The Impact of Quantitative Easing on Liquidity Creation

We study the effects of the US Federal Reserve’s large-scale asset purchase programs during 2008-2014 on bank liquidity creation. Banks create liquidity when they transform the liquid reserves resulted from quantitative easing into illiquid assets. As the composition of banks’ loan portfolio affects the amount of liquidity it creates, the impact of quantitative easing on liquidity creation is not a priori clear. Using a difference-in-difference identification strategy, we find that banks that were more exposed to the policy increased lending relative to a control group. However, while the increase in lending was present across all three rounds of quantitative easing, we only find a strong effect on liquidity creation during the last round. This suggests, that during the first two rounds, affected banks transformed the reserves created through the asset purchase program into less illiquid assets, such as real estate mortgages, pointing to a weaker impact of the policy on the real economy.


Introduction
Starting with the 2008-09 Global Financial Crisis, a growing number of central banks have included large-scale asset purchase programs (LSAPs) in their toolkit of unconventional monetary policies. The US Federal Reserve, in particular, implemented several rounds of quantitative easing (QE) through which they purchased both agency mortgage-backed securities (MBS) and Treasuries securities. The scale and unprecedented use of these unconventional policies has led to a large interest into understanding their effect on the banking sector and real economy. Empirical evidence thus far points to an effect of LSAPs on medium to longterm interest rates, through a signaling or portfolio-rebalancing channel (Krishnamurthy & Vissing-Jorgensen 2011, Gagnon et al. 2011, D'Amico et al. 2012, Maggio et al. 2016 Quantitative easing can also lead to an increase in credit supply through a classical bank lending channel, as banks gain new reserves and/or customer deposits, which are a relatively cheaper source of funding, and could result in a shift in the loan supply (Kashyap & Stein 2000, Butt et al. 2014, Kandrac & Schlusche 2017. 2 Yet, evidence on the impact of QE on bank lending is more confounded. Rodnyansky & Darmouni (2017) and Luck & Zimmermann (2020) find that banks increased overall lending after the first and third round of quantitative easing, with the first round corresponding to mostly an increase in mortgage origination, while the third round to an increase in both real-estate, as well as commercial and industrial loans. 3 Chakraborty et al. (2019), on the other hand, find that the increase in mortgage lending crowded-out the origination of commercial and industrial loans, the latter actually decreasing as a result of the Fed's asset purchase programs.
In this paper, we study the implications of this heterogeneous impact of QE on different types of loans for bank liquidity creation, one of the most important raison d'être of financial intermediaries. 4 Banks create liquidity in the economy by financing relatively illiquid assets 1 Under these channels, the central bank affects the relative supply of different assets, thereby lowering their yields and increasing the prices of current asset holdings of banks. The strength of the effect generally depends on the type of assets the central bank is purchasing. For instance, Maggio et al. (2016) find that, while loan interest rates decreased on average as a result of the policy, the decrease was substantially larger for assets that were conforming with the Government Sponsored Enterprises (GSEs)-guaranteed mortgages that the Fed was purchasing.
2 Regardless of whether a bank or a bank customer is the ultimate seller of the securities purchased by the Federal Reserve through QE, the reserves created by the policy will be held by banks. If the seller is a bank, securities are simply swapped for reserves on the bank's balance sheet. If the seller is a non-bank entity, bank deposits will also increase by the amount of securities sold to the Fed. 3 The Federal Reserve implemented three rounds of QE: the first (QE1) started in November 2008, the second (QE2) in November 2010 and third (QE3) in September 2012. 4 Modern theory of financial intermediation argues that banks exist to perform two central roles in the economy: create liquidity and transform risk (Diamond 1984, Ramakrishnan & Thakor 1984, Boyd & Prescott 1986. While risk transformation and liquidity creation sometimes coincide -for example when riskless liquid liabilities are transformed into risky illiquid assets-, bank liquidity creation is often seen as a distinct function such as business loans with relatively liquid liabilities such as deposits (Bryant 1980, Berger & Bouwman 2009, Berger & Sedunov 2017. This important role of banks was the main focus of policymakers at the peak of the 2008 Global Financial Crisis, when large and explicit government support was granted to banks to support liquidity provision (Acharya & Mora 2015, Bai et al. 2018. However, while the impact of early measures such as the Troubled Asset Relief Program (TARP) in supporting market liquidity is well established, the role of later unconventional policies is less clear (Acharya & Mora 2015).
QE initially makes banks' balance sheets more liquid by purchasing assets and crediting the reserves account of banks with the Fed. Banks can then use this liquidity injection to invest in relatively more illiquid assets such as loans to businesses and individuals, thereby creating new liquidity in the economy. Crucial to our analysis are the types of loans given by banks, as their liquidity differs. For instance, classical measures of liquidity creation like Berger & Bouwman (2009) assume that loans that can be securitized and sold off the balance sheet such as real estate mortgages are less illiquid and, as such, lead to less liquidity creation in the economy. Similarly, banks can also "destroy" liquidity if the new reserves or deposits resulted from QE are not transformed into illiquid loans. Hence, the amount of liquidity created in the banking sector depends on the composition of the asset side of banks' balance sheets as a result of this policy intervention.
