paid by peers (row 2, columns 4 through 6). Pre-crisis, if peers paid on average 10 basis points higher for funds, an individual borrower might only pay about 4 basis points higher. By the late crisis period, rates moved nearly in tandem, with the coefficient on the deviation from the target approaching 1. Importantly, none of the network effects nor the contextual effects appear to significantly predict rates paid in the late crisis period (column 6). Part of the explanation for these results may be the hollowing out of the network. As this occurred, only the best borrowers remained in the market.27 As such, the amount borrowed, number of nodes, and individual counterparty risk ceased to influence rates paid. 5.3. How do monetary policy surprises interact with the network? Table 6 displays the results of the same specification used in equation 4, but supplements it with interactions of the key network variables with the monetary policy surprises used in section 4. The results in the top part of the table (rows 1 through 6) are similar to those in Table 5. Looking at the rows in the bottom half of the table, three results stand out. Easing surprises dampened peer hoarding effects. In the early part of the crisis, the median monetary policy surprise was roughly -2.5 basis points. Coupled with the coefficient on peer effects for reserve balances (column 2, row 8), an easing surprise dampened the peer effect of hoarding reserve balances. The coefficient implies that in very rough terms, a typical easing surprise was associated with an average of $1 billion less in reserve balances. This is consistent with the general idea that central banks can ease crisis strains either by providing more reserve balances or lowering rates. In addition, increased borrowing in the network is associated with fewer reserve balances (column 2, row 3), but this effect is dampened by easing surprises (column 2, row 10). Easing surprises led borrowers to retain more lenders, and borrow at lower rates. In the early part of the crisis, counterparty credit concerns were one of the factors that prompted lenders to trim their “acceptable” borrower lists. For the borrowers that remained, rates were generally higher (column 5, rows 4 and 6). The interaction of the easing surprise with the network effect of the number of nodes (column 5, row 11) and borrower effect of the number of nodes (column 5, row 13), suggest that, on net, an easing surprise is associated with a slight increase in the number of lenders, leading to a slight decrease in rates paid, relative to the baseline effect. Taken together, assuming a 2.5 basis point easing surprise that passes through to federal funds rates completely, banks would be able to capture at least one more lender–an increase from 7 to 8 lenders, at the median number of lenders for the early crisis period. This effect is also present in the late crisis period (column 6, row 11), although to a lesser degree. Dispersion in rates fell through the crisis. Pre-crisis, peer effects interacted with monetary policy shocks led to a dampening of pass-through (column 4, row 9). For example, an easing shock presumably passes through to lower federal funds rates. However, the negative coefficient on the peer effect for deviation from the target has the opposite sign, suggesting that the collective action of the peers dampens the shock. As the crisis wore on, this effect subsided and rates in the federal funds market generally converged (columns 5 and 6, rows 2 and 9). While a monetary policy surprise may have led to a general reshuffling of positions and rates pre-crisis, this effect was less pronounced early and late crisis. One reason for this effect is likely the observed general homogeneity of borrowers and lenders remaining by the late crisis, as discussed above. More broadly, other reasons could plausibly be the increase in the level of reserve balances more generally, the effect of the zero lower bound on rates, or the introduction of IOER. 27 In fact, average z-scores for borrowers went up in the late crisis period, with the mean z-score in the late crisis closer to the 75th percentile of the overall distribution.

