the full story of runs - fdic: federal deposit insurance ... › bank › analytical › cfr ›...
TRANSCRIPT
The Full Story of Runs∗
Jun Kyung Auh† and Hayong Yun‡
July 2017
ABSTRACT
We examine the dynamics of credit supply during a repo run. Upon the receipt of an initialbad signal, lenders relax their requirements to help the borrower avoid default. As morenegative news arrives, the lenders’ patience is exhausted and they rapidly reduce the creditsupply until the borrower defaults. Lenders with a greater exposure and a closer lendingrelationship have a longer period of patience, and their credit subsequently contracts at amore gradual pace. The initial patience of lenders is consistent with a feedback channelbetween endogenously-chosen collateral requirements and lenders’ decisions about collateralwhen supplying credit. Substantial collateral provides lenders a higher recovery rate uponliquidation, but it simultaneously incentivizes lenders to abandon their borrower prema-turely. The feedback exacerbates a run, as lenders further tighten collateral requirementsupon a weak signal about borrowers’ fundamentals. Lenders factor in this destabilizing roleof collateral when setting ex-ante collateral requirements. Our findings suggest a novel roleof collateral: a medium of lender (mis)coordination. They also highlight the importance oflenders’ incentives in collateralized loans.
Keywords: Bank Run, Collateral Requirement, Lender Coordination, Lending Relationship
JEL classification: D86, G33, G34, K22.
∗We thank Jennie Bai, Sandeep Dahiya, Charles Hadlock, Naveen Khanna, and seminar participants at GeorgetownUniversity, and Michigan State University for helpful comments; and Hojong Shin and Chris Yun for data assistance.All errors are our own. Please address correspondence to the authors via email.†Finance Department, McDonough School of Business, Georgetown University. 37th and O Streets NW, Hariri
583, Washington, DC 20037. Phone: 202-687-2783. Email: [email protected].‡Department of Finance, Eli Broad College of Business, Michigan State University, 645 North Shaw Lane, Room
339, East Lansing, MI 48824, Phone: 517-884-0549. Email: [email protected].
How do lenders behave (e.g., adjust their prices and non-price contractual features) following a
sequence of negative signals about their borrower’s ability to repay the debt? Especially, when the
loan is secured by specific collateral, how do lenders dynamically adjust their credit supply before
a run is triggered? Several episodes in the repurchase agreement (repo) market from the 2008-2009
financial crisis call for a better understanding of these questions. However, due to a lack of available
data, little has been documented on the ex-ante dynamic behavior of lenders with collateralized
loans. In this paper, we fill this gap by examining lenders’ decisions on credit contraction during a
period of borrower stress, using micro-level data from repo contracts. Repo is a type of short-term
collateralized borrowing, and is a dominant source of short-term liquidity for financial and non-
financial firms. In addition to its systematic importance, repo contracts provide an ideal laboratory
in which to study lenders’ behavior because of its short-term nature of repo: lenders make rollover
decisions and reset contract terms at a high frequently.1
Economists pay attention to a potential destabilizing mechanism of the collateral requirement in
a repo contract: the haircut, or the degree of overcollateralization. Lenders use the haircut to
adjust their credit supply and risk exposure. Several studies show that raising haircut during
stress periods can destabilize the credit market empirically (Copeland et al., 2014; Gorton and
Metrick, 2012; Krishnamurthy, Nagel, and Orlov, 2014), as well as theoretically (Martin, Skeie, and
Thadden, 2014a,b). These works document a sharp increase in average aggregate haircuts during
the 2008-2009 crisis, which eventually led to the collapse of the repo market.2 These run studies
mostly focus on aggregate post-run behavior (or the onset of runs). However, these papers do not
consider an important ex-ante aspect of runs: what do lenders do when faced with a possible run in
the near future? Lenders could immediately jump to a run, or they could try to diffuse a potential
run by providing financial slack, such as relaxed covenants or reduced collateral requirements.
Moreover, such decisions on whether to run or to attempt to rescue borrowers could be driven
by lenders’ heterogeneous incentives. Our study focuses on dynamic lender behavior the pre-run
period, especially the impact of lender incentives on their credit contract decisions.1Prior to the 2008-2009 crisis, aggregate outstanding repo activities were estimated to be around $5 trillion to
$10 trillion (Copeland, Martin, and Walker, 2014; Gorton and Metrick, 2012). Currently, they are estimated to bearound $5 trillion (Baklanova, Copeland, and McCaughrin, 2015).
2For example, Figure 4 in Gorton and Metrick (2012) shows a steep increase in the repo-haircut index (equallyweighted average haircut) during 2007-2009.
1
Using daily repo contract data for a large hedge fund (we refer to it as Fund X hereafter) that
eventually collapsed, we find that haircut trends exhibited a U-shaped pattern during the three-
month period prior to the end of this fund’s life. Despite the negative news regarding the fund, we
observe an initial decrease in haircuts: at its lowest, the reduction in the average haircut was around
2% (out of an average haircut of 3.7%). However, as negative news continue to accumulate, the
pattern flipped: haircuts began to increase rapidly, which eventually triggered the fund’s failure.
Our results contrast with the monotonic increase in haircuts in the U.S. repo market in the 2007-
2009 period (e.g., Gorton and Metrick (2012)). This increase in aggregate average haircuts reflect
both changes in the composition of collateral (e.g., the liquidation of high quality assets associated
with low haircuts during adverse market conditions) and the tightening of lending conditions (e.g.,
the increase in lenders’ risk aversion for the same collateral). To examine the latter (i.e., pure
haircut changes for rolled-over loans), we focus on micro (lender and loan) level haircut changes in
a matched loan sequence.
The relaxation of credit supply restrictions upon receiving negative news about the borrower is con-
sistent with lender coordination to collectively avoid exacerbating runs. One of the key trade-offs
in the rollover decision is the payoff from forcing the borrower to foreclose the project versus the
payoff from taking a risk by rolling over and hoping for higher returns when the project succeeds.
That is, collateral provides an outside option for lenders when they make rollover decisions. Ex-
cessively high and liquid collateral, therefore, may cause runs even when economic conditions are
strong. Given this trade-off, lenders internalize the knowledge that a larger collateral size increases
the likelihood of a run, and use that knowledge when setting a collateral amount ex-ante.
We then examine whether there is variation in the timing of haircut changes across lenders, and
if this variation is explained by certain lender characteristics. Regarding decision timing, we focus
on two particular points: the moment when lenders start decreasing haircuts after negative news
on the borrower surfaces (Initial Response), and the duration of the decreased haircuts before the
pattern reverses (Lender Patience). We find that each lender significantly drops haircuts at different
times, and there is a substantial variation in the duration of the relaxed credit supply via lowered
haircut. Out of the sixteen lenders that Fund X had lending relationships with during our sample
period, eight had a U-shape haircut pattern during the last three months of this fund’s life. Four
2
lenders monotonically decreased haircuts till the end, and four did not significantly change their
haircuts. News that Fund X might be in trouble because of its sub-prime mortgage investments
began to spread approximately three months prior to the eventual shutdown of the fund. Lenders
start decreasing their haircut as early as 70 days prior to the collapse and as late as 32 days prior
to the collapse (excluding those who did not significantly change their haircuts). The duration of
Lender Patience ranged from 0 to 57 days.
To examine how the timing of these actions are associated with lender properties, we consider two
lender-specific characteristics: loan exposure to Fund X and the strength of the lending relationship
(measured by the number of days since the lender’s first trade with Fund X). These variables
measure lenders’ vested interest in a borrower’s financial health and its expected impact on lenders’
haircut decisions. In four different specifications (linear, Cox, Weibull, and accelerated failure time
(AFT) models), we find that loan exposure (measured by the log of the total loan amount) shortens
the Initial Response and lengthens Lender Patience. For example, in the linear specification, a one
standard deviation increase in loan exposure leads to a three week faster Initial Response and
two more weeks of Lender Patience. Similarly, the strength of the lender’s relationship with the
borrower tends to shorten the Initial Response delay and lengthen the duration of the haircut-
reducing period: a one standard deviation increase in the strength of the lending relationship leads
to one week faster Initial Response and one additional week of Lender Patience.
We further investigate variations in Initial Response and Lender Patience at the loan level, tracking
rollover points within each holding position. In addition to the two lender-level incentive variables
(lender-level principal and the strength of the lending relationship), we add loan-level characteris-
tics: (log of) the principal of each loan, and a short-term loan indicator (below median maturity).
Even at the loan level, the total principal aggregated at the lender level is significantly and neg-
atively related to Initial Response and significantly and positively related to Lender Patience.
However, at the loan level, the lender-level strength of relationship (while still the same sign as the
lender-level regressions) loses its statistical significance. Instead, we find that loan-level variables
are significantly related to Initial Response and Lender Patience. We find earlier reduction in hair-
cut and longer Lender Patience period when loan principals are larger and when loan maturities
are shorter. These findings are consistent with the notion that lenders with a strong incentive to
3
support the survival of the borrower (larger loans and a long relationship with borrower) are more
willing to give favorable lending terms to the borrower in the face of a possible run, up to a certain
point. They step in earlier and continue the rescue effort longer.
Finally, we show that collateral type also plays an important role in lenders’ actions. While the
dominant asset class in our sample is structured finance securities, government and corporate bonds
take some fraction of the collateral. Setting government bonds as a benchmark, we show that lenders
step in earlier and continue the rescue attempt longer when the collateral is from a less liquid asset
class, such as structured finance. Further, structure finance, the main asset class in the sample, can
be broken into two categories: collateralized debt obligation (CDO) and pass-through mortgage-
backed security (MBS). Due to additional layers of securitization, CDOs are typically more opaque
and illiquid than vanilla MBSs. Using this liquidity-related variation within the structured finance
category, we find a similar pattern: loans against CDOs are associated with longer Lender Patience
than those against MBSs. We repeat this analysis with the contemporaneous credit ratings of
each structured finance asset as of t0, and find that loans against lower ratings are associated with
greater rescuing attempts from lenders. These results consistently suggest that lenders who have
hard-to-sell assets as collateral are more incentivized to save the borrower by relaxing the credit
supply. In other words, if lenders’ loans have relatively liquid collateral, the collateral provides
better liquidation options, and they have less interest in the borrower’s survival.
Our paper relates to several strands of the literature and makes at least two new contributions.
First, this is the first paper to document non-monotonic (U-shaped haircut trends) behavior during
creditor runs. Prior studies on repo market failure document a unidirectional rundown deterioration
during the 2008-2009 financial crisis (Gorton and Metrick, 2012). Our work shows that at the
individual lender level, the “run” phenomenon is more subtle: upon the arrival of bad news about
a borrower, lenders initially decrease their haircut, hoping to revive the borrower. However, after
a period of patience with accumulating negative signals, lenders give up hope, and they rapidly
increase their haircuts until default. That initial period of credit relaxation and patience has never
before been documented in the repo contract literature.
Second, this paper identifies a distinct role of collateral that has not been studied in the extant lit-
erature. The conventional view of collateral is that it complements or substitutes for loan covenants
4
and thereby influences banks’ monitoring/screening incentives. Unlike conventional collateral of-
fered by industrial borrowers, collateral for the co-funding of financial assets is not indispensable to
the borrower’s whole production. Therefore, repo collateral provides an outside option for lenders
during rollovers, which can cause coordination failure among lenders. Despite the prominent role
of collateral (haircuts) in repo borrowing and the large impact of the repo market on financial
markets, the way in which lenders choose haircuts has not been extensively studied.3 Our study
provides new insight on lenders’ choice of the amount of collateral (e.g., haircuts in repo loans),
while accounting for its destabilizing impact during rollovers, i.e., the likelihood of repo runs.
The road map of the paper is as follows. Before our empirical analysis, Section I provides institu-
tional details for the repo market, a description of the extant literature on the subject, and state
hypotheses for the empirical tests. Section II describes the sample and variables used for this study.
Section III presents empirical findings on lender behavior via collateral requirements during periods
of disstress for the borrower. Section IV concludes.
I. Background
A. The Repo Market and Haircutss
Repo it the primary financing channel for financial institutions. It is composed of two inter-
connected markets: the tri-party and the bilateral repo markets. Through the tri-party repo
market, money market mutual funds (and security lenders) lend cash to dealer banks against
a particular set of collateral (typically safe assets such as U.S. Treasury securities). Using the
proceeds from the tri-party repo market, dealer banks make loans (in the bilateral repo market) to
hedge funds who use these funds to finance acquisition of risky assets. One of the distinguishing
feature between these two markets is the existence of a centralized clearing house: the tri-party
repo market consists of three parties (a lender, a borrower and a clearing house) whereas bilateral
repo market only involves a lender and a borrower without any clearing system.
Repo is a type of secured lending contract. A repo borrower borrows cash from a repo lender by3Exceptions are Martin et al. (2014a) and Martin et al. (2014b) who, in a dynamic general equilibrium model,
link collateral constraints to conditions on a run.
5
pledging collateral. At that time, the borrower promises to purchase back the collateral from the
lender (thus it is called a repurchase agreement or repo) at a predetermined price after a certain
time period. For example, suppose a lender borrows $90 by pledging an asset whose market value
is $100. The borrower makes a promise to buy back the asset by paying $90.45 after 1 month.
As seen in this illustration of a 1-month loan, the loan price can be calculated by the difference
between the lending amount ($90) and the promised amount to repurchase ($90.45), yielding 50
bps per month or a 6% per annum repo rate. Also, the value of the collateral at the time of loan
initiation ($100 in our example) is typically larger than the loan amount ($90). Such a difference
measures the degree of overcollateralization: we say the haircut is 10% ($100/$90-1) in this case.
As a repo contract rolls over, the lending terms can also dynamically change. In particular, the
lender can effectively run on the borrower by increasing the haircut. Deciding not to roll over is a
special case of making the haircut 100%.
In the middle of 2007, pressured by big losses in the subprime mortgage market, Fund X collapsed.
Table 1 summarizes the major news regarding Fund X during this period. About 90 days prior
to the failure, negative news on this fund started to emerge, and by about 45 days prior to the
collapse, the market became widely aware of the potential risk of the fragile subprime mortgage
market. In the last month of the fund’s life, negative news about it accumulated. In the final
week before the fund’s collapse, multiple news items appeared regarding the details of the fund’s
losses, lender negotiations, and restructuring plans. However, the hedge fund could not recover,
and collapsed in the middle of 2007.
[Place Table 1 about here]
B. Hypotheses
What is unique about this episode is that the overcollateralization (the haircut on the repo) ex-
hibited a non-monotonic change over time during the final days of the fund.4 Figure 1 shows the
changes in overall haircut (Panel A), spread (Panel B), and duration (Panel C) during the final4According to Table 2 of Gorton and Metrick (2012), the average haircut of the U.S. repo market was constant
during the first half of 2007, which coincides with our sample period.
