QE, Bank Liquidity Risk Management, and Non-Bank Funding: Evidence from U.S. Administrative Data

Finance and Economics Discussion Series Federal Reserve Board, Washington, D.C. ISSN 1936-2854 (Print) ISSN 2767-3898 (Online) QE, Bank Liquidity Risk Management, and Non-Bank Funding: Evidence from U.S. Administrative Data Darst, R. Matthew, Sotirios Kokas, Alexandros Kontonikas, Jose-Luis Peydro, and Alexandros P. Vardoulakis 2025-030 Please cite this paper as: Darst, R. Matthew, Sotirios Kokas, Alexandros Kontonikas, Jose-Luis Peydro, and Alexandros P. Vardoulakis (2025). “QE, Bank Liquidity Risk Management, and Non-Bank Funding: Evidence from U.S. Administrative Data,” Finance and Economics Discussion Series 2025-030. Washington: Board of Governors of the Federal Reserve System, https://doi.org/10.17016/FEDS.2025.030. NOTE: Staff working papers in the Finance and Economics Discussion Series (FEDS) are preliminary materials circulated to stimulate discussion and critical comment. The analysis and conclusions set forth are those of the authors and do not indicate concurrence by other members of the research staff or the Board of Governors. References in publications to the Finance and Economics Discussion Series (other than acknowledgement) should be cleared with the author(s) to protect the tentative character of these papers.

QE, Bank Liquidity Risk Management, and Non-Bank Funding: ∗ Evidence from U.S. Administrative Data R. Matthew Darst Sotirios Kokas Alexandros Kontonikas Federal Reserve Board University of Essex University of Essex Jose-Luis Peydro Alexandros P. Vardoulakis Imperial College London & CEPR Federal Reserve Board April 21, 2025 Abstract We show that the effectiveness of unconventional monetary policy is limited by how banks adjustcreditsupplyandmanageliquidityriskinresponsetofragilenon-bankfunding. For identification, we use granular U.S. administrative data on deposit accounts and loan-level commitments, matched with bank-firm supervisory balance sheets. Quantitative easing increases bank fragility by triggering a large inflow of uninsured deposits from non-bank financial institutions. In response, banks that are more exposed to this fragility actively manage their liquidity risk by offering better rates to insured deposits, while cutting uninsured rates. Doing so, they shift away from uninsured to insured deposits. Importantly, on the asset side, these banks also reduce the supply of contingent credit lines to corporate clients. This tightening of liquidity provision has real effects, as firms reliant on more exposed banks experience a reduction in liquidity insurance stemming from credit lines, leading to lower investment. Our analysis reveals that the fragility of deposit funding can disrupt the complementarity between deposit-taking and the provision of credit lines. Keywords: Bank fragility, Liquidity risk, Liquidity Insurance, Deposits, Credit lines, Quantitative Easing, Quantitative Tightening, Non-banks ∗WearegratefultoAnilKashyap,BethKlee,ViralAcharya,andseminarparticipantsattheFederalReserve Board,King’sCollegeLondon,LoughboroughUniversity,Toulouse,andVirginiaTech. WethankAraziLubisfor excellent analytical support. The views expressed in this paper are those of the authors and do not necessarily represent those of the Federal Reserve Board, or anyone in the Federal Reserve System. All errors are our responsibility. Emails: matt.darst@frb.gov; skokas@essex.ac.uk; a.kontonikas@essex.ac.uk; jose.peydro@ gmail.com;alexandros.vardoulakis@frb.gov.