11

6. Conclusion We use a network approach to better understand how changes in monetary policy implementation and counterparty credit risk interacted to shape monetary policy transmission in the wake of the financial crisis. The federal funds market provides a good laboratory for our analysis because the fed funds market is an unsecured overnight market, and it is one of the first steps in the transmission of monetary policy. We find evidence that monetary policy surprises result in an increase in fed funds activity, and in the number of lenders relative to borrowers. Furthermore, the transmission of monetary policy is stronger when the network is more complete, and is weaker when the system is saturated with reserves. At the bank level, peer effects are strong and influence the propagation of these factors through the network. References References Acharya, V. V., Merrouche, O., 2012. Precautionary hoarding of liquidity and interbank markets: Evidence from the subprime crisis. Review of Finance, rfs022. Afonso, G. M., Kovner, A., Schoar, A., 08 2011. Stressed, Not Frozen: The Federal Funds Market in the Financial Crisis. Journal of Finance 66 (4), 1109–1139. Angrist, J., 2014. The perils of peer effects. Labour Economics 30, 98–108. Ashcraft, A., McAndrews, J., Skeie, D., October 2011. Precautionary Reserves and the Interbank Market. Journal of Money, Credit and Banking 43, 311–348. Ashcraft, A. B., Duffie, D., May 2007. Systemic illiquidity in the federal funds market. American Economic Review 97 (2), 221–225. Bai, J., Perron, P., 2003. Computation and analysis of multiple structural change models. Journal of Applied Econometrics 18 (1), 1–22. Bech, M. L., Atalay, E., 2010. The topology of the federal funds market. Physica A: Statistical Mechanics and its Applications 389 (22), 5223–5246. Bech, M. L., Klee, E., 2011. The mechanics of a graceful exit: Interest on reserves and segmentation in the federal funds market. Journal of Monetary Economics 58 (5), 415–431. Board of Governors of the Federal Reserve System, 2008. Conference call of the federal open market committee on october 7, 2008. http: //www.federalreserve.gov/monetarypolicy/files/FOMC20081007confcall.pdf. Bramoulle, Y., Djebbari, H., Fortin, B., 2009. Identification of Peer Effects through Social Networks. Journal of Econometrics 150 (3), 41–55. Chen, K., Fleming, M. J., Jackson, J., Li, A., Sarkar, A., 2011. An analysis of cds transactions: Implications for public reporting. Staff Reports 517, Federal Reserve Bank of New York. Furfine, C., 1999. The Microstructure of the Federal Funds Market. Financial Markets, Institutions, and Instruments 8 (5), 24–44. Gertler, M., Karadi, P., January 2015. Monetary policy surprises, credit costs, and economic activity. American Economic Journal: Macroeconomics 7 (1), 44–76. Giorgi, G. D., Pellizzari, M., Redaelli, S., April 2010. Identification of Social Interactions through Partially Overlapping Peer Groups. American Economic Journal: Applied Economics 2 (2), 241–275. Jackson, M. O., 2008. Social and Economic Networks. Princeton University Press, Princeton, NJ, USA. Kim, K., Martin, A., Nosal, E., December 2018. Can the us interbank market be revived? FEDS working paper 2018-088, Federal Reserve Board. Kroeger, A., Sarkar, A., 06 2016. Monetary Policy Transmission before and after the Crisis. Liberty street economics (blog), Federal Reserve Bank of New York. URL http://libertystreeteconomics.newyorkfed.org/2016/06/monetary-policy-transmission-before-and-after-the-crisis. html Kuo, D., Skeie, D. R., Vickery, J., Youle, T., March 2013. Identifying term interbank loans from Fedwire payments data. Staff Reports 603, Federal Reserve Bank of New York. URL https://ideas.repec.org/p/fip/fednsr/603.html Kuttner, K., 2001. Monetary policy surprises and interest rates: Evidence from the fed funds futures market. Journal of Monetary Economics 47 (3), 523–544. Laeven, L., Levine, R., August 2009. Bank governance, regulation and risk taking. Journal of Financial Economics 93 (2), 259–275. Leary, M. T., Roberts, M. R., February 2014. Do peer firms affect corporate financial policy? Journal of Finance 69 (1), 139–178. Manski, C., 1993. Identification of endogenous social effects: The reflection problem. The Review of Economic Studies 93 (3), 531–542. Moody’s Investor Service, 2007. Moody’s downgrades subprime first-lien RMBS. Jul 10, 2007. Newey, W. K., West, K. D., May 1987. A Simple, Positive Semi-definite, Heteroskedasticity and Autocorrelation Consistent Covariance Matrix. Econometrica 55 (3), 703–708. Reichardt, J., Bornholdt, S., Jul 2006. Statistical mechanics of community detection. Phys. Rev. E 74, 016110. Reis, R., Sep. 2016. Funding Quantitative Easing to Target Inflation. Discussion Papers 1626, Centre for Macroeconomics (CFM). URL https://ideas.repec.org/p/cfm/wpaper/1626.html Silva, A., November 2017. Strategic liquidity mismatch and financial sector stability. Working paper, Cass Business School. Soramaki, K., Bech, M. L., Arnold, J., Glass, R. J., Beyeler, W., 03 2006. The Topology of Interbank Payment Flows. Staff Reports 243, Federal Reserve Bank of New York. URL http://ideas.repec.org/p/fip/fednsr/243.html

12

Standard & Poor’s, 2007. 612 u.s. subprime rmbs classes put on negative watch; methodology revisions announced. Ratings Announcement: July 10, 2007. Williams, J. C., Taylor, J. B., January 2009. A Black Swan in the Money Market. American Economic Journal: Macroeconomics 1 (1), 58–83.

