TAF Effect on Liquidity Risk Exposures∗
Stefano Puddu†
Andreas Wälchli
‡
First Version: March 2011
This Version: May 2012
Abstract
Using a unique dataset, based on micro-banking data, we study banks’ liquidity and
liability features depending on whether banks received credit from the Term Auction
Facility (TAF) program. Moreover, we assess the impact of the TAF program on
banks’ liquidity risk by employing a treatment-effect model with a binary dependent
variable. The results suggest that, on average, banks that benefited from the TAF
program exhibit ex ante higher levels of liquidity risk proxies and illiquid collateral.
These banks drastically reduce their funding liquidity risk positions in the periods
following the first time they received the financial support. Finally, TAF banks exhibit
larger liquidity exposures reduction. This reduction is bigger, the larger the amount
of reserves received. Several robustness checks confirm the main results. Our findings
document that the TAF program was effective at the bank level in decreasing funding
liquidity risk.
Keywords
TAF, Liquidity Risk, Financial Crisis
JEL Classification
G21, G28, G32
∗
This paper previously circulated under the title “Too TAF towards the risk”. The authors are grateful to
Christian Castro, Luca Deidda, Michel Dubois, John V. Duca, Luigi Infante, Rafael Lalive, Antoine Martin,
Cyril Monnet, Judit Montoriol-Garriga, Climent Quintana-Domeque, Florian Pelgrin, Antoni Rubı́ Barceló,
Klaus Schaeck, Pascal St-Amour and Javier Suarez for their useful comments.
†
University of Lausanne and Université de Neuchâtel, e-mail: stefano.puddu@unine.ch
‡
University of Lausanne and Study Center Gerzensee, e-mail: andreas.waelchli@szgerzensee.ch
1
1
Introduction
The bursting of the housing bubble in 2007 led to the most severe financial crisis since the
Great Depression. As banks were forced to write down billions of dollars in bad loans, the
interbank market for short-term funding froze, leaving several banks with severe liquidity
problems. Although these banks were not able to roll over their short-term debt, they
were also reluctant to use the Fed’s traditional channel of the discount window (DW) credit
programs. Notably, banks’ aversion was due to the fact that this strategy could have been
interpreted by the market as a signal of being in financial trouble, therefore intensifying the
pressure on the financial institution.
In this context, the Federal Reserve was directly involved in promoting several extraordinary actions, including the creation of a number of new facilities for auctioning short-term
credit, with the general aim of supporting the financial sector and ensuring that financial
institutions have adequate access to liquidity. One of these programs was the Term Auction
Facility (TAF), which operated from December 2007 to April 2010, when the last loans were
repaid. According to the Fed’s definition, “[The TAF program] could help ensure that liquidity provisions can be disseminated efficiently even when the unsecured interbank markets
are under stress”1 . More explicitly, Taylor (2009) claims, “The main aim of the TAF was
to reduce the spreads in the money markets and thereby increase the flow of credit and
lower interest rates”. The Fed, through the TAF program, was injecting liquidity into the
market, effectively substituting the interbank credit market, and therefore affecting liquidity
risk. TAF-like programs are criticized regarding the quality of the banks that benefited from
the funds and with respect to their effectiveness in decreasing liquidity risk. Among others,
Taylor (2009) claims that “government actions and interventions prolonged and worsened the
financial crisis [...] by focusing on liquidity rather than risk [...] and by providing support
for certain financial institutions but not others.”.
1
http://www.federalreserve.gov/monetarypolicy/files/TAFfaqs.pdf.
2
Our paper aims to assess which type of banks benefited from the program and quantitatively answer the question of whether the TAF program was successful in helping banks
reduce their liquidity risk positions by compensating for the temporary collapse of the interbank credit market. More precisely, by using a unique dataset, which we constructed
by merging TAF program information with balance sheet data of the banks, we compare
liquidity and liability features of banks that received TAF reserves with those that did not.
For banks that received the financial support, we document their funding liquidity risk behaviour before and after the period they received the funds for the first time. Finally, we
assess the impact of the TAF program on liquidity risk changes, measured by the difference
of the liquidity risk before the beginning and after the end of the TAF program. We adopt
a treatment-effect model with a binary endogenous regressor. The answer to these questions
can help in clarifying which type of banks benefited from the TAF program and in assessing
how the TAF program affected banks’ liquidity risk.
We find that banks that benefited from the TAF program exhibit ex ante higher levels
of liquidity risk proxies. These measures indicate that banks with a more severe maturity
mismatch are most exposed to the freezing of the interbank market and are unable to roll
over their short-term liabilities during the crisis, so they are more likely to participate to
the TAF program. Furthermore, we find that the banks benefiting from the reserves reduce
their funding liquidity risk positions in the quarters after receiving the financial support
for the first time. Moreover, the results highlight that ex ante levels of liquidity risk and
illiquid collateral positively affect the probability of receiving funds and that TAF program
participation implies a larger contraction of funding liquidity risk. The bigger the reduction
in liquidity risk, the larger the amount of reserves received. These findings are robust to
the specification of the model, the alternative measures employed to proxy liquidity risk,
the sample period considered, the econometric approach employed, the collapse of Lehman
Brothers, and the uneven composition of the sample due to the relatively small number of
3
banks that obtained the reserves.
Importantly, our findings challenge Taylor’s criticism, by stressing the fact that the TAF
program was useful for those banks with important funding liquidity mismatches. The results highlight that the TAF program was effective because it decreased, at the individual
bank level, funding liquidity risk, alleviating banks’ short-term financing exposure. The TAF
program supported the depository institutions during the time intensive period of restructuring their liability side and helped them in not converting their liquidity issues into solvency
problems. In this sense the Fed, through the TAF program, acted as lender of last resort,
providing liquidity to banks in liquidity distress, avoiding the worsening of the crisis. In
terms of policy implications, our results support the opinion of those that consider the TAF
program as a new countercyclical tool to be included among the other instruments, such as
the discount window credit programs, employed by the Fed to provide liquidity to depository
institutions2 .
Our contribution shows several differences with respect to previous studies. This is the
first study in this topic based on micro-banking data. This allows us to analyse the effect
of the TAF program on liquidity distress from a new perspective. Indeed, we focus on bank
funding liquidity instead of market liquidity. Funding liquidity refers to the ability of a
bank to raise funds by selling assets to meet financial obligations on short notice. Market
liquidity, instead, refers to the ability of a bank to execute transactions in financial markets
without affecting significantly the correspondent prices3 . Specifically, we concentrate on the
effect of TAF program on bank quantities instead of on liquidity risk spreads. The reason is
that prices are also impacted by other factors (e.g., see Michaud and Upper, 2008). During
the financial crisis interest rates rose due to increased uncertainty and higher dispersion of
2
The two elements that distinguish the TAF program from the existing credit programs in normal periods
are that the information about institutions benefiting from the program is not publicly available, and that
the funds are ascribed by using a market mechanism, i.e., an auction system.
3
For further details on funding versus market liquidity, see Brunnermeier and Pedersen (2009), Fontaine
and Garcia (2009), and Allen, Babus and Carletti (2010).
4
credit quality. Moreover, as stressed by Drehmann and Nikolaou (2009) “The spread between
interest rates in the interbank market and a risk free rate is purely a price measure and it
does not reveal anything about market access, which maybe severely impaired during crisis,
nor the volume of net-liquidity demand [...]”. Finally, our choice is also motivated by the
fact that, as recently documented4 , there is the suspicion that the London Interbank Offered
Rate (LIBOR), during the crisis, was misreported by banks such that its informative power
was negatively affected. In general, it is important to note that the LIBOR is not a market
interest rate, but rather the average of the answers of large banks to the question, “At what
rate could you borrow funds, were you to do so by asking for and then accepting interbank
offers in a reasonable market size just prior to 11 a.m.?”5 . Moreover, with respect to previous
contributions, we are not subject to the critique of Taylor and William (2008) concerning
the persistence effect of the TAF program on liquidity risk, that refers to the definition of
the variable capturing the TAF effect. In our approach, we distinguish between banks that
received at some point the reserves provided by the TAF program and the rest of the banks,
and we also take into account the amount of funds received by each bank. Our results,
contrary to the findings of previous studies, are robust to the specification, the liquidity
measure, the sample period, the sample composition, and the methodologies employed.
The literature reports mixed results about the effectiveness of the TAF program on liquidity risk, measured by the spread between the LIBOR and the overnight indexed swap (OIS).
Taylor and William (2009) and McAndrews, Sarkar, and Wang (2008) obtain different results
by employing the same set of explanatory variables6 , but using, as a dependent variable, the
level and the first difference of the liquidity risk measure, respectively. Specifically, the former study finds no impact, while, according to the latter contribution, the TAF program a
4
See for instance, Abrantes-Metz et. al. (2012).
http://www.bbalibor.com/bbalibor-explained/the-basics
6
The variables included refer to the asset-backed commercial paper spread, the credit default swaps for
major banks, the Tibor-Libor spread, the Libor-Repo spread, and a TAF dummy variable, which is one on
each of the TAF bid submission dates and zero elsewhere.
5
5
has negative impact on the liquidity risk spread.
Wu (2008) expands the specification employed in previous contributions by adding a new
set of explanatory variables and assuming that the TAF program has a permanent effect
on LIBOR-OIS spreads. Wu shows that the TAF program decreases liquidity risk spreads;
however, these findings are subject to the Taylor and William (2008) criticisms that the TAF
program has not a permanent effect on spreads.
Cui and Maharaj (2008) distinguish between short-run and long-run TAF effects. They find
that the LIBOR-OIS spread decreases when the TAF is announced, but the effect is not maintained over time. Moreover, according to their results, TAF only affects 3-month spreads.
Sarkan and Shrader (2010) study the impact of TAF changes on 3-month LIBOR-OIS spread
changes by augmenting the specification employed in previous contributions on this topic.
Their results show that changes in the TAF issuance volumes have a negative impact on the
changes in the LIBOR-OIS spread. Moreover, they find that spread changes depend on the
amount of reserves provided.
Our findings may differ from those of previous studies for several reasons. First, we
adopt a micro-data approach, which allows us to avoid potential aggregation effects that can
affect the results. Second, we employ a cross-section dataset instead of using a time-series
perspective, and that we focus on quantities instead of prices, as done instead in previous
contributions. Finally, our study assesses the impact of the TAF program once it was already
over (that is, when all loans were paid back). This is not the case for previous contributions
which investigate the short run effects.
The rest of the paper is organized as follows: In Section 2 we discuss the TAF program
and other similar programs promoted by the Fed during the last financial crisis. The dataset
is analysed in Section 3, while the econometric model is described in Section 4. In Sections
5 we discuss the results. Section 6 concludes.
6
2
Fed facilities during the last financial crisis
During the last financial crisis the Federal Reserve undertook several extraordinary actions,
including the creation of a number of new facilities for auctioning short-term credit, with the
general aim of supporting the financial sector and of ensuring adequate access to liquidity by
financial institutions. In this section, we discuss in detail the TAF program as well as the
other programs launched by the Fed during this period in order to underline their common
points and their main differences.
2.1
Term Auction Facility program: how it works
According to the Fed’s definition, “The TAF is a credit facility that allows a depository
institution to place a bid for an advance from its local Federal Reserve Bank at an interest
rate that is determined as the result of an auction”7 . The aim of the TAF was to compensate
for the collapse of the short-term funding market by ensuring liquidity provisions when the
inter-bank credit market was under stress.
All banks eligible for the discount window credit programs at the moment of the auction
and during the term of TAF loans were also eligible for the TAF8 . The reserves provided in the
TAF program had maturity term of 28 or 84 days, and they had to be fully collateralized.
Banks were allowed to have at the same time more than one loan so that facilities with
different maturities could overlap. The information about banks bidding and receiving funds
was private. Only in December 2010, the Fed was forced to disclose documentation on the
financial aids, after Bloomberg filed a federal lawsuit and won. For each auction the Fed
fixed the total amount to supply, the maximum amount an individual bank was allowed to
obtain, and the minimum bid interest rate (rF ed ). For each auction, eligible banks had the
7
http://www.federalreserve.gov/monetarypolicy/taffaq.htm
The definition employed by the Fed refers to “[banks] in sound financial conditions”. However, this
definition is opaque in the sense that there are not details about it. The soundness of a particular bank has
to be certified by its local Reserve Bank. It refers to bank solvency, liquidity, and profitability.
8
7
possibility to make two rate-amount offers. Specifically, the bid was characterized by the
amount asked by the bank and a repayment interest rate. Bids were ordered according to
the repayment interest rate bid (rBank i ). The Fed then began to accept the bids starting
from those associated with the highest interest rate. It would continue to do so until the
offered amount was reached, otherwise all the bids were accepted. In the former case, the
interest rate that had to be paid by all successful bidders was determined by the stop-out
rate, i.e., by the interest rate of the last accepted bid. If the supply exceeded the demand, the
equilibrium interest rate would simply be equal to the minimum bid rate. The equilibrium
interest rate r∗ was therefore
r∗ =



r
F ed


rbBank i
if Supply > Demand
(1)
if Supply ≤ Demand
where rbBanki is the lowest interest rate that was accepted by the Fed.