6
three months of the fund. In the figure, each panel also displays the number of negative news items
about the fund (bar chart) and the outstanding total loan amount (dotted line) over the same
period. Panel A shows that the haircut significantly decreases in response to the initial negative
news. However, after a period with favorable margin requirements (the period of lender patience),
lenders rapidly increase the margin requirement until the borrower’s default, which results in a U-
shaped pattern for the haircut. This haircut raise is contemporaneously accompanied by complete
withdrawals of lending (rollover failures), as seen by the rapid drop in the loan balance. This kind
of loan withdrawal would correspond to a haircut of 100%, but would also disappear from the data.
Therefore, the increasing part of the U-shaped pattern is substantially understated. As opposed
to the clear pattern in the haircut, spread and duration do not show any comparable dynamics.
This stark contrast implies that, in a period of extreme borrower stress, lenders use this particular
non-price term, the haircut, rather than other lending terms, to manage their risk exposure.5
[Place Figure 1 about here]
The relaxation of credit supply restrictions through a reduction in haircut upon receiving negative
news about the borrower is consistent with lenders’ effort to avoid exacerbating runs.6 The amount
of collateral is a (mis)coordination device: when the liquidation of collateral (or declining to roll
over a loan) before the completion of an investment project is inefficient, greater collateral exacer-
bates the coordination failure and helps trigger a run. In other words, excessively strict collateral
requirements raise the risk of a coordination failure among lenders and a premature termination of
the loan.
The key trade-off in the rollover decision is the payoff from forcing the borrower to foreclose the
project versus the potential payoff from taking a risk by rolling the loan over and hoping for higher
returns when the project succeeds.7 In this setting, collateral provides an outside option for lenders
when they make roll over decisions. Excessively high collateral, therefore, may cause runs even5Prior works on credit rationing (e.g., Stiglitz and Weiss (1981)) also emphasize the prominent role of non-price
terms (i.e., credit rationing) when prices (i.e., interest rates) cannot fully resolve market imperfections.6The observed lender patience, as shown in Figure 1 is also consistent with lenders postponing a run (an increase
in haircut) in order to gather more information on the borrower’s bad news. This reason is not mutually exclusiveand can coexist with the collateral-lender coordination channel. However, this information acquisition channel cannotexplain why only the haircut (not prices nor maturity) is used for learning. The collateral-lender coordination channelcan provide a more direct explanation for why mainly haircut is adjusted during the lender patience period.
7A strategic-complementarity-based model to illustrate this tradeoff is provided in the Appendix.
7
when economic conditions are strong. Given this trade-off, lenders internalize the knowledge that
a greater collateral size increases the likelihood of a run, and use that knowledge when setting a
collateral amount ex-ante.
Because of the results showns in Figure 1, we focus on haircut dynamics in the following analysis.
Moreover, beyond this overall pattern, we examine how the timing and pace of the haircut changes
differ by lenders and collateral types.
Prior studies on the lending relationship, such as Petersen and Rajan (1994, 1995) show that lenders
with a vested interest in a borrower’s continuation tend to offer more favorable lending terms when
the borrower is in trouble. For this study, we measure lenders’ vested interest in the borrower using
their outstanding loan amount and the length of their lending relationship with the borrower. The
impact of lenders’ incentives on lender patience is summarized in the following hypothesis:
Hypothesis 1. Lenders with larger loans and a longer lending history with the borrower offer
more favorable lending terms when the borrower is in trouble.
The dominant asset types of collateral in repo contracts are financial securities, which are more eas-
ily separable from the borrower’s operation than non-financial assets. This characteristic contrasts
with corporate debt contracts (e.g., corporate bonds), which typically pledge indivisible production
assets as collateral. Seizing or auctioning them off can, therefore, lead to a disruption in the pro-
duction process and have a negative impact on the borrowing firm’s credit quality. Through this
channel, conventional collateral constrains creditors to stay with the firm and mitigates liquidation
bias. However, repo is secured by easily separable assets that could be sold without foreclosing the
whole operation. Together with the short-term nature of repo contracts, the separability impedes
coordination among lenders and makes repo collateral very different from corporate loan collateral.
While greater collateral provides increased recovery upon foreclosure, it makes it harder for lenders
to coordinate with each other to avoid premature foreclosure; in other words, the probability of a
self-fulfilling run increases with the size of the collateral. In this sense, repo collateral plays a role
similar to the one that credit default swap (CDS) plays in corporate debt contracts: it provides
outside options for lenders and thereby weakens incentives for lenders to stay with the borrower in
8
times of financial distress (Bolton and Oehmke, 2011). The impact of asset separability on lender
coordination (and lender patience) is summarized in the following hypothesis:
Hypothesis 2. The asset separability (liquidity) of collateral is inversely proportional to lender
patience.
In Section III, we explain what drives this heterogeneity in haircut dynamics across lenders upon
receiving negative news about the borrower, based on these two hypotheses.
C. Extant Literatures
Bank run, or running on the borrower, is one of the central themes in economics, especially in
the financial intermediation literature. Early theories describe the run as a bistable equilibrium
(Diamond and Dybvig, 1983). Recently, Acharya, Gale, and Yorulmazer (2011), Brunnermeier and
Oehmke (2013), Gorton and Ordonez (2014), He and Xiong (2012), and Bebchuk and Goldstein
(2011) make further theoretical contributions, with models that explain a run in a dynamic context.
Empirically, however, observing such dynamic runs at an individual loan level has been difficult due
to a scarcity of run events and data limitations. Several prior studies have made an attempt to do
this. Schmidt, Timmermann, and Wermers (2016) document a run in the money market. Iyer and
Puri (2014) show a link between geographic ties and runs in local banks. Hertzberg, Liberti, and
Paravisini (2011) consider lender coordination during credit registry expansions. Using repo data,
Copeland et al. (2014), Krishnamurthy et al. (2014), and Gorton and Metrick (2012) document
the repo run during the 2007 to 2008 crisis, showing that repo lenders actively increased the
haircut during the run. Most theories about and empirical evidence on runs describe a monotonic
relationship between borrowers’ creditworthiness and lending terms: deterioration in a borrowers’
financial health leads to rapidly worsening lending conditions (fewer loans and/or higher loan prices)
until the borrower collapses.
On the other hand, theories on collateral have mainly been motivated by the substitution effect
of lender screening (Bester, 1985; Inderst and Mueller, 2007; Rajan and Winton, 1995).8 In these8Prior literature also extensively studies other contractual features, such as maturity (Berglof and Thadden, 1994),
9
models, more collateral reduces the incentive for a lender to screen a borrower because the ex-ante
loss upon default is low when the loan is sufficiently collateralized. In relation to these two strands of
literature, this paper provides empirical evidences that are consistent with the interaction between
collateral and lenders’ run behavior from the perspective of strategic complementarity, which was
pioneered by Carlsson and van Damme (1993) and further developed by Morris and Shin (1998).9
II. Data
A. Sample Construction
We use proprietary data that consists of the complete repo position of multiple fixed-income hedge
funds. From 2004 to 2007, the funds actively traded mostly securitized bonds and other structured
finance securities, typically backed by underlying loans such as mortgages (mortgage backed secu-
rities or MBSes), commercial loans (collateralized loan obligations or CLOs) and other securitized
products themselves (collateralized debt obligation or CDO). The funds took leverage positions on
such assets by borrowing money from multiple dealer banks via secured loans, specifically bilateral
repo contracts. In other words, those funds were borrowers and the dealer banks were secured
lenders. They themselves were the in-house hedge funds of a large dealer bank, and they differ only
in leverage ratios. Therefore, we treat them as a single borrower without explicitly distinguishing
them. The raw data covers 2004 to 2007 and consists of 297,606 daily observations (16,807 repo
agreements, 54 lenders, and 1590 unique CUSIPs). While the full data spans the three-year time
period, we limit our focus to the last three-month period leading up to the collapse of the fund.
Hence, the data that we use in this paper is a subset of the data used by Auh and Landoni (2016),
except that we include loans from the affiliated lender that are excluded from their study because
they focus on interactions among contract terms in a general context. The selection of the three-
month study period is based on news flows. The negative news regarding this borrower started
and covenants (Rajan, 1992). Recent empirical studies on collateral include Calomiris, Larrain, Liberty, and Sturgess(2017), Campello and Larrain (2016), and Vig (2013).
9Other notable works on global games include Angeletos and Werning (2006), Angeletos, Hellwig, and Pavan(2007), Campos (2013), Chen, Goldstein, and Jiang (2010), Ciccone and Costain (2004), Dasgupta (2007), Frankel(2012), Frankel, Morris, and Pauzner (2003), Goldstein and Pauzner (2005), Guimaraes and Morris (2007), He andManela (2016), Izmalkov and Yildiz (2010), Morris and Shin (2006), and Sakovics and Steiner (2012).
10
appearing about 90 days prior to the fund’s collapse. In order to identify rollover contracts, we
follow a procedure proposed by Auh and Landoni (2016): We sequentially match contracts that
finance the same asset position, where the later contract is a roll-over of an earlier one. In addition,
we require the later contract to follow the earlier one in rapid succession (within one day). These
matched rollover contracts have the same lender and the fundamental loan risk at succession is
nearly identical (they are at most one day apart with the same borrower, lender, and underlying
collateral). The final sample leaves 584 loans from 16 lenders.10 We refer to Auh and Landoni
(2016) for detailed information regarding the general data structure, variables and construction
methods.
During the sample period, this fund took out repo loans from sixteen different lenders, including
the affiliated lender. On average, each lender made loans against 89 different items of collateral
to finance the fund’s investment positions. Each position is a set of consecutive individual loans
because, as a current loan matures, the fund rolls over the loan into the subsequent one to keep
its holding position. During the sample period, on average, each position consists of 5 loans, i.e.,
the fund rolled over each loan for 4 times on average, and the initial loan is also included in the
sample. While the loan amount changes over time, as of the start date of our sample, the average
loan size was $28.4 million per each position.
B. Description of Variables
The key variable of our study is the timing of events. The starting point of our sample is when the
first significant negative news item on this firm surfaces. As shown in Table 1, the first negative
news appeared 88 days before the collapse. Hence, we set out a reference point (t0) as the day three
months prior to the failure date. This date is common to all lenders. From this reference date,
the period of main interest includes moments that lenders actively decreased and subsequently
increased their haircuts. Our result shows that the earliest date of the initial haircut drop occurred
about two weeks after t0. We refer to all dates during our sample period as days relative to the
reference point t0.10The composition of collateral types is CDO (55.31%), corporate bond (12.84%), government security (2.05%),
MBS (13.18%), and other structured finance (16.61%).
11
There are three key event dates that characterize the run dynamics: the first date that each lender
decreases its haircuts after t0 is defined as t1; the date that each lender increases its haircuts after
t1 is defined as t2; and the date that loans by all lenders are terminated is defined as t3 (hence, t3
is common to all lenders). These four dates are illustrated in Figure 2.
[Place Figure 2 about here]
Table 2 shows event dates for each lender. These dates are computed based on the average daily
haircuts of each lender. To pin down the timing, we rely on the following procedures. For each day,
we compare the haircuts of new repo contracts with those of the maturing contracts that the new
contracts replace. This methodology has the merit of controlling for the potential effect of asset
composition changes on haircut dynamics. Under adverse financing conditions, a borrower may
liquidate high quality assets, which are likely to be associated with lower haircuts. In this case,
without any action by lenders, the haircuts would appear to increase, because the remaining assets
are financed with a higher haircut. Therefore, we need to be able to shut off this channel to observe
a pure haircut move as a consequence of the lender’s reaction. The initial response of a haircut
drop has great variation. t1 ranges from t0 + 19 at the earliest to t0 + 55 at the latest (excluding
no change case). The haircuts start increasing again after the dropping period: t2 occurs mostly
ranging from t0 + 55 to t0 + 68. Loan termination occurred mostly on the last day (t3), but some
lenders terminated as early as t0 + 74.
[Place Table 2 about here]
The lender-driven haircut dynamics during this period of borrower stress can be classified by four
patterns: (i) U-shaped (haircuts initially decrease and then subsequently increase), (ii) monotonic
decrease until the end, (iii) monotonic increase until the end, and (iv) no significant change. As
shown in Table 2, most lenders exhibit the U-shaped haircut pattern, while four lenders have no
change, and three lenders show a monotonic decrease in haircuts until the end of the sample period.
For the U-shaped pattern, starting from t0, all three dates t1, t2, and t3, are well defined. For a
monotonically decreasing pattern, t1 and t3 are well defined, and we set t2 = t3 (or t2 = t3−ε more
rigorously), implying that a lender attempts to increase their haircut in the last minute before the
12
fund’s collapse. For the pattern without a significant change in haircuts, we set t1 = t2 = t3 (or,
more rigorously, t1 = t2 − ε; t2 = t3 − ε), indicating that the lender attempts to decrease and then
immediately increase their haircuts, but it is too late, since by the time of the attempt, the loan is
terminated by the fund’s bankruptcy.
We now construct two key variables of interest corresponding to the timing variables: Initial Re-
sponse, which is defined as the time it takes for a lender to make a haircut drop, i.e., t1 − t0, and
Lender Patience, which is defined as the length of the period that a lender gives favorable lending
terms, i.e., t2 − t1. To explain variations in these two variables, we consider several lender-specific
variables. First, we measure lenders’ loan exposure to the borrower by taking the natural logarithm
of their total principal outstanding as of t0. Also, we measure the lending relationship between a
lender and the borrower using the natural logarithm of the number of days since the first loan to
the borrower by each lender (as of t0), assuming that if a lender made loans to the borrower for
a longer period, it has a closer relationship with the borrower (Petersen and Rajan (1994), and
Petersen and Rajan (1995)). Given that our original data starts with this fund’s inception point,
we have full information to calculate this variable.
Column V of Table 2 shows the pattern of Initial Response, sorted by the outstanding loan amount
as of t0. Lenders with a larger exposure to the borrower appear to have a smaller value for Initial
Response, indicating that they respond more quickly to negative news by decreasing their haircuts
at some point after t0. Column VI of Table 2 shows the pattern of Lender Patience, sorted by
outstanding loan size as of t0. Lenders with a greater loan exposure tend to have a larger value
for t2 − t1. This result implies that lenders with a greater loan exposure maintain their reduced
haircuts for longer before they give up and raise haircuts until the borrower fails.