13

14

Aggregate measures (2) (3) Number Degree of links completeness -5034.6∗∗∗ 0.16159∗∗ (1734.7) (0.07225) 5364.2∗∗∗ -0.18662∗∗ (1848.2) (0.08237) 4786.9∗∗∗ -0.14003 (1751.7) (0.09925) 23.577 0.00062 (18.533) (0.00081) ∗ 45.557 -0.00066 (23.591) (0.00089) ∗ -35.586 -0.0035∗∗∗ (19.343) (0.00125) 3.5928∗∗∗ -0.00005 (0.98573) (0.00005) ∗∗∗ -4.0195 0.00007 (0.9836) (0.00005) ∗∗∗ -3.6628 0.00007 (0.98522) (0.00005) ∗∗ 762.1 -0.0281∗ (319.65) (0.01584) -599.26∗ 0.02214 (355.23) (0.01674) ∗∗ -739.64 0.02009 (327.86) (0.02069) 0.98 0.97 (4) N.borrowers/ N.lenders -0.74814 (1.6685) 1.8632 (1.7839) 1.6887 (1.7856) -0.00045 (0.0121) -0.03199∗∗ (0.0159) 0.11343∗∗∗ (0.01943) -0.00399∗∗∗ (0.00083) 0.00418∗∗∗ (0.00084) 0.0039∗∗∗ (0.00083) 0.0098 (0.28753) -0.17102 (0.337) -0.47602 (0.30852) 0.94

(5) Log dollar GIN to core -9.5163∗∗∗ (3.6159) 11.819∗∗∗ (4.1554) 7.0083∗ (3.8126) 0.11612∗∗∗ (0.03088) -0.04446 (0.04063) -0.31019∗∗∗ (0.04418) 0.00205 (0.00165) -0.00206 (0.00165) -0.00203 (0.00166) 0.67686 (0.4528) -1.131∗ (0.62554) 0.32262 (0.50043) 0.6

Network components (6) (7) Avg. in-degree Avg. in-degree GSCC GOUT -115.41 46.713 (79.857) (47.371) 161.97∗∗ -31.645 (81.968) (49.316) 96.016 -47.426 (80.524) (48.134) 0.96763 0.33469 (0.71655) (0.39815) 0.69942 0.27509 (0.84908) (0.53978) ∗∗∗ -2.0296 -1.4978∗∗∗ (0.73472) (0.43609) 0.0933∗∗∗ 0.04729∗∗ (0.02851) (0.02073) -0.09487∗∗∗ -0.04727∗∗ (0.02854) (0.02075) -0.09386∗∗∗ -0.047∗∗ (0.02851) (0.02074) 6.4264 -9.958 (7.023) (8.9913) -5.8008 15.64∗ (9.1951) (9.4509) -4.0798 13.212 (7.2309) (9.2583) 0.85 0.58 (8) Size Tendrils 184.96 (161.55) -152.17 (181.98) -241.84 (166.12) -1.9832 (1.2409) 3.1328 (1.9512) 10.188∗∗∗ (1.6002) 0.00893 (0.08135) 0.00289 (0.08151) -0.00831 (0.08154) -37.986 (55.006) 41.446 (55.979) 56.681 (57.142) 0.5

Heteroscedasticity and autocorrelation consistent standard errors (Newey and West, 1987) in parentheses. *** denotes significance at α = 0.01, ** at α = 0.05, and * at α = 0.10. All specifications include a constant, early (7/2/2007-12/8/2008) and late (12/9/2008,8/31/2012) period dummies, lagged dependent variable, effective federal funds to repo rate spread (also interacted with early and late dummies), trend (also interacted with early and late dummies), effective federal funds to target rate spread, log dollar volume transacted in the S&P500, and dummies for: maintenance period day, day before/after holiday, first/middle/last day of month, last day of quarter, last day of year, and FOMC announcement day. Estimation sample is from 1/4/2005-8/31/2012 (1927 observations).