During the last financial crisis the normal instruments, such as the discount window
credit programs, employed by the Fed to provide liquidity to depository institutions were
less effective because of the “stigma effect”. Depository institutions were concerned about
the fact that the market would have interpreted as a bad signal (“stigma”), regarding their
financial conditions, the fact of benefiting from the loans provided by the Fed by the normal
discount window credit programs9 . In order to avoid or minimize the stigma effect the Fed
decided to keep private the information regarding the institutions that benefited from the
loans in the framework of the TAF program, and, at the same time, it adopted an auction
mechanism to determine which institutions would obtain the reserves and to establish the
9
The “stigma effect” related to the last financial crisis has been discussed and measured by Armantier
et. al. (2008) and Armantier et. al. (2011) by using TAF program banks bids. They find that in the third
quarter of 2008 banks preferred to pay on average at least 34 basis points more to borrow from the TAF
program than from the DW.
8
repayment interest rate. An auction mechanism such as that described above has several
important advantages in decreasing the potential stigma effect. First, the interest rate is
determined through a market mechanism instead of being imposed by the authorities; second,
banks approach the Fed collectively instead of individually.
−4.1
−4.4
−4.3
−4.2
St Liabilities / St Assets
Billions of $
500
1000
1500
Figure 1: TAF reserves, market events, policy measures and liquidity risk
AIG
Facilities extendend
TAF
Lehman and AIG
TAF 84
Lehman
−4.5
Bear Stearns
0
TAF 28
No TAF
2006q1
2007q1
2008q1
2009q1
2010q1
2011q1
date
TAF supplied
TAF loans
Notes: The figure shows (left-hand scale) the reserves offered and effectively provided in the context of the
TAF program. Moreover, squares (triangles) refer to market (policy) events. Finally, the figure shows (righthand scale) the average level of short-term liabilities over short-term assets, distinguishing by banks group:
TAF and NO TAF.
Figure 1 shows the reserves supplied by the Fed and those effectively provided to the
depository institutions in each quarter under the TAF program (left-hand scale). The graph
highlights that before the collapse of Lehman Brothers, the auctions were competitive. This
is no more the case for the auctions following Lehman Brothers’ collapse: all depository
institutions interested in TAF facilities obtained them, since the Fed doubled the amounts
supplied.
Moreover, Figure 1 reports several market events (squares) and policy measures related to the
TAF program (triangles). The program was announced on December 12, 2007. Specifically,
the initial reserves had a maturity of 28 days. The amount provided was increased in the
9
first quarter of 2008, after Fannie Mae and Freddie Mac requirements were eased to allow
for increases in lending and Bear Stearns received emergency loans from the Fed. Reserves
with longer maturities were established in 2008:Q2, after Lehman Brothers reported losses
of $2.8b. The amount of reserves provided kept rising after Lehman Brothers’ bankruptcy
and the downgrade of AIG debt. The maximum amount was supplied during 2009:Q1, when
Fannie Mae and Freddie Mac reckoned a need for $51b to continue operations and AIG
announced large losses. From 2009:Q2 on, new facilities decreased and lasted until March 8,
2010, when the last auction took place.
The graph also shows (right-hand scale) the average level of short-term liabilities over shortterm assets10 for two group of banks: TAF banks are those that received reserves at least
once, while NO TAF banks have not. Before the beginning of the program the two groups
of banks report different levels of liquidity risk, although for the two groups of banks the
patterns of the series are similar. The differences get smaller after the collapse of Lehman
Brothers. This trend characterizes the series after the end of the TAF program as well. These
features of the series hold when comparing the quartiles (25th, 50th and 75th) of the series
for the two groups of banks, as highlighted by Figure 7.a reported in the Appendix.
2.2
Other facilities
Since 2003, depository institutions have had access to primary credit, secondary credit, and
seasonal credit, three types of discount window (DW) facilities. As for the TAF program,
all regular discount window loans must be fully collateralized with an appropriate haircut
applied to the collateral such that the collateral must exceed the value of the loan.
Before the crisis the primary credit maturity was overnight. With the strengthening of the
crisis the maturity was extended up to 90 days. In February 2010, the maturity was again
10
Short-term liabilities over short-term assets is the main measure of liquidity risk employed in this study.
More details are provided in the following sections.
10
reduced to overnight. Those depository institutions that are not eligible for the primary
credit can ask for secondary credit DW at the cost of being restricted in the uses of the credit
received, a higher haircut to be applied to the value of the collateral, and a closer monitoring
activity. Usually, the secondary credit maturity is overnight. Finally, the seasonal credit DW
has been conceived for small depository institutions with significant seasonal swings in their
loans and deposits. In order to be eligible for this type of DW, depository institutions have
to be located in agricultural or tourist areas11 .
In March 2008 two additional programs were launched by the Fed. The first one was
the Term Securities Lending Facility (TSLF). It was a weekly loan facility, with the aim of
promoting the functioning of financial markets, by offering “Treasury general collateral (GC)
to the Federal Reserve Bank of New York’s primary dealers in exchange for other programeligible collateral”12 . Its maturity term was 28 days. The main difference between the TSLF
and the TAF lies in the fact that the former offered Treasury GC to the New York Fed’s
primary dealers in exchange for other program-eligible collateral, while the latter offered term
funding to depository institutions. Both programs were based on an auction system.
The second program opened in March 2008 was the Primary Dealer Credit Facility
(PDCF). As with the previous program its goal was to promote the functioning of financial markets by providing funding to the primary dealer through overnight loan facilities in
exchange for any tri-party-eligible collateral. The difference of the PDCF program with respect to the TAF program refers to the institutions that benefited from the program (primary
dealers versus depository institutions), the maturity term of the loan (overnight versus 28
or 84 days), and the type of mechanism employed for allocating the credit (exchange versus
auction).
Two other programs, less related to the TAF program but nevertheless important, were
launched by the Fed between October and November 2008. The Commercial Paper Funding
11
12
http://www.federalreserve.gov/monetarypolicy/bst lendingdepository.htm
http://www.newyorkfed.org/markets/tslf faq.html
11
Facility (CPFF) had the goal “of enhancing the liquidity of the commercial paper market
by increasing the availability of term commercial paper funding to issuers and by providing
greater assurance to both issuers and investors that firms will be able to roll over their
maturing commercial paper”13 . Finally, the Term Asset-Backed Securities Loan Facility
(TALF) has been designed “to increase credit availability and support economic activity by
facilitating renewed issuance of consumer and business asset-backed securities at more normal
interest rate spreads. Under the TALF, the New York Fed will provide non-recourse funding
to any eligible borrower owning eligible collateral”14 .
0
100
Billions of $
200
300
400
500
Figure 2: Fed lending during the financial crisis
01jul2007
01jul2008
CPFF
TALF
01jul2009
Primary Credit
TAF
01jul2010
PDCF
TSLF
Notes: The figure shows weekly outstanding lending by the Fed to financial institutions through the different
programs operating during the last financial crisis.
Table 11 of the Appendix provides a detailed analysis of the different types of facilities
programs operated by the Fed during the last financial crisis. Specifically, reported are the
loans maturity, the period when the program was operated, the mechanisms to provide funds,
13
http://www.newyorkfed.org/markets/cpff faq.html. During this period, another program, the Money
Market Investor Funding Facility (MMIFF) was created to complete the CPFF program, but the funding
available through this program has never been distributed.
14
http://www.newyorkfed.org/markets/talf faq.html
12
the overall amounts provided, and the type of eligible institutions. Moreover, Figure 2 shows
the weekly outstanding lending by the Fed to financial institutions through the different
programs. The graph highlights the importance of the TAF program in the context of the
measures launched by the Fed during the financial crisis both with respect to the amounts
employed and the length of the period when the program was operated.
3
Data and descriptive analysis
In this section we report a detailed analysis of the dataset employed in this study and we
summarize the main results found.
3.1
The dataset
We obtained the data employed in this paper by merging information from different sources.
The data concerning banks’ balance sheets is a combination from the Report of Condition
and Income (generally referred to as Call Report) and the Uniform Bank Performance Report
(UBPR). US banks are required to hand in these reports by the Federal Financial Institutions
Examination Council (FFIEC). The specific reporting requirements depend on the size of
the bank and whether it has foreign offices. We accessed the Call Report data through the
website of the Federal Reserve of Chicago and the UBPR data through the website of the
FFIEC15 . The period taken into account goes from 2001:Q1 to 2010:Q3. The data on the
TAF auctions comes from the Federal Reserve Board. The sample covers the period from
2007:Q4 to 2010:Q1. We merged the datasets and transformed it so that we work with a
cross-section dataset, with the control variables measured in 2007:Q3.
Our final sample includes 8017 banks. Among them, 273 banks obtained TAF program
15
A known issue of the Call Report data we cannot control for, is the so-called “window dressing” effect.
Specifically, the day before the report, banks adopt a virtuous behaviour so that their balance sheets look
particularly good on the day of the report.
13
reserves at least once. These banks represent approximately 3.4% of the total number of the
banks in the sample16 . From the final sample we excluded 67 US branches of foreign bank
and agencies of foreign banks that were initially included in the TAF dataset, because we did
not have comparable balance sheet data for these banks. Moreover, the sample includes both
failed and survived banks, so the results do not suffer from a potential survivorship bias17 .
Specifically, between 2007:Q3 and 2010:Q3, 912 banks failed. Among them 28 obtained TAF
program reserves.
3.2
Description of the variables
Since we are interested in the TAF program’s effect on the change of banking funding liquidity
risk, we distinguish between banks that obtained reserves through the TAF program at least
once, and those which did not. The dummy variable labelled T AF takes value 1 if a bank
received TAF reserves at least once and 0 otherwise. We also focus on funds received by
each bank through the TAF program. Specifically, we define T AF AM OU N T 1 as the log
of one plus the overall amount of TAF funds received by each bank, and T AF AM OU N T 2
as the log of one plus the ratio of the overall amount of TAF funds received by each bank
and the total loans measured in 2007:Q3. Finally, AV G T AF AM OU N T is defined as the
ratio between the overall amount of TAF funds received by each bank over the corresponding
number of times the bank received TAF reserves18 .
In the baseline analysis, we approximate the liquidity risk of funding by the log of the
short-term liabilities over short-term assets (ST LIABST ASS). Larger values of this ratio
16
In the robustness checks part, we control for the fact that the dataset is characterized by the uneven
distribution of the number of banks in the two groups.
17
The survivor bias could arise if the sample included only survived banks, disregarding those that failed
during the period the program was operated by the Fed. If this were the case the results would not take into
account the information associated with failed institutions, leading to biased results.
18
The impact of the amounts received on liquidity risk should be studied at the margin; that is, by
considering banks’ liquidity needs at the moment of receiving the funds. Unfortunately, the dataset precludes
this type of analysis. In the cross-section context, we think that the alternative measures proposed above are
the best approximation to capture the effect of TAF amounts on liquidity risk.
14
imply a higher level of funding liquidity risk. This choice is consistent with the definition
provided by the Basel Committee of Banking Supervision, which defines liquidity as “the
ability to fund increases in assets and meet obligations as they come due”.
In the robustness checks we employ different measures of liquidity risk. These include
the ratio of short-term liabilities to total liabilities (ST LIAB T LIAB), the short-term net
liabilities (ST N ET LIAB), and the ratio of log of short-term liabilities to risk-free assets
(ST LIAB/P F RISK 0). These proxies show how important short-term liabilities are with
respect to different measures of liquid assets or with respect to the total volume of liabilities.
Control variables include bank’s liquidity capacity, portfolio composition, loans structure,
loan losses, different type of collateral assets, capital capacity, and profitability. As a proxy for
liquidity capacity we employed two alternative measures. LIQU IDIT Y is defined as the sum
of total trading assets, total available-for-sale securities, and total held-to-maturity securities
over total assets, while CASH is determined by cash and balances due from depository
institutions over total assets.
We also consider as controls some additional bank features regarding capital capacity and
profitability. Specifically, CAP BU F F ER is obtained by taking the difference between the
tier 1 capital ratio and the minimum requirement established by the banking authorities19 ,
the return on assets (ROA) is equal to the ratio of the income before income taxes and
extraordinary items and other adjustments over total assets, SIZE is measured by the log of
total assets, the ratio of non-performing loans to total loans (N P T L) are defined as loans that
are past due at least 30 days or are on non-accrual basis, and provisions for non-performing
loans (P ROV ) equal the ratio of loan-loss provisions over total loans.
To account for the portfolio composition of bank assets, we calculate the ratio of riskweighted assets to total assets (P F RISK)20 . This measure can be interpreted as a proxy of
19
In the period under analysis the minimum capital requirement was equal to 6%.
The weights (0, 20, 50, or 100%) are ascribed according to Basel I accords. On- and off-balance sheet
items have been summed when calculating total assets.
20
15
the portfolio risk: the higher this ratio, the higher the fraction of assets that are considered
risky by the regulatory authorities. Moreover, another control variable is the fraction of each
asset risk category, according to Basel I accords.
We include as explanatory variables measures of bank loans. We consider total loans
over total assets (T LOAN S), as well as the ratio of different loan types over total loans.
Specifically, we focus on commercial and industrial, real estate, individual, and agricultural
loans (CI LOAN S, REST LOAN S, IN DIV LOAN S, and AGRI LOAN S, respectively).
Finally, we add variables that serve as proxies for the amount of illiquid collateral. They
are Mortgage-Backed (pass-through) Securities, other types of Mortgage-Backed securities,
and Asset-Backed Securities. They are defined as the ratio between Mortgage-Backed (passthrough) Securities and total assets (M BS P ASS), the ratio of other type of Mortgage
over total assets (M BS OT HER), and the ratio of Asset-Backed Securities over total assets
(ABS). These measures assume that securities are held to maturity or are available-for-sale
at their fair value. A detailed analysis of the sources and definitions of the variables are
reported in Table 4 in the Appendix.