1 Introduction Assetpurchasesbycentralbanksviaquantitativeeasing(QE)andtheirreversalviaquantitative tightening(QT)haveplayedanimportantroleintheconductofmonetarypolicyoperationssince the Great Recession (Bernanke, 2022). Central banks fund these purchases by issuing central bank liabilities, known as reserves. The exchange of reserves for securities alters the portfolio composition of the private sector and the risk premium investors require to hold long-duration securities (Vayanos and Vila, 2021). Yet, the effects of quantitative policies on the financial system and the economy extend beyond the change in long-term yields and securities prices. Asset purchases can affect the size and composition of financial institutions’ liabilities, resulting in an expansion of more fragile forms of funding for banks (Acharya and Rajan, 2024; Acharya, Chauhan, Rajan, and Steffen, 2023). We show that the effectiveness of unconventional monetary policy is limited by how banks adjustcreditsupplyandmanageliquidityriskinresponsetofragilenon-bankfunding. Ourmain contribution shows that banks who are more exposed to QE-induced funding fragility actively, and simultaneously, manage their deposit liabilities and loan commitments to reduce liquidity risk. On the liabilities side, more exposed banks increase (decrease) the rates offered on insured (uninsured) deposits, which facilitates a shift from uninsured to insured deposits. On the asset side, they reduce the supply of contingent credit lines to firms. Importantly, the relatively lower liquidity insurance provided to firms via committed lines of credit has real effects and results in less firm investment. To that extent, our analysis uncovers novel results on the documented complementarity between deposit funding and bank-provided liquidity insurance, highlighting unintended consequences of quantitative policies. To support our empirical results, we extend themodelofKashyap,Rajan,andStein(2002)byintroducingrunnabledepositsakintoDiamond and Kashyap (2016) and show that deposit fragility can disrupt the complementarities between deposit-taking and credit-line issuance. To the best of our knowledge, this is the first evidence showing that the Kashyap et al. (2002)’s documented complementarity may break down when deposit inflows come from fragile funding, such as uninsured NBFI deposits. We use two administrative datasets that provide confidential information on U.S. deposits and lending. First, we utilize data from the Complex Institution Liquidity Monitoring Report 1

(FR 2052a), a component of the Federal Reserve’s supervisory surveillance program for liquidity riskmanagement. FR2052adatahaveuniqueadvantages,intermsofgranularityandfrequency, compared to publicly available regulatory bank filings. The data are daily or monthly, provide information about deposit counterparty-types, including NBFIs, and indicate whether deposits are insured or uninsured as well as their maturity. Second, we use granular information about bank loan commitments from FR Y-14Q, quarterly collected by the Federal Reserve as part of the Comprehensive Capital Assessment and Review (CCAR) stress testing process. The data include the type of loan (term loan or credit line), total loan commitment and utilized amounts, pricing information as well as information about firms’ investment, which allows us to examine therealeffectsoftheQE-inducedfragility. Importantly, ourdatacoversbothpublicandprivate firms. The deposits and lending datasets are supplemented with Call Reports information on bankcharacteristicsanddepositsratedatafromRateWatch. Theresultingrichgranulardataset is combined with a multi-stage empirical approach to estimate the response of deposits and lending outcomes to funding fragility, stemming from unconventional monetary policy. How does an increase in uninsured deposit funding influence bank strategies for managing assetsandliabilities? Answeringthisquestionischallengingduetotheendogenouslinksbetween bank assets and liabilities. Banks simultaneously originate loans and create uninsured demand deposits, particularly when loan sizes exceed deposit insurance limits. Moreover, when issuing credit lines, they generate contingent claims on liquidity. Our novel identification strategy exploits the fact that COVID-driven QE led to a surge in nonbank deposits, altering the funding composition of banks that were more exposed to nonbanks before the pandemic. This variation, exogenous to both COVID and the QE response, allows us to isolate the impact of an external funding shock. Crucially, these deposit inflows were not inherently related to banks’ loan origination or liability management decisions, but stemmed from changes in non-bank liquidity holdings. This enables us to study how banks adjusted their balance sheets in response to an exogenous shock to funding fragility. Toensurethatourempiricalstrategyisnotconfoundedbyexistingdifferencesbetweenbanks withdifferentlevelsofexposuretoNBFIfunding,weassessbalancestatisticscomparabilityacross key characteristics before the onset of QE (Roberts and Whited, 2013; Imbens and Wooldridge, 2009). Our analysis confirms that, apart from differences in total uninsured deposits and total 2