Adj. R-squared

Late

Early

Six month eurodollar

Late

Early

Excess reserves

Late

Early

Log CDS (residuals)

Late

Early

Log change ABCP

(1) Log dollar volume -2.1139 (1.7651) 0.77929 (2.2361) 1.0367 (2.0809) 0.01158 (0.01904) 0.02193 (0.02379) -0.10027∗∗∗ (0.02417) 0.00111 (0.00084) -0.0011 (0.00085) -0.00105 (0.00085) 0.51617 (0.36387) -0.48031 (0.42399) -0.08915 (0.38024) 0.8

Table 1: The fed funds network during the crisis

Table 2: Transmission of monetary policy

Six month eurodollar Lagged six month eurodollar Late Lagged six month eurodollar * deg.complt. Late Lagged six month eurodollar * ex.res. Late

(1) ∆ eff.fed funds rate 0.13401 (0.41563) 6.6926∗ (3.4394) -7.2029∗∗ (3.4401) -465.27∗ (256.32) 485.77∗ (255.31) 0.02174∗ (0.01226) -0.02251∗ (0.01221)

(2) ∆ 1m.Libor OIS spread -0.62335∗∗∗ (0.2012) 5.7403∗∗∗ (1.998) -2.7813 (2.077) -392.16∗∗∗ (144.25) 360.73∗∗ (145.8) 0.01705∗∗ (0.0067) -0.01723∗∗ (0.00676)

∆ (eff.-target ff) spread Late ∆ (eff.-target ff) spread * deg.complt. Late ∆ (eff.-target ff) spread * ex.res. Late Adj. R-squared

0.15

0.21

(3) ∆ repo target spread 0.49812 (0.80939) 0.00163 (0.77626) -0.00898 (0.79384)

1.101 (0.75322) -2.6058∗∗∗ (0.79843) -51.923 (46.582) 89.32∗ (46.053) 0.0037∗∗∗ (0.00131) -0.00414∗ (0.00213) 0.3

Heteroscedasticity and autocorrelation consistent standard errors (Newey and West, 1987) in parentheses. *** denotes significance at α = 0.01, ** at α = 0.05, and * at α = 0.10. All specifications include a constant, 4 lags of the dependent variable, degree of completeness, level of excess reserves, late period dummy, as well as dummies for calendar effects: FOMC meeting day, maintenance period day, end-of-month, day after holiday, twenty-fifth day of the month, and year-end. Estimation sample is from 1/6/2005-8/31/2012 (1925 observations).

15

Table 3: Summary statistics

Variable

N

Mean

Std. Dev.

p25

p50

p75

Network characteristics Number of nodes Log(amount)

82,205 82,205

14.47 3.84

19.30 0.88

4 3.35

8 3.84

18 4.38

Reserves and rates Log(reserve balances) Deviation from target

80,680 82,205

10.63 -0.04

3.08 0.18

8.70 -0.09

10.24 -0.02

12.24 0.03

Borrower characteristics Log(assets) 81,538 Log(z-score) 82,205

17.16 3.09

1.73 0.72

15.93 2.53

17.18 3.14

18.29 3.59

Peer effects Number of nodes Log(amount) Log(reserve balances) Deviation from target Log(z-score)

23.66 17.92 8.81 -0.03 2.52

18.22 0.54 1.56 0.18 0.32

12.40 17.65 7.77 -0.08 2.34

19.04 17.78 8.30 -0.01 2.52

26.02 18.08 9.64 0.03 2.70

82,205 82,205 82,180 82,205 82,192

Observations are borrower-date. Estimation sample is from 1/3/2005 to 8/31/2012. Log(amount) is log($ millions), log(reserve balances) is log($ thousands), deviation from target is percent, log(assets) is log($ thousands).