Table 3 shows summary statistics on the variables considered in this paper. On average, lenders
waited for 6 weeks (42.5 days) following t0 before decreasing their haircuts. More than half of the
lenders waited 31 days before decreasing their haircuts. One reason for such a long delay is that we
set the time of the initial bad news outbreak conservatively, to account for potential private leaks
to lenders before the information appeared in news available to the public. Informal discussions
with repo market participants suggest that, a little before the first public news surfaced, market
participants were already speculating that the fund’s heavy exposure to subprime mortgages could
13
cause some trouble. Based on the news summary shown in Table 1, negative news articles started
appearing more frequently one month after t0. Also, note that there is substantial heterogeneity
in Initial Response, with a standard deviation of 23.85 days. This large variation appears to be
driven by lender characteristics related to their vested interest in the borrower: lenders with a
greater interest in the borrower’s survival are incentivized to offer more favorable lending terms
(by reducing their haircuts early). Lender Patience lasts, on average, almost a month (28.81 days),
and there is also a large variation here, with a standard deviation of 18.72 days. Most banks start
reducing their haircuts by t0+31, and negative news starts coming out more frequently after t0+45.
It is remarkable that lenders continue to lower the haircut and keep it at this lower level despite
the negative news for a substantial period of time.
[Place Table 3 about here]
Among other variables, the average loan amount outstanding is above $1 billion (with a mean of
$1.03 billion and a median of $0.772 billion). However, there are lenders with very small exposure
and lenders with very large exposure: a 5th percentile loan amount is $15 million and 95th percentile
loan amount is $2.5 billion. Despite truncation before t0, there is a great variation in relationship
length as well. The mean of relationship length is 930 days, with a standard deviation of 214 days.
The 5th percentile is 140 days, whereas the 95th percentile is 1000 days, which is the same as the
median (suggesting truncation). Panel B shows loan-level summary statistics. The average loan
size is $28 million, with great variation: the bottom 5th percentile is $1.31 million and the top 5th
percentile is $95.04 million. The mean maturity is 25.78 days, with a bottom 5th percentile of 3.22
days and top a 5th percentile of 40.11 days.
Table 4 shows correlations among the key variables used in this paper. Initial Response is negatively
related to the other key variables. The very strong negative correlation (-90%) between t1− t0 and
t2 − t1 suggests that lenders stepping in early to reduce haircuts are also likely to maintain the
decreased haircuts for longer before giving up and running on the fund. The correlation between
t1 − t0 and log(principal) is -0.81, suggesting that lenders with a larger loan exposure tend to
step in earlier to decrease their haircuts despite negative news against the borrower. Similarly,
the correlation between t1 − t0 and log(relationship) is -0.41, demonstrating that lenders with a
14
longer history with the borrower tend to help the borrower during hard times by reducing haircuts.
Lender Patience (t2−t1) is positively related with log(principal) and with log(relationship), showing
a correlation of 0.7 and 0.38, respectively. This suggests that lenders with a greater vested interest
(in terms of a larger loan exposure and longer lending history) tend to wait more patiently for the
borrower’s revival. Finally, log(principal) and log(relationship) have a very low correlation, which
indicates that these two variables measure two distinct aspects of lenders’ vested interest in the
borrower. As shown in Panel B, correlations of loan-level variables are consistent with those at the
lender level, shown in Panel A.
[Place Table 4 about here]
III. Determinants of Lender Behavior Dynamics
In this section, we explore lenders’ decisions about supplying credit when they know there is the
potential for a run. We investigate their dynamics though their haircut changes because the degree
of overcollateralization (or collateral requirements) characterize the lenders’ willingness to extend
credit against a unit of collateral. Since Fund X eventually experienced failure at the end of the
sample period, one may expect, at the aggregate level, lending terms to be tightened at some point.
However, our high frequency and micro-level data allow us to observe the full dynamics of the credit
rationing pattern by each lender, rather than just the aggregate eventual outcome. Furthermore,
such granularity facilitates investigation into the sources of the cross-lender variation in haircut
dynamics.
A. Lender-Specific Analysis
We first hypothesize that a lender with a larger vested interest in a borrower’s continuation is
incentivized to offer more favorable lending terms when the borrower is in trouble, as predicted by
Petersen and Rajan (1994, 1995). For this study, we measure lenders’ vested interest in the borrower
by outstanding loan amount and the length of their lending relationship with the borrower. Figure
3 shows survival rates on Initial Response (Panel A) and Lender Patience (Panel B) for lenders of
15
large (above median) and small (below median) loan amounts. Survival rates for Initial Response
are lower for lenders with a large loan amount (i.e., those lenders have shorter t1 − t0), and those
of Lender Patience are higher for lenders with a large loan amount (i.e., those lenders have longer
t2 − t1). Both figures, therefore, suggest that lenders with a vested interest in the borrower (as
measured by the size of loan) tend to act early to offer relief when the borrower is in trouble and
tend to have more patience before abandoning the borrower and running on their repo loans.
[Place Figure 3 about here]
For more formal analysis, we consider various empirical specifications, and examine the effect of
lenders’ vested interest on their behavior upon receiving negative signals about the borrower. The
dependent variables that describe lender behavior are Initial Response and Lender Patience. The
key explanatory variables we use to measure lenders’ vested interest in the borrower are loan
exposure (i.e., the natural logarithm of the principal amount) and the strength of the lending
relationship (i.e., the natural logarithm of the number of days since the first day of loaning to the
borrower from the inception of the fund).
For each lender behavior (Initial Response and Lender Patience), we examine four different empirical
specifications: linear, Cox proportional hazard, Weibull, and accelerate failure time (AFT) models
(Cleves, Gutierrez, Gould, and Marchenko (2010)). The regression specification for the linear model
is as follows.
∆tj = α+ β · xj + εj , (1)
where ∆tj is Initial Response (t1 − t0) or Lender Patience (t2 − t1) of lender j, xj is a vector of
lender j’s characteristics, and εj is the error term. Although linear models are very intuitive, one
of their shortcomings is that they assume normally distributed residuals, which is unlikely to be
satisfied. In order to account for more realistic distributional assumptions for the residuals, we
consider the Cox proportional hazard model, which fits a nonlinear model of hazard function, as
specified below:
h (t|xj) = h0 (t) exp (xjβx) , (2)
where h0 (t) is the base hazard (not estimated), xj is the explanatory variable (loan exposure and
16
strength of lending relationship) for lender j, and βx is the hazard rate for variable x. The Cox
regression model does not make parametric assumptions on base hazard, h0 (t). As an alternative,
Weibull regressions do make parametric assumptions on the base hazard function, with the following
specification:
h (t|xj) = ptp−1exp (xjβx) , (3)
where xj , and βx are defined the same way as in the Cox model. Parameter p determines the shape
of the base hazard function (h0 (t) = ptp−1). While Weibull regression requires further parametric
assumptions, it permits explicit estimates on failure time. For both Cox and Weibull specifications,
a positive hazard rate (βx) indicates increasing hazard (reduced time to failure) for increasing x,
and a negative hazard rate indicates decreasing hazard (increased time to failure) for increasing
x. Finally, we consider an AFT specification because the Weibull model has an accelerated failure
time interpretation, i.e., there is a one-to-one mapping between these two models (Cleves et al.
(2010)). AFT estimates the following specification:
log (tj) = xjβx + β0 + uj , (4)
where uj follows the extreme value distribution. The advantage of this specification is that param-
eter estimates measure marginal effects on log-failure time, i.e., βx measures the increase in the
expected value of log(time to failure) due to an increase in underlying variable xj (accelerated stop-
ping time is τ j = tje−xjβx , so that when an acceleration parameter is larger than 1, e−xjβx > 1, time
passes more quickly for the lender and failure is expected to occur sooner). For duration models
(Cox, Weibull, and AFT), robust standard errors are used.
Table 5 reports the results of the Initial Response and Lender Patience regressions specified in
Equations 1 to 4. Columns I through IV show the results of Initial Response regressions for each
respective specification. The linear model in Column I shows that both loan exposure and the
strength of the lending relationship have a significant and negative effect on Initial Response: one
standard deviation increase in log(principal) (1.49 from Table 3) causes a nearly three week (19.1
days) decrease in the time of the lender’s Initial Response (to reduce haircuts), and a one standard
deviation increase in log(relationship) (0.49 from Table 3) speeds up Initial Response by more
17
than a week (9.3 days). Columns II and III report results from the hazard rate model in the
Cox (Equation (2)) and Weibull (Equation (3)) specifications. They show that both lender vested
interest variables, log(principal) and log(relationship), have significant and positive effects on the
hazard of failure, i.e., a larger principal amount and a longer relationship leads to an earlier Initial
Response. In order to make a time-dimensional interpretation, we show the results of Weibull-AFT
regressions (Equation (4)) in Column IV. Estimates on both explanatory variables are significantly
negative, indicating that a lender’s large vested interest contributes to an earlier Initial Response:
a one standard deviation increase in log(principal) leads to an earlier haircut reduction by the
estimated acceleration parameter of 0.67 (e−1.49 ×-0.265). This result means the Initial Response
(42.5 days) will be shortened to 28.6 days (0.67×42.5), resulting in a two week (42.5-28.6) reduction
from the mean. Similarly, a one standard deviation increase in log(relationship) leads to an earlier
reduction in haircut by an acceleration parameter of 0.85 (e−0.49 ×-0.33). This indicates that the
mean Initial Response (42.5 days) will be shortened to 36.23 days, resulting in a one week (42.5-
36.23) reduction from the mean. A comparison of the AFT results - two weeks for loan exposure
and one week for relationship - with those from the linear model - three weeks for loan exposure
and one week for relationship - shows that they are in the same direction with similar size.
[Place Table 5 about here]
Columns V to VIII of Table 5 show the results of Lender Patience regressions for each respective
specification. In Column V, the estimation of the linear model (Equation (1)) shows that both
lender interest variables, the log of the principal amount and log of the relationship, are significantly
positive: a one standard deviation increase in loan exposure leads to a two week or 12.9 days
(1.49×8.67) longer period of patience, t2 − t1 and a one standard deviation increase in lending
relationship leads to a one week or 6.9 days (0.49×14.05) longer period of patience, t2−t1. Columns
VI and VII show that hazard ratios estimated by the regression specifications in Equation (2) and
(3), respectively, are significantly negative for both lender interest variables (loan exposure and
the strength of the lending relationship), i.e., a lower hazard of failure or longer Lender Patience
period. For a time-dimensional interpretation, we estimate the AFT regression (Equation (4)) in
Column VIII, and show a significant and positive acceleration parameter of 3.07 (e1.49 ×0.75) for
loan exposure and 2.16 (e0.49 ×1.57) for lending relationship per one standard deviation increase in
18
each lender interest variable. These quantities imply that, from the mean of Lender Patience (28.81
days), a one standard deviation increase in loan exposure leads to 88.44 total days (28.81×3.07) of
Lender Patience, resulting in a 59.63 day (88.44-28.81) longer Lender Patience period. Similarly,
a one standard deviation increase in the strengths of the lending relationship leads to 62.23 days
(28.81×2.16) of Lender Patience, resulting in a 33.42 day (62.23-28.81) longer Lender Patience
period. These estimates are larger than those from the linear model, suggesting that the Weibull
model can be sensitive to parametric assumptions. Nevertheless, all regression specifications (linear,
Cox, Weibull, and AFT) strongly suggest that the lender’s increased vested interest in the borrower
leads to increased effort to rescue the troubled borrower via decreased Initial Response and longer
Lender Patience period.
B. Loan-Level Analysis
To fully exploit the granularity of our data, we further examine the impact of loan-level character-
istics on Initial Response and Lender Patience beyond the explanations provided by lender-specific
variables. The first four columns of Table 6 show the results on Initial Response (t1 − t0), using
combinations of lender-level and loan-level characteristics. As before, we consider four specifica-
tions from Equations (1) to (4): linear (Column I), Cox (Column II), Weibull (Column III), and
AFT (Column IV), respectively. As with the lender-level regressions, we find earlier reduction in
haircut (shorter Initial Response) when lender’s aggregate principal amount is larger (in log-scale).
For example, a one standard deviation increase in the log of the lender level aggregate principal
amount (1.25) leads to an Initial response that is 6 days (-4.848×1.25) earlier. In the AFT regression
(Column IV), a one standard deviation increase in log(Principal) (1.25) leads to an earlier haircut
reduction associated with the acceleration parameter of 0.94 (e1.25 ×-0.051). This result implies
that, from the mean (54.89 days), the Initial Response will be reduced to 51.37 days (0.936×54.89),
lowering the Initial Response by a half of a week (54.89-51.37).
[Place Table 6 about here]
Columns I to IV of the same table show that the parameter estimates of loan-level relationship
(the log of relationship length) variables have the same sign as those in the lender-level regressions
19
(Columns V to VIII in Table 5), but they are statistically insignificant. The impact of the loan-level
variables on Initial Response are weaker than those from the lender-level regressions. This might be
due to the fact that loan-level characteristics explain some of the lender-level effects. In other words,
the Initial Response of each loan may not be entirely driven by lender-specific variables, and the
lender may make different choices for each loan based on individual loan characteristics, ensuring
that aggregate changes across all loans for each lender meets its (lender-level) strategic goal. For
example, among loan-level (therefore collateral-specific) variables, the parameter estimates of loan
principal are significantly positive. In the AFT specification (Column IV), a one standard deviation
increase in the log of loan principal (1.37) leads to an earlier Initial Response by the acceleration
parameter of 0.92 (e−0.064 4×1.37). We can interpret this result as follows: the Initial Response
will be reduced to 50.5 days (54.89×0.92) from its mean (54.89 days). Overall, lenders start
dropping their haircuts (upon initial acknowledgment of bad news on Fund X) earlier for larger
loans (principal amount) and higher quality (investment grade) loans.
Columns V to VIII of Table 6 show the impact of lender-level and loan-level attributes on Lender
Patience (t2− t1). While lender-level loan exposure (the log of the lender-level aggregate principal)
is positively related to Lender Patience, the strength of the lender’s relationship with the borrower
is positive but statistically insignificant. For example, in the AFT specification, a one standard
deviation increase in lender loan exposure (1.25) is associated with a 1.78 (e0.461 ×1.25) times
longer Lender Patience period. This suggests that Lender Patience will be increased 20 days
(25.45×1.78-25.45) from its mean of 25.45 days. On the other hand, all three loan-level attributes
have a significantly positive impact on Lender Patience: larger loans, and shorter loans tend to have
longer Lender Patience periods. This result can be translated as follows: a one standard deviation
increase in the log of the loan principal (1.37) is associated with a Lender Patience period that is
1.18 (e0.123 ×1.37) times longer, and short-term loans (below the median maturity) are associated
with a Lender Patience period that is 1.36 (e0.309) times longer. These correspond to a Lender
Patience period that is 4.6 days (25.45×0.18) and 9 days (25.45×0.36) longer, respectively, for a
one standard deviation increase in the log of the loan principal and shortness of term from the
mean of 25.45 days.
Overall, the results in Tables 5 and 6 show that there is a large variation in lenders’ dynamic
20
patterns of credit supply during a period of borrower distress, and that can be largely explained
by lenders’ vested interests, which make each lender’s incentives different from the others’.