3.3
Main facts at a glance
Table 1 reports descriptive statistics on the variables employed. We distinguish along two
dimensions. On the one hand, columns (5) and (11) refer to the average values of the variables
measured in 2007:Q3 (before) and in 2010:Q3 (after) for all the banks in the sample21 . These
two periods correspond to the quarter just before the beginning of the program and two
quarters after its conclusion, respectively. On the other hand, columns (1), (3), (7), and
(9) report the average values of the variables by distinguishing between banks that received
TAF program reserves and other banks in each of the two periods. Focusing on liquidity risk
measures, the main findings highlight that before the beginning of the program (2007:Q3),
21
For failed banks, the after period has been defined as the last quarter in which the bank was operating.
16
TAF banks report levels of funding liquidity risk higher than those of other banks (column
(3) vs. column (1)), and that these differences become smaller once the program is over
(column (9) vs. column (7)). Liquidity risk volatility is higher for TAF banks than for the
rest of the banks. This is true for the two periods analysed. The other relevant result is that
although all banks lowered their funding liquidity exposures, TAF banks did more. CASH
is the only measure that does not follow this pattern. Specifically, banks that did not receive
reserves under the TAF program increased CASH more than the other banks. A plausible
explanation for this result is that NO TAF banks would have employed cash as a substitute
for TAF reserves. In order to meet their liquidity needs they would have increased cash, given
that they chose not to benefit from alternative financial aid. In Table 2, we test whether on
average there exist differences within groups across time and within time across groups. The
results confirm previous intuitions: ex ante, TAF banks exhibit higher levels of liquidity risk.
Moreover, these differences get smaller after the end of the program.
TAF banks are on average bigger than the NO TAF banks, have a lower level of capital
buffer, and exhibit a higher level of ROA and larger of provisions for future loan losses.
These features do not depend on the period analysed. Interesting, there are not differences
in the non-performing loans. This variable shows higher values after the end of the program
for all banks with no differences between TAF and NO TAF banks.
Total loans as a percentage of total assets and portfolio risk are higher for TAF banks than
for the rest of the banks. These patterns hold also after the end of the program even if the
differences are smaller.
The descriptive analysis highlights that both groups of banks adjust the quantities that
refer to liquidity risk. This is true by looking at liabilities and liquidity indicators. Moreover,
in the majority of the cases TAF banks change these amounts more than the NO TAF banks.
These changes imply as well that the differences between groups are smaller or disappear once
the program is over.
17
These patterns are illustrated in Figure 3. On average banks’ liquidity risk levels between
the groups are different just before the program began, while these differences were smaller
after the program ended. By distinguishing between the quartiles (25th, 50th and 75th),
Figure 7 in the Appendix reports the behaviour of the different measures of liquidity risk
per group. The graphs confirm previous findings: for each quartile, TAF banks show higher
levels of liquidity risk measures before the beginning of the program. These differences are
reduced or eventually reverted once the program is over.
.003
−4.5
St Liabilities / Total Liabilities
.0035
.004
Log(St Liabilities / St Assets)
−4.4
−4.3
−4.2
−4.1
.0045
Figure 3: Per-quarter-group banks average liquidity risk measures
2006q1
2007q1
2008q1
TAF
2009q1
2010q1
2011q1
2006q1
2007q1
No TAF
2008q1
TAF
2010q1
2011q1
No TAF
(b) ST Liabilities / Total Liabilities
0
−3
St Net Liabilities
5
10
Log(St Liabilities / PF Risk 0\%)
−2.5
−2
−1.5
15
−1
(a) ST Liabilities / ST Assets
2009q1
2006q1
2007q1
2008q1
TAF
2009q1
2010q1
2011q1
No TAF
2006q1
2007q1
2008q1
TAF
(c) ST Net Liabilities / Total Assets
2009q1
2010q1
2011q1
No TAF
(d) ST Liabilities / PF Risk 0%
Notes: We document the average behaviour of the following measures of liquidity risk: ST LIAB / ST ASS,
ST LIAB / TLIAB, ST NET LIAB, and ST LIAB / PF RISK ZERO, distinguishing by bank group (TAF
and NO TAF). In grey the period when the TAF program was operating.
18
.25
−4.5
Log(St Liabilities / St Assets)
−4.4
−4.3
−4.2
St Net Liabilities / Total Liabilities
.3
.35
.4
−4.1
.45
Figure 4: Per-quarter average banks liquidity risk behaviour
−20
−10
0
10
−20
0
10
(b) ST Liabilities / Total Liabilities
0
−3.5
5
St Net Liabilities
10
Log(St Liabilities / 0% Risk Assets)
−3
−2.5
−2
−1.5
15
−1
(a) ST Liabilities / ST Assets
−10
−20
−10
0
10
−20
(c) ST Net Liabilities / Total Assets
−10
0
10
(d) ST Liabilities / PF Risk 0%
Notes: For TAF banks only, we document the average behaviour of the following measures of liquidity risk
ST LIAB / ST ASS, ST LIAB / TLIAB, ST NET LIAB and ST LIAB / PF RISK ZERO from 20 quarters
before to 10 quarters after the the first time the banks obtained the reserves. In grey the period when the
TAF program was operated.
Figure 4 plots different measures of liquidity risk between 20 quarters before and 10
quarters after the first time banks received the reserves under the TAF program22 . For all
measures of liquidity risk, on average, banks decreased their funding liquidity risk positions
once they received the reserves. The graphical analysis suggests that the TAF program
was effective and useful and that it especially improved the funding liquidity exposures of
recipient banks.
Table 3 reports the pairwise correlation between the log of short-term liabilities over shortterm assets, the alternative measures of liquidity risk, along with the main controls employed
in the econometric analysis of this study, measured in 2007:Q3. The correlations with the
22
The two bounds have been established by considering the length of the program (10 quarters) so that it
is possible, also for a bank that received the funds in the last quarter the program was operating (2010:Q1),
to show its liquidity exposures for at least 10 quarters before the beginning of the program.
19
alternative measures of liquidity risk are positive and range between .238 (for short-term
liabilities over risk-free assets) and .901 (for the short-term net liabilities). Focusing on the
additional covariates, we have a negative correlation between the log of short term liabilities
over short term assets and CASH (−.187) and BU F F ER (−.362), while the correlation with
P F RISK is negative although it is quite small, around −.043. The correlation between net
liabilities and the rest of the variables, SIZE (.200), T OT LOAN S (.137), ROA (.061), and
N P T L (.0339) are positive even if the correlation coefficients show a lot of variability.
4
Econometric analysis
This section describes the econometric model employed to answer the main questions addressed in this study, and discusses the associated econometric issues.
4.1
The econometric model
To assess the impact of the TAF program on the change in funding liquidity risk, we employ
a treatment-effects model with a binary endogenous regressor. This approach is motivated
by the fact that the traditional OLS methods might lead to biased results because the ex
ante unobservable features of the TAF banks likely affect their decision to participate to the
TAF program, generating a potential bias in the effect of the TAF program on the change in
funding liquidity risk23 .
23
Technical analysis of the selection bias issue is provided in Appendix A.
20
More specifically, we are interested in fitting the following treatment-effects model:
∆ LIQ RISKi = T AFi β1 + x0i β2 + ξi
T AFi =



1
if T AFi∗ > 0


0
otherwise
(2)
(3)
where
T AFi∗ = x0i π1 + zi0 π2 + νi
(4)
x and z are two vectors of exogenous explanatory variables, and ξi and νi are assumed to
have a bivariate normal distribution with covariance matrix


2
 σ ρσ 


ρσ 1
In the outcome equation (2), the change of funding liquidity risk, ∆ LIQ RISK, depends on a set of explanatory variables x and on T AF , a binary endogenous regressor that
captures the TAF program’s impact on the dependent variable. Moreover, in equation (4)
the latent variable determines the values of the binary variable T AF , according to equation
(3). Equations (3) and (4) represent the participation part of the model. The T AF dummy
can be interpreted as a participation indicator: it equals one if bank i received at least once
the funds, and zero otherwise.
The above model is estimated simultaneously using Maximum Likelihood (ML) that pro-
21
vides consistent, efficient, and asymptotically normal estimators, under the assumption that
the error terms are bivariate normally distributed. If this is not the case, then consistency
is no longer guaranteed. One way to control for this potential issue is to estimate the model
using a two-step semi-parametric approach (SML)24 , which guarantees consistent estimates.
We adopt the following specification of our model, as described in equations (2), (3), and
(4):
∆ ST LIAB /ST ASSi = β0 + β1 T AFi + β2 CASHi + β3 ROAi + β4 CAP BU F F ERi +
(5)
β5 SIZEi + β6 Risk 0i + β7 Risk 20i + β8 Risk 50i + β9 Risk 100i + ξi
T AFi =



1
if T AFi∗ > 0


0
otherwise
(6)
where the unobserved latent variable follows the specification below:
T AFi∗ = π0 + π1 ST LIAB /ST ASSi + π2 CASHi +
(7)
π3 T LOAN Si + π4 M BSP ASSi + π5 M BSOT HERi + π6 ABSi + νi
The econometric model defined by equations (6) and (7) captures the probability of
obtaining the reserves. The variables included are measured in 2007:Q3, before the beginning
of the program. Specifically, we focus on funding liquidity risk, cash, and illiquid collateral
assets such as ABS, MBS, and TLOANS. We expect that banks with higher levels of funding
24
More details about this approach are reported in Appendix C, with robustness results reported and
discussed below.
22
liquidity risk were more likely to participate in the program. The same is also true for banks
with a high level of illiquid collateral or a high percentage of loans with respect to total
assets, reflecting a greater maturity mismatch. These banks were solvent, but temporarily
illiquid, because they were unable to increase liquidity by selling some of their assets. Due to
a lack of trust in the inter-bank credit market, these banks could not obtain liquidity from
other banks, and only the Fed accepted their illiquid collateral assets in exchange of reserves.
CASH is expected to negatively affect the probability of receiving funds because banks with
a sufficient level of cash were better able to manage liquidity distress.
All the variables included in equation (5) are measured in 2007:Q3, prior to the beginning
of the program. The dependent variable in equation (5) refers to the change in funding
liquidity risk between 2010:Q3 and 2007:Q3. Once controlled for selection bias, the T AF
variable is expected to negatively affect the change of the funding liquidity risk. If this is the
case, it implies that the TAF program is effective in the sense that it allows banks in funding
liquidity distress to adjust and improve their funding liquidity exposures.
Several additional controls are added to equation (5). A first set of variables captures
liquidity ability, capital cushion, size, and profitability. More precisely, we focus on CASH,
the level of CAP BU F F ER, the SIZE of the banks, and the ROA. CASH captures potential liquidity distress associated with banks’ liquidity needs. The higher the level of cash, the
smaller the change in funding liquidity risk. The inclusion of CAP BU F F ER is useful for
assessing the impact of capital cushions on the level of liquidity risk. More precisely, higher
capital buffer implies that banks are prone to adopt more aggressive investment strategies,
so we expect that capital buffer positively affects the change of funding liquidity risk. We
explicitly take into account the SIZE of the banks, because banks of different sizes have
different abilities to manage liquidity risk. In particular, big banks can more easily adjust
funding liquidity mismatches, so SIZE is expected to have a negative impact on the change
in liquidity risk. Finally, return on assets is a measure of investment returns. Higher ROAs
23
reflect that some banks more efficiently invest and might reduce the funding liquidity exposure. A second set of explanatory variables included in the baseline specification regard the
different types of assets25 held by banks. In this way, it is possible to assess the effects of
portfolio composition on the change of funding liquidity risk. We do not have an a priori expected sign for the effect of this second set of explanatory variables on the change of funding
liquidity risk.
The potential effect on liquidity risk of other Federal Reserve programs is not taken into
account in the specifications. The reason is that the financial institutions that benefited from
the other programs were not depository institutions26 . The only programs directly affecting
depository institutions are the primary, secondary, and seasonal credit discount window, but,
as previously mentioned, during the crisis these programs were less effective due to the fact
that depository institutions were concerned about the stigma effect. Accordingly the TAF
program effects captured in our analysis appears unlikely to be driven by other programs
not explicitly taken into account in the specifications. This view is also supported by the
information highlighted by Figure 2, which plots the weekly outstanding lending by the Fed
to financial institutions through the different programs operating during the last financial
crisis. The figure shows the relevant role played by the TAF program in terms of the amounts
provided as well as the period length when the program was operated by the Fed.
As previously mentioned, the Maximum Likelihood approach employed in the baseline
model requires the assumption that the error terms as being bivariate normally distributed.
Following Lee (1984), Pagan and Vella (1989), and Bera, Jarque, and Lee (1984), we test for
joint normality of the error terms27 . As shown by Table 10, in all the cases we cannot reject
the null hypothesis of joint normality28 . We can conclude that the estimates computed using
25
The type of the assets refers to their riskiness consistently with Basel I accords.
More details are provided in Table 11 of the Appendix.
27
Further details about testing joint normality assumption are provided in Appendix B.
28
The results refer to the baseline specification. Results referring to the other specifications are available
by request.
26
24
Maximum Likelihood are consistent.
5
Results
This section presents the results of the effect of participating in the TAF program as well
as the effect of the amount received on the liquidity exposures of the banks. Moreover, we
also check the robustness of our results in a specific subsection. Finally, we discuss the main
implications of our findings from a policy point of view.