NBFI deposits, banks were otherwise similar in terms of size, capital levels, loan composition, and asset holdings. This comparability strengthens the validity of our approach, ensuring that theobservedresponsestoQE-inducedfragilityreflectasystematicreactiontofundingriskrather than pre-existing structural differences across banks. In addition, we control for different sets of fixedeffects,takingadvantageofthegranularityinourdata,totackleunobservedheterogeneity. We provide four key results. First, we show that the more exposed banks experience a higher inflow of uninsured NBFI deposits during QE. This result is robust to controlling for, among others, (i) bank size and the presence of Global Systemically Important Banks (GSIBs); (ii) for other policy interventions during this period, namely the relaxation and re-activation of the Supplementary Leverage Ratio (SLR); and (iii) for draw-downs of credit lines by NBFIs, which would mechanically push their deposits up. Moreover, we confirm there was no pre-trend difference in NBFI deposits between more and less exposed banks. Second, we show that more exposed banks actively manage the liquidity risk of their deposit liabilities. Relativetolessexposedbanks,moreexposedbanksreducebothnon-NBFIuninsured depositsandtotaluninsureddeposits. Hence,theyovercompensatefortheinfluxoffragileNBFI funding by reducing other sources of fragile funding. In addition, more exposed banks increase their insured deposits more, making up for the decrease in uninsured deposits. Importantly, we showthattheshiftfromuninsuredtoinsureddepositconstituteactiveliquidityriskmanagement by the exposed banks. Results from the analysis of deposit rates suggest that more exposed banks increase the deposit rates offered for insured deposits, while decreasing the remuneration of uninsured deposits, consistent with an effort to reduce exposure to funding fragility. Thus, the active reshuffling and repricing of deposit liabilities suggest that banks strategically manage their funding structure, consistent with a bank-driven adjustment to mitigate liquidity risk. Third, we show that the more exposed banks decrease the credit lines to firms relative to less exposed banks. Note that credit-line commitments increased for both types of banks during the QE and the inflow of reserves. However, our granular data and the novel identification of QE exposure allows us to capture the differential effect. By contrast, there is no significant difference in the term loans offered by more and less exposed banks. Zooming in the credit-line sub-components, the reduction is associated with the undrawn credit line amount, while there is no difference with respect to credit-line utilization between the more and less exposed banks. 3

Hence, the more exposed banks effectively manage the liquidity risk on their loan exposures by reducing the claims to future liquidity and, thus, decreasing the possibility of double runs whereby both depositors withdraw their deposits and firms draw down on their credit lines.1 This result is intuitive but may not appear to be in line with existing results on the complementaritiesbetweendeposittakingandtheissuanceofcreditlines. Wecorroborateourempirical findings by extending the theoretical model in Kashyap, Rajan, and Stein (2002) to introduce runnable deposits akin to Diamond and Kashyap (2016). The intuition is simple. Liquidity risk management with runnable deposits requires considering off-equilibrium withdrawals, not just withdrawals expected in equilibrium. Thus, a bank needs to guarantee it has enough liquidity also in off-equilibrium paths with more expensive non-deposit funding. Doing so may not be profitable under high deposit fragility resulting in a reduction in the issuance of credit lines. Fourth, we show that the relative reduction in liquidity insurance offered by the more exposed banks has aggregate implications. Although firms’ access to current credit is not affected, those firms that have more lending relationships before the Pandemic QE with exposed banks experience a reduction in the amount of liquidity insurance they enjoy against future shocks. This reduction results in relatively lower investment by exposed firms.2 Related literature. Our main contributions to the literature are (i) to demonstrate that limits in the effectiveness of unconventional monetary policy can arise due to an increase in deposit fragility and the associated liquidity risk management of (more exposed) banks both on theirdepositsandcreditsupply,and(ii)toshowhowthedocumentedcomplementaritybetween deposit-taking and the provision of liquidity to firms may break down. Our paper relates to three main strands of the literature. First, we show that bank liquidity risk management limits the effects of unconventional monetary policy. In addition to the aforementioned seminal paper by Kashyap et al. (2002), Hanson, Shleifer, Stein, and Vishny (2015) examine how funding fragility interacts with the holdings of liquid assets in financial institutions, focusing on the distinction between banks with insured deposits and non-banks with runnable liabilities.3 Instead, we study the effect of 1SeeIppolito,Peydr´o,Polo,andSette(2016). 2SeeHolmstr¨omandTirole(1998)forthelinkbetweenliquidityinsuranceviacreditlinesandfirm’sinvestment. 3Empirical support for such complementarities comes from studies showing that during episodes of market stress, deposit inflows and credit line drawdowns are negatively correlated (Gatev and Strahan, 2006; Gatev, 4