C. Collateral-Level Analysis
In this section, we study the effect of the asset separability (or liquidity) of collateral on Initial
Response and Lender Patience. When collateral can easily be sold in the secondary markets, we
expect lenders to be less incentivized to rescue a borrower from financial difficulties, because they
can liquidate the collateral without incurring much loss.
[Place Table 7 about here]
Table 7 shows results using Weibull accelerated failure time models. Columns I to III show Initial
Response based on collateral characteristics. Column I considers all 584 loans included in our
sample. Relative to loans against Treasury securities, ones pledged using structured finance assets
(CDO, MBS, and other structured finance products) and corporate bonds are associated with a
faster Initial Response. For example, loans against structured finance have an Initial Response that
is 24 days earlier than the sample mean (54.89 days × (1− e−0.578)). Loans against corporate bonds
have an Initial Response that is 20 days earlier (54.89 days × (1− e−0.453)). This result implies
that the time in takes for the lender to begin its attempt to rescue the borrower is associated with
illiquidity of collateral. Specifically, we find that loans against structured finance, followed by those
against corporate bonds, lead to an earlier Initial Response than the much more liquid case of
Treasury securities.
Since the dominant asset class of Fund X is structured finance, we focus only on this class of asset,
which mostly contains CDOs and pass-through MBSs.11 Relative to pass-through vanilla MBSs,
CDOs are much more opaque and illiquid due to additional layers of securitization (i.e., a CDO is a
pool of MBSs). Column II shows no significant variation of Initial Response based on the collateral
type within the structured finance class.
We repeat the previous analysis using the credit ratings of the structured finance assets as of t0 in11There is a small fraction of hard-to-classify sub-asset classes, such as asset-backed securities (ABS).
21
Column III. We expect collateral with higher credit ratings to be more liquid and, hence, lenders to
have fewer reasons to be borrower-friendly (so, a slower Initial Response). As expected, collateral
with a high (AAA or AA) credit rating has a slower Initial Response. For example, AAA-rated
collateral has an eight day slower Initial Response (54.89 days × (e0.140 − 1)).
Columns IV to VI show Lender Patience based on collateral characteristics. Consistent with our
previous results on Initial Response, lenders are more patient and maintain their reduced haircut
for longer for loans against less liquid collateral. Column IV considers Lender Patience by collat-
eral types for all loans. Loans with structured finance (CDO, MBS, and other structured finance)
collateral have longer Lender Patience periods than ones pledged with Treasury securities or cor-
porate bonds. In contrast to our results on Initial Response (in Column II), the results in Column
(V) show that, within structured finance, CDO-pledged loans are associated with greater Lender
Patience: CDO-pledged loans have an 11 day (25.45 × (e0.373 − 1)) longer Lender Patience period
than vanilla MBS pledged loans. Also, Column VI indicates that loans pledged with collateral with
high credit ratings receive less Lender Patience. For example, loans with AAA-rated collateral
have an 8.6 day shorter Lender Patience period than those pledged with A-rated collateral (25.45
× (1− e−0.413)).
Overall, the results in Table 7 show that lenders start decreasing their haircuts earlier and keep
decreasing haircuts for longer for loans with less liquid collateral.
IV. Conclusion
In this paper, we show that the amount of collateral (or haircuts in repo contracts) has a destabi-
lizing feedback effect by increasing the likelihood of a run. Because of this destabilizing feedback,
lenders tend to be patient when deciding on the amount of collateral: when choosing a high level
collateral, lenders must balance the potential for better recovery should the borrower default with
their increased likelihood of prematurely abandoning the project (by running the borrower). Dur-
ing the last months of a large hedge fund, we observe a U-shaped pattern in the haircut trend:
lenders initially decrease haircuts to avoid exacerbating the borrower’s default risk. After a favor-
able period with a relaxed credit supply (a reduced haircut), as more bad news about the borrower
22
accumulates, lenders fail to coordinate and rapidly increase their haircuts until the borrower de-
faults. Closer examination shows that the incentive of lenders to keep the borrowing firm alive is
critical in their haircut choices: lenders with a larger loan exposure and a stronger prior lending
relationship with the borrower extend the relaxed credit supply more quickly, and they are more
patient with credit contraction, raising their haircut later than those lenders that had less loan
exposure to the borrower and a weaker relationship with them.
While this paper exploits empirical evidence in the repo market, our insights have general impli-
cations. The risk that insuring against a future downside can create time-inconsistency problems
(e.g., a premature run) once misfortune (e.g., borrower default) occurs is applicable to a wide vari-
ety of financial contracting circumstances. Also, the new role of collateral as a lender coordination
mechanism requires further study. Casual observation suggests that the impairment of a borrower’s
production by collateral liquidation is tied to the separability of the collateral, which is determined
by the nature of asset type. This, in turn, influences the degree to which collateral requirements
interact with lender coordination regarding rollover decisions. Further investigation is needed to
clarify interrelations among these factors. We leave it for future works.
23
REFERENCES
Acharya, Viral V, Douglas Gale, and Tanju Yorulmazer, 2011, Rollover Risk and Market Freezes,Jounal of Finance 66, 1177–1209.
Angeletos, George-Marios, Christian Hellwig, and Alessandro Pavan, 2007, Dynamic Global Gamesof Regime Change : Learning , Multiplicity , and the Timing of Attacks, Econometrica 75, 711–756.
Angeletos, George-Marios, and I. Werning, 2006, Crises and Prices: Information Aggregation,Multiplicity, and Volatility, American Economic Review 96, 1720–1736.
Auh, Jun Kyung, and Mattia Landoni, 2016, Lender Protection versus Risk Compensation : Evi-dence from the Bilateral Repo Market, Georgetown University Working Paper .
Baklanova, Viktoria, Adam Copeland, and Rebecca McCaughrin, 2015, Reference Guide to U.S.Repo and Securities Lending Markets, Technical report.
Bebchuk, Lucian A, and Itay Goldstein, 2011, Self-fulfilling Credit Market Freezes, Review ofFinancial Studies 24, 3249–2555.
Berglof, Erik, and Ernst-ludwig Von Thadden, 1994, Short-Term Versus Long-Term Interests: Cap-ital Structure with Multiple Investors, Quarterly Journal of Economics 109, 1055–1084.
Bester, Helmut, 1985, Screening vs . Rationing in Credit Markets with Imperfect Information,American Economic Review 75, 850–855.
Bolton, Patrick, and Martin Oehmke, 2011, Credit Default Swaps and the Empty Creditor Problem,Review of Financial Studies 24, 2617–2655.
Brunnermeier, Markus K, and Martin Oehmke, 2013, The Maturity Rat Race, Jounal of FinanceLXVIII, 483–521.
Calomiris, Charles W., Mauricio Larrain, Jose M. Liberty, and Jason Sturgess, 2017, How CollateralLaws Shape Lending and Sectoral Activity, Journal of Financial Economics 123, 163–188.
Campello, Murillo, and Mauricio Larrain, 2016, Enlarging the Contracting Space: Collateral Menus,Access to Credit, and Economic Activity, Review of Financial Studies 29, 349–383.
Campos, Rodolfo G, 2013, Risk-sharing and Crises: Global Games of Regime Change with En-dogenous Wealth, Journal of Economic Theory 148, 1624–1658.
Carlsson, Hans, and Eric van Damme, 1993, Global Games and Equilibrium Selection, Economet-rica 61, 989–1018.
Chen, Qi, Itay Goldstein, and Wei Jiang, 2010, Payoff Complementarities and Financial Fragility:Evidence from Mutual Fund Outflows, Journal of Financial Economics 97, 239–262.
Ciccone, Antonio, and James Costain, 2004, On Payoff Heterogeneity in Games with StrategicComplementarities, Oxford Economic Papers 56, 701–713.
Cleves, Mario, Roberto G. Gutierrez, William Gould, and Yulia V. Marchenko, 2010, An Introduc-tion to Survival Analysis Using Stata (Cambridge University Press).
24
Copeland, Adam, Antoine Martin, and Michael Walker, 2014, Repo Runs: Evidence from TheTri-Party Repo Market, The Journal of Finance 69, 2343–2380.
Corsetti, Giancarlo, Amil Dasgupta, Stephen Morris, and Hyun Song Shin, 2004, Does One SorosMake a Difference ? A Theory of Currency Crises with Large and Small Traders, Review ofEconomic Studies 71, 87–113.
Dasgupta, Amil, 2007, Coordination and Delay in Global Games, Journal of Economic Theory 134,195–225.
Diamond, Douglas W, and Philip H Dybvig, 1983, Bank Runs , Deposit Insurance , and Liquidity,Journal of Political Economy 91, 401–419.
Frankel, David M, 2012, Recurrent Crises in Global Games, Journal of Mathematical Economics48, 309–321.
Frankel, David M, Stephen Morris, and Ady Pauzner, 2003, Equilibrium Selection in Global Gameswith Strategic Complementarities, Journal of Economic Theory 108, 1–44.
Goldstein, Itay, and A D Y Pauzner, 2005, Demand-Deposit Contracts and the Probability of BankRuns, Jounal of Finance LX, 1293–1327.
Gorton, Gary, and Andrew Metrick, 2012, Securitized Banking and the Run on Repo, Journal ofFinancial Economics 104, 425–451.
Gorton, Gary, and Guillermo Ordonez, 2014, Collateral Crises, American Economic Review 104,343–378.
Guimaraes, Bernardo, and Stephen Morris, 2007, Risk and Wealth in a Model of Self-fulfillingCurrency Attacks, Journal of Monetary Economics 54, 2205–2230.
He, Z., and Asaf Manela, 2016, Information Acquisition in Rumor-Based Bank Runs, Journal ofFinance 71, 1113–1158.
He, Z., and W. Xiong, 2012, Dynamic Debt Runs, Review of Financial Studies 25, 1799–1843.
Hertzberg, Andrew, Jose Maria Liberti, and Daniel Paravisini, 2011, Public Information and Co-ordination : Evidence from a Credit Registry Expansion, Jounal of Finance LXVI, 379–412.
Inderst, Roman, and Holger M Mueller, 2007, A Lender-based Theory of Collateral, Journal ofFinancial Economics 84, 826–859.
Iyer, Rajkamal, and Manju Puri, 2014, Understanding Bank Runs : The Importance of Depositor-Bank Relationships and Networks, American Economic Review 102, 1414–1445.
Izmalkov, Sergei, and Muhamet Yildiz, 2010, Investor Sentiments, American Economic Review 2,21–38.
Khanna, Naveen, and Richmond D Mathews, 2015, Posturing and Holdup in Innovation, Reviewof Financial Studies 29.
Khanna, Naveen, and Mark Schroder, 2010, Optimal Debt Contracts and Product Market Compe-tition with Exit and Entry, Journal of Economic Theory 145, 156–188.
25
Khanna, Naveen, and Sheri Tice, 2000, Strategic Responses of Incumbents to New Entry: TheEffect of Ownership Structure, Capital Structure, and Focus, Review of Financial Studies 13,749–779.
Khanna, Naveen, and Sheri Tice, 2005, Pricing , Exit, and Location Decisions of Firms: Evidenceon the Role of Debt and Operating Efficiency, Journal of Financial Economics 75, 397–427.
Krishnamurthy, Arvind, Stefan Nagel, and Dmitry Orlov, 2014, Sizing Up Repo, The Journal ofFinance 69, 2381–2417.
Martin, A, D Skeie, and E Thadden, 2014a, Repo Runs, Review of Financial Studies 27, 957–989.
Martin, Antoine, David Skeie, and Ernst-ludwig Von Thadden, 2014b, The Fragility of Short-termSecured Funding Markets, Journal of Economic Theory 149, 15–42.
Morris, Stephen, and Hyun Song Shin, 1998, Unique Equilibrium in a Model of Self-FulfillingCurrency Attacks, American Economic Review 88, 587–597.
Morris, Stephen, and Hyun Song Shin, 2004, Coordination Risk and the Price of Debt, EuropeanEconomic Review 48, 133–153.
Morris, Stephen, and Hyun Song Shin, 2006, Global Games : Theory and Applications (CambridgeUniversity Press).
Petersen, Mitchell A., and Raghuram G Rajan, 1994, The Benefits of Lending Relationships :Evidence from Small Business Data, Journal of Finance 49, 3–37.
Petersen, Mitchell A., and Raghuram G Rajan, 1995, The Effect of Credit Market Competition onLending Relationships, Quarterly Journal of Economics 110, 407–443.
Rajan, Raghuram, and Andrew Winton, 1995, Covenants and Collateral as Incentives to Monitor,Journal of Finance L, 1113–1146.
Rajan, Raghuram G, 1992, Insiders and Outsiders : The Choice between Informed and Arm’s-Length Debt, Journal of Finance XLVII, 1367–1400.
Sakovics, By Jozsef, and Jakub Steiner, 2012, Who Matters in Coordination Problems?, AmericanEconomic Review 102, 3439–3461.
Schmidt, By Lawrence, Allan Timmermann, and Russ Wermers, 2016, Runs on Money MarketMutual Funds, American Economic Review 106, 2625–2657.
Stiglitz, Joseph E., and Andrew Weiss, 1981, Credit Rationing in Markets with Imperfect Informa-tion, American Economic Review 71, 393–410.
Vig, Vikrant, 2013, Access to Collateral and Corporate Debt Structure: Evidence from a NaturalExperiment, Journal of Finance 106, 881–2928.
26
Figure 1. Average Change in Haircut by Repo Lenders Until Default: This figure shows the relativechanges in haircut (Panel A), spread (Panel B), and duration (Panel C) by lenders, Fund X’s outstanding loan amount(dotted line), and the number of negative news articles (bar graph) about Fund X. Haircut (spread, duration) as oft0 is normalized to zero for each lender. Specifically, we estimate the following regression:
Terml,t = αl + Σ0i=−88βi · Ii + εl,t,
where Term refers to each lending term (Haircut, Spread and Duration), αl is lender fixed effects and It is a dummyvariable that is 1 only on each corresponding day. The number of news articles is computed by counting the numberof negative news articles on Fund X found in a Factiva search.
05
1015
20Nb
r. of
Arti
cle /
Out
stan
ding
Loa
n Am
t. (in
bn.
)
-6
-4
-2
0
2Ha
ircut
Cha
nges
-88 -80 -72 -64 -56 -48 -40 -32 -24 -16 -8 0Time From Failure
95% CI Mean # Articles Loan Amt.
(a) Haircut
05
1015
20Nb
r. of
Arti
cle /
Out
stan
ding
Loa
n Am
t. (in
bn.
)
-.05
0
.05
.1
.15
Spre
ad C
hang
es
-88 -80 -72 -64 -56 -48 -40 -32 -24 -16 -8 0Time From Failure
95% CI Mean # Articles Loan Amt.
(b) Spread
05
1015
20Nb
r. of
Arti
cle /
Out
stan
ding
Loa
n Am
t. (in
bn.