5.1
Participation effect on liquidity risk
Table 5, column (1), reports the results of the baseline model. In columns (2) and (3), we
replace the different type of assets by a P F RISK that represents an overall measure of
portfolio risk and by the different types of loans held by the banks, respectively. In column
(4) we estimate a reduced form of the baseline specification by dropping the different types
of assets, while in column (5) we augment the previous specification by adding P ROV and
N P T L in order to capture the impact on the change of funding liquidity risk of expected
future and current distress due to bad loans on liquidity risk. On the one hand, a higher
level of P ROV can induce banks to employ more prudent investment strategies in the future.
Therefore, provisions for future non-performing losses are expected to decrease funding liquidity risk. On the other hand, if banks already have a high level of losses, they may be forced
to take riskier strategies (gambling for resurrection strategy, as pointed out by Kane (1989))
in the future, so we expect that N P T L affects positively the level of the funding liquidity
risk. Finally, in column (6) we replace the CASH of column (1) with LIQU IDIT Y . The
participation part of the model is the same for the different specifications, and it includes
the level of short-term liabilities over short-term assets, cash, or liquidity, depending on the
case, the total loans over total assets, and finally different measures of illiquid collaterals
25
approximated by MBS and ABS.
In the outcome equation several regularities arise. First, regardless of the specification,
the coefficient of the T AF dummy is always negative and statistically significant. This means
that banks that received TAF reserves decrease funding liquidity exposures more than those
that did not. This effect is not only statistically significant, but it is also economically
substantial. The fact of receiving TAF loans has an average extra effect on the quarterly
growth rate of the funding liquidity exposures of about 6.3 percentage points29 . These results
support the intuition that TAF reserves were crucial for decreasing exposures, and to control
for the funding liquidity risk of those banks with a more severe maturity mismatch, which
were most exposed to the freezing of the interbank market and unable to roll over their
short-term liabilities during the crisis.
Two additional results refer to the positive impact of capital buffer on the change in
liquidity risk and the negative relationship between the SIZE of the bank and the dependent variable. These findings confirm our intuition about the impact of CAP BU F F ER and
SIZE on the dependent variable. ROA is never statistically significant, while CASH shows
the expected positive sign only in one case, column (3). All asset types except P F Risk 50,
column (1), positively affect the change in liquidity risk, suggesting a U-shaped relationship
between the risk asset types and the dependent variable. However, aggregating the information and computing the ratio of risk-weighted assets to total assets, P F RISK, does not
affect the change in liquidity risk (see column (2)). Column (5) confirms our intuition about
the impact of N P T L on the change in funding liquidity risk, while the P ROV coefficient
is not different from zero, though it reports the expected negative sign. Finally, in column
(3), commercial and industrial loans as well as real estate loans positively affect the change
in liquidity risk, while neither individual loans nor agricultural loans has an impact on the
dependent variable. Focusing on the participation part of the model, funding liquidity risk
29
In order to interpret the dependent variable as a quarterly growth rate we have to divide the estimated
coefficient of the dummy variable TAF by 13, the number of quarters between 2010:Q3 and 2007:Q3.
26
positively affects the probability of receiving TAF reserves. The same is true for illiquid
collateral and total loans. CASH is never significant from zero, while this is not the case
for liquidity, column (6), which shows a negative impact on the probability of receiving TAF
reserves.
According to the specification of the model we are able to test whether we are facing
a selection bias, implying that, if not controlled, the T AF dummy would capture spurious
effects. Formally, this information is provided by the estimated coefficient on lambda (that
is defined in the following way: λ ≡ ρσ )30 . In all the cases, we can reject the null hypothesis
that the estimated coefficient is zero (see the corresponding χ2 statistic). Therefore, it follows
that there exists a selection bias that we have to control for.
5.2
Amount effect on liquidity risk
In the previous section we discussed the effect of the binary participation in the program on
the liquidity risk behaviours of the banks. Another element that can affect the dependent
variable is the amount of reserves that banks received in the framework of the TAF program.
In order to assess the amount impact on the change in liquidity risk we employ three different measures referring to the total amount of reserves received by each bank. Specifically,
we focus on the amount received, the amount received weighted by the level of total loans
measured in 2007:Q3, and the average amounts received by each bank.
Due to the nature of these alternative measures, specifically that they are continuous and
left-censored to zero, we modified the econometric model described in Section 4.1 that refers
to the binary participation in the program. More precisely, equations (3) and (4) are estimated using a Tobit model instead of a Probit model. The explanatory variables used do
not change with respect to the baseline model.
The results are reported in Table 6. More precisely, in columns (1), (3), (4), and (5)
30
Further information is provided in Appendix A.
27
T AF AM OU N T 1 has been employed to measure the TAF program’s impact on the liquidity risk behaviour, approximated by the different measures employed in this paper, while in
columns (2) and (6) T AF AM OU N T 2 and AV G T AF AM OU N T , respectively, have been
used to assess the effect of the TAF program on liquidity risk, measured by ST LIAB /ST ASS.
The findings highlight a negative relationship between the amount of reserves received and
the adjustment of the funding liquidity risk. According to the results of column (1) a 1%
increase in the reserves received leads to a drop in the quarterly liquidity risk growth rate
of .115%. A similar drop in the quarterly liquidity risk growth rate follows to an increase
by 1% in the fraction of amounts received with respect to the total loans the bank holds, as
reported in column (2). Finally, as highlighted in column (6), an increase of 1% of the average
amount of the reserves received decreases by 1% the quarterly growth rate of the liquidity
risk measure. These findings suggest that participation as well as the amount received play
an important role in decreasing funding liquidity exposures.
5.3
Robustness
We perform several robustness checks by using as a baseline the specification employed in
column (1) of Table 5, which has been also reported for comparative reasons in column (1) of
Table 7. Our result could suffer from omitted variable bias due to the fact that other events
occurred contemporaneously to the TAF program and they have not been explicitly taken
into account31 . One relevant episode was the failure of Lehman Brothers in 2008:Q3. We
have already eliminated one potential consequence of the “Lehman event” by dropping all
banks that had a large fraction of their credit lines co-syndicated with Lehman Brothers, as
reported by Ivashina and Scharfstein (2010). Moreover, in column (2) we verify the baseline
31
As previously discussed, we do not control for the effect of the other extraordinary programs promoted
by the Fed because their “target institutions” were not depository institutions.
28
results also by limiting our sample to the pre-September 2008 TAF auctions. The results
document that the TAF coefficient is statistically different from zero and it has the expected
negative sign. In the competitive auctions, however, the impact on the quarterly growth rate
of the dependent variables is larger (more than double) than when the entire sample is taken
into account (15 percentage points versus 6.5 percentage points). Therefore, we can conclude
that the direction of the impact is not driven by the collapse of Lehman Brothers, while after
this event the impact of the program on the quarterly growth rate of the dependent variable
decreases. This finding can be justified by the fact that depository institutions in need of
liquidity participated in the program before the collapse of Lehman Brothers and that the
additional banks that obtained the facilities after the Lehman Brothers collapse were in a
better situation from a liquidity risk perspective32 . To support this intuition, in Figure 5 we
report the average amounts of reserves received by all the banks since the first time, by banks
that received for the first time the reserves before the collapse of Lehman Brothers, and by
banks that received the reserves for the first time after the collapse of Lehman Brothers.
The figure shows important differences. First, the majority of the funds have been ascribed
to depository institutions that received the facilities for the first time before the collapse
of Lehman Brothers. Second, for these banks, the maximum average amount received is
attained after three periods, and they benefited from the program for a longer period than
the depository institutions that obtained the reserves for the first time after the collapse of
the Lehman Brothers.
Another event that could affect our results is the Troubled Asset Relief Program (TARP)33
32
Two events could have motivated some banks to participate in the auctions only after the collapse of
Lehman Brothers. First, they potentially experienced an increase in the marginal benefit, related to the
collapse of Lehman Brothers, in participating in the auctions. Second, they got a substantial reduction in the
marginal cost, represented by the repayment interest rate of the auctions. As shown in Figure 2 the interest
rate fell due to the fact that the Fed doubled the amounts supplied.
33
In October 2008, the US Treasury launched the Troubled Asset Relief Program. One part of the TARP
program was the Capital Purchase Program (CPP), an equity infusion program made by the US Treasury in
favour to credit institutes. Specifically, the US Treasury bought preferred non-voting stocks of U.S. financial
institutions for a total value of $250 billion.
29
0
Millions of $
100000
200000
300000
Figure 5: Average TAF amount since the first time
0
2
4
6
8
Periods after first taken TAF
Before Lehman
After Lehman
All
Notes: The figure shows the average amounts of reserves received since the first time by all the banks, by
banks that received for the first time the reserves before the collapse of Lehman Brothers, and by banks that
received the reserves for the first time after the collapse of Lehman Brothers.
promoted by the US Treasury in October 2008. Among the banks in our sample, only one
received both TAF funds and benefited from the TARP program. However, we found that
88 bank holding companies of TAF banks participated to the TARP program. In order to
control for the TARP effect on liquidity risk, we exclude from the sample the 88 banks whose
holding companies participated to the TARP program. The results, reported in column (7)
of Table 7, are unchanged. Therefore, we can conclude that our findings are not driven by
the TARP program.
The consistency of the results based on the Maximum likelihood relies on the joint normality assumption of the error terms. The tests reported in Table 10 indicates that the
assumption holds. However, for comparison purposes in columns (3), we relax the joint
normal distribution assumption about the error terms, and report the semi-non-parametric
two-step estimation results by adapting the approach followed by Martins (2001) to our spe-
30
cific case34 . The results about the direction of the TAF program’s effect on the change of
the liquidity risk are consistent with the baseline findings, even if they differ in terms of the
size of the impact. These differences may be due to two reasons. On the one hand, in the
two-step estimation approach the model is not estimated simultaneously; on the other hand,
the approximation of the term capturing the unobservable features of the banks in the two
cases is not the same: the inverse Mills ratio is employed in the ML approach, while a higher
order series, based on the predicted values of the participation part of the model, has been
used in the semi-non-parametric approach.
Since our sample includes all commercial banks that handed in Call reports, and only a
small fraction of those banks received TAF funding, we face a potential problem from the
uneven distribution of the number of banks in the two groups that could drive the main
results. More precisely, only 273 out of around 8000 banks (3.4%) received the TAF reserves.
In order to alleviate this potential problem we run a bootstrapping exercise, repeated 1000
times, consisting of generating subsamples of banks. The subsamples include all banks that
participated to the program and a randomly chosen subset of banks that did not receive
TAF funding. Each estimation is based on approximately 1000 observations. In this way, it
is possible to construct for each estimate a distribution based on the 1000 estimations35 . As
column (5) of Table 7 shows, the results are largely unchanged compared to our benchmark
case, even if the TAF effect is now larger than in the results of the baseline model. In Figure
6, the distribution of the estimate of the TAF variable obtained from the bootstrapping
exercise is provided, as well as the bounds of the 95% confidence interval.
Alternatively, we balance the sample by employing a matching exercise. We use propensity
score matching with 3 neighbours and match TAF and NO TAF banks with respect to LIQ.
RISK, CAPBUFFER, PF RISK, ROA, SIZE, CASH and LIQUIDITY. We estimate the
34
More details about testing joint normality assumption and semi-parametric estimation are provided in
Appendices B and C.
35
More details about the bootstrapping exercise are provided in Appendix D.
31
model by including all the TAF banks, and the correspondent 604 matched NO TAF banks
(we use matching with replacement). As shown in column (6) Table 7, the results do not
change.
In order to check whether the results are robust to the variable measurement period
chosen before the beginning of the program, in column (4) we measure the variables in
2005:Q3, instead of in 2007:Q3; that is, two years before the beginning of the program. The
results do not differ with respect to those of the baseline model: the larger value of the
estimate is compensated for the larger period taken into account. Rescaling the estimate
appropriately and dividing it by 21 periods we obtain a value of 5 percentage points, in line
with the baseline results.
Throughout our paper we have used short-term liabilities over short-term assets as the
measure for bank liquidity riskiness. The literature suggests various other measures of liquidity risk, which include, among others, the ratio between short-term liabilities over total
liabilities, the short-term net liabilities and the short-term liabilities over risk free assets. Table 9 compares the estimation results when different measures of liquidity risk are employed.
Column (1) reports the baseline results using the ratio of short-term liabilities over shortterm assets as a proxy for liquidity risk, while columns (2) to (4) report the results referring
to the above-mentioned measures of funding liquidity risk. The estimation results for the
T AF dummy in the outcome equation are negative and statistically significant, confirming
the baseline findings.
During the last financial crisis, systemically important commercial banks were not allowed
to fail. Being “Too Big To Fail” (TBTF) might lead to a moral hazard problem36 , a potential
source of attenuation bias for our finding. In order to assess the TBTF effect on the impact of
the TAF program on liquidity risk, we focus on the 75, 90 and 95 size percentiles. As shown
in columns (1) to (3) of Table 8, the TAF (NO TAF) banks included in the sample in the
36
If banks know that they are always saved with taxpayer money, they could adopt riskier strategies.
32
three cases are 190 (1928), 145 (769) and 111 (381), respectively. The results are unchanged
with respect to the baseline findings. In particular, the TAF effect is larger, the larger the
bank size.
Finally, throughout the paper we focused on liquidity issues disregarding solvency aspects
related to banks participation to the TAF program. In particular, it could be that banks participated to the TAF program because of solvency problems instead of maturity mismatches.