bank deposit fragility on bank liquidity risk management and credit supply to firms.4 Ippolito et al. (2016) study banks’ liquidity risk management in the presence of double runs due to a joint withdrawal of interbank funding and credit-line draw-downs during the 2007 freeze in the European interbank market. They find that banks with higher interbank borrowing before the shock also extended fewer credit lines to firms. Acharya and Mora (2015) present similar dynamics when banks get run upon: banks with higher liquidity risk in the onset of the GFC experiencedlowerdepositgrowthandcutbankonnewcreditoriginations. Wedifferbystudying how banks actively manage their assets and liabilities in response to a quasi-exogenous shock in their funding fragility. Moreover, our paper studies the interaction of quantitative monetary polices and bank fragility. Cooperman, Duffie, Luck, Wang, and Yang (2023) study how banks adjust their provision of credit lines when the effective cost of funding them goes up: banks are lesswillingtoprovidecreditlinesexantewhenthelendingrateuponwithdrawalisnotalsorisk sensitive, which is an increase in effective funding costs. Our mechanism is different because we focus on the impact of funding fragility rather than effective funding cost. Moreover, we also examine how banks actively adjust their balance sheets to manage liquidity risk.5 Second, werelatetotheliteratureontheeffectsofunconventionalmonetarypolicy. Acharya and Rajan (2024) and Acharya et al. (2023) link QE to persistent bank fragility via the creation ofuninsureddeposits,whichisasteppingstoneforouranalysis(seealsoJoyce,Miles,Scott,and Vayanos,2012). Weshowthattheeffectsofunconventionalmonetarypolicyarelimitedthrough deposit risk management and the supply of new credit. We use granular administrative data to show how uninsured deposits—particularly from NBFIs—are heterogeneously injected in the banking system and how banks actively manage their deposit liabilities and loan commitments inresponsetothisfragility;thisisotherwisehardtoteaseoutfrommoreaggregateddatadueto Schuermann,andStrahan,2009). 4Ourpaperalsoanalyzescreditsupplyandtheassociatedrealeffects,hencecontributingtothelargeliterature on the real effects of credit supply. For example, Chodorow-Reich (2013) studies how an adverse shock in bank capitalaffectscreditsupplyandsubsequentrealoutcomes. Wedifferintwoways. First,westudytheeffectsof the ex ante build-up in fundingfragilityrather thanan ex post shock to capital. Second, weholistically explore howbanksmanageliquidityriskonbothsidesoftheirbalancesheet. 5We also contribute to recent studies on the behavior of credit lines and deposits during the Pandemic. Li, Strahan,andZhang(2020)andAcharya,Engle,Jager,andSteffen(2024)showthatfirmsmassivelydrewdown on their lines of credit at the outbreak of the pandemic keeping the funds as deposits at banks, while Levine, Lin, Tai, and Xie (2021) suggest that the increase in deposits also accrued from a flight-to-safety motive. We complement this analysis by showing that the QE-induced fragility did not differentially affect the draw-downs of credit lines and total deposits across exposed banks, but rather affected the undrawn amounts and the mix betweenuninsuredandinsureddeposits. 5

a simultaneous increase in deposits and credit lines across banks. Importantly, our data allows us to distinguish between utilized and undrawn credit-lines at the bank-firm level, which is not possible with publicly available regulatory data. This distinction allows us to control for credit demand and isolate the credit supply effect on bank provided (contingent) liquidity insurance. Pre-PandemicstudiesofQEfocusontheassetswapchannel—exchangingreservesforsecuritiesontheassetsideofbanksbalancesheets—thatdonotinvolvecreatingfragilebankdeposits. For example, Rodnyansky and Darmouni (2017) shows that banks with higher ex ante holdings of the QE-purchased securities increase lending relatively more after QE. Di Maggio, Kermani, and Palmer (2019) shows how QE facilitated the refinancing of mortgage debt by households, which reduced interest expenses and supported aggregate consumption. We also relate to papers studying the unintended consequences of QE. Chakraborty, Goldstein,andMacKinlay(2020)demonstratehowbanksmayshifttheirportfoliostowardssecurities purchasedbycentralbanks, suchasmortgages, andawayfromC&Iloans. Thisparticularprofit seeking mechanism is mitigated in our analysis by two facts: First, prior to QE, there is no significant difference in mortgage and C&I lending among banks that are more or less exposed to NBFI uninsured deposits. Second, most pandemic-QE purchases were Treasuries rather than mortgages. Diamond, Jiang, and Ma (2024) show that large injection of central bank reserves has the unintended consequence of crowding out bank loans, due to bank balance sheet costs. Third,ourworkcontributestoagrowingstrandoftheliteraturethathighlightstheincreasing interdependence between banks and NBFIs. Relative to banks, NBFIs have grown significantly since the GFC but remain lightly regulated (Acharya, Cetorelli, and Tuckman, 2024; Irani, Iyer, Meisenzahl,andPeydro,2021). TheconnectionsbetweenbanksandNBFIscanoperatethrough both assets and liabilities. From a lending perspective, several studies show that NBFIs act as shock absorbers, by filling the space left by banks during periods of monetary policy tightening (Elliott, Meisenzahl, and Peydr´o, 2024; Chen, Ren, and Zha, 2018). Our paper contributes to thisstrandoftheliteraturebyanalyzingadifferentchannelofinteraction,focusingonthebanks’ funding dependency on NBFIs. 6