We thus investigate the impact of the Fed's quantitative easing programs on bank liquidity creation using a sample of US bank-holding companies during 2006-2014. In doing so, we study the distributional effects of QE within the balance sheet of financial intermediaries using a difference-in-differences identification strategy that follows Rodnyansky & Darmouni (2017) and Luck & Zimmermann (2020). This strategy exploits the cross-sectional variation in banks' exposure to the Fed's large-scale asset purchase programs. The underlying argument is that banks with a higher share of mortgage-backed securities in total assets benefited more from the program. 5 We employ several definitions based on the share of MBS-to-total assets prior to QE to classify banks into treated and control groups and investigate the differential effect of the policy across banks.
of banks (Gorton & Pennacchi 1990, Gorton & Winton 2003. 5 There are several reasons why banks that held more mortgage-backed securities benefited more from the large scale asset programs. First, during the three waves of QE, Fed focused on easing the deterioration in the MBS market by lowering yields and increasing the prices of banks' current asset holdings, thereby improving the balance sheets of banks that held higher shares of mortgage-backed securities. Second, banks with more MBS sold to the Fed saw a higher increase in reserves, which should have shifted their loan supply (Kandrac & Schlusche 2017). Third, banks with higher MBS holdings might have a different business model and will particularly increase their real estate lending as their liquidity position improves. Finally, since the QE programs were largely unanticipated, especially the third round, banks that held more MBS witnessed a prompt recovery in stocks and an improved capital position (see Washington Post 2012).
We first study the impact of QE and bank lending. Similar to previous work, we find that banks with a higher MBS-to-total assets ratio had a disproportionally larger increase in lending. This differential effect is present across all three rounds of QE, when treated banks had a larger increase in both real estate and commercial loans. However, while the increase in lending was present across all three rounds of QE, we only find a robust effect on liquidity creation during the third round, when the Fed purchased a large amount of MBS securities. During this last round, banks with a higher MBS-to-total assets ratio created 4.1% more liquidity relative to their size as compared to the control group. This implies that during the first two rounds treated banks transformed the reserves created by QE into less illiquid assets such as real estate mortgages, pointing to a weaker impact of the policy on the real economy.
Our main measures of liquidity creation follow those proposed by Berger & Bouwman (2009) and Bai et al. (2018), however our results are robust to different definitions of liquidity creation.
Our findings also survive a battery of other robustness tests including various definitions of the treated and control groups, as well as controlling for bank-level variables. Furthermore, we include alongside bank fixed effects, year-quarter fixed effects to mitigate potential demandside factors that can influence the composition of banks' loan portfolio and the amount of liquidity created on their balance sheets. Our work provides a novel and robust channel through which unconventional monetary policy can affect the function of the banking sector and its impact on the real economy.
The remainder of the paper is organized as follows. Section 2 presents related literature.
Section 3 describes our conceptual framework and the mechanism we investigate. Section 4 discusses the data and identification strategy. Section 5 presents our results, while section 6 concludes.

Relation to literature
There is a growing empirical literature that studies the channels through which unconventional monetary policies such as quantitative easing are transmitted through the economy. These include the signaling channel (Krishnamurthy & Vissing-Jorgensen 2011, Bauer & Rudebusch 2014, portfolio-rebalancing channel (Gagnon et al. 2011, D'Amico & King 2013, Brunnermeier & Sannikov 2016 or reserves accumulation (Kandrac & Schlusche 2017, Butt et al. 2014. A more recent literature looks at the effects of QE on bank lending. Rodnyansky & Darmouni (2017) exploit the cross-sectional variation of banks' exposure to mortgage-backed securities to show that banks with larger MBS holdings expanded lending more than their counterparts. This disproportional increase in lending appears to come from both real estate lending as well as corporate lending. Similarly, Luck & Zimmermann (2020) find that find that the first round of QE led to mostly an increase in mortgage origination, while in the third round both mortgage lending, as well as commercial and industrial loans increased. Chakraborty et al. (2019) also find that high-MBS banks increased mortgage origination disproportionally more.
However, they also find that these banks reduced commercial lending suggesting a crowding out effect of QE. The main difference between Chakraborty et al. (2019) andRodnyansky &Darmouni (2017) rests in the way QE is defined. We use both definitions in this paper.
Furthermore, Maggio et al. (2016) shows that the type of assets purchased through QE has an impact on the type of loans originated. For example, QE1, which involved significant purchases of GSE-guaranteed mortgages, increased GSE-guaranteed mortgage originations significantly more than the origination of non-GSE mortgages. Kandrac & Schlusche (2017) show that reserves created by the Fed as a result of the first two QE programs led to higher total loan growth and an increase in the share of riskier loans within banks' portfolios. Butt et al. (2014), on the other hand, find little effect of QE on lending in the UK, since they show that the increase in deposits created by the policy was short-lived.
Our work complements these findings by focusing on a distinct channel through which QE might affect the real economy, i.e. liquidity creation. Liquidity creation is a key role of financial intermediaries that has been robustly linked to real economic growth (Berger & Sedunov 2017).