We address this issue by adopting the following strategy. First, we calculate the median of
the variables CAPBUFFER, PF RISK, CASH, and ST LIAB / RISK FREE assets of those
banks that participated to TAF, but that nevertheless failed. Since these banks had access
to the liquidity, it is highly likely that these banks failed due to solvency problems. Second,
we only consider banks that had “better fundamentals” by considering the before mentioned
variables. The new sample includes 51 TAF banks and 3023 NO TAF banks. As reported in
column (4) of Table 8, the results are consistent with those of the benchmark.
5.4
Discussion
We show that banks in major funding liquidity distress benefited from the reserves auctioned
in the context of the TAF program. Moreover, we find that the TAF program has an impact
on the reduction of funding liquidity risk. This impact is stronger, the higher the amount of
reserves received. The results highlight that, for banks with a more severe maturity mismatch
–and therefore most exposed to the freezing of the interbank market and unable to roll over
their short-term liabilities during the crisis– the TAF reserves are crucial for decreasing their
exposures and to control for their funding liquidity risk. Moreover, our findings confirm the
opinion that TAF-like programs are appropriate during situations similar to the last crisis.
In particular, our results support the view of those who consider the TAF program as an
additional countercyclical monetary policy instrument useful to mitigating banks’ liquidity
concerns during economic busts.
33
Our study stresses the importance of banking liability term structure as a source of the
banking soundness. In this perspective, our contribution provides empirical justification to
those arguments in favour of the introduction of measures for liquidity risk in international
financial regulations. In particular, the new measures, implemented in the Basel III accords,
such as the liquidity coverage ratio and the net stable funding ratio, go in the right direction
of focusing on liquidity management for the proper functioning of the banking sector and
financial markets.
Finally, our results shed light on the behaviour of a particular group of banks. Specifically,
we document that only banks in funding liquidity distress obtained loans through TAF. This
has been the case even if TAF loans were provided at favourable conditions (the minimum
interest rate proposed by the Fed was the discount window interest rate; the information
related to the banks benefiting from the TAF program was private), and despite the fact that
after Lehman Brothers’ collapse all bids were accepted. This result raises the question of why
the “good” banks decided not to participate in the TAF auctions. One potential explanation
is that, even if the information about the participation was, at least theoretically, private,
they were still concerned about the “stigma effect”.
6
Conclusion
During the last financial crisis the Federal Reserve promoted several extraordinary actions,
including the creation of a number of new facilities for auctioning short-term credit, with the
general aim of supporting the financial sector and ensuring that financial institutions have
adequate access to liquidity. One of these programs was the Term Auction Facility (TAF).
The goals of this paper were to assess which type of banks benefited from the program
and to quantitatively answer the question of whether the TAF program was successful in
helping banks reduce their liquidity risk positions. Acquiring this information was relevant
34
to confirming or challenging important criticisms addressed to the TAF-like programs. In
particular, these criticisms referred to the type of banks that benefited from the reserves and
the programs’ effectiveness in improving banks’ liquidity positions.
By using a unique dataset, which we constructed by merging TAF program information
with balance sheet data of the banks, we analysed the characteristics of the banks that
received TAF reserves and compared them with those of the other banks. Moreover, we
assessed the impact of the TAF program on the change of bank liquidity risk measured
before the beginning and after the close of the program. We employed a treatment-effects
model with a binary explanatory variable in order to avoid potential selection biases. The
probability of obtaining TAF program reserves is expressed as a function of control variables,
measured before the beginning of the TAF program. Simultaneously, we measure the impact
of TAF reserves on the change of bank liquidity risk.
The results suggested that banks that obtained program reserves showed, on average,
higher levels of funding liquidity risk indicators prior to the beginning of the program. Moreover, we found that the level of funding liquidity risk, as well as the level of illiquid collateral
measured in 2007:Q3, positively affected the probability of receiving program reserves. Finally, our results highlighted the fact that banks that obtained at some point TAF reserves
exhibited larger contraction of liquidity risk exposures, and that this effect was larger, the
higher the amount of reserves received. Several robustness checks confirmed the main results.
Our results accorded with the intuitions that the TAF program was employed by banks with
higher levels of funding liquidity risk and that it was effective in the reduction of funding liquidity risk at the bank level. Our findings did not support the criticisms that arose regarding
the TAF program’s effectiveness.
In this contribution we focused on the effect of the program on the liabilities side of
banks’ balance sheets. It could be interesting to assess how banks modified their portfolio
risk depending on whether they received the reserves associated with the TAF program.
35
Before the last financial crisis the two concepts of liquidity and solvency were clear, but this
was no more the case after 2007. It could be of relevance to analyse these two concepts
by assessing the behaviour of the banks that benefited from the TAF program designed to
solve liquidity mismatches, and of those that profited of the TARP funds, which were more
oriented toward solvency issues.
36
References
[1] R. M. Abrantes-Metz, M. Kraten, A. D. Metz, and G. S. Seow, “LIBOR manipulation?,”
Journal of Banking and Finance, vol. 36, no. 1, pp. 136 – 150, 2012.
[2] F. Allen, A. Babus, and E. Carletti, “Financial connections and systemic risk,” NBER
Working Papers 16177, National Bureau of Economic Research, Inc, July 2010.
[3] O. Armantier, E. Ghysels, A. Sarkar, and J. Shrader, “Stigma in financial markets:
evidence from liquidity auctions and discount window borrowing during the crisis,” FRB
of New York Staff Report, no. 483, 2011.
[4] O. Armantier, S. Krieger, and J. McAndrews, “The Federal Reserve’s Term Auction
Facility,” Current Issues in Economics and Finance, vol. 14, July 2008.
[5] A. K. Bera, C. M. Jarque, and L.-F. Lee, “Testing the normality assumption in limited dependent variable models,” International Economic Review, vol. 25, pp. 563–78,
October 1984.
[6] M. Bruche and J. Suarez, “Deposit insurance and money market freezes,” Journal of
Monetary Economics, vol. 57, pp. 45–61, January 2010.
[7] M. K. Brunnermeier and L. H. Pedersen, “Market liquidity and funding liquidity,” Review of Financial Studies, vol. 22, no. 6, pp. 2201–2238, 2009.
[8] A. C. Cameron and P. Trivedi, Microeconometrics Using Stata, Revised Edition. StataCorp LP, 2010.
[9] J. H. E. Christensen, J. A. Lopez, and G. D. Rudebusch, “Do central bank liquidity
facilities affect interbank lending rates?,” WPS, Fed of San Francisco, no. 13, 2009.
37
[10] G. D. Luca, “SNP and SML estimation of univariate and bivariate binary-choice models,” Stata Journal, vol. 8, pp. 190–220, June 2008.
[11] J.-S. Fontaine and R. Garcia, “Bond liquidity premia,” Review of Financial Studies,
vol. 25, no. 4, pp. 1207–1254, 2012.
[12] A. R. Gallant and D. W. Nychka, “Semi-nonparametric Maximum Likelihood estimation,” Econometrica, vol. 55, pp. 363–90, March 1987.
[13] W. H. Greene, Econometric Analysis. Prentice Hall, Aug. 2002.
[14] R. Huang and L. Ratnovski, “The dark side of bank wholesale funding,” Journal of
Financial Intermediation, vol. 20, no. 2, pp. 248 – 263, 2011.
[15] V. Ivashina and D. Scharfstein, “Bank lending during the financial crisis of 2008,” Journal of Financial Economics, vol. 97, pp. 319–338, September 2010.
[16] E. J. Kane, The S&L insurance crisis: How did it happen. Washington, DC: Urban
Institute Press, 1989.
[17] R. W. Klein and R. H. Spady, “An efficient semiparametric estimator for binary response
models,” Econometrica, vol. 61, pp. 387–421, March 1993.
[18] L.-F. Lee, “Tests for the bivariate normal distribution in econometric models with selectivity,” Econometrica, vol. 52, pp. 843–63, July 1984.
[19] M. F. O. Martins, “Parametric and semiparametric estimation of sample selection models: an empirical application to the female labour force in Portugal,” Journal of Applied
Econometrics, vol. 16, no. 1, pp. 23–39, 2001.
[20] A. Mian and A. Sufi, “The consequences of mortgage credit expansion: Evidence from
the US mortgage default crisis,” The Quarterly Journal of Economics, vol. 124, pp. 1449–
1496, November 2009.
38
[21] J. McAndrews, A. Sarkar, and Z. Wang, “The effect of the Term Auction Facility,”
[22] W. K. Newey, “Two-step series estimation of sample selection models,” Econometrics
Journal, vol. 12, pp. S217–S229, 2009.
[23] A. Pagan and F. Vella, “Diagnostic tests for models based on individual data: A survey,”
Journal of Applied Econometrics, vol. 4, pp. S29–59, Supplemen 1989.
[24] J. B. Taylor, Getting Off Track - How Government Actions and Interventions Caused,
Prolonged, and Worsened the Financial Crisis. No. 3 in Books, Hoover Institution,
Stanford University, 2009.
[25] J. B. Taylor and J. C. Williams, “A black swan in the money market,” American Economic Journal: Macroeconomics, vol. 1, pp. 58–83, January 2009.
[26] J. Taylor and J. Williams, “Further results on a black swan in the money market,”
Discussion Papers 07-046, Stanford Institute for Economic Policy Research, May 2008.
[27] D. L. Thornton, “The effectiveness of unconventional monetary policy: the Term Auction Facility,” tech. rep.
[28] F. Vella, “Estimating models with sample selection bias: A survey,” The Journal of
Human Resources, vol. 33, no. 1, pp. pp. 127–169, 1998.
[29] T. Wu, “On the effectiveness of the Federal Reserve’s new liquidity facilities,” Working
Papers 0808, Federal Reserve Bank of Dallas.
39
Appendices
A
The selection bias issue
The selection bias issue occurs when one or more explanatory variables are correlated with
the residuals. Therefore, these covariates are capturing “pure” effects that can be ascribed
directly to them, but at the same time they capture the effects referring to the residual term.
As a consequence we cannot interpret the estimated coefficient of these variables as their
effect on the dependent variable.37 In the case analysed in this paper, if banks participate
in the TAF program because they have an unobservable higher propensity to risk, then TAF
participation effect on risk could be overstated.
Assume that we are interested in assessing the impact of the TAF program on the liquidity
risk exposure of a group of banks. The econometric specification takes the following form:
Yi = β1 Di + Xi0 β + ξi
(8)
where X is a vector of explanatory variables and D is a dummy variable that takes value 1
if a individual i attends the TAF program and zero otherwise. Moreover, assume as well that
the fact of attending the program is driven by a set of an unobservable characteristic, so the
regression results could be potentially biased. First, we do not take explicitly into account
the bias issue and a linear regression is estimated. The expected values of Y if D = 1 and
when D = 0 take the following forms:
E(Yi |Di = 1) = β1 + Xi0 β + E(ξ|Di = 1)
37
See also Cameron and Trivedi (2010).
40
(9)
and
E(Yi |Di = 0) = Xi0 β + E(ξ|Di = 0)
(10)
respectively. Therefore the effect on the average value of Y is given by
E(Yi |Di = 1) − E(Yi |Di = 0) = β1 + E(ξ|Di = 1) − E(ξ|Di = 0)
(11)
The estimated coefficient is capturing both the “pure” effect βb1 that can be ascribed to
the fact of attending the TAF program as well as the effects related to unobservable features
E(ξ|Di = 1) − E(ξ|Di = 0). One way to solve for this potential issue is to estimate a
treatment-effect model with a binary endogenous regressor. The treatment-effect model is
based on a simultaneous estimation of two regressions.
On the one hand, a probit model is estimated in order to compute the predicted probability
of participating in the program controlling for a set of potential explanatory variables.
Di∗ = Zi0 θ + i
(12)
where Di∗ is a latent variable, Zi is the vector of the observable features affecting the fact
of participating, and i is the residual. We assume that the error terms of the probit and the
linear model, i and ξi , respectively, are bivariate normally distributed with zero mean and
covariance matrix


 1 ρσξ 


ρσξ σξ
Finally,
Di =



1
if Di∗ > 0


0
if Di∗ ≤ 0
41
It follows that
P (Di = 1) = Φ(Zi0 θ) and P (Di = 0) = 1 − Φ(Zi0 θ)
and from the joint density of the bivariate normally distributed variables, equations (9) and
(10) can be written as
E(Yi |Di = 1) = β1 + Xi0 β + ρσξ
φ(Zi0 θ)
Φ(Zi0 θ)
φ(Zi0 θ)
E(Yi |Di = 0) = βXi − ρσξ
1 − Φ(Zi0 θ)
The average participation effect is therefore the difference,
E(Yi |Di = 1) − E(Yi |Di = 0) = β1 + ρσξ
φ(Zi0 θ)
Φ(Zi0 θ)[1 − Φ(Zi0 θ)]
(13)
where ρ is the correlation between the two error terms and σξ is the noise term standard
error of the linear regression. By using the treatment-effect model with a binary endogenous
regressor we are able to capture the effects of unobservable features captured by the participation variable and therefore to exactly measure the “pure” effect of participating in the
program. The “cost” of adopting this approach is the strong assumption of the distribution
of the error terms. An alternative approach that does not require the previous assumption
is to run a two-step estimation, computing robust standard error.
B
Testing joint normality
Assume that we want to estimate the following model:
Yi = Di β1 + Xi0 β + ξi
42
(14)
Di =



1
if Di∗ > 0


0
otherwise
(15)
where
Di∗ = Zi0 θ + νi
(16)
the X and Z are two vectors of exogenous explanatory variables, ξi and νi are the error
terms in the two equations. Estimating the model, represented by equations (14), (15), and
(16), by Maximum Likelihood requires the joint normality assumption of the error terms.