)
-60
-40
-20
0
20
Dura
tion
Chan
ges
-88 -80 -72 -64 -56 -48 -40 -32 -24 -16 -8 0Time From Failure
95% CI Mean # Articles Loan Amt.
(c) Duration
27
Figure 2. Timeline for a Repo Run: The starting point of our sample is one day before the first significantnegative news on this firm surfaced (88 days before default), which we set as t0. After this date, there are three keyevent dates during the repo run: the first date on which each lender decreased their haircut after t0 is defined as t1,the date on which each lender increased their haircut after t1 is defined as t2, and the date when all loans from eachlender were terminated is defined as t3.
t0 t1 t2 t3
28
Figure 3. Kaplan-Meier Plot Across Fund X’s Lenders: This figure shows the Kaplan-Meier survivalestimates for Initial Response and Lender Patience for two groups based on loan exposure (below and above themedian loan amount against Fund X).
����
����
����
����
����
6XUYLYDO�5DWHV
� �� �� �� ��,QLWLDO�5HVSRQVH��'D\V
6PDOO�/RDQ/DUJH�/RDQ
.DSODQ�0HLHU�6XUYLYDO�(VWLPDWHV
(a) Initial Response (t1 − t0)
0.00
0.25
0.50
0.75
1.00
Surv
ival
Rat
es
0 20 40 60Lender Patience: Days
Small LoanLarge Loan
Kaplan-Meier Survival Estimates
(b) Lender Patience (t2 − t1)
29
Tab
le1.
Neg
ativ
eN
ews
Art
icle
son
Fund
X:
Thi
sta
ble
show
sth
eda
tes
(Col
umn
I),n
umbe
r(C
olum
nII
),an
dco
nten
ts(C
olum
nII
I)of
new
sar
ticle
sre
late
dto
Fund
Xar
ound
the
perio
dw
hen
itco
llaps
ed(8
8to
0ev
entd
ates
).N
ews
artic
les
are
colle
cted
from
Fact
iva
usin
ga
keyw
ord
sear
chfo
rFu
ndX
’spa
rent
com
pany
’sna
me
inm
ajor
new
sou
tlets
(e.g
.,W
SJ,N
YT
,WP,
and
Forb
es).
We
drop
new
sar
ticle
sth
atar
eno
tre
late
dto
Fund
Xits
elf,
such
asFu
ndX
anal
yst’s
com
men
tson
othe
rst
ocks
.
Eve
ntD
ate
No.
ofA
rtic
les
Con
tent
(I)
(II)
(III
)
871
Fund
Xm
akes
loss
from
lend
ing
tom
ortg
age
inve
stm
ent
firm
.85
1C
omm
ent
byFu
ndX
onsu
bprim
em
ortg
age
trou
ble.
671
Fund
Xst
ock
drop
sdu
eto
subp
rime
mor
tgag
etr
oubl
e.52
1Fu
ndX
com
men
tsas
subp
rime
mor
tgag
etr
oubl
esu
rfac
es.
492
Fund
Xin
volv
edin
insi
der
trad
ing.
452
Fund
Xw
illw
rite
dow
ndu
eto
deva
luat
ion
ofon
eof
itsdi
visi
on.
391
Fund
Xm
arke
ttim
ing
scan
dal.
351
Fund
Xea
rnin
gsdo
wng
rade
.24
1Fu
ndX
sued
infix
ing
trad
ing
cost
.22
1Fu
ndX
subp
rime
mor
tgag
etr
oubl
e.17
1Fu
ndX
hurt
bysu
bprim
em
ortg
age
loan
.15
3Fu
ndX
isfa
cing
mor
tgag
elo
sses
.14
4Fu
ndX
drop
inne
tin
com
e.12
1Fu
ndX
incu
rbi
glo
ss.
111
Fund
Xle
nder
sm
eet
toke
epfu
ndafl
oat.
92
Fund
Xfa
cesh
utdo
wn.
87
fund
X’s
pare
ntco
mpa
nyst
aves
offco
llape
ofits
fund
s;Fu
ndX
shar
edr
ops.
71
Fund
Xsu
rviv
esfo
rno
w.
66
Fund
X’s
pare
ntco
mpa
nyw
illpu
t$3
.2B
tosa
veits
hedg
efu
nd;n
egat
ive
impa
cton
stoc
k&
junk
bond
mar
ket.
53
Fund
Xre
scue
plan
for
one
ofits
hedg
efu
nd;l
ende
rsse
llco
llate
ral.
41
Mor
em
eltd
owns
like
Fund
Xex
pect
ed.
35
Fund
Xne
gativ
ely
impa
cts
stoc
km
arke
t;IP
Oca
ncel
latio
nby
Fund
Xre
late
dfir
m.
26
SEC
Fund
XC
DO
prob
e;Fu
ndX
’spa
rent
com
pany
won
’tpr
ovid
ead
ditio
nalf
undi
ngfo
rfa
iling
hedg
efu
nd.
16
Fund
Xco
llaps
ed.
01
Rep
utat
ion
dam
age
topa
rent
com
pany
ofFu
ndX
30
Tab
le2.
Sum
mar
yof
Eve
ntD
ates
Dur
ing
aR
epo
Run
in20
07:
Thi
sta
ble
sum
mar
izes
maj
orev
ent
date
s.C
olum
nI
show
sth
eto
tal
amou
ntof
prin
cipa
lout
stan
ding
atth
ebe
ginn
ing
ofth
esa
mpl
epe
riod
(88
days
prio
rto
defa
ult)
.C
olum
nII
show
sth
eda
teon
whi
chea
chle
nder
star
tsde
crea
sing
thei
rha
ircut
.C
olum
nII
Ish
ows
whe
nea
chle
nder
star
tsin
crea
sing
thei
rha
ircut
.C
olum
nIV
show
sw
hen
each
lend
erte
rmin
ated
thei
rlo
ans.
Col
umn
Vsh
ows
Initi
alR
espo
nse,
whi
chis
the
num
ber
ofda
ysbe
twee
nth
ebe
ginn
ing
ofsa
mpl
epe
riod
(whe
nne
gativ
ene
ws
first
com
esou
t)an
dw
hen
each
lend
erst
arts
decr
easi
ngha
ircut
.C
olum
nV
Ish
ows
Lend
erPa
tienc
e,w
hich
isth
enu
mbe
rof
days
betw
een
whe
nea
chle
nder
star
tsde
crea
sing
thei
rha
ircut
,and
whe
nea
chle
nder
star
tsin
crea
sing
thei
rha
ircut
agai
n.C
olum
nV
IIis
the
hairc
utpa
tter
n,w
hich
isU
-sha
peif
the
hairc
utin
itial
lyde
crea
ses
and
then
incr
ease
sag
ain,
noch
ange
ifth
ere
isno
sign
ifica
ntch
ange
sin
the
hairc
utun
tillo
ante
rmin
atio
n;an
dde
crea
seif
the
hairc
utde
crea
ses
until
loan
term
inat
ion.
Lend
erC
hara
cter
istic
sE
vent
date
sH
airc
utC
hara
cter
istic
s
Lend
erP
rinci
pal
Hai
rcut
Hai
rcut
Loan
Initi
alR
espo
nse
Lend
erPa
tienc
eH
airc
utPa
tter
nID
($M
illio
n)D
ecre
ase
(t1)
Incr
ease
(t2)
Term
inat
ion
(t3)
t 1−t 0
(Day
s)t 2−t 1
(Day
s)
(I)
(II)
(III
)(I
V)
(V)
(VI)
(VII
)
1424
95.3
758
21
3056
Dec
reas
e13
2453
.07
5928
329
31U
-sha
pe5
2183
.70
7029
018
41U
-sha
pe9
2085
.74
7030
018
40U
-sha
pe4
1610
.26
5928
129
31U
-sha
pe8
901.
9565
231
2342
U-s
hape
188
0.07
6529
823
36U
-sha
pe17
803.
1152
21
3650
Dec
reas
e10
741.
1435
1711
5318
U-s
hape
1162
0.18
5222
836
30U
-sha
pe3
590.
8458
1211
3046
Dec
reas
e12
496.
089
87
791
No
chan
ge6
344.
9016
1514
721
No
chan
ge16
300.
4452
1615
3636
Dec
reas
e2
23.4
92
10
861
No
chan
ge15
15.2
56
54
821
No
chan
ge
31
Table 3. Summary Statistics: This table shows number of observations (Column I), mean (Column II), standarddeviation (Columns III), and 5/50/95-percentiles (Columns IV to VI) of the key variables used in this study. PanelA shows lender-level summary statistics for the 16 lenders of this study. Panel B shows loan-level summary statistics.Event time variables are Initial Response and Lender Patience. Initial Response is the first time each lender reducestheir haircut after the bad news began minus the time when the bad news first came out (t1− t0). Lender Patience isthe time when each lender increases their haircut minus the time when each lender decreases their haircut (t2 − t1).Variables on lender characteristics are loan amount, natural logarithm of the principal, relationship (in days), andthe natural logarithm of relationship. Variables on loan characteristics are the principal amount and its logarithmand maturity (in days). Loan amount is the total outstanding principal from each lender as of the beginning of thesample period (88 days prior to default or event date 88). Relationship is the number of days (as of event date 88)since a lender first had dealings with the borrower in the database, which starts in 2004.
Panel (a)
Variables N Mean Std.Dev. p5 p50 p95
(I) (II) (III) (IV) (V) (VI)
Event timesInitial Response, t1− t0 (Days) 16 42.5 23.85 18 33 86Lender Patience, t2− t1 (Days) 16 28.81 18.72 1 33.5 56
Lender characteristicsPrincipal ($Million): Lender 16 1030 849 15 772 2500Log(Principal): Lender 16 20.17 1.49 16.54 20.46 21.64Relationship (Days) 16 930 214 140 1000 1000Log(Relationship) 16 6.77 0.49 4.94 6.91 6.91
Panel (b)
Variables N Mean Std.Dev. p5 p50 p95
(I) (II) (III) (IV) (V) (VI)
Event timesInitial Response, t1− t0 (Days) 584 54.89 25.97 21 52 86Lender Patience, t2− t1 (Days) 584 25.45 23.56 1 28.5 63
Lender characteristicsPrincipal ($Million): Lender 584 1511.48 792.22 23.49 1610.26 2495.37Log(Principal): Lender 584 20.76 1.25 16.97 21.2 21.64Relationship (Days) 584 982.3 102.14 950 1000 1000Log(Relationship) 584 6.87 0.23 6.86 6.91 6.91
Loan characteristicsPrincipal ($Million): Loan 584 28.28 57.22 1.31 11.25 95.04Log(Principal): Loan 584 16.24 1.37 14.09 16.24 18.37Maturity (Days) 584 25.78 13.43 3.22 30.26 40.11
32
Table 4. Correlations: This table shows correlations among major event periods during the repo run and keydeterminants. Panel A shows the correlations for the lender-level variables. Panel B shows correlations among theloan-level variables. Major event periods are Initial Response (t1 − t0) and Lender Patience (t2 − t1). In PanelA, correlations among event times, the natural logarithm of principal and the natural logarithm of relationship areshown. In Panel B, correlations among event times, the log of principal (lender), the log of relationship (lender), logof principal (loan), and the short-term repo (below median) indicators are shown.
Panel (a)
Initial Response Lender Patience Log(Principal)t1− t0 (Days) t2− t1 (Days) Lender
Lender Patience, t2-t1 (Days) -0.90Log(Principal): Lender -0.81 0.70Log(Relationship) -0.41 0.38 0.02
Panel (b)
Initial Lender Log- Log-Response Patience Principal: Log- Principal:
t1 − t0 (Days) t2 − t1 (Days) Lender Relationship Loan Short-term
Lender Patience (t2 − t1) -0.91Log(Principal): Lender -0.24 0.25Log(Relationship) -0.04 0.06 0.08Log(Principal): Loan -0.14 0.04 0.06 -0.11Short-term -0.02 0.06 0.19 0.05 -0.18
33
Tab
le5.
Det
erm
inan
tsof
Init
ialR
esp
onse
and
Len
der
Pat
ienc
eat
the
Len
der
Lev
el:
Thi
sta
ble
show
sth
ede
term
inan
tsof
the
maj
orev
ent
perio
dsdu
ring
the
repo
run
inth
ele
nder
-leve
ldat
a.D
epen
dent
varia
bles
are
maj
orev
ent
perio
ds:
Initi
alR
espo
nse
(t1−t 0
)an
dLe
nder
Patie
nce
(t2−t 1
).Fo
rbo
thde
pend
ent
varia
bles
we
use
four
diffe
rent
spec
ifica
tions
:or
dina
ryle
ast
squa
res
(Col
umn
Ifo
rt 1−t 0
and
Col
umn
Vfo
rt 2−t 1
),th
eC
oxpr
opor
tiona
lhaz
ard
mod
el(C
olum
nII
fort 1−t 0
and
Col
umn
VI
fort 2−t 1
),W
the
eibu
llre
gres
sion
(Col
umn
III
fort 1−t 0
and
Col
umn
VII
fort 2−t 1
),an
dth
eW
eibu
llre
gres
sion
inth
eac
cele
rate
dfa
ilure
time
(AFT
)m
etric
(Col
umn
IVfo
rt 1−t 0
and
Col
umn
VII
Ifo
rt 2−t 1
).In
the
Cox
prop
ortio
nalh
azar
dm
odel
s(C
olum
nsII
and
VI)
,w
efit
h(t|x
j)
=h
0(t
)exp
(xjβ
x),
whe
reh
0(t
)is
base
haza
rd,x
jis
the
expl
anat
ory
varia
ble
for
lend
erj,
andβ
xis
the
haza
rdra
tefo
rva
riabl
ex
.C
olum
nsII
and
VI
repo
rtha
zard
rate
.A
posi
tive
haza
rdra
tein
dica
tes
anin
crea
sing
haza
rd(r
educ
edtim
eto
failu
re)
for
incr
easi
ngx
,and
ane
gativ
eha
zard
rate
indi
cate
sa
decr
easi
ngha
zard
(incr
ease
dtim
eto
failu
re)
for
incr
easi
ngx
.In
the
Wei
bull
regr
essi
ons,
we
estim
ate,h
(t|x
j)
=ptp−
1exp
(xjβ
x).
We
repo
rtha
zard
rate
san
dlogâĄ
ą(p)in
Col
umns
III
and
VII
.The
acce
lera
ted
failu
retim
etr
ansf
orm
sth
eW
eibu
llha
zard
mod
elin
tolo
g(t
j)=
xjβ
x+β
0+u
j,w
hereu
jfo
llow
sth
eex
trem
eva
lue
dist
ribut
ion.
For
dura
tion
mod
els,
robu
stst
anda
rder
rors
are
used
.St
anda
rder
rors
are
show
nin
pare
nthe
ses
and
are
sign
ifica
ntat
the
1%(*
**),
5%(*
*),
and
10%
(*)
leve
ls.