A simple measure of liquidity creation is proposed in Deep & Schaefer (2004) as the difference between liquid assets and liquid deposits. The ability to honour the obligations associated with liquid deposits while having assets that are mainly illiquid is the classic liquidity transformation mechanism associated with modern fractional reserve banking. If, for instance, banks had to hold liquid assets to fully back every dollar of liquid deposits, then they would not really be involved in liquidity creation. Effectively, they would be acquiring liquid assets and holding them on behalf of their deposits, in a similar manner to a money market mutual fund. So measuring the gap between liquid deposits and liquid assets describes the extent of liquidity creation is occurring via banks.
A more sophisticated measure of liquidity creation is described in Berger & Bouwman (2009). 6 Similar to the Deep & Schaefer (2004) measure, liquidity is created when liquid deposits are used to finance illiquid assets such as loans. The measure thus assigns positive weights to all illiquid assets and liquid liability on and off the balance sheet of each bank, suggesting that banks that use liquid liabilities to finance illiquid assets create liquidity. Similarly, banks can also destroy liquidity when illiquid liabilities and equity are transformed into liquid assets. As such, these balance sheet items are assigned a negative weight. Moreover, since the degree of "liquidity" of a balance sheet item depends on how easily it can be sold or its maturity, Berger & Bouwman (2009) assign different positive and negative weights to various balance sheet items. 7 An even more sophisticated measure of liquidity creation is proposed in Bai et al. (2018), where these weights are time-varying and depend on market conditions that affect the liquidity of different asset classes. Given these classifications of assets and liabilities proposed by different measures of liquidity creation and the heterogeneous impact of QE on different types of loans suggested by previous research, the impact of the policy on liquidity creation is not obvious.
Finally, our work is also related to a recent literature that looks at how banks' liquidity positions affects lending, in particular during periods of bank distress. For instance, Cornett et al. (2011) find that banks with more illiquid asset portfolios, i.e., those banks that held more loans and securitized assets, increased their holdings of liquid assets and decreased lending following the collapse of Lehman Brothers in 2008. Similarly, Dagher & Kazimov (2015) find that banks more exposed to market liquidity shocks cut credit more for less liquid loans. They exploit the threshold above which a loan cannot be securitized and purchased by a government sponsored enterprises (GSE) as a cut-off for loan liquidity. Our work takes a new approach to understand how banks create liquidity by looking at the effect of policy interventions on this essential feature of financial intermediation.

Transmission mechanism
The Federal Reserve implemented three rounds of QE during 2008-2012 through which it purchased mortgage-backed and/or Treasury securities by crediting the reserves accounts of banks who sold (or whose customers sold) securities to the Fed. 8 If the final seller is a bank, 7 As such the Berger & Bouwman (2009) metric is a more comprehensive measure of the liquidity transformation gap in Deep & Schaefer (2004). First, Berger & Bouwman (2009) includes off-balance sheet activities that are considered to be important contributors in liquidity creation by banks (Holmstrom & Tirole 1997, Kashyap et al. 2002. Second, Berger & Bouwman (2009) also classifies bank loans based on categories, rather than maturity. securities are simply swapped for reserves on the bank's balance sheet. If the seller is a nonbank entity, bank deposits will also increase by the amount of securities sold to the Fed.
Thus, regardless who the ultimate seller of securities is, large scale asset programs result in an increase in bank reserves. This is evident in Figure 1 for our sample of banks. A notably sharper increase can be observed after QE3, which entailed the largest volume of purchased assets and, as a result, reserves creation. This significant injection of reserves should affect banks' optimal portfolio allocation by changing their liquidity profile and duration of assets (Joyce & Spaltro 2014, Kandrac & Schlusche 2017). This might, in turn, induce banks to engage in additional lending (see Bianchi & Bigio 2014, for a general equilibrium model). However, from the point of view of the amount of overall liquidity created in the banking sector, the composition of this increase in lending is important. Specifically, banks create liquidity when they transform liquid assets or liabilities into illiquid assets. However, different categories of assets have different degrees of illiquidity.
For example, business loans are generally more illiquid than residential mortgages as the latter At the end of the three rounds, the balance of the Fed contained $1.75 trillion MBS and $1.68 trillion Treasury bills.
can often be more easily securitized and sold to meet liquidity needs (see Berger & Bouwman 2009). As we will show below, this has non-obvious implications for the amount of liquidity creation.
We use a simple example to illustrate the confounding effects of QE on the liquidity created by banks. Our main measure of liquidity creation follows Berger & Bouwman (2009) and classifies assets and liabilities into three categories: liquid, semi-liquid and illiquid. For assets, this depends on how easy and fast a bank can sell them to meet liquidity demands, while for liabilities, on how easy customers can withdraw their funds from the bank. Weights are then assigned to reflect the idea that liquidity creation occurs when the bank finances relatively illiquid assets with relatively liquid liabilities. Therefore, a weight of 1/2 is applied to illiquid assets and liquid liabilities. Conversely, a weight of -1/2 is applied to liquid assets and illiquid liabilities and a weight of 0 is assigned to semi-liquid assets and liabilities. Appendix B discusses in detail the construction of this liquidity index.