Following Lee (1984), Pagan and Vella (1989), and Bera, Jarque, and Lee (1984), we test
joint normality in the following way.
We rewrite equation (14), including an extra term that in the traditional participation approach represents an estimate of the inverse Mills ratio:
Yi = Di β1 + Xi0 β + µ(Zi0 θ) + ξi
(17)
We compute the participation part of the model, represented by equations (15) and (16),
by assuming that the error terms are normally distributed and, alternatively, following a
semi-parametric approach based on Gallant and Nychka (1987)38 . The linear prediction of
the model (Zi0 θ̂) are computed and then employed to construct an approximation of µ(Zi0 θ).
Using a parametric approach the term µ(Zi0 θ̂) is proportional to the inverse Mills’ ratio,
P
µ(Zi0 θ̂)p = κj=1 (Zi0 θ̂)j M ills, while in the semi-parametric approach it is a function of Zi0 θ̂,
P
µ(Zi0 θ̂)sp = κj=1 (Zi0 θ̂)j .
Equation (17) is estimated considering the two approximations for µ(Zi0 θ). The joint
38
More details are provided in Appendix C.
43
normality is tested by running a F-test on the higher-order terms added to equation (14). If
we can reject the null hypothesis, this implies that there is not evidence joint normality. The
specific results referring to the model employed in this study are reported in Table 10.
C
Semi-parametric estimations
Estimating the model, represented by equations (14), (15), and (16), when the joint normality
assumption of the error terms does not hold leads to inconsistent estimates. One way to solve
this potential issue is to adopt a two-steps semi-parametric approach39 . Specifically, assume
that
0
P (Di = 1|Zi ) = E[Di |Zi ] = G(Zi θ)
(18)
where G is an unknown function. Due to the fact that we cannot invoke a distributional
assumption regarding the error term of the binary choice model, in the first step we estimate
the vector θ by using a semi-parametric estimation. Specifically, we employ two alternative
ways to obtain the estimates for θ. On the one hand we adopt the approach proposed by
Klein and Spady (1993). The idea behind their approach is to maximize a pseudologlikelihood function in which the unknown probability functions are locally approximated by
non-parametric kernel estimators. On the other hand, we employ the approach proposed by
Gallant and Nychka (1987), where the unknown densities of the latent regression error is
approximated by Hermite polynomial expansions and the approximations are then employed
to derive a pseudo-ML estimator for the model parameters40 .
In the second step, the vector θ̂ is employed to find an approximation of the single
0
index µ(Zi θ) included in 17 and necessary to adjust for the individual unobservable charac39
40
See for instance Vella (1998) and Martins (2001).
For more information about these two approaches see De Luca (2008).
44
teristics that can potentially bias the results. Following Newey (2008), the single index is
approximated by
0
µ(Zi θ̂) =
κ
X
0
ακ (Z θ̂)κ
(19)
i=1
where κ expands as the sample size increases. The advantage of this approach is that, as
√
shown by Newey, the estimates of the first step are n and the second step can be easily
computed by OLS.
45
D
Bootstrapping approach
0
1
2
Density
3
4
5
Figure 6: TAF estimated coefficient obtained from a bootstrapping approach
−1.2
−1.1
−1
−.9
−.8
−.7
TAF
In order to alleviate the potential problem of the uneven distribution of TAF and NO
TAF banks, we ran a bootstrapping exercise. In each iteration, the sample includes all TAF
banks and a randomly chosen subset of NO TAF banks.
The graph in Figure 6 shows the distribution of the estimate of TAF reserves as well
as the bounds of the corresponding confidence interval at 95% obtained by repeating the
estimation 1000 times and by using a sample of around 1000 observation randomly. Before
the estimation we check whether the mean of all used variables of the chosen subsample are
within a narrow band around the mean of the entire sample (we use 0.2 times the standard
deviation as a threshold).
46
E
Tables
Table 1: Summary statistics
No TAF
mean
sd
Liquidity and Liabilities
NET LIAB ASS
Before
TAF
mean
sd
All
mean
sd
No TAF
mean
sd
After
TAF
mean
sd
Total
mean
sd
(1)
(2)
(3)
(4)
(5)
(6)
(7)
(8)
(9)
(10)
(11)
(12)
4.286
18.11
8.326
19.14
4.434
18.17
2.676
15.57
2.793
15.64
2.680
15.57
LIAB PF RISK 0
-1.427
1.277
-1.201
1.613
-1.419
1.291
-2.728
1.368
-3.045
1.428
-2.740
1.372
ST LIAB TLIAB
.00404
.00131
.00413
.00142
.00404
.00131
.00338
.00119
.00311
.00127
.00337
.00119
.446
ST LIAB TASS
-5.735
.460
-5.760
.481
-5.736
.461
-5.906
.444
-6.037
.498
-5.910
ST LIAB ASS
-4.408
.749
-4.252
.849
-4.402
.753
-4.453
.688
-4.442
.800
-4.453
.692
ST LIAB LIQ
-3.858
1.111
-3.547
1.279
-3.847
1.119
-3.985
1.173
-3.936
1.165
-3.984
1.172
LIQUIDITY
.208
.139
.158
.118
.206
.138
.206
.145
.173
.124
.205
.144
CASH
.0377
.0397
.0288
.0355
.0374
.0395
.0841
.0778
.0636
.0619
.0833
.0774
Other banks features
CAPBUFFER
.0541
.0654
.0476
.0835
.0538
.0662
.0424
.0411
.0371
.0316
.0422
.0408
SIZE
11.90
1.248
14.08
2.136
11.98
1.354
12.05
1.215
14.15
2.086
12.13
1.319
ROA
.00840
.0102
.0102
.00976
.00846
.0102
.00407
.0116
.00132
.0162
.00397
.0118
NPTL
.0239
.0245
.0173
.0165
.0236
.0243
.0448
.0464
.0560
.0487
.0452
.0466
PROV
.00142
.00353
.00197
.00318
.00144
.00351
.00459
.00678
.0101
.0108
.00480
.00705
TLOANS
.645
.149
.681
.143
.646
.149
.616
.142
.656
.133
.618
.142
CILOANS
.148
.106
.176
.130
.149
.107
.136
.0979
.158
.114
.137
.0986
Loans
RTESTLOANS
.685
.194
.700
.202
.686
.194
.710
.189
.720
.209
.710
.189
INDIVLOANS
.0768
.0902
.0731
.145
.0767
.0928
.0663
.0880
.0753
.178
.0667
.0930
AGRILOANS
.0734
.126
.0174
.0543
.0714
.124
.0731
.125
.0183
.0560
.0711
.123
Portfolio Assets
PF RISK
.691
.125
.758
.115
.694
.126
.656
.116
.703
.108
.658
.116
PF RISK 0
.0260
.0488
.0250
.0615
.0260
.0493
.0759
.0831
.0850
.0878
.0762
.0833
PF RISK 20
.251
.143
.188
.118
.249
.142
.224
.139
.171
.105
.222
.139
PF RISK 50
.162
.120
.133
.0965
.161
.119
.171
.116
.134
.0844
.170
.115
PF RISK 100
.561
.170
.655
.154
.565
.171
.530
.156
.610
.137
.533
.156
Observations
7520
271
7791
6832
252
7084
Notes: We can distinguish along two dimensions. On the one hand, columns (5) and (11) refer to the
average values of the variables measured in 2007:Q3 (before), just before the beginning of the program, and
in 2010:Q3, two quarters after its conclusion (after). On the other hand, columns (1), (3), (7), and (9) report
the variables, average values by distinguishing between banks that received TAF program reserves and the
others banks in each of the two periods.
47
Table 2: Average differences tests: Before and After
Liabilities and Liquidity
Before
After
No TAF
TAF
Diff in Diff
(1)
(2)
(3)
(4)
(5)
∗∗∗
NET LIAB ASS
5.279
( 1.143)
0.144
( 0.096)
0.165∗∗∗
( 0.055)
1.984∗∗∗
( 0.131)
2.199∗∗∗
( 0.137)
-0.049∗∗∗
( 0.008)
-0.016∗∗∗
( 0.002)
0.000
( 0.000)
-0.071
( 0.043)
0.209∗∗
( 0.082)
LIAB PF RISK 0
ST LIAB ASS
ST ASS
ST LIAB
LIQUIDITY
CASH
ST LIAB TLIAB
ST LIAB TASS
ST LIAB LIQ
∗
1.723
( 1.034)
-0.384∗∗∗
( 0.095)
0.040
( 0.062)
1.976∗∗∗
( 0.130)
2.030∗∗∗
( 0.135)
-0.032∗∗∗
( 0.008)
-0.031∗∗∗
( 0.004)
-0.000∗∗
( 0.000)
-0.157∗∗∗
( 0.048)
-0.030
( 0.081)
∗∗∗
-1.304
( 0.301)
-1.282∗∗∗
( 0.023)
-0.034∗∗
( 0.015)
0.037
( 0.024)
0.001
( 0.025)
0.002
( 0.003)
0.046∗∗∗
( 0.001)
-0.001∗∗∗
( 0.000)
-0.174∗∗∗
( 0.014)
-0.159∗∗∗
( 0.023)
∗∗∗
-4.860
( 1.512)
-1.810∗∗∗
( 0.133)
-0.160∗
( 0.082)
0.029
( 0.183)
-0.168
( 0.191)
0.019∗
( 0.011)
0.032∗∗∗
( 0.004)
-0.001∗∗∗
( 0.000)
-0.261∗∗∗
( 0.063)
-0.398∗∗∗
( 0.113)
-3.556∗∗
( 1.542)
-0.528∗∗∗
( 0.135)
-0.125
( 0.083)
-0.008
( 0.185)
-0.169
( 0.193)
0.017
( 0.011)
-0.015∗∗∗
( 0.004)
-0.000∗∗∗
( 0.000)
-0.086
( 0.065)
-0.239∗∗
( 0.116)
Notes: Columns (1) and (2) test whether on average there exists a difference within groups across time
(Before period: 2007:Q3. After period: 2010:Q3). Columns (3) and (4) test whether on average there exists
a difference within time across groups (TAF and NO TAF). Finally, column (5) tests whether there are
differences in differences.
Table 3: Correlation Matrix
(1)
(1)
(2)
(3)
(4)
(5)
(6)
(7)
(8)
(9)
(10)
(11)
ST LIAB / ST ASSETS
1
ST LIAB / TOT LIAB
.453
ST NET LIAB
.901
ST LIAB / PF RISK 0% .238
SIZE
.200
CASH
-.187
PF RISK
-.0430
TOT LOANS
.137
CAPBUFFER
-.362
ROA
.0610
NPTL
.0339
(2)
(3)
1
.547
1
.460
.228
.00860 .198
-.102
-.208
.185 -.0177
.210
.134
-.0727 -.339
-.0260 .0384
.0751 .0468
(4)
(5)
(6)
(7)
(8)
(9)
(10)
(11)
1
.0270
- .157
.380
.333
-.0869
.00107
.0118
1
-.222
.234
.166
-.267
.189
-.0926
1
-.254
-.250
.124
-.0159
.00203
1
.818
-.190
.111
-.0256
1
-.296
.128
.0178
1
-.362
-.0862
1
-.0498
1
Notes: The correlations have been computed measuring the variables in 2007:Q3.
48
−5
.002
.003
st_liab_ass
−4.5
−4
st_liab_totliab
.004
.005
.006
−3.5
Figure 7: Liquidity risk measures, percentiles
2006q3
2007q3
TAF 25
NO TAF 25
2008q3
2009q3
TAF 50
NO TAF 50
2010q3
2006q3
TAF 75
NO TAF 75
2007q3
TAF 25
NO TAF 25
2009q3
TAF 50
NO TAF 50
2010q3
TAF 75
NO TAF 75
(b) ST Liabilities / Total Liabilities
−4
−10
−3
st_liab_riskz
−2
st_net_liab_ass
0
10
−1
20
0
(a) ST Liabilities / ST Assets
2008q3
2006q3
2007q3
TAF 25
NO TAF 25
2008q3
TAF 50
NO TAF 50
2009q3
2010q3
2006q3
TAF 75
NO TAF 75
2007q3
TAF 25
NO TAF 25
(c) ST Net Liabilities / Total Assets
2008q3
TAF 50
NO TAF 50
2009q3
2010q3
TAF 75
NO TAF 75
(d) ST Liabilities / PF Risk Zero
Notes: For the four measures of liquidity risk employed in this paper (ST LIAB / ST ASS, ST LIAB /
TLIAB, ST NET LIAB, and ST LIAB / PF RISK ZERO), Figure 7 documents the by-group behaviour of
the 25th, 50th and 75th percentile. In grey is the period when the TAF program was operating.
49
50
Dummy variable. It takes value 1 if a bank received TAF reserves at least once, and 0 otherwise.
Overall amount of TAF funds received by each bank.