Dep
ende
ntVa
riabl
esIn
itial
Res
pons
e(t
1−t 0
)Le
nder
Patie
nce
(t2−t 1
)
(I)
(II)
(III
)(I
V)
(V)
(VI)
(VII
)(V
III)
Log(
Prin
cipa
l)-1
2.78
988*
**1.
3434
2***
0.95
823*
**-0
.265
14**
*8.
6747
9***
-0.6
0383
***
-1.2
1017
***
0.75
367*
**[1
.944
40]
[0.2
8481
][0
.148
37]
[0.0
4822
][2
.131
33]
[0.1
4232
][0
.344
51]
[0.1
9534
]Lo
g(R
elat
ions
hip)
-19.
0419
3***
1.26
958*
**1.
1765
5***
-0.3
2554
***
14.0
5126
*-0
.976
48**
*-2
.524
03**
*1.
5719
2***
[5.9
3764
][0
.380
26]
[0.2
2609
][0
.043
91]
[6.5
0845
][0
.267
83]
[0.5
6409
][0
.116
31]
Con
stan
t42
9.35
038*
**-4
0.85
056*
**11
.303
14**
*-2
41.2
5901
***
36.5
7800
***
-22.
7801
3***
[55.
5937
8][4
.407
17]
[0.9
6606
][6
0.93
820]
[8.2
6205
][3
.724
85]
Log(
p)1.
2848
4***
1.28
484*
**0.
4735
6**
0.47
356*
*[0
.127
22]
[0.1
2722
][0
.185
29]
[0.1
8529
]
Mod
elLi
near
Cox
Wei
bull
Wei
bull
(AFT
)Li
near
Cox
Wei
bull
Wei
bull
(AFT
)Lo
gps
eudo
likel
ihoo
d-2
2.49
-3.4
5-3
.45
-27.
95-1
7.63
-17.
63W
ald
chi2
22.2
866
.34
99.7
018
.03
24.5
528
5.44
Pro
b>
chi2
0.00
0.00
0.00
0.00
0.00
0.00
R-s
quar
ed0.
8074
60.
6246
0O
bser
vatio
ns16
1616
1616
1616
16
34
Tab
le6.
Det
erm
inan
tsof
Init
ialR
esp
onse
and
Len
der
Pat
ienc
eat
the
Loa
nL
evel
:T
hist
able
show
sde
term
inan
tsof
maj
orev
entp
erio
dsdu
ring
the
repo
run
inth
elo
an-le
veld
ata.
Dep
ende
ntva
riabl
esar
em
ajor
even
tper
iods
:In
itial
Res
pons
e(t
1−t 0
)and
Lend
erPa
tienc
e(t
2−t 1
).Fo
rbot
hde
pend
entv
aria
bles
we
use
four
diffe
rent
spec
ifica
tions
:or
dina
ryle
asts
quar
es(C
olum
nIf
ort 1−t 0
and
Col
umn
Vfo
rt2−t 1
),th
eC
oxpr
opor
tiona
lhaz
ard
mod
el(C
olum
nII
fort
1−t 0
and
Col
umn
VI
fort 2−t 1
),th
eW
eibu
llre
gres
sion
(Col
umn
III
fort 1−t 0
and
Col
umn
VII
fort 2−t 1
),an
dth
eW
eibu
llre
gres
sion
inth
eac
cele
rate
dfa
ilure
time
(AFT
)m
etric
(Col
umn
IVfo
rt 1−t 0
and
Col
umn
VII
Ifort 2−t 1
).In
the
Cox
prop
ortio
nalh
azar
dm
odel
s(C
olum
nsII
and
VI)
,we
fith
(t|x
j)=
h0
(t)exp
(xjβ
x),
whe
reh
0(t
)is
base
haza
rd,x
jis
the
expl
anat
ory
varia
ble
for
lend
erj,
andβ
xis
the
haza
rdra
tefo
rva
riabl
ex
.C
olum
nsII
and
VIr
epor
tha
zard
rate
.A
posi
tive
haza
rdra
tein
dica
tesi
ncre
asin
gha
zard
(red
uced
time
tofa
ilure
)for
incr
easi
ngx
,and
ane
gativ
eha
zard
rate
indi
cate
sdec
reas
ing
haza
rd(in
crea
sed
time
tofa
ilure
)fo
rin
crea
sing
x.
InW
eibu
llre
gres
sion
s,w
ees
timat
e,h
(t|x
j)
=ptp−
1exp
(xjβ
x).
We
repo
rtha
zard
rate
san
dlog(p
)in
Col
umns
III
and
VII
.Acc
eler
ated
failu
retim
etr
ansf
orm
sth
eW
eibu
llha
zard
mod
elin
tolo
g(t
j)=
xjβ
x+β
0+u
j,w
hereu
jfo
llow
sth
eex
trem
eva
lue
dist
ribut
ion.
The
colla
tera
lass
etcl
ass
(Gov
ernm
ent,
Cor
pora
teB
ond,
MB
S,C
DO
,Oth
erSt
ruct
ured
Fina
nce)
fixed
-effe
cts
are
incl
uded
,but
not
repo
rted
inth
eta
ble.
For
dura
tion
mod
els,
robu
stst
anda
rder
rors
are
used
.St
anda
rder
rors
are
show
nin
pare
nthe
ses
and
are
sign
ifica
ntat
the
1%(*
**),
5%(*
*),a
nd10
%(*
)le
vels
.
Dep
ende
ntVa
riabl
esIn
itial
Res
pons
e(t
1−t 0
)Le
nder
Patie
nce
(t2−t 1
)
‘(I
)(I
I)(I
II)
(IV
)(V
)(V
I)(V
II)
(VII
I)
Log(
Prin
cipa
l):Le
nder
-4.8
4840
***
0.13
044*
**0.
1226
3***
-0.0
5141
***
4.62
031*
**-0
.199
94**
*-0
.353
24**
*0.
4608
2***
[0.8
2811
][0
.018
45]
[0.0
1821
][0
.007
67]
[0.7
5678
][0
.034
07]
[0.0
6975
][0
.081
15]
Log(
Rel
atio
nshi
p)-4
.459
04-0
.127
320.
1573
9-0
.065
995.
0296
4-0
.141
78-0
.176
210.
2298
7[4
.469
33]
[0.0
9101
][0
.110
06]
[0.0
4615
][4
.084
35]
[0.1
7966
][0
.269
21]
[0.3
5205
]Lo
g(P
rinci
pal):
Loan
-4.3
8985
***
0.17
739*
**0.
1530
5***
-0.0
6417
***
1.81
863*
*-0
.051
84*
-0.0
9465
***
0.12
347*
**[0
.812
05]
[0.0
2959
][0
.032
78]
[0.0
1369
][0
.742
10]
[0.0
2973
][0
.028
41]
[0.0
3770
]Sh
ort-
term
-5.1
6587
**0.
1313
60.
0967
1-0
.040
546.
4495
2***
-0.4
1266
***
-0.2
3701
***
0.30
920*
**[2
.304
25]
[0.0
9078
][0
.096
06]
[0.0
4037
][2
.105
77]
[0.0
8702
][0
.083
61]
[0.1
0887
]C
onst
ant
256.
6672
0***
-15.
8819
4***
6.65
850*
**-1
33.5
0409
***
7.77
598*
**-1
0.14
426*
**[3
7.10
338]
[1.1
1024
][0
.442
50]
[33.
9073
7][2
.164
32]
[2.7
4054
]
Log(
p)0.
8692
9***
0.86
929*
**-0
.265
87**
*-0
.265
87**
*[0
.024
60]
[0.0
2460
][0
.036
22]
[0.0
3622
]
Mod
elLi
near
Cox
Wei
bull
Wei
bull
(AFT
)Li
near
Cox
Wei
bull
Wei
bull
(AFT
)Lo
gps
eudo
liklih
ood
-326
0.03
78-4
12.5
3521
-412
.535
21-3
183.
1178
-109
6.50
19-1
096.
5019
Wal
dch
i213
1.62
171.
3314
7.99
238.
7834
6.12
1076
.7P
rob>
chi2
00
00
00
R-s
quar
ed0.
1487
60.
1364
3O
bser
vatio
ns58
458
458
458
458
458
458
458
4
35
Table 7. The Effect of Collateral on Initial Response and Lender Patience: This tables shows theeffect of the separability of the assets used as collateral on Initial Response and Lender Patience. All regressions usethe Weibull accelerated failure time (AFT) model, where robust standard errors are used. In Weibull regressions,we estimate, h (t|xj) = ptp−1exp (xjβx). The accelerated failure time transforms the Weibull hazard model intolog (tj) = xjβx +β0 +uj , where uj follows the extreme value distribution. Structured Finance is an indicator variablethat is one if the collateral is a CDO, MBS, or other structured finance product (CLO or ABS, etc.). Corporate isone if the collateral is a corporate bond. CDO is one if the collateral type is a collateralized debt obligation. MBS isone if the collateral type is mortgage backed security. Each credit rating indicator (AAA, AA, BBB, BB, and B) isone if the collateral has the corresponding credit rating as of t0. Log(Principal):Loan is the natural logarithm of theloan amount. Short-term is one if the loan maturity is below the median, and zero otherwise. We report log(p) forthe hazard rate estimate of the Weibull model. Standard errors are shown in parentheses and are significant at the1% (***), 5% (**), and 10% (*) levels.
Dependent Variables Initial Response (t1 − t0) Lender Patience (t2 − t1)
(I) (II) (III) (IV) (V) (VI)
Structured Finance -0.57793*** 3.91747***[0.06105] [0.16019]
Corporate -0.45295*** 3.41386***[0.06516] [0.19259]
CDO 0.00199 0.37254***[0.05098] [0.12820]
AAA 0.13995** -0.41298***[0.06071] [0.14542]
AA 0.10884* -0.14962[0.06503] [0.15373]
BBB -0.05597 -0.04055[0.09674] [0.16424]
BB -0.15481 0.34959[0.23186] [0.28790]
B -0.65455*** 0.38069***[0.05623] [0.10601]
Log(Principal): Lender -0.05205*** -0.07058*** -0.04883*** 0.45522*** 0.63742*** 0.46795***[0.00757] [0.00876] [0.00847] [0.08272] [0.04347] [0.08407]
Log(Relationship) -0.08591** -0.11431** -0.12705*** 0.23334 0.97181 0.48052[0.04376] [0.05821] [0.03763] [0.35044] [0.65883] [0.46473]
Log(Principal): Loan -0.06097*** -0.04985*** -0.06516*** 0.13895*** 0.05091 0.14267***[0.01345] [0.01636] [0.01490] [0.03501] [0.03739] [0.04119]
Short-term -0.04921 0.01368 -0.02887 0.28779*** 0.0681 0.15544[0.03844] [0.04551] [0.04045] [0.10374] [0.10885] [0.10801]
Constant 7.35287*** 7.12746*** 6.94956*** -14.26366*** -17.99945*** -12.05268***[0.47526] [0.57300] [0.42170] [2.83307] [4.65337] [3.39690]
Log(p) 0.86750*** 0.84110*** 0.85452*** -0.26945*** -0.16351*** -0.24110***[0.02441] [0.02617] [0.02574] [0.03597] [0.04790] [0.04003]
Log pseudoliklihood -413.75 -287.95 -350.75 -1098.20 -723.20 -909.89Wald chi2 147.90 85.58 1871.64 1013.51 260.15 299.19Prob >chi2 0.00 0.00 0.00 0.00 0.00 0.00Observations 584 400 491 584 400 491
36
Appendix-A. Model
In this section, we develop a theoretical framework to offer an insight into collective lender behav-
ior. Specifically, in a similar spirit to Morris and Shin (2004), we propose a lender coordination
channel through which lenders exhibit borrower-friendly behavior through collateral requirement
determination, i.e., the haircut. While data limits us to providing direct empirical evidence that
the observed haircut dynamics are driven by the strategic interaction of lenders, the model provides
general implications consistent with our empirical findings presented in later sections.
A. Model Setup
At date 0, a firm borrows from a continuum of lenders (total mass of 1) to finance a project. At date
2, the debt contract pays 1 +R if the debt rolls over and the project succeeds, and pays zero if the
project fails. If a lender chooses not to rollover at date 1, he collects the collateral, K0 ∈ [0, 1].12
The notation of K0 emphasizes the fact that the collateral level was predetermined at date 0, before
the intermediate (date 1) signal is realized.13 The success of the project depends on the number of
lenders rolling over and the underlying economic fundamental, θ. Each lender independently and
simultaneously decides whether to rollover or run (foreclose) the loan. We denote the fraction of
lenders deciding to not to rollover as l1 ∈ [0, 1]. The project succeeds if enough lenders choose to
roll over the loan, θ ≥ l1, and fails if too many lenders choose to foreclose the loan, θ < l1. Note
that 0 ≤ l1 ≤ 1, and hence, the project will always succeed if the economic fundamental is very
strong (θ ≥ 1) and will always fail when the economic fundamental is very weak (θ ≤ 0).
We assume that prior beliefs about the current period’s economic fundamental, θ, are normally
distributed with mean y0 and standard deviation σ0 (i.e., precision is τ0 = 1(σ0)2 ). While lenders
cannot directly observe the economic fundamental, they receive a private signal, x1 = θ + σ1 · ε1,
where ε1 follows a standard normal distribution (i.e., the precision of σ1·ε1 is τ1 = 1(σ1)2 ). We assume
12Figure A.1 shows the relationship between haircut and collateral in this model.13Here we assume that, upon deciding not to rollover, the lender seizes the collateral and liquidates it. In reality,
the borrower can liquidate the pledged collateral and repay the lender. However, these two scenarios are economicallyequivalent, as long as termination of the loan requires the borrower to foreclose the pledged asset. In other words,under the assumed restriction that the borrower cannot get any additional capital injection in the interim period,these two are effectively identical.
37
Figure A.1. Haircut and Collateral: At date 0, a borrower pledges asset, A, as collateral. The liquidation ofthe asset at an interim date 1 is inefficient and will only be worth L1, The asset at full maturity (date 2) can generateeither V or L2 depending on the realized state of economy at date 2. Assuming L2 < L1 < A forces lenders to facerisky choices between seizing higher value of collateral (L1) at date 1, or take risk for a gamble at date 2, whose worstcase will yield a lower payoff, L2. Note that majority of collateral in our data are non-government bond (98% suchas CDO, MBS) and incur liquidity cost. Given that the date 1 liquidation value is less than the amount of originalcollateral value (A), lenders may decide to request haircut (H), where 0 ≤ H ≤ A − L1, to reduce risk exposure oftheir loan (L), i.e., H +L = A. Lending L for collateral A(> L) implies overcollateralization at date 0 due to future(dates 1 and 2) possible depreciation of collateral value. Rescaling L − L2 = 1, L2 = 0, and K = L1−L2
L−L2transform
haircut choice into an equivalent collateral choice (K) of a unit loan.