The example in Table 1 shows how liquidity creation following the definition above can be affected by QE. In this example a "Treated Bank" is one which sells MBS to the Fed for a value of, say, 100, which results in a corresponding increase in Reserves by 100. The "Control bank" is not affected by the asset purchase program, but we assume all banks have an increase in deposits of 10. Suppose the "Control bank" invests the 10 additional deposits in commercial and industrial (C&I) loans. This leads to a liquidity creation of 10 by transforming the most liquid liabilities (deposits), which have a weight of 1/2 in the Berger & Bouwman (2009) index, into the most illiquid assets (loans to enterprises), which are also assigned a weight of 1/2. We then analyze three different scenarios, where the Treated banks also invest the additional deposits of 10 in C&I loans, but differ in how they invest the new reserves created by the LSAP.
In Case 1, the Treated bank simply keeps the reserves on its balance sheet. Since the bank has substituted a relatively illiquid assets for a very liquid one, it is "destroying" liquidity according to the Berger & Bouwman (2009) measure, as the newly added reserves on the balance sheet of the bank are assigned a weight of -1/2. The total amount of liquidity destroyed is -40, as the bank created liquidity in the amount of 10 (by transforming 10 of liquid deposits into 10 C&I loans, as in the benchmark control bank) and destroyed liquidity in the amount of 50 (-1/2 × 100) by substituting an illiquid asset with a very liquid one. Thus, QE, by making the asset side of banks' balance sheets more liquid, results in less liquidity being created in the financial sector if banks do not engage in additional lending.

Assets Liabilities
Reserves +0 Deposits +10 MBS -100 C&I Loans +0 RE lending +110 In Case 2, we assume the bank uses the reserves to fund mostly Real Estate (RE) loans. Since RE lending can be securitized and sold, it is considered a semi-liquid asset and is assigned a weight of 0. As such, in Case 2, the bank does not create any liquidity in the system. In Case 3, we assume that the bank uses all reserves to invest in both RE lending and C&I lending in equal shares. In this case, the level of liquidity created is greater than that of the control bank. Finally, Case 4 assumes that the QE program crowds out C&I lending by making real estate loans more appealing. In this case, the liquidity created is lower than the one of the control bank.
As this simple example shows, whether banks exposed to QE create more liquidity in the banking sector depends crucially on the distribution of assets on their balance sheet after the policy. If QE crowded out C&I lending as shown in Chakraborty et al. (2019), we should expect that treated banks created less liquidity as compared to the control ones. If banks increase both real estate and industrial lending, the amount of liquidity created depends on the relative size of each asset class. As such, the effect of QE on liquidity creation is not a priori clear.

Data and identification strategy
We obtain bank-level data from the Consolidated Financial Statements for Bank Holding Companies (BHC), FR Y-9C quarterly reports that are filled by BHC with at least $500 million in total assets. 9 The FR Y-9C reports provide not only balance sheet data, but also capital positions, risk-weighted assets, securitization activities and off-balance sheet exposures, 10 Consolidated Report of Condition and Income (FR Y-9C) contains separate reporting for the parent company of large BHCs (FR Y-9LP) and parent company of small BHCs (FR Y-9SP). The number of observations varies from quarter to quarter because the Y-9SP is collected on a semiannual basis (in June and December). Since holding companies that file this report are included in those quarters, there is a significant increase in the number of observations for June and December. The first and third quarter only include banks that file the Y-9C and Y-9LP.
11 Banks create liquidity off the balance sheet through guarantees that allow customers to draw-down liquid funds when needed (Kashyap et al. 2002). includes off-balance sheet components and finally, (iv) mat non-fat includes a classification by maturity but excludes off-balance sheet items. As the authors argue, the most comprehensive measure is the cat fat one. This will also be our main measure of liquidity creation, but we will employ some of the other measures in robustness checks.
We also consider an additional liquidity measure based on Bai et al. (2018) Bai et al. (2018) and use their data on repo market haircuts and spreads (price-based measures) to construct the index. The measure is constructed to capture a maturity mismatch, i.e., how much cash the bank can raise against its balance sheet to withstand the cash withdrawals in case of a stress event in which all claimants seek to extract the maximum liquidity. Since our goal is to employ an index of liquidity creation and not mismatch, we change the signs of the weights accordingly. A description of the weights and construction of the index is presented in Appendix B.
Our identification strategy follows Rodnyansky & Darmouni (2017) and exploits the crosssectional variation in MBS holdings across banks. This methodology relies on the assumption that banks that held more MBS on their balance sheet were more likely to be affected by the Fed's asset purchases. Several arguments support this claim. First, during the three waves of QE, the Fed focused on easing the deterioration in the MBS market by lowering yields and increasing the prices of banks' current asset holdings, thereby improving the balance sheets of banks that held higher shares of mortgage-backed securities. Second, banks with more MBS sold to the Fed saw a higher increase in reserves, which should have shifted their loan supply (Kandrac & Schlusche 2017). Third, since the QE programs were largely unanticipated, more affected banks witnessed an improvement in their market capitalization (see Washington Post 2012).