2 Data and Empirical Strategy Thissectiondescribesthedatasetsusedinouranalysis,providesbackgroundontheinstitutional context, and presents key descriptive statistics. In turn, we introduce our empirical strategy. 2.1 Datasets Our analysis relies mainly on two administratively matched datasets. To document the effects of QE on NBFI uninsured deposits and bank funding fragility, we use granular data on deposit accounts at the counterparty-bank level for all large U.S. BHCs. We supplement this data with information on bank balance sheets. To investigate how deposit inflows affect bank lending, we use U.S. administrative bank-firm matched data at the loan-level containing firm-level balance sheet information. In sum, our deposit dataset comprises monthly observations of individual deposit accounts reported by 29 banks, covering January 2016 to February 2023.6 Our credit dataset consists of quarterly observations of term loans and credit lines extended by the same 29 banks to 120,797 non-financial firms, spanning from 2016Q1 through 2022Q4.7 For brevity, throughout the paper, we refer to bank holding companies (BHCs) simply as banks. This subsection describes each dataset and outlines the main sample selection criteria. Deposit data. Our primary dataset for deposits is the Complex Institution Liquidity Monitoring Report, commonly referred to as the FR 2052a, which monitors the liquidity profiles of significant U.S.BHCs. The FR2052a datacollection beganin December2015, initially covering Global Systemically Important Banks (GSIBs) and foreign banking organizations (FBOs) with substantialU.S.broker-dealeroperations. InJuly2017,thedatasetexpandedtoincludealarger set of banks. This dataset offers two distinct advantages over publicly available regulatory filings such as the FR Y-9C. First, it provides granular breakdowns of banks’ assets and liabilities by maturity, collateral, and depositor type (counterparty), allowing us to document previously unexplored aspects of U.S. banks’ funding structures and depositor exposure. Second, it offers higher-frequency reporting: banks with $700 billion or more in total consolidated assets or $10 6OursampleperiodendsinFebruary2023toexcludethepotentialdistortionsfromtheMarch2023banking turmoilintheUnitedStates. 7ThecompletedetailsofthedatacleaningprocedurecanbefoundinAppendixC. 7

trillionormoreinassetsundercustodysubmitdailyreports,whereasbankswithassetsbetween $50 billion and $700 billion report monthly. To ensure consistency across banks with different reporting frequencies, we harmonize the data by aggregating daily observations into monthly averages, aligning them with the reporting frequency of the remaining banks. In Appendix C, Table OA2 provides a detailed list of banks along with their respective reporting schedules. Additionally, FR 2052a explicitly identifies insured versus uninsured deposits, facilitating a precise analysis of liquidity risk stemming from banks’ funding sources.8 WefurthersupplementourdepositdatasetwithdepositrateinformationfromRatewatch–S&P Global, which provides detailed interest rates offered by banks across various deposit categories. This complementary dataset enables us to directly examine how banks adjust deposit pricing strategies in response to changing liquidity conditions. Loan-Level Data Our analysis of bank lending utilizes detailed loan-level data from the Federal Reserve’s FR Y-14Q H.1, collected quarterly as part of the Comprehensive Capital Analysis and Review (CCAR). FR Y-14Q collects detailed information on bank holding companies’ (BHCs), savings and loan holding companies’ (SLHCs), and U.S. intermediate holding companies’ (IHCs) of foreign bank organizations (FBOs) on a quarterly basis.9 We use the Corporate Loan H.1. Schedule comprising two sections: (1) the Loan and Obligor Description section, providing detailed characteristics of each loan and borrower; and (2) the Obligor Financial Data section, which includes borrowers’ balance sheets and income statements. Facility-level data include, among much more, total committed and utilized amounts, pricing and spread details, origination and maturity dates, and collateral information. 2.2 Institutional Context The Pandemic QE, which commenced in March 2020 and ended in March 2022, was the largest expansion in the Federal Reserve’s history. Moreover, it led to significant changes in the balance sheetsizeandcompositionofboththeFedandthebankingsystem. Ouranalysisstartswiththe observationthatnotallfinancialinstitutionscanholdreserves,whichhasimportantimplications 8AppendixBexplainstheselectionrulesweimposetoavoidbiasesinoursample. 9Data are collected for BHCs, SLHCs and IHCs with at least $50 billion ($100 billion starting from 2019) in totalassets. BanksthatsubmitFRY-14Qcompriseover85percentofthetotalassetsintheU.S.bankingsector. 8