A
H
L
L2
L1
K
1
Original scale(Loan size=L)
Modified scale(Loan size=1)
Le
that ε1 is identical and independently distributed across lenders. After receiving the private signal,
each agent’s updated posterior belief on the economic fundamental follows a normal distribution
with mean y1 = τ0y0+τ1x1τ0+τ1
and precision τ0 + τ1.
B. Endogenous Choice of Collateral (K0)
Once news on the economic fundamental arrives, each lender updates their belief on the economic
fundamental from their prior one, specified by a normal distribution, with a mean of y0 and τ0.
Based on their belief and the information structure at the moment that the rollover decision (rollover
or run) has to be made, each lender sets their collateral amount (K0) to maximize the expected
38
payoff from the rollover decision and subsequent payoff upon completion of the project. When the
collateral level is endogenously determined, similar steps as in Proposition 3 show that the critical
state is proportional to collateral amount, θ∗0 = K01+R (see Corollary 1 and its proof for details). Since
the critical state increases in collateral level (K0), lenders will endogenously determine the collateral
requirement (K0) by taking into account that a higher collateral level makes a run (because of a
higher critical state) more likely and reduces the chance of receiving a reward upon success (1+R).
Proposition 1 characterizes the optimal choice of collateral by lenders using their prior belief as
follows:
PROPOSITION 1: [Optimal Collateral Choice] The endogenous choice of collateral takes extreme
values, i.e., either K0 = 0 or K0 = 1. When R = 0, lenders’ optimal choice of collateral is K0 = 1.
When R > 1, lenders’ optimal choice of collateral is K0 = 0. When 0 < R < 1, there is a switching
state (economic fundamental), θ0, at which lenders are indifferent between either of the (optimal)
extreme collateral choices, K0 = 0 or K0 = 1. Below this switching state, lenders prefer K0 = 1,
and above it lenders prefer K0 = 0.
The first part of Proposition 1 describes the convex nature of this optimization problem: the
expected payoff in case of default increases with collateral, and expected payoff in case of non-
default decreases with collateral. Figure A.2 shows an example of the expected payoff for different
choices of collateral (K0). When the lender’s prior belief is that the economic fundamental (y0) is
low, a payoff from default is more likely, and full collateral maximizes the expected payoff (solid
line in the figure). On the contrary, when the lender’s prior belief is that the economic fundamental
(y0) is high, the payoff from non-default is more likely, and no collateral becomes the optimal
choice. The gradual change in collateralization can be expressed in terms of switching state, θ0,
which is the value of the economic fundamental at which lenders are indifferent between having no
collateral (K0 = 0) and full collateral (K0 = 1). When the true economic fundamental falls below
this switching state, lenders prefer K0 = 1; otherwise lenders prefer K0 = 0.
The second part of Proposition 1 shows that such a switching state exists when 0 < R < 1. 14 Panel14The proof shows that there is a unique switching state if 0 < R < 1, because V (y0, K0 = 1) − V (y0, K0 = 0)
starts above zero and has a negative slope at y0 = 0 and eventually becomes zero at y0 = ∞. The optimalityconditions of this function is the intersection of two scaled normal probability distribution functions, which intersectwhen 0 < R < 1, implying that the curve will have a minimum at a finite y0. The only possible shape that starts
39
Figure A.2. Expected Payoff vs. Collateral Size: This figure shows the lender’s ex-ante expected payoff(with no cost of collateral) as a function of collateral level (K0). A pessimist, with a low prior on the economicfundamentals (y0 = 0), is shown with the solid line, and the optimist, with a high prior (y0 = 1), is shown with thedashed line.
A in Figure A.3 shows the difference in the expected payoff from full collateral, V (y0, K0 = 1), and
zero collateral, V (y0, K0 = 0), with varying prior beliefs on the economic fundamental, y0. When
R = 0 (the solid line), then the payoff upon the successful completion of the project is the one that
is the same amount as the loan. Any positive probability of project failure (with a payoff that is
less than one) makes lenders lose money on average. Hence, full collateralization is optimal. When
R > 1 (the dash-dot line), the lender’s gain from continuing (and receiving the reward, 1+R, upon
success) far outweighs the immediate seizure of collateral. Hence, no collateralization is optimal.
With the intermediate value of the reward upon success, 0 < R < 1 (dashed line), the expected
payoff from full collateral outweighs that of no collateral when the economic fundamental (or the
prior belief about the economic fundamental, y0) is zero. As the economic fundamental improves,
the payoff from no collateral becomes larger and, at some value of y0, the expected payoffs from
full and zero collateral become identical, i.e., V (y0, K0 = 1)− V (y0, K0 = 0) = 0.
Furthermore, Panel B of Figure A.3 compares switching states for a given prior belief with less
above zero, and that has the minimum value at a finite value, and that approaches zero at infinity is the dashed curvein Figure A.3.
40
Figure A.3. Difference in Expected Payoff from Full and Zero Haircuts: This figure shows the differencein expected payoff from choosing full collateral and zero collateral. Panel A shows the effect of R (R = 0, 0 < R <1, R > 1). Panel B shows the effect of uncertainty (σ = 0.25 and σ = 0.5).
𝑉 𝑦#$, 1 −𝑉 𝑦#$, 0
R=0
0<R<1
R>1
𝜃
(a) Effect of R (R = 0, 0 < R < 1, R > 1) (b) Effect of Uncertainty (σ = 0.25 and σ = 0.5)
(σ = 0.25) and more (σ = 0.5) uncertainty. It shows that, with higher uncertainty in the lender’s
prior belief, the switching state is larger. Even moderately negative news, y0, can lead lenders to
switch from zero to full collateral. This is because with great uncertainty, precision (τ t−1) becomes
small, and hence it needs a larger value of y0 to obtain the same value of√τ0(y0 − 1
1+R
)or√τ0y0,
which determines expected payoffs.
Finally, we show that the destabilizing feedback of collateral (i.e., higher collateral level leads to
a higher probability of a run) makes lenders collectively patient. Without the feedback, lenders
would have required equal or higher level of collateral. We summarize this in the next proposition:
PROPOSITION 2: [Patient Lenders] An optimal level of collateral with the feedback effect be-
tween collateral and critical state θ∗0 = K01+R is less than, or equal to, the case without the feedback
effect in which θ∗0 is independent of K0.
The intuition of Proposition 2 is straightforward: if θ∗0 is independent of collateral K0, i.e., there is
no feedback from collateral to the critical state, then, as shown in the proof, full collateralization
is always optimal (K0 = 1). In other words, the feedback effect of collateral (K0) on the critical
state (θ∗0 = K01+R) lowers the optimal collateral level if beliefs about the economic fundamental are
sufficiently strong (i.e., K0 = 0 if y0 > θ0), and lenders become more lenient and more patient upon
receiving of negative news.
41
C. The Model’s Implications for Action Timing
We summarize the rollover game: Proposition 1 shows that lenders endogenously choose collateral
requirements depending on their prior beliefs about the economic fundamental y0: if y0 < θ0, then
lenders set the optimal collateral level to K0 = 1 (full collateral); otherwise, lenders set the optimal
collateral level to K0 = 0 (no collateral). After observing their private signal x1, lenders follow the
triggering strategy (θ∗1 and x∗1), according to Proposition 3.
The key insight of this result is that knowing that larger collateral can exacerbate a self-fulfilling
run (the feedback channel), lenders are incentivized to keep the haircut at a lower level, rather
than immediately require a higher haircut upon the first arrival of negative news. Proposition 2
formalizes this observation, that the feedback between collateral and critical state tend to decrease
the ex-ante collateral choice relative to the case without the feedback channel.
The model has several implications for action timing. For this task, we relate the switching state (θ0)
to the timing of two lender actions as suggested by Figure 1. In the context of the coordination game,
we are interested in a condition under which lender coordination to lower collateral requirements
can begin. To ease the interpretation, we define the following variables:
• Initial Response: The time it takes for a lender to switch to a zero collateral requirement
from a full collateral requirement, starting from an arbitrary point in time.
• Lender Patience: The time it takes for a lender to switch to a full collateral requirement from
a zero collateral requirement, starting from an arbitrary point in time.
Furthermore, we consider sequential repo contracts (each with a duration of ∆t) with an inde-
pendent random draw of the economic fundamental (θ) from a cumulative distribution, F (θ0) =
Prob(θ < θ0) for each repo contract duration.15
C.1. Expected Initial Response and Expected Lender Patience
To estimate the expected Initial Response, we start with a state of full collateral and then compute
the expected time until lenders switch to zero collateral when a new repo contract rollover is15This section heuristically links the switching theta and action timing. Formal analysis involving repeated games
requires further analysis similar to Angeletos et al. (2007), which is beyond the scope of this paper.
42
repeated with a random draw of the economic fundamental. When the economic fundamental is
above the switching state (θ0), lenders will switch from full collateral (K0 = 1) to zero collateral
(K0 = 0), and the Initial Response can be measured by the survival time of the full collateral state.
Note that the collateral decision is repeated every ∆t. Denoting the probability of maintaining full
collateral as
p = F (θ0) = Prob(θ < θ0),
we can derive the expected survival time of full collateral (or Initial Response) as follows:
E [Initial Response] = ∆t×{
1 + p+ p2 + · · ·}
= ∆t1− p. (A.1)
We follow similar steps to estimate the expected Lender Patience. Here, we start with a state of
zero collateral and compute the expected time until lenders switch to full collateral when a new
repo contract rollover is repeated with a random draw of the economic fundamental. When the
economic fundamental falls below the switching state (θ0), lenders will switch from zero collateral
(K0 = 0) to full collateral (K0 = 1), and Lender Patience can be measured by the survival time
of the zero collateral state. The resulting expected survival time of the zero collateral state (or
Lender Patience) is obtained as follows:
E [Lender Patience] = ∆t×{
1 + (1− p) + (1− p)2 + · · ·}
= ∆tp. (A.2)
C.2. Empirical Implications
Interpretation of the theoretical results with respect to the time domain provides implications that
are potentially observable. Without the feedback channel, full collateral is always optimal. This
implies that θ0 is infinitely large, i.e., p = 1. The Initial Response will therefore also be infinity
according to Equation (A.1), implying that a drop in the haircut will never happen. Also, Lender
Patience will be ∆t by Equation (A.2), indicating that lenders would immediately require full
collateral after the current contract with zero collateral requirement matures (after ∆t).
On the other hand, Proposition 2 predicts a stark contrast to the benchmark case when the feedback
43
effect is taken into the consideration. The proposition shows that, in the presence of destabilizing
feedback between collateral and the critical state, the switching state (θ0) is finitely smaller, ob-
taining 0 < p < 1. In this case, there will be a point in time when lenders require zero collateral
(i.e., Initial Response < ∞), and lenders will not immediately require full collateral (i.e., Lender
Patience > ∆t).
44
Appendix-B. Proofs
A. The Effect of Collateral on Lender Coordination
Once collateral is set at date 0 toK0, lenders make rollover decisions based on the private signal (x1).
Following steps similar to those in Morris and Shin (2004), the equilibrium can be characterized by
a triggering strategy, which is summarized in the following proposition:
PROPOSITION 3: Suppose that the lender’s signal satisfies τ0√τ1<√
2π. For a given collateral
level, K0, there exists a unique Bayesian Nash Equilibrium in which all lenders with a signal larger
than x∗1 roll over the loan and all others foreclose. The investment project succeeds if and only if
the economic fundamentals are above the threshold, 0 ≤ θ∗1 ≤ 1, which is given by
θ∗1 = Φ(τ0√τ1
(θ∗1 − y0 −
√τ0 + τ1τ0
Φ−1(
1− K01 +R
))). (B.1)
The corresponding critical private signal, x∗1, is given by
Φ (√τ1 (x∗1 − θ∗1)) = θ∗1. (B.2)
In addition, the trigger state is increasing in collateral amount (K0) and decreasing in payoff upon
success (R), i.e., ∂θ∗1(K0)∂K0
≥ 0 and ∂θ∗1(R)∂R ≤ 0.
The first part of Proposition 3 states that there is a critical state (economic fundamental) above
which all lenders expect the project to succeed. Accounting for the noise in the signal each lender
receives, lenders will rollover the repo loan when they receive a signal above x∗1. The critical state, θ∗1
(given by Equation (B.1), and illustrated in Figure B.1) is central to understanding the equilibrium
of the economy, and it is closely related to collateral level K0 and reward upon success (R), the
dependence of which are described in the second part of the proposition.
The second part of Proposition 3 indicates that the critical state below which a run occurs is
increasing in the level of collateral (K0) and decreasing in the amount of reward upon success
(R). The key insight from this proposition arises from a lender’s tradeoff when making a rollover
decision. The state at which lenders are indifferent between foreclosure and rollover is reached
45
Figure B.1. An Illustration of Determining θ∗1 in Equation (B.1): This figure shows the procedure tosolve Equation (B.1) for the critical state, θ∗1, which is found by the intersection of the 45o line (solid line) and thecumulative normal distribution function (dashed line).
y 𝜃# = Φ��(𝜏#� 𝜃#− 𝑦,( −
��( +𝜏#�
��(Φ.# 1 −
𝐾(1 +𝑅
y 𝜃# = 𝜃#
𝜃#∗
Economic Fundamentals (𝜃#)
y(𝜃#)
when the expected payoff from liquidating the collateral (foreclosing) is the same as the expected
payoff from rolling over and receiving payoff 1 +R when the project succeeds. When the collateral
is set very high, then even with a strong economic fundamental, lenders are tempted to seize the
collateral rather than take the risk to receive the reward upon the project’s success. This tradeoff is
described in the equilibrium outcome: a high collateral level (K0) pushes the critical state upward.
Likewise, when the payoff upon success (1 + R) is large, lenders will tolerate the risk of rolling
over even with a low economic fundamental. That is, a large R pushes the critical state downward.
Lenders choose the collateral to maximize their payoff from lending (or minimize their losses from
the borrower’s failure). Therefore, when the economic condition worsens, lenders may choose to
raise collateral levels. However, the fact that high collateral levels also elevate the critical state (i.e.,
runs are more likely to occur even when economic fundamentals are strong) makes the inefficient
run more likely. To avoid such self-fulfilling runs, lenders may choose to set ex-ante collateral lower
than a case without such destabilizing feedback, which is explored in the next subsection.16
16Endogenous contracts accounting for interim period hold up problems have been considered in other applicationssuch as product market competition (Khanna and Schroder, 2010; Khanna and Tice, 2000, 2005) and venture capitals(Khanna and Mathews, 2015).