Number of times a bank received TAF funds.
log of one plus the overall amount of TAF reserves received
log of one plus the ratio of the overall amount of TAF reserves received and the total loans
AMOUNT/NUM
Short term assets
On- and Off-Balance Sheet assets
Short term assets over short term liabilities
Total liabilities
Short term liabilities
log of Short term liabilities over short term assets
log of Short term liabilities over total liabilities
Short term liabilities - Short tern assets over Total assets
log of Short term liabilities over Risk Free assets
Liquid assets over total assets
Cash and balances due from depository institutions over total assets
Ratio of the risk-weighted assets to total assets
Assets with a risk weight 0% over total assets
Assets with a risk weight 20% over total assets
Assets with a risk weight 50% over total assets
Assets with a risk weight 100% over total assets
Total loans and Leases, Gross over total assets
Commercial and Industrial Loans over total loans
Real Estate Loans over total loans
Loans to Individuals over total loans
Agricultural Loans over total loans
Ratio of Asset-Backed Securities* over Total Assets
Ratio of Mortgage* Backed (pass-through) over Total Assets
Ratio of other type of Mortgage* over Total assets
Tier 1 capital ratio minus 6%**
Ratio of the income before income taxes and extraordinary items and other adjustments over total assets
Log of banks total asset
Loans that are past due at least 30 days or are on non-accrual basis over total loans
Ratio of loan loss provision over total loans
TAF
AMOUNT
NUM
TAF AMOUNT 1
TAF AMOUNT 2
AVG TAF AMOUNT
ST ASS
TOTAL ASSETS
ST ASS / ST LIAB
TLIAB
ST LIAB
ST LIAB / ST ASS
ST LIAB / TLIAB
ST NET LIAB
ST LIAB / PF RISK 0
LIQUIDITY
CASH
PF RISK
PF RISK 0
PF RISK 20
PF RISK 50
PF RISK 100
TLOANS
CI LOANS
REST LOANS
INDIV LOANS
AGRI LOANS
ABS
MBS
MBS OTHER
CAPBUFFER
ROA
SIZE
NPTL
PROV
U.S.
U.S.
U.S.
U.S.
U.S.
U.S.
U.S.
U.S.
U.S.
U.S.
U.S.
U.S.
U.S.
U.S.
U.S.
U.S.
RCFD8274-.06
RIAD4301/TOTAL ASSETS
log(TOTAL ASSETS)
(RCFD1403 + RCFD1406 + RCFD1407)/RCFD1400
RIAD4230/RCFD1400
U.S.
U.S.
U.S.
U.S.
U.S.
Call
Call
Call
Call
Call
Call
Call
Call
Call
Call
Call
Call
Call
Call
Call
Call
Call
Call
Call
Call
Board
Board
Board
Board
Board
Board
Reports
Reports
Reports
Reports
Reports
Reports
Reports
Reports
Reports
Reports
Reports
Reports
Reports
Reports
Reports
Reports
Reports
Reports
Reports
Reports
Reports
Reports
Reports
Reports
Reports
Reports
Reports
Reports
Reports
Reserve
Reserve
Reserve
Reserve
Reserve
Reserve
Call
Call
Call
Call
Call
Call
Call
Call
Call
Federal
Federal
Federal
Federal
Federal
Federal
Source
RCFD1400/TOTAL ASSETS
U.S.
RCFD1600/RCFD1400
U.S.
RCFD1410/RCFD1400
U.S.
RCFD1975/RCFD1400
U.S.
RCFD1590/RCFD1400
U.S.
(RCONC988+RCON027)/TOTAL ASSETS
U.S.
(RCON1699+RCON1702+RCON1705+RCON1707+RCON1710+RCON1713)/TOTAL ASSETS U.S.
(RCON1734+RCON1736)/TOTAL ASSETS
U.S.
(RCFD3545 + RCFD1773 + RCFD1754) /TOTAL ASSETS
RCFD0010/TOTAL ASSETS
RCFDA223/ TOTAL ASSETS
RCFDB696 / TOTAL ASSETS
RCFDB697 / TOTAL ASSETS
RCFDB698 / TOTAL ASSETS
RCFDB699 / TOTAL ASSETS
UBPRE583
RCFDB696 + RCFDB697 + RCFDB698 + RCFDB699
UBPR598
RCFD2950
UBPRE583/UBPR898
log(1/UBPR598)
log(ST LIAB/RCFD2950)
UBPRE599
log(ST LIAB /PF RISK 0)
log(1+AMOUNT)
log[1+(AMOUNT/TLOANS)]
Chicago Fed Label
Notes: *Securities held to maturity or available-for-sale at their fair value. **The minimum requirement established by the banking authorities.
Variable definition
Variable Label
Table 4: Sources and definition of the variable
Table 5: Baseline model
(1)
(2)
(3)
-.315
(.438)
2.466***
(.255)
-.021
(.017)
-.038***
(.007)
-.820***
(.09)
.663**
(.305)
.464***
(.109)
-.075
(.114)
.388***
(.102)
-.165
(.443)
2.614***
(.254)
-.021
(.017)
-.037***
(.007)
-.860***
(.086)
.446*
(.254)
2.759***
(.258)
.009
(.023)
-.044***
(.009)
-.810***
(.089)
Outcome equation
Dependent variable:
CASH
CAPBUFFER
ROA
SIZE
TAF
PF Risk 0
PF Risk 20
PF Risk 50
PF Risk 100
(4)
(5)
(6)
.421
(.258)
2.740***
(.27)
-.003
(.024)
-.032***
(.007)
-.843***
(.088)
2.612***
(.251)
-.021
(.017)
-.037***
(.007)
-.825***
(.087)
∆ ST LIAB ASS
PF RISK
-.189
(.436)
2.607***
(.254)
-.021
(.017)
-.036***
(.007)
-.860***
(.086)
.041
(.082)
CILOANS
1.036***
(.33)
.808***
(.301)
.399
(.377)
.515
(.331)
RESTLOANS
INDIVLOANS
AGRILOANS
PROV
-.001
(.036)
.689**
(.319)
NPTL
LIQUIDITY
Constant
-.282***
(.069)
.370***
(.082)
.290***
(.11)
-.431
(.355)
.313***
(.096)
.196**
(.086)
.379***
(.07)
-1.891
(1.505)
1.194***
(.277)
.109
(.645)
1.891***
(.542)
19.135***
(4.434)
.416***
(.07)
-1.611
(1.457)
1.310***
(.28)
-.03
(.649)
1.804***
(.554)
19.114***
(4.439)
.398***
(.069)
-1.136
(1.487)
1.293***
(.254)
-.02
(.622)
1.830***
(.528)
19.276***
(4.44)
.418***
(.07)
-1.625
(1.456)
1.288***
(.283)
-.028
(.649)
1.805***
(.554)
19.096***
(4.433)
.426***
(.072)
-1.296
(1.458)
1.151***
(.254)
-.214
(.643)
1.658***
(.547)
19.361***
(4.58)
Constant
-.903**
(.407)
-.808**
(.406)
-.897**
(.379)
-.785*
(.407)
-.663*
(.401)
-.001
(.369)
1.209*
(.720)
3.034***
(.658)
18.768***
(3.991)
-2.608***
(.498)
.623
(.466)
Obs.
ρ
λ
7662
.524
.329
(.0391)
52.97
7662
.558
.353
(.0373)
63.42
7643
.54
.331
(.0387)
52.85
7662
.558
.353
(.0373)
63.38
7643
.554
.343
(.0385)
56.04
7662
.536
.338
(.0372)
61.02
Participation equation
ST LIAB ASS
CASH
TLOANS
MBS PASS
MBS OTHER
ABS
LIQUIDITY
χ2
.467***
(.066)
Notes: Joint estimation of Treatment-effects model with binary dependent variable T AF , using Maximum
Likelihood. Robust s.e. in parentheses. *** = p < .01, ** = p < .05, * = p < .1.
51
Table 6: Different measures for capturing the TAF program effect
Outcome equation
Dependent variable:
(1)
(2)
(3)
(4)
(5)
(6)
∆ ST LIAB ASS
∆ ST LIAB ASS
∆ ST LIAB TLIAB
∆ NET LIAB
∆ ST LIAB PF RISK 0
∆ ST LIAB ASS
-.272
(.167)
2.464***
(.093)
-.022***
(.005)
-.034***
(.006)
.638***
(.163)
.424***
(.079)
-.155*
(.086)
.336***
(.077)
-.098***
(.009)
.362**
(.176)
2.615***
(.095)
-.008
(.006)
-.027***
(.006)
.244
(.163)
.353***
(.078)
-.273***
(.085)
.209***
(.076)
.068**
(.029)
.090***
(.016)
.001
(.001)
-.006***
(.001)
.078***
(.029)
.040***
(.014)
-.035**
(.015)
.006
(.013)
-.015***
(.002)
-17.804***
(2.995)
31.956***
(1.839)
-.178***
(.034)
-.887***
(.146)
3.811
(3.628)
13.676***
(1.915)
-4.848**
(2.129)
11.556***
(1.875)
-3.050***
(.206)
.262
(.398)
.222
(.226)
.057***
(.015)
-.193***
(.013)
9.429***
(.364)
1.349***
(.179)
1.586***
(.196)
.320*
(.175)
-.033
(.032)
-.266
(.167)
2.468***
(.093)
-.022***
(.005)
-.032***
(.006)
.616***
(.163)
.401***
(.079)
-.175**
(.087)
.314***
(.077)
CASH
CAPBUFFER
ROA
SIZE
PF Risk 0
PF Risk 20
PF Risk 50
PF Risk 100
TAF AMOUNT 1
TAF AMOUNT 2
-.093***
(.008)
AVG TAF AMOUNT
-.138***
(.012)
Participation equation
ST LIAB ASS
CASH
TLOANS
MBS PASS
MBS OTHER
ABS
Constant
Obs.
4.131***
(.548)
-26.269**
(13.355)
15.111***
(3.286)
4.417
(7.027)
28.076***
(7.267)
250.592***
(39.222)
-15.805***
(3.789)
4.488***
(.597)
-23.978*
(14.383)
15.018***
(3.532)
3.479
(7.555)
29.350***
(7.817)
270.368***
(42.192)
-15.993***
(4.116)
2.788***
(.550)
-29.267**
(13.952)
14.489***
(3.447)
8.421
(7.121)
27.137***
(7.623)
208.108***
(40.796)
-22.069***
(4.102)
4.734***
(.519)
-29.215**
(11.894)
11.949***
(3.032)
8.515
(6.474)
23.421***
(6.888)
220.066***
(36.846)
-10.335***
(3.515)
1.260**
(.508)
-29.064**
(14.337)
16.154***
(3.528)
11.707
(7.344)
34.683***
(7.573)
220.742***
(42.306)
-30.538***
(4.127)
3.062***
(.396)
-19.142**
(9.742)
11.072***
(2.398)
3.123
(5.127)
20.415***
(5.303)
183.661***
(28.551)
-11.365***
(2.75)
7662
7642
7662
7790
7662
7662
Notes:
Joint estimation of Treatment-effects model with left-censored dependent variable
(T AF AM OU N T 1, T AF AM OU N T 2, and AV G T AF AM OU N T ) using Maximum Likelihood.
Robust s.e. in parentheses. *** = p < .01, ** = p < .05, * = p < .1. T AF AM OU N T 1 is the log of one
plus the overall amount of TAF reserves received; T AF AM OU N T 2 is the log of one plus the ratio of
the overall amount of TAF reserves received and total loans; AV G T AF AM OU N T is the ratio of overall
amount of TAF funds received by each bank over number of times a bank received TAF funds.
52
Table 7: Methodologies and subsamples
MLE
(1)
Lehman
(2)
SNP Two-Step
(3)
Outcome equation
Dependent variable:
CASH
CAPBUFFER
ROA
SIZE
PF Risk 0
PF Risk 20
PF Risk 50
PF Risk 100
TAF
Early
(4)
Bootstrap Matching W/out TARP
(5)
(6)
(7)
∆ ST LIAB ASS
-.315
(.438)
2.466***
(.255)
-.021
(.017)
-.038***
(.007)
.663**
(.305)
.464***
(.109)
-.075
(.114)
.388***
(.102)
-.820***
(.090)
-.355
(.442)
2.492***
(.257)
-.020
(.016)
-.040***
(.008)
.664**
(.305)
.491***
(.112)
-.128
(.117)
.396***
(.105)
-.636***
(.122)
-2.963***
(.892)
2.465***
(.259)
-.012
.017)
-.029***
(.007)
.625**
(.318)
.436***
(.102)
.091
(.128)
.588***
(.113)
-.098**
(.049)
.264
(.35)
2.077***
(.34)
-.021***
(.009)
-.0124
(.008)
.027
(.276)
.0124
(.123)
.056
(.013)
.247**
(.115)
-1.118***
(.083)
-.013
(.824)
1.802**
(.764)
-.044
(.050)
-.034**
(.017)
.891
(.557)
.43
(.2669)
.554*
(.303)
.612***
(.232)
-1.035***
(.138)
3.937***
(1.212)
1.638**
(.654)
-.044
(.052)
-.043**
(.017)
.457
(.530)
.469
(.336)
.406
(.320)
.681***
(.241)
-1.101***
(.116)
-.255
(.438)
2.468***
(.255)
-.022
(.017)
-.038***
(.008)
.667**
(.313)
.467***
(.112)
-.103
(.118)
.385***
(.106)
-.801***
(.117)
.379***
.271***
(.070)
(.073)
-1.891
-7.901*
(1.505)
(4.218)
1.194***
.274
(.277)
(.418)
.109
-1.425
(.645)
(.982
1.891*** 1.891***
(.542)
(.595)
10.97*** 18.456***
(4.434)
(4.416)
-.903**
-1.146**
(.407)
(.481)
.018
(.025)
-7.806***
(1.18)
1.11***
(.156)
1.11***
(.31)
1.82***
(.52)
29.88***
(5.99)
-2.32
(fixed)
.458***
(.054)
-3.187**
(1.389)
1.158***
(.264)
.138
(.563)
1.89***
(.5205)
16.112***
(4.035)
-.4323
(.358)
.559***
(.097)
-1.335
(1.713)
1.182***
(.378)
-.384
(.801)
1.484***
(1.101)
40.28***
(13.206)
1.101*
(.583)
.424***
(.089)
2.141
(1.638)
-.156
(.360)
-.768
(.824)
1.083
(.989)
14.042**
(6.493)
1.383***
(.511)
.357***
(.083)
-.310
(1.238)
1.082***
(.311)
.122
(.744)
1.827***
(.596)
17.780***
(4.438)
-1.144**
(.467)
7662
7932
≈ 1000
873
7576
.594
.707
(.182)
93.95
.784
.616
(.086)
6.354
.786
.674
(.0705)
92.61
.458
.286
(.0434)
34.80
Participation equation
ST LIAB / ASS
CASH
TLOANS
MBS PASS
MBS OTHER
ABS
Constant
Obs.