We measure a bank's exposure to QE by the ratio of MBS-to-total assets. Following Rodnyansky & Darmouni (2017), we define as the treatment group banks in the highest 25% of the MBS-to-total assets distribution, while those in the lowest 25% are included in the control group. To minimize endogeneity, banks are classified according to their MBS-to-total assets ratio in 2007:Q4, which is more than half a year before QE1. We also consider several alternative definitions for the assignment to treatment and control groups. First, we classify banks in the top decile of the distribution of MBS-to-total assets into the treatment group, while those in the bottom decile in the control. Second, we employ the ratio of MBS-to-total assets in 2007:Q4, which allows for an analysis of the entire sample of banks. The table shows correlations between the treatment condition and bank characteristics in 2007Q4. T reat i is a dummy that takes the value one for banks in the 75 th percentile of the MBS-to-total assets ratio, and zero for banks in the 25 th percentile. T reat D i is a dummy that takes the value one for banks in the 90 th percentile of the MBS-to-total assets ratio, and zero for banks in the 10 th percentile. M BS Assets i is the ratio of MBS to Total assets in 2007:Q4. Robust standard errors in brackets. ***, **, * represent significance at the 1%, 5% and 10%, respectively.
As shown in Rodnyansky & Darmouni (2017), the classification of banks into treatment and control is rather stable over time, as the level of MBS-to-total assets is fairly sticky. This alleviates the concern that banks might respond strategically to the LSAPs by increasing their holdings of mortgage-based securities. Nonetheless, it might be that banks in the treatment  Table 3, where T reat i is the treatment definition based on quartiles (column 1), T reat D i the one based on deciles (column 2), and M BS Assets i is the ratio of MBS-to-total assets in 2007:Q4 (column 3).
These simple correlations suggest that banks that hold more mortgage backed securities tend to be different than control banks along several characteristics, which include size (log of assets), Tier 1 Capital ratio, the ratio of securities to total assets, and the log of net income.
As such, treated banks are typically larger, more leveraged, hold more securities and have lower net income. We will thus control for these bank characteristics throughout our analysis.
The underlying argument behind our identification strategy is that banks with a higher share of mortgage-backed securities in total assets prior to QE (treated banks) benefited more from the program. Figure 2 shows the reserve accumulation by treated and control banks throughout the sample period. Clearly, we observe that banks in the treatment group witnessed a higher surge in reserves relative to control banks, potentially as a result of QE. deposits on the liabilities side of banks' balance sheets (Choulet 2015). Figure 4 shows that customer deposits did increase in the sample of treated banks, especially after QE2. That being said, it is clear that no single mechanism explains why banks with higher MBS were more affected by the Fed MBS purchases, rather this can be explained through a variety of distinct direct and indirect purchase mechanisms.
Our identification strategy exploits the cross-sectional variation in banks' exposure to the Fed's large-scale asset purchases via difference-in-differences regressions, as follows: capital, profitability and level of securities, which were the main variables correlated to all treatment definitions. We control for bank fixed effects to remove all time-invariant differences across banks. Bank fixed effects also capture the average difference in liquidity creation between treated and control banks across the sample period. We also add year-quarter fixed effects to control for unobserved macroeconomic conditions that might affect both the demand and supply of bank loans.

Results
This section examines the impact of Federal Reserve's LSAP on lending behaviour of banks, and liquidity creation. First, we consider the effects of the three rounds of QE on lending, distinguishing between total lending, real estate (RE) loans and commercial and industrial (C&I) loans. Second, we present our main results pertaining to liquidity creation. Lastly, we present a series of robustness tests of our main results. The results are presented in Table 4. Columns (1)-(2) pertain to Total lending, while columns (3)-(4) to RE lending and (5)-(6) to C&I loans, respectively. Across both definitions of treated  (1)-(2) is log of Total lending, in Columns (3)-(4) is the log of real estate loans and in Column (5)-(6) is the log of commercial and industrial loans. T reat i is a dummy that takes the value one for banks in the 75 th percentile of the MBS-to-total assets ratio, and zero for banks in the 25 th percentile. M BS Assets i is the ratio of MBS to Total assets in 2007Q4. QE1, QE2, QE3 are dummies for each QE wave. Bank-level controls include the log of Total Assets, Tier 1 Capital Ratio, the log of Net Income and the log of Securities. Constant terms included but not reported. Robust standard errors in parentheses. ***, **, * represent significance at the 1%, 5% and 10%, respectively. and control banks, we find that treated banks expanded total lending more than control banks.

The impact of QE on bank lending
These effects are present across all three rounds of QE. With the natural logarithm of lending as the dependent variable, our estimates in column (1) suggest that QE1 boosted lending of treated banks by 3.2% relative to the control group, QE2 by 4%, while QE3 by 8.2%, respectively. These results are mostly consistent across both RE and C&I lending, and show a larger quantitative impact of the third round of QE.