46
B. Proof of Proposition 3
The first part: The proof follows arguments similar to those in Morris and Shin (2004); The
equilibrium of this economy can be determined by the following two conditions:
(i) Critical Mass Condition:
Conditional on the (true) economic fundamental, θ, the probability that a lender will receive a
signal below x1 < x∗1 is
Prob(x1 < x∗1|θ) = Φ(x∗1 − θσ1
),
where Φ (·) is a standard normal cumulative distribution function. Since the total mass of lenders
is one, this probability is equal to the fraction of lenders foreclosing the loan (l1). The project
succeeds when this fraction, l1, is less than the economic fundamental, θ. Hence, the critical level
of the economic fundamental that is exactly enough to succeeds is l1 (θ∗) = θ∗1, i.e.,
Φ (√τ1 (x∗1 − θ∗1)) = θ∗1. (B.3)
This condition is often referred to as the critical mass condition in the global games literature
(Corsetti, Dasgupta, Morris, and Shin, 2004).
(ii) Optimal Trigger Condition:
For a given economic fundamental threshold, θ∗1, a lender who receives a signal, x1, has a conditional
probability of successful rollover,
Prob (θ < θ∗1|x1) = 1− Prob (θ ≥ θ∗1|x1) = 1− Φ(θ∗1 − x1σ1
),
and hence will only decide to roll over the loan if the expected payoff from the rollover (i.e., the
probability of a successful rollover, in which the project succeeds in time and the lender receives
the payoff that results from the project’s success, 1 + R) is at least as large as the payoff from
foreclosure (K0). The critical threshold signal (x∗1) is the point above which a lender is indifferent
between rolling over and foreclosing to collect the collateral. For a given economic fundamental
threshold, θ∗1, a lender who receives a signal, x1, has a posterior distribution of θ being a normal
47
distribution with mean y1 = τ0y0+τ1x1τ0+τ1
and precision τ0 + τ1. Also, each lender knows that the
project succeeds if and only if the economic fundamental is above θ∗1. This leads to the following
optimal trigger condition:
{1− Φ
(√τ0 + τ1
(θ∗1 −
τ0y0 + τ1x1τ0 + τ1
))}· (1 +R) = K0.
This condition can be written as
θ∗1 = τ0y0 + τ1x1τ0 + τ1
+Φ−1
(1− K0
1+R
)√τ0 + τ1
which obtains
(τ0 + τ1) θ∗1 = τ0y0 + τ1x1 +√τ0 + τ1Φ−1
(1− K0
1 +R
)
that leads
θ∗1 − x∗1 = −τ0 (θ∗1 − y0)τ1
+√τ0 + τ1Φ−1
(1− K0
1+R
)τ1
.
Substituting this into the critical mass condition gives,
θ∗1 = Φ(τ0√τ1
(θ∗1 − y0 −
√τ0 + τ1τ0
Φ−1(
1− K01 +R
))). (B.4)
As shown in Figure 5, the LHS (θ∗1) is a 45 degree line passing through the origin. The RHS will
cross this 45 degree line once if the slope of the LHS is less than one. The slope of the RHS is
given by τ0√τ1φ (·) where the maximum value of φ (·) is 1√
2π . Hence, the RHS’s slope is less than
one everywhere if τ0√2πτ1
< 1.
The second part: According to Equation (B.4) (and illustrated in Figure B.1), the trigger state
is determined by the intersection of a 45 degree line (θ∗1) and the normal distribution function. The
48
argument inside the cumulative distribution function is
z1 = τ0√τ1
(θ∗1 − z3)
z2 = 1− K01 +R
z3 = y0 +√τ0 + τ1τ0
Φ−1 (z2) .
Note that z2 is decreasing in collateral and increasing in payoff upon success, i.e., ∂z2(K0)∂K0
= − 11+R <
0 and ∂z2(R)∂R = K0
(1+R)2 > 0. Since the inverse cumulative distribution function is smooth, continuous,
and monotonically increasing between -1 and 1, Φ−1 (z2) decreases in collateral (K0) and increases
in payoff upon success (R), i.e., ∂Φ−1(z2)∂K0
= ∂Φ−1(z2)∂z2
∂z2(K0)∂K0
< 0 and ∂Φ−1(z2)∂R = ∂Φ−1(z2)
∂z2∂z2∂R > 0.
Therefore, z3 is decreasing in collateral and increasing in payoff upon success:
∂z3∂K0
=√τ0 + τ1τ0
∂Φ−1 (z2)∂K0
< 0
∂z3∂R
=√τ0 + τ1τ0
∂Φ−1 (z2)∂R
> 0.
Now, Equation (B.4) takes the form
θ∗t = Φ(τ0√τ1
(θ∗1 − z3)).
If there is an increase in collateral decreases z3, which shifts the cumulative distribution function,
Φ(τ0√τ1
(θ∗1 − z3)), to the left as shown in Figure 5. As a result, the intersection of the 45 degree
line and the cumulative distribution function (θ∗1) shifts to the right, i.e., ∂θ∗1(K0)∂K0
≥ 0. Similarly, an
increasing payoff upon success (R) increases z3, which shifts the cumulative distribution function,
Φ(τ0√τ1
(θ∗1 − z3)), to the right as shown in Figure 5. As a result, the intersection of the 45 degree
line and the cumulative distribution function (θ∗1) shifts to the left, i.e., ∂θ∗1(R)∂R ≤ 0. QED �
49
COROLLARY 1: Trigger strategies based on prior beliefs are as follows:
θ∗0 = K01 +R
x∗0 = K01 +R
−Φ−1
(1− K0
1+R
)√τ0
.
Proof of Corollary 1
Denoting the trigger strategy and trigger state using prior beliefs as x∗0 and θ∗0, we can derive the
equilibrium as follows:
(i) The critical mass condition (given the true economic fundamental, θ, the probability of lenders
below x∗0 must be less than or equal to θ. This holds with equality at the critical state)
Φ(√
τ0(x∗0 − θ
∗0
))= θ
∗0
from which we obtain
x∗0 = θ∗0 +
Φ−1(θ∗0
)√τ0
. (B.5)
(ii) The optimal trigger condition (the probability that the economic fundamental is above the
trigger state, which is normal with mean y0 and precision τ0)
{1− Φ
(√τ0(θ∗0 − x∗0
))}· (1 +R) = K0.
This condition can be written as
θ∗0 = x∗0 +
Φ−1(1− K0
1+R
)√τ0
.
Plugging that into critical mass condition in Equation (B.5) gives
θ∗0 = Φ
((−Φ−1
(1− K0
1 +R
)))= 1− Φ
((Φ−1
(1− K0
1 +R
)))= 1−
(1− K0
1 +R
)
50
and
θ∗0 = K0
1 +R.
QED �
C. Proof of Proposition 1
The first part: The optimal collateral choice at date 1 maximizes the expected payoff for the
next period, date 2, based on information held before observing the private signal, i.e., using only
prior beliefs.
The expected payoff can be broken down into two terms: the expected payoff in case of default (if
the creditors run), VD, and in case of non-default (if the project succeeds), VND, which are shown
as:
VD = K0Φ(√
τ t−1(θ∗0 − y0
))VND =
{1− Φ
(√τ0(θ∗0 − y0
))}· (1 +R) ,
where Φ(√
τ0(θ∗0 − y0
))is the probability that the true economic state (based on prior beliefs;
this is normally distributed with mean y0 and precision τ0) is below trigger state (θ∗0), i.e., it is the
probability of failure.
The first term (VD) is increasing in collateral because
∂θ∗0
∂K0= ∂
∂K0
(K0
1 +R
)= 1
1 +R> 0,
and∂Φ
(√τ0(θ∗0 − y0
))∂K0
=∂Φ
(√τ0(θ∗0 − y0
))∂θ∗0
∂θ∗0
∂K0> 0,
and∂VD
∂K0= Φ
(√τ0(θ∗0 − y0
))+ K0
∂Φ(√
τ0(θ∗0 − y0
))∂K0
> 0. (B.6)
51
The second term (VND) is decreasing in collateral because
∂VD
∂K0= − (1 +R)
∂Φ(√
τ0(θ∗0 − y0
))∂K0
< 0. (B.7)
Adding these two terms in Equation (B.6) and (B.7) leads to a convex payoff function whose
maximizing choice is one of the extreme collateral choices. That is, K0 = 0 if the second (non-
default payoff) is larger, or K0 = 1 if the first (default payoff) is larger.
The second part: The expected payoff for a given K0 as a function of the economic fundamental
based on prior belief (y0) is
V(y0, K0
)= K0Φ
(√τ0(θ∗0 − y0
))+{
1− Φ(√
τ0(θ∗0 − y0
))}· (1 +R)
= K0 ·{
1− Φ(√
τ0(y0 − θ
∗0
))}+ Φ
(√τ0(y0 − θ
∗0
))· (1 +R)
or, substituting for the prior-belief-based critical economic fundamental that triggers run, θ∗0 = K01+R ,
is
V(y0, K0
)= K0 ·
{1− Φ
(√τ0
(y0 −
K01 +R
))}+ Φ
(√τ0
(y0 −
K01 +R
))· (1 +R) .
Considering both optimal choices of haircuts, when K0 = 0 (zero haircut), it becomes,
V (y0, 0) = Φ(√
τ0·y0
)· (1 +R) , (B.8)
and when K0 = 1 (full haircut), it becomes
V (y0, 1) ={
1− Φ(√
τ0
(y0 −
11 +R
))}+ Φ
(√τ0
(y0 −
11 +R
))· (1 +R) . (B.9)
Using Equation (B.8) and (B.9), we define the difference in the expected payoff from these haircut
52
choices as
∆V1−0 (y0) = V (y0, 1)− V (y0, 0)
={
1− Φ(√
τ0
(y0 −
11 +R
))}+ Φ
(√τ0
(y0 −
11 +R
))· (1 +R)
−Φ(√
τ0·y0
)· (1 +R)
= 1 +R · Φ(√
τ0
(y0 −
11 +R
))− Φ
(√τ0·y0
)· (1 +R) .
and its derivative with respect to the economic fundamental, y0, is
d (∆V1−0 (y0))dy0
=√τ0R · φ
(√τ0
(y0 −
11 +R
))−√τ0φ
(√τ0·y0
)· (1 +R)
The first order condition (in this case, the value of y that minimizes the expected payoff) becomes
R · φ(√
τ0
(y0 −
11 +R
))− φ
(√τ0·y0
)· (1 +R) = 0.
For R = 0, the first order condition (FOC) is
−φ(√
τ0·y0
)< 0,
i.e., FOC has no solution, and ∆V1−0 (y0) monotonically decreases. For R > 0, FOC has one
solution as shown in Figure A.1 (R = 0.5, τ0 = 2). This is because the first term is a rescaled
(by R < 1) bell-shaped function shifted by 11+R from the origin, and the second term is a rescaled
(by 1 + R > 1) bell-shaped function centered at origin. The latter is a monotonically decreasing
function from 1+R√2π to zero at range yt−1 > 0, and the former has a positive value everywhere with
a peak value of R√2π hence, both terms will cross each other at a point as long as R > 0. Note that
as R approaches zero, the peaks of both bell-shaped curves will approach each other and the first
term will be centered near zero. Hence, in the limit (of R = 0), they will coincide each other.
The slope at zero economic fundamental is negative,
d (∆V1−0 (y0))dy0
=√τ0R · φ
(√τ0
(− 1
1 +R
))−√τ0φ (0) · (1 +R) < 0
53
because φ (·) is at its maximum at zero (i.e., φ (0) > φ(√
τ0(− 1
1+R
))) and R < 1 +R. Hence, the
second term is larger than the first term.
The difference in the expected payoff, ∆V1−0 (y0), at y0 = 0 is positive if 0 < R < 1, which means
that the expected payoff from setting a full haircut (K0 = 1) is larger than that of setting zero
collateral (K0 = 0). To see this, at y0 = 0, we have
∆V1−0 (y0 = 0) = 1 +R · Φ(√
τ0
( 11 +R
))− Φ (0) · (1 +R) .
Since Φ (0) = 12 , we have
∆V1−0 (y0 = 0) = 12 +R ·
{Φ(√
τ0
( 11 +R
))− 1
2
}.
For 0 < R < 1, ∆V1−0 (y0 = 0) is positive because
−12 < Φ
(√τ0
( 11 +R
))− 1
2 < 0.
For R > 1, it is possible for ∆V1−0 (y0 = 0) to be negative which implies that the expected payoff
from zero collateral is higher than for full collateral, even with weak low economic fundamentals
(i.e., y0 = 0).
At the other extreme of y0 → ∞, the difference in the expected payoff approaches zero because
both Φ(√
τ0(y0 − 1
1+R
))and Φ
(√τ0·y0
)approach unity, and we get limyt−1→∞∆V1−0 (y0) =
1 +R− (1 +R) = 0.
For values of 0 < R < 1, ∆V1−0 (y0) starts at a positive value at y0 = 0. Then it monotonically
decreases below zero as y0 increases. When the FOC is zero (which is the case when R > 0),
the trend reverses and ∆V1−0 (y0) monotonically increases as y0 increases. Eventually, when y0 →
∞, ∆V1−0 (y0) approaches zero. For values of R > 1, ∆V1−0 (y0) starts at a negative value at
y0 = 0. Then it monotonically decreases below zero as y0 increases, and reverses the trend whend(∆V1−0(y0))
dy0= 0. Eventually, it approaches zero as y0 → ∞ (Figure 7). As shown in Panel A of
Figure 7, when R = 0, V1−0 (y0) is always positive and K0 = 1 is optimal. When 0 < R < 1,
54
V1−0 (y0) is initially positive where K0 = 1 is optimal. But for strong economic fundamentals
(y0 > θ), V1−0 (y0) is negative and K0 = 0 is optimal. The switching economic fundamental (θ) is
where V1−0 (y0) = 0, and lenders are indifferent between K0 = 0 and K0 = 1. For R > 1, V1−0 (y0)
is always negative and K0 = 0 is optimal. QED�
D. Proof of Proposition 2
If θ∗0 is independent of collateral, K0 , the expected payoff from zero collateral is
V(y0, K0 = 0
)={
1− Φ(√
τ0(θ∗0 − y0
))}· (1 +R) ,
and the expected payoff from full collateral is
V(y0, K0 = 1
)= Φ
(√τ0(θ∗0 − y0
))+{
1− Φ(√
τ0(θ∗0 − y0
))}· (1 +R) .
The difference between these two expected payoffs is always positive, that is,
∆V1−0 (y0) = V (y0, 1)− V (y0, 0) = 1− Φ(√
τ0(y0 − θ
∗0
))> 0.
Proposition 2 shows that full collateralization is always optimal without feedback effects. In other
words, the feedback effect of collateral (K0) on the critical state (θ∗0 = K01+R) lowers the optimal
collateral level if (for a sufficiently large R and low uncertainty (a large τ0) to have θ < 1) the
belief about the economic fundamental are sufficiently strong (i.e., K0 = 0 if y0 > θ), and lenders
become patient. QED�
55