ρ
λ
χ2
2
Radj
7662
7457
.524
.329
(.0391)
82.81
.359
.218
(.0410)
16.61
0.1624
Notes: Column (1), (2), (4)-(7): Joint estimation of Treatment-effects model with binary dependent variable
T AF , using Maximum Likelihood. Column (3): Semi-Non-Parametric two-step estimation. Robust s.e. in
parentheses. *** = p < .01, ** = p < .05, * = p < .1. Column (1) repeats the results of the baseline
model using MLE. Columns (2) and (5)-(7) use different subsamples: Lehman excludes TAF auctions after
the collapse of Lehman Brothers; column (5) reports the result of the bootstrap exercise; in column (6)
TAF banks are matched with NO TAF banks; column (7) excludes banks whose bank holding companies
participated to TARP. Finally, in column (4) the variables before the beginning of the program are measured
in 2005:Q3.
53
Table 8: Too-big-to-fail and solvency
Full
(1)
75% perc. 90% perc.
(2)
(3)
Outcome equation
Dependent variable:
CASH
CAPBUFFER
ROA
PF Risk 0
PF Risk 20
PF Risk 50
PF Risk 100
TAF
95% perc.
(4)
Solvent
(5)
∆ ST LIAB ASS
-.145
(.432)
2.592***
(.251)
-.024
(.017)
.157
(.277)
.041
(.069)
-.521***
(.065)
-.084**
(.033)
-.907***
(.087)
-.998
(2.039)
2.477**
(1.142)
-.047
(.038)
-.371
(.337)
-.247
(.163)
-.474***
(.146)
.092
(.087)
-1.040***
(.087)
2.942**
(1.208)
3.387***
(1.211)
-.091
(.097)
-.454
(.349)
-.450*
(.235)
-.292
(.247)
.072
(.148)
-1.156***
(.154)
2.185
(1.751)
3.589***
(.904)
.089
(.078)
-.249
(.394)
-.227
(.382)
-.024
(.361)
-.270
(.172)
-1.231***
(.135)
.092
(.377)
2.991***
(.438)
-.032
(.025)
.091
(.295)
.334**
(.159)
-.430***
(.161)
.243
(.152)
-.656***
(.211)
-.026**
(.011)
.389***
(.070)
-1.915
(1.521)
1.182***
(.278)
.131
(.641)
2.000***
(.543)
19.586***
(4.539)
-.850**
(.410)
.441***
(.090)
-3.761
(2.929)
.070
(.327)
-1.822**
(.774)
-.776
(.698)
16.408***
(5.109)
.756*
(.454)
.440***
(.105)
1.047
(2.263)
-.068
(.369)
-2.301**
(.984)
-.268
(.857)
17.407***
(5.635)
1.086*
(.560)
.387***
(.107)
.787
(2.850)
-.382
(.524)
-2.706**
(1.053)
-.460
(1.083)
22.615***
(7.220)
1.393**
(.709)
.271*
(.157)
.238
(1.304)
.712
(.554)
-.486
(1.143)
2.356**
(1.132)
20.060***
(6.013)
-1.389*
(.832)
7662
1928
769
381
3074
.532
.335
(.0388)
55.14
271
.680
.469
(.0481)
61.89
190
.711
.560
(.0853)
45.05
145
.801
.699
(.0841)
54.91
111
.428
.248
(.0875)
6.518
51
SIZE
Participation equation
ST LIAB ASS
CASH
TLOANS
MBS PASS
MBS OTHER
ABS
Constant
Obs.
ρ
λ
χ2
No. of TAF banks
Notes: Joint estimation of Treatment-effects model with binary dependent variable T AF , using Maximum
Likelihood. Robust s.e. in parentheses. *** = p < .01, ** = p < .05, * = p < .1. Column (1) repeats the
results of the baseline model using MLE, excluding SIZE. Columns (2)-(4) only consider banks that are
larger (in terms of SIZE of all banks) than the 75, 90 and 95% quantiles. Column (5) includes only banks
with all fundamentals (CAPBUFFER, PF RISK, CASH, and ST LIAB / RISK FREE ASSETS) “better”
than the median of the fundamentals of failed TAF banks.
54
Table 9: Different dependent variables
(1)
Outcome equation
Dependent variable:
CASH
CAPBUFFER
ROA
SIZE
PF Risk 0
PF Risk 20
PF Risk 50
PF Risk 100
TAF
(2)
(3)
∆ ST LIAB ASS ∆ ST LIAB TLIAB ∆ NET LIAB
(4)
∆ ST LIAB PF RISK 0
-.315
(.438)
2.466***
(.255)
-.021
(.017)
-.038***
(.007)
.663**
(.305)
.464***
(.109)
-.075
(.114)
.388***
(.102)
-.820***
(.090)
.057
(.049)
.092***
(.028)
.001
(.001)
-.007***
(.001)
.085***
(.029)
.051***
(.015)
-.023
(.017)
.022
(.015)
-.149***
(.015)
-18.393***
(5.263)
32.079***
(3.400)
-.177
(.114)
-.941***
(.157)
4.307
(5.562)
14.070***
(2.277)
-2.848
(2.290)
12.588***
(2.062)
-28.960***
(1.202)
.087
(.462)
.291
(.371)
.063***
(.022)
-.190***
(.015)
9.191***
(.682)
1.326***
(.204)
1.607***
(.216)
.364*
(.196)
-1.728***
(.192)
-1.891
(1.505)
1.194***
(.277)
.109
(.645)
1.891***
(.542)
19.135***
(4.434)
.379***
(.070)
-2.582*
(1.418)
.878***
(.243)
1.291***
(.447)
2.434***
(.418)
14.940***
(4.027)
-2.354**
(1.009)
.622***
(.214)
.282
(.425)
1.288***
(.445)
14.998***
(3.045)
-2.184
(1.748)
.902***
(.296)
1.461***
(.554)
2.317***
(.439)
15.935***
(4.021)
Participation equation
CASH
TLOANS
MBS PASS
MBS OTHER
ABS
ST LIAB ASS
ST LIAB TLIAB
1.831***
(.303)
NET LIAB
.027***
(.002)
ST LIAB PF RISK 0
Constant
Obs.
ρ
λ
χ2
-.903**
(.407)
-3.138***
(.205)
-2.216***
(.166)
.169***
(.038)
-2.194***
(.239)
7662
7662
7790
7382
.524
.329
(.0391)
52.97
.513
.0565
(.00491)
104.7
.765
12.47
(.428)
409.8
.529
.733
(.0797)
60.40
Notes: Joint estimation of Treatment-effects model with binary dependent variable T AF , using Maximum
Likelihood. Robust s.e. in parentheses. *** = p < .01, ** = p < .05, * = p < .1.
55
Table 10: Testing joint normality assumption of the error terms
(1)
Dependent variable: ∆ ST LIAB ASS
CASH
CAPBUFFER
ROA
SIZE
PF Risk 0
PF Risk 20
PF Risk 50
PF Risk 100
TAF
MILLS
-.681*
(.399)
2.403***
(.270)
-.020
(.018)
-.030***
(.007)
.571**
(.246)
.425***
(.105)
-.018
(.106)
.454***
(.090)
-4.041***
(.719)
1.763***
(.310)
FIRST
(2)
(3)
(4)
(5)
(6)
(7)
(8)
∆ ST LIAB ASS
∆ ST LIAB ASS
∆ ST LIAB ASS
∆ ST LIAB ASS
∆ ST LIAB ASS
∆ ST LIAB ASS
∆ ST LIAB ASS
-.684
(.567)
2.409***
(.187)
-.020
(.015)
-.029***
(.009)
.567**
(.270)
.420***
(.128)
-.018
(.131)
.452***
(.118)
-4.120***
(.951)
1.911***
(.741)
.065
(.210)
-.774*
(.414)
2.428***
(.256)
-.020
(.015)
-.024***
(.008)
.509*
(.266)
.373***
(.089)
.012
(.124)
.456***
(.118)
-5.554***
(1.104)
2.188***
(.530)
-.618
(.572)
-.264
(.176)
-.808**
(.411)
2.415***
(.229)
-.020
(.014)
-.024***
(.007)
.503
(.358)
.373***
(.095)
.017
(.119)
.456***
(.105)
-5.747***
(1.035)
2.601***
(.850)
-.634
(1.526)
-.529
(.927)
-.083
(.190)
-3.132***
(.707)
2.363***
(.275)
-.023
(.019)
-.032***
(.007)
.690***
(.268)
.495***
(.094)
.116
(.123)
.623***
(.099)
-.095**
(.048)
-.354***
(.069)
-2.056**
(1.044)
2.460***
(.244)
-.017
(.016)
-.030***
(.006)
.616**
(.282)
.422***
(.088)
.010
(.120)
.520***
(.114)
-.098*
(.056)
-.279**
(.118)
-.024
(.024)
-3.209***
(.772)
2.429***
(.293)
-.015
(.017)
-.028***
(.007)
.634**
(.264)
.434***
(.105)
.117
(.123)
.616***
(.124)
-.097*
(.051)
-.431***
(.085)
.005
(.014)
.005*
(.003)
-3.363***
(.809)
2.429***
(.271)
-.014
(.020)
-.029***
(.006)
.654**
(.268)
.449***
(.100)
.133
(.129)
.630***
(.106)
-.098*
(.051)
-.437***
(.091)
.018
(.026)
.005
(.004)
-.000
(.001)
.09
(.759)
4.26
(.1189)
5
(.172)
1.01
(.3145)
3.36
(.1964)
2.36
(.5)
7662
.171
7662
.175
7662
.176
7662
.160
7662
.162
7662
.162
SECOND
THIRD
χ2
p-value
Observations
2
Radj
7662
.171
7662
.159
Notes: Bootstrap s.e. in parentheses. *** = p < .01, ** = p < .05, * = p < .1. Columns (1) to (4) report
the estimate based on a parametric approach to computing the treatment model, while columns (5) to (8)
refer to the semi-parametric two-step approach based on Gallant and Nychka (1987). The row MILLS refers
to the inverse Mills ratio or to the predicted value of the treatment model, while the rows FIRST, SECOND,
and THIRD refer to terms of higher order included to take into account potential non-normality of the error
terms. The row F-test refers to the joint test of the higher-order terms included in the specification. Null
hypothesis: the error terms are jointly normally distributed. χ2 and (p-value) reported. The results refer to
the baseline model represented respectively by equations (5), (6), and (7).
56
57
Maturity
Amounts
Institutions benefiting of the program
Depository Institutions (DI)
Primary Dealer
Primary Dealer
Commercial paper issuers
Any U.S. company†
DI
DI*
Small DI**
Notes: In order to make the programs comparable, one needs to consider both amount and maturity. ∗ The loans have been provided by a limited liability company (LLC), specially created by
the Federal Reserve Bank of New York (Fed NY). The LCC was dissolved on August 30, 2010. ∓ Five special purpose vehicles (SVP) received senior secured funding from the Fed NY in order
to finance the purchase of certain money market instruments from eligible investors. †An entity is a U.S. company if it is (1) a business entity or institution that is organized under the laws of
the United States or a political subdivision or territory thereof (U.S.-organized) and conducts significant operations or activities in the United States, including any U.S.-organized subsidiary of
such an entity; (2) a U.S. branch or agency of a foreign bank (other than a foreign central bank) that maintains reserves with a Federal Reserve Bank; (3) a U.S. insured depository institution;
or (4) an investment fund that is U.S.-organized and managed by an investment manager that has its principal place of business in the United States.‡Before and after the crisis the loans had an
overnight maturity. *DI eligible for Secondary Credit are not eligible for primary credit. **DI with significant seasonal swings.
Mechanism and Tools
TAF
12.2007-03.2010 Auctioned loans
Between 28 and 84 days $3.81 trillion
TSLF 03.2008-02.2010 Auctioned treasury general collateral
28 days
$2 trillion
PDCF 03.2008-02.2010 Loans available in exchange of collateral
overnight
$8.95 trillion
CPFF 10.2008-02.2010 Loans available in exchange of commercial papers? 90 days
$738 billion
TALF 11.2008-06.2010 Loans available in exchange of collateral
Between 3 and 5 years
$71 billion
since 2003
Between 28 and 90 days‡
since 2003
Overnight
since 2003
Period
Label
Term Auction Facility
Term Securities Lending Facility
Primary Dealer Credit Facility
Commercial Paper Funding Facility
Term Asset-Backed Securities Loan Facility
Primary Credit
Secondary Credit
Seasonal Credit
Table 11: Facilities programs operated by the Fed during the last financial crisi
Name of the program