Our results complement those in Rodnyansky & Darmouni (2017), however in their case, the effects appear stronger for QE1 and QE3 and more robustly related to an increase in RE lending as opposed to C&I loans. Our robustly estimated impact of QE on lending across all rounds of QE is most likely the result of our smaller sample of banks (we have 964 BHC with data on MBS/Assets in 2007Q4, as opposed to their sample of 3,949). Moreover, since our sample mainly includes the right tail of the bank distribution by assets, our results suggest that the lending effects might be stronger among these larger banks. This points to important heterogeneous effects of the policy across the banking sector and might shed some light on the mixed evidence in previous research (see Rodnyansky & Darmouni 2017, Chakraborty et al. 2019).  (1) T reat i is a dummy that takes the value one for banks in the 75 th percentile of the MBS-to-total assets ratio, and zero for banks in the 25 th percentile. T reat D i is a dummy that takes the value one for banks in the 90 th percentile of the MBS-to-total assets ratio, and zero for banks in the bottom 10 th percentile. M BS Assets i is the ratio of MBS-to-total assets in 2007Q4. QE1, QE2, QE3 are dummies for each QE wave. Bank-level controls include Tier 1 Capital Ratio, the log of Net Income and the log of Securities. Robust standard errors in parentheses. ***, **, * represent significance at the 1%, 5% and 10%, respectively.

QE and bank liquidity creation
We now turn our main empirical specification that estimates Equation 1 for the two main measures of liquidity creation we employ in this paper, namely the Berger & Bouwman (2009) cat fat measure and Bai et al. (2018) LMI index. We scale both dependent variables by total assets. The results are presented in Table 5. We employ three treatment variables that classify banks based on quantiles, deciles and the continuous measure of MBS-to-total assets.
As before, the main variable of interest is the interaction term between the QE time dummies and banks' treatment status. Our results are consistent across the two measures of liquidity creation we employ, suggesting they capture similar bank behavior. Overall, we find that treated banks created a disproportionally larger amount of liquidity in the banking sector, but mainly during the third round of QE. The interaction term between QE3 and the treatment status is the only one that is robustly estimated across the three definitions of treated banks.
With the ratio of liquidity creation to total assets as the dependent variable, the estimates in  (1)-(3) is ratio of Berger & Bouwman (2009) measure of liquidity creation to total assets, while in Columns (4)-(6) it is the Bai et al. (2018) LMI index. T reat i is a dummy that takes the value one for banks in the 75 th percentile of the MBS-to-total assets ratio, and zero for banks in the 25 th percentile. T reat D i is a dummy that takes the value one for banks in the 90 th percentile of the MBS-to-total assets ratio, and zero for banks in the bottom 10 th percentile. QE1, QE2, QE3 are dummies for each QE wave. Bank-level controls include Tier 1 Capital Ratio, the log of Net Income and the log of Securities. Robust standard errors in parentheses. ***, **, * represent significance at the 1%, 5% and 10%, respectively. column (1) suggest that, during QE3, treated banks created 4.1% more liquidity relative to their size as compared to the control group.
In our most conservative definition of treated banks that includes banks in the 90 th percentile of MBS/Assets versus those in the bottom 10 th (columns (3) and (6)), we find, as expected, a strong difference in liquidity creation across all the three rounds of QE. However, this includes only a small percentage of banks. These results, coupled with the ones in Table 4, suggest a strong heterogeneous impact of the LSAPs depending on bank characteristics, chiefly those related to size and securities holdings. Moreover, the results in Tables 4 and 5 suggest that, while banks with higher MBS/Total Assets were characterized by a disproportionally higher level of lending during all three rounds of QE, this increase in leading resulted in a higher liquidity creation only during QE3. This implies that, during the first two rounds, treated banks transformed the reserved created by QE into less illiquid assets such as RE loans, which points to a less important impact of the policy on the real economy.
As bank size seems to matter, we check the robustness of our results by employing a matching procedure that matches our treated and control groups by size, measured by the log of total assets. Banks are matched using propensity scores based on a logit model in 2007Q4 that relates the probability of being assigned to the treated group to their level of total assets. We consider the definitions of treatment groups based on the 75 th and 90 th percentile, respectively.
We then employ this propensity score to re-weight treatment and control groups such that the distribution of bank size looks the same in both groups. This is done using the conditional probability of being in the treated group,λ, to compute a weight as the odds ratioλ/(1 −λ) (see Nichols 2007). We re-estimate the model in Equation 1 using the weighted data based on propensity scores. The results are presented in Table 6. As before, columns (1) index. The estimations yields consistent results, with a strong differential impact on liquidity creation mainly present during QE3. As expected, the strong differences between control and treatment groups we found when comparing the 10 th versus the 90 th percentile, are less robust when we match banks by size, suggesting the effects in Table 5 were largely driven by large banks. Nonetheless, we still find a statistically stronger impact during QE3, in particular for the Berger & Bouwman (2009) measure.