16

Table 4: The network, reserve balances, and rates

Lender characteristics 1. GSE Early Late Borrower characteristics 2. Domestic Early Late 3. Log(assets) Early Late 4. Log(z-score) Early Late Contextual characteristics 5. Log(z-score) Early Late 6. Constant Early Late N Number of borrowers R2 Within Between Overall

(1)

(2)

(3)

(4)

Log(amount)

Number of nodes

Log(reserve balances)

Deviation from target

1.589*** (0.223) 0.0585 (0.137) 0.161 (0.282)

-0.174 (0.240) -0.0760 (0.245) 0.315 (0.364)

0.295 (0.403) -0.532 (0.327) -0.239 (0.780)

0.0251 (0.0143) -0.0795** (0.0279) -0.105*** (0.0198)

0.179* (0.0901) 0.350*** (0.0644) -0.157 (0.103) 0.197** (0.0620) -0.00813 (0.0175) 0.117* (0.0571) -0.0517 (0.0662) -0.0309 (0.0416) 0.0680 (0.100)

0.206 (0.179) 0.275* (0.120) -0.213 (0.192) 0.237*** (0.0639) -0.0538 (0.0286) -0.211*** (0.0418) -0.0480 (0.114) 0.0979 (0.0606) -0.0289 (0.148)

-0.0666 (0.106) 0.375 (0.217) 0.532** (0.180) 0.0772* (0.0344) 0.594*** (0.149) 0.0217 (0.274) -0.0773 (0.104) 0.00794 (0.405)

-0.00439 (0.0153) -0.0814** (0.0251) -0.0101 (0.0182) -0.0110* (0.00473) -0.0258*** (0.00485) 0.00135 (0.00260) 0.0102 (0.00526) -0.0130 (0.0101) -0.00661 (0.00717)

0.0228 (0.0511) -0.00179 (0.0517) -0.160 (0.0849)

0.0228 (0.0846) 0.111 (0.0820) 0.339* (0.155)

-0.153 (0.106) -0.115 (0.123) -0.179 (0.214)

-0.0135** (0.00476) -0.00969 (0.0123) 0.0156 (0.00826)

-0.636 (0.761) 2.041 (1.216) 81,514 144

0.828 (3.371) 0.982 (0.866) -7.621* (3.317) 80,202 150

0.198* (0.0825) 0.284** (0.0849) -0.0902 (0.0573) 81,525 155

0.515 0.598 0.648

0.401 0.245 0.420

0.0542 (0.964) -0.304 (0.303) -1.816 (0.970) 81,525 155 0.249 0.581 0.465

0.697

This table displays results from estimating equation 3. The dependent variable is listed at the top of each column and the controls are listed as rows. The ”Early” and ”Late” rows indicate the estimated coefficient when the control variable is interacted with a dummy variable indicating the early crisis subsample and the late crisis subsample. Each uses a fixed effects panel regression framework, except for column (2), which uses a Poisson panel fixed effects framework. All specifications include borrower fixed effects, year-quarter fixed effects, and calendar effects. Standard errors are clustered at the borrower-date level.∗ significant at the 5 percent level; ∗∗ significant at the 1 percent level; ∗∗∗ significant at the 0.01 percent level.

17

Table 5: Peer effects—Reserve balances and rates Log(reserve balances)

Peer effects 1. Log(reserve balances)t−1

Deviation from target

Base (1)

Early (2)

Late (3)

0.0790 (0.113)

0.953*** (0.171)

0.416* (0.189)

2. Deviation from targett−1 Aggregate network effects 3. Log(amount)t−1 4. Number of nodest−1 Contextual characteristics 5. Log(z-score) Lender characteristics 6. GSE Borrower characteristics 7. Log(amount)t−1 8. Number of nodest−1 9. Log(assets) 10. Log(z-score) 11. Constant N adj. R-sq Cragg-Donald Wald F-stat Kleinbergen-Paap rk Wald F-stat

Base (4)

Early (5)

Late (6)

0.354*** (0.0183)

0.606*** (0.0216)

0.920*** (0.0133)

-0.110 (0.150) 0.00779* (0.00332)

-1.459*** (0.429) 0.0363* (0.0161)

0.156 (0.512) -0.0655 (0.118)

0.0132*** (0.00208) -0.000801*** (0.000237)

-0.00519 (0.0264) 0.00588*** (0.00173)

-0.000574 (0.00564) 0.00112 (0.00101)

0.0586 (0.0551)

-0.0213 (0.131)

-0.222* (0.111)

-0.0144*** (0.00301)

0.00801 (0.0114)

0.00181 (0.00152)

0.304 (0.173)