Alternative identification strategy
An alternative identification strategy is proposed in Chakraborty et al. (2019) who investigate the impact of the Fed's LSAPs on bank lending and firm investment. They employ as main The dependent variable in columns (1)-(4) is the (Berger & Bouwman 2009) ratio of liquidity to total assets, while in columns (5)- (8) it is the Liquidity Mismatch Index (LMI) to total assets in (Bai et al. 2018). M BS t−1 and T SY t−1 are the of log amount of mortgage-backed securities Treasury securities purchased by the Fed during 2008-2014. T reat i is a dummy equal 1 for banks in the 75 th percentile of MBS-to-assets ratio in 2007Q4, and zero for those in the 25 th percentile. T reat D i is a dummy equal 1 for banks in the 90 th percentile of MBS-to-assets ratio in 2007Q4, and zero for those in the 10 th percentile. Bank-level controls include the log of bank's net income, the log of securities, and Tier 1 risk-based capital ratio. Constant term included but not reported. Robust standard errors in parentheses. ***, **, * represent significance at the 1%, 5% and 10%, respectively. independent variable the actual amount of MBS and treasury securities purchased as opposed to time dummies corresponding to the introduction of each QE episode. Figure  The most robust evidence points to an impact on liquidity creation following MBS purchases, and less so following purchases of T-bills, which mainly occurred during QE2. This is in line with our previous results and Chakraborty et al. (2019), who also find an impact on lending mainly following MBS purchases. Yet, unlike Chakraborty et al. (2019), who find that real estate mortgages crowded out commercial loans, we find a consistently positive impact on liquidity creation.

Other robustness checks
We perform a series of further robustness checks of our main results. First, we introduce a new treatment variable based on the mean values of MBS holdings to total assets. This dummy variable takes the value of 1 if a bank is in the top 50% of mortgage backed securities to total assets in 2007Q4 and 0 if it lies in the bottom 50 th percentile. The results are presented in Appendix Table 10 and are qualitatively similar to the ones obtained in our main specification, albeit less robustly estimated. We still find a stronger support for an differential increase in lending and liquidity creation during QE3.
Second, we conduct a sub-sample analysis by dropping observations in the first and third quarter in each year in which small BHCs that only file the FR Y-9SP do not report data.
The results employing the different lending categories as dependent variable are only robustly estimated during QE3, while for liquidity creation we only find a significant result when employing the Berger & Bouwman (2009) measure of liquidity creation (see Appendix Table   11.) Next, in Appendix Table 12, we consider alternative proxies for liquidity creation, namely the cat nonfat measure in Berger & Bouwman (2009) and the liquidity transformation gap proposed by Deep & Schaefer (2004). First, we construct the Berger & Bouwman (2009) catnonfat index (scaled by total assets) that includes loans based on category (cat) and excludes off-balance sheet items (nonfat). Figure 6 shows that both measures of liquidity creation follow similar trends at the aggregate level: there in a spike just prior to the start of the 2008 Global Financial Crisis and followed by a sharp decline at the start of the crisis and a gradual increase afterwards. The increase is more pronounced after 2012 which corresponds to the start of QE3, particularly for the cat fat measure, which is the main one employed in our analysis. Nonetheless, we obtain consistent results when we employ the alternative dependent variable, suggesting that there was not a significant liquidity destruction off-balance sheet among treated banks (see Appendix Table 12). Second, we construct the measure of liquidity transformation in Deep & Schaefer (2004) as the difference between liquid liabilities and liquid assets, normalized by total assets. A higher liquidity transformation gap occurs when banks are largely financed by liquid deposits and hold mostly illiquid loans. Results in Appendix Table 12 suggest a significantly higher increase in liquidity transformation among treated banks across all three rounds of QE. This is not surprising considering the simpler concept of liquidity creation captured by this measure and the fact that lending was found to be disproportionally higher in all rounds of QE among treated banks (see Table 4).

Conclusions
We study the effects of large scale asset purchases on bank liquidity creation. While existing evidence shows how LSAPs can affect bank lending, our work takes a new approach by looking at whether banks that benefited more from the Fed's three rounds of QE have also contributed more to the creation of liquidity in the economy.
We show that banks with higher share of assets in mortgage-backed securities prior to the start of program have increased both real estate and commercial loans disproportionally more following all three rounds of QE. However, not all types of loans contribute the same to liquidity creation, which increases more when banks give out more illiquid loans such as commercial lending. As such, we find evidence that treated banks contribute more to liquidity creation only the last round of QE which started in 2012 and when the Fed bought large amounts of MBS. Similar to previous evidence, our work points to important asymmetric effects of this unconventional monetary policy across banks and suggests that its impact on liquidity creation, as one of the main functions of the banking sector, was not strong across the entire duration of the program.

B Liquidity Creation:
One of the primary reasons that banks exist is because they create liquidity, through balance sheet activities, such as provisioning of loans to businesses and individuals or through offbalance sheet activities: loan commitments to their customers, extending letters of credit etc. In our analysis, we employ two indices of liquidity creation, namely, Berger & Bouwman (2009) measure and the Liquidity Mismatch Index (LMI) propose by Bai et al. (2018). Both the liquidity measures take into account the components of on and off-balance sheet including assets, liabilities, equity and off-balance sheet items such as loan commitments and derivatives.