0.105 (0.224)

-0.719 (0.517)

0.0173 (0.0108)

-0.0560* (0.0285)

-0.0145** (0.00463)

0.0790 (0.0464) -0.00555* (0.00281) 0.903*** (0.102) 0.00547 (0.0738) -5.036 (2.594) 23,904 0.076 109.355 52.057

0.895*** (0.165) -0.0228* (0.00902) 0.650*** (0.186) -0.118 (0.0759) 12.96 (7.593) 10,296 0.3130 23.156 20.698

-0.361 (0.199) 0.00450 (0.0442) 0.141 (0.233) 0.265 (0.790) 5.212 (7.592) 13,665 0.0814 20.706 14.628

0.00564** (0.00193) 0.000482*** (0.000111) -0.00821*** (0.00201) 0.00175 (0.00247) -0.0591 (0.0521) 24,576 0.4462 71.969 50.073

0.0687*** (0.0194) -0.00220* (0.000938) -0.0140 (0.00955) 0.00752 (0.00697) -0.216 (0.477) 10,368 0.5840 21.683 18.969

-0.000852 (0.00256) -0.000600 (0.000500) -0.00189 (0.00119) -0.000398 (0.00390) 0.0280 (0.0836) 13,664 0.9085 28.089 19.276

This table displays results from estimating equation 4. The dependent variable is listed at the top of each column and the controls are listed as rows. The “Base”, “Early” and “Late” columns indicate estimated parameters on the respective subsample. Each specification uses a two-stage instrumental variable fixed effects panel regression framework. The “Peer effects”, “Network effects”, and the first two borrower characteristics are instrumented as described in the main text; all other control variables are considered exogenous. All specifications include borrower fixed effects, year-quarter fixed effects, and calendar effects. Standard errors are clustered at the borrower-date level.∗ significant at the 5 percent level; ∗∗ significant at the 1 percent level; ∗∗∗ significant at the 0.01 percent level.

18

Table 6: Monetary policy surprises, the network, and the market Log(reserve balances)

Peer effects 1. Log(reserve balances)t−1

Deviation from target

Base (1)

Early (2)

Late (3)

0.0786 (0.112)

0.948*** (0.170)

0.429 (0.222)

2. Deviation from targett−1 Aggregate network characteristics 3. Log(amount)t−1 -0.108 (0.148) 0.00768* 4. Number of nodest−1 (0.00332) Borrower characteristics 5. Log(amount)t−1 6. Number of nodest−1

0.0788 (0.0464) -0.00543* (0.00277)

Interactions with monetary policy surprises 7. Six-month eurodollar 50.57 (50.67) Peer effects 8. Log(reserve balances)t−1 1.394 (1.234) 9. Deviation from targett−1 Network effects 10. Log(amount)t−1 11. Number of nodest−1 Borrower characteristics 12. Log(amount)t−1 13. Number of nodest−1 14. Constant N adj. R-sq

Base (4)

Early (5)

Late (6)

0.391*** (0.0204)

0.608*** (0.0225)

0.920*** (0.0134)

-1.470*** (0.417) 0.0386* (0.0161)

0.187 (0.589) -0.0762 (0.153)

0.0124*** (0.00201) -0.000788*** (0.000237)

-0.00519 (0.0261) 0.00603*** (0.00175)

-0.000952 (0.00564) 0.00120 (0.00102)

0.907*** (0.165) -0.0234* (0.00914)

-0.380 (0.250) 0.00819 (0.0562)

0.00535** (0.00181) 0.000460*** (0.000111)

0.0688*** (0.0199) -0.00230* (0.000955)

-0.000733 (0.00255) -0.000626 (0.000501)

206.5** (72.21)

6.144 (98.34)

0.0978 (4.988)

-2.106 (5.311)

-0.955 (1.342)

9.769*** (2.279)

-0.972 (2.513) -4.692*** (0.770)

-0.202 (0.860)

-0.501 (0.394)

-3.732 (3.380) -0.0706 (0.0649)

-16.97*** (5.062) 0.220 (0.126)

-0.0376 (7.407) -0.270 (1.295)

-0.0251 (0.298) -0.00338 (0.00430)

0.00588 (0.336) 0.0495** (0.0163)

0.0355 (0.0827) 0.0255* (0.0129)