Both the liquidity creation measures take into account all bank activities (all assets including different types of loans based on category, all liabilities, equity capital and all off-balance sheet activities). Both measures also recognize that banks create liquidity, but can also destroy liquidity (Berger & Bouwman 2015). In addition, LMI also includes price-based measures such as, haircuts, spreads etc. LMI aims to measure the liquidity imbalances in the system and the amount of liquidity the Fed would have to provide to BHCs during crisis. Table 8 presents the weights employed in the creation of these two indices. In Berger & Bouwman (2009), assets and liabilities are classified as liquid, semi-liquid and illiquid. The classification of loans is done through categories, and this measure also includes off-balance sheet items. The first step in LMI calculation involves assigning weights to the market liquidity of assets, which range from 0 (hard or time-consuming to sell, such as fixed assets) to 1 (very liquid items such as cash). The second step multiplies each initial weight by one minus the repo haircut of the asset class. The calculation of asset side weights includes haircuts as it measures how much cash can be borrowed against the asset. Then, haircut adjusted weights are multiplied by each asset category. Similarly, the same steps are repeated for funding liquidity of liabilities. The key difference occurs in the liability weights where they are assigned based on maturity. Each initial weight for liability is multiplied liquidity premium (spread between the overnight index swapped rate and Treasury bill rate). Since LMI is an indicator that measures mismatch of liquidity between assets and liabilities, we revise its weights to convert it into a liquidity creation measure by changing the sign to match that of the Berger & Bouwman (2009) index (see Column 5 in Table 8). Net participation acquired (Liquid) -1/2 Notes: 1. All securities regardless of maturity are taken as liquid assets under Berger-Bouwman index 2. Loans secured by real estate is a sum of residential and commercial real estate loans 3. Unused commitments include revolving, open-end loans, unused credit card lines, to fund commercial real-estate related loans, to provide liquidity to ABCP conduit structures, to provide liquidity to securitization structures, other unused commitments 4. Credit lines include financial standby letters of credit, performance standby letters of credit, commercial and similar letters of credit. 5. Haircut is the difference between asset's collateral value and its sale price. 6. Overnight index swaps (OIS) enable financial institutions to exchange fixed rate interest payments for floating rate payments based on specified principal amount.  (1) and (2) is the log of total lending, Column (3) and (4) is the log of real estate loans and Column (5) and (6) is the log of corporate loans. MBS purchases is the lagged of log amount of mortgage-backed securities purchased by the Fed and TSY purchases is the lagged of log amount of treasury securities purchased by the Fed. We take Treated as the bank treatemnt status defined by top 75 th percentile of MBS-to-assets ratio in 2007Q4, while control group belongs in the bottom 25 th percentile. We include controls such as log of total assets (proxy for bank size), log of bank's net income and Tier 1 risk-based capital ratio (proxy for bank capitalization). Constant term included but not reported. Robust standard errors in parentheses *** p<0.01, ** p<0.05, * p<0.1  (1) is the log of Total lending, in column (2) the log of real estate loans, in column (3) corporate loans, in column (4) it is the ratio of Berger & Bouwman (2009) cat fat to total assets and finally in column (5) the ratio of Bai et al. (2018) LMI to total assets, respectively. T reat M i is a dummy that takes the value one for banks that lie above the 50 th percentile of the MBS-to-total assets ratio, and zero for banks below the 50 th percentile. QE t is a dummy variable for each QE wave, where t= 1,2,3. Bank-level controls include the log of Assets (in columns (1)-(3)), Tier 1 Capital Ratio, the log of Net Income and the log of Securities. Constant terms included but not reported. Robust standard errors in parentheses. ***, **, * represent significance at the 1%, 5% and 10%, respectively.  (1)-(2) is log of Total lending, in Columns (3)-(4) is the log of real estate loans and in Column (5)-(6) is log of corporate loans. T reat i is a dummy that takes the value one for banks in the 75 th percentile of the MBS-to-total assets ratio, and zero for banks in the 25 th percentile. M BS Assets i is the ratio of MBS-to-total assets in 2007Q4. QE t is a dummy variable for each QE wave, where t= 1,2,3. Bank level controls include the log of Total assets (in columns (1)-(6)), the log of Net income and securities and Tier 1 Capital ratio. Constant terms included but not reported. Robust standard errors in parentheses. ***, **, * represent significance at the 1%, 5% and 10%, respectively.  (1)-(3) is ratio of Berger & Bouwman (2009) cat-nonfat liquidity measure to total assets (TA), while in Columns (4)-(6) liquidity transformation gap in Deep & Schaefer (2004). T reat i is a dummy that takes the value one for banks in the 75 th percentile of the MBS-to-total assets ratio, and zero for banks in the 25 th percentile. T reat D i is a dummy that takes the value one for banks in the 90 th percentile of the MBS-to-total assets ratio, and zero for banks in the bottom 10 th . M BS Assets i is the ratio of MBS-to-total assets in 2007Q4. QE t is a dummy variable for each QE wave, where t= 1,2,3. Bank-level controls include: Tier 1 capital ratio, log of net income and log of securities. Constant term included but not reported. Robust standard errors in parentheses. ***, **, * represent significance at the 1%, 5% and 10%, respectively.