1.377 (1.374) 0.127** (0.0423) -5.059* (2.571) 23,904 0.076

3.620* (1.691) -0.145* (0.0659) 13.17 (7.416) 10,296 0.3142

0.0648 (2.756) 0.386 (0.521) 4.751 (8.581) 13,665 0.0811

0.0642 (0.110) -0.00230 (0.00201) -0.0544 (0.0503) 17,401 0.4470

-0.257 (0.274) -0.0265** (0.0101) -0.213 (0.471) 10,368 0.5923

-0.000885 (0.0472) -0.00791 (0.00558) 0.0334 (0.0835) 13,664 0.9083

This table displays results from estimating equation 4, and adds interactions with six-month Eurodollar surprises to the specification. The dependent variable is listed at the top of each column and the controls are listed as rows. The “Base”, “Early” and “Late” columns indicate estimated parameters on the respective subsample. Each specification uses a two-stage instrumental variable fixed effects panel regression framework. The “Peer effects”, “Network effects”, and the first two borrower characteristics are instrumented as described in the main text; all other control variables are considered exogenous. All specifications include borrower fixed effects, year-quarter fixed effects, and calendar effects. Standard errors are clustered at the borrower-date level.∗ significant at the 5 percent level; ∗∗ significant at the 1 percent level; ∗∗∗ significant at the 0.01 percent level.

19

GWCC 7HQGULOV IURP*,1

7HQGULOV ERWK

DC

GSCC "core" GIN

GOUT

7HQGULOV WR*287

7HQGULOV UHPRWH

Figure 1: Components of the Federal funds network Top: Figure adapted from Soramaki et al. (2006). Bottom-left: Network on September 15, 2008. Bottom-right: Network on December 8, 2008. Arrows denote direction of lending.

20

Exhibit 1

2-14-2019

Network statistics Volume and number of nodes 800 700

Number

Daily

Volume (right scale) Number of nodes (left scale)

Bear Stearns

Lehman

Billions of dollars

350 300

600

250

500

200

400

150

300

100

200

50

100

2000

2001

2002

2003

2004

2005

2006

2007

2008

2009

2010

2011

Number of links and degree of completeness 1750 1500

Number

Bear Stearns

Daily

2012

Number Lehman

1250 1000 750

Number of links (left scale) Degree of completeness (right scale)

500 250 0

2000

2001

2002

2003

2004

2005

2006

2007

2008

2009

2010

2011

2012

0

0.15 0.14 0.13 0.12 0.11 0.10 0.09 0.08 0.07 0.06 0.05 0.04 0.03 0.02 0.01 0.00

Figure 2: Measures of Federal funds network activity Thick lines denote 1-month moving averages. “Bear Stearns” line corresponds to March 17, 2008, when it was acquired by J.P. Morgan. “Lehman” line corresponds to September 15, 2008, when it filed for bankruptcy. Source: Authors’ calculations using Fedwire data.

21

Exhibit 1

2-21-2019

Size of Network Fed Funds Network Component Size

Number of nodes Bear Stearns

Quarterly average

Lehman

700 600

Total nodes GSCC GOUT Tendrils GIN

500 400 300 200 100

2000 2001 2002 2003 2004 2005 2006 2007 2008 2009 2010 2011 2012

Fed Funds Volume Shares by Network Component Quarterly average

GIN --> GSCC GSCC <--> GSCC GSCC --> GOUT GIN --> GOUT Other*

Bear Stearns

Lehman

Percent

0

70 60 50 40 30 20 10

2000 2001 2002 2003 2004 2005 2006 2007 2008 2009 2010 2011 2012 * Other comprises transactions involving tendrils.

0

Figure 3: Fed funds network component size and volume share * Other comprises transactions involving tendrils and disconnected components. “Bear Stearns” line corresponds to March 17, 2008, when it was acquired by J.P. Morgan. “Lehman” line corresponds to September 15, 2008, when it filed for bankruptcy. Source: Authors’ calculations using Fedwire data.

22

Credit Author Statement Daniel Beltran Conceptualization, Formal analysis, Writing - Original Draft Valentin Bolotnyy – Conceptualization, Formal analysis, Writing - Original Draft Elizabeth Klee – Conceptualization, Formal analysis, Writing - Original Draft

23