for the conduct of quantitative policies. Suppose first that the central bank purchases securities directly from banks that can hold reserves. Then, QE is purely an asset swap (reserves for securities). Nowsupposethatthecentralbank’scounterpartyisanon-bankfinancialinstitution (NBFI) that cannot hold reserves outright. In this case, the trade between the central bank and the NBFI is intermediated by banks. Banks source the securities from NBFIs to sell to the central bank, use the proceeds to credit NBFIs’ deposit accounts, and receive reserves from the central bank. In practice, NBFIs exchange securities for bank deposits. Given the scale of QE, the resulting NBFI deposits are uninsured and hence more flighty.10 Figure1Ashowsbanks’andNBFIs’holdingsofTreasuryandAgencysecurities. Asubstantial portionofTreasuriesandAgencieswereheldbyNBFIsattheonsetoftheCOVID-19pandemic.11 Inresponsetothepandemic,theFederalReserveexpandeditsbalancesheetbyabout$3trillion from mid-March to early June 2020, while asset purchases stabilized at $120 billion per month until the end of QE in March 2022. A large component of the securities purchased by the Federal Reserve were offloaded by NBFIs. As the figure shows, NBFIs’ holdings declined during the Pandemic QE, while banks’ holdings continued to increase. QT commenced in June 2022 with a balance sheet reduction of $ 47.5 billion per month. Figure 1B uses administrative data for uninsured deposits held by NBFIs, which spike immediately after QE commenced, continued increasing, and finally stabilized at a higher level into the QT period. 2.3 Descriptive statistics Table 1 presents aggregate-level descriptive statistics across deposit categories and bank lending activities. Panel A reports statistics for banks’ primary deposit categories across distinct monetary policy periods. Both uninsured NBFI and insured retail deposits expanded significantly during QE and continued to rise through QT, with the most pronounced changes occurring in QE. Specifically, uninsured NBFI deposits rose by approximately 40% from $699.7 billion pre- 10See Joyce et al. (2012) and Leonard, Martin, and Potter (2017) for details about the accounting operations of QE in the presence of NBFIs. In the United States, commercial banks, government-sponsored enterprises, clearing houses, credit unions, and branches of foreign banking organizations are the main financial institutions with reserve accounts at the Federal Reserve. Certain other institutions may have access to Federal Reserve liabilities, other than reserves, such as the ON RRP facility. It is conceivable that those NBFIs withdraw the newly issued bank deposits to deposit directly at the central bank, but the scope of this operation is limited to eligiblenon-banks(seeAfonso,Cipriani,andLaSpada,2022). 11Althoughthemechanismhighlightedaboveisalwaysoperational,wefocusonthepandemicQEforwhichwe havedetailedadministrativedataonbankdeposits. 9

Figure 1: Security holdings and NBFI deposits (A) Banks and NBFIs holdings of Treasury and (B)NBFIsuninsureddepositsinbillionUSD AgencysecuritiesinbillionUSD Note: Panel (A) reports quarterly data from the Financial Accounts-Z.1. NBFIs include Insurance Companies, Pension Funds, Open- and Closed-ended Funds, REITs, ETFs, Money Market Funds, Broker Dealers, Hedge Funds,andotherFinancials. Panel(B)reportsmonthlyadministrativeFR2052adataforthebiggestbanks. QE to $978.7 billion during QE. Similarly, insured retail deposits increased by roughly 27%, expanding from $3.28 trillion pre-QE to $4.16 trillion during QE. Notably, uninsured deposits consistently account for between 95% and 98% of total NBFI deposits, underscoring the inherently risky nature of banks’ exposure to these institutions. Collectively, these patterns highlight meaningful shifts in banks’ deposits composition driven by monetary policy adjustments. Panel B documents substantial heterogeneity in banks’ deposit composition based on their exposuretoNBFIdepositsovertotaldepositfundingasofFebruary2020,priortothepandemic QE. These NBFI deposit shares serve as key cross-sectional measures of bank exposure to NBFI deposits, which we discuss below in our empirical methodology section 2.4. Banks with high NBFIsharesexhibitsignificantlygreaterrelianceonuninsuredNBFIdeposits(22.84%)compared to banks with low NBFI exposure (2.12%). Likewise, the overall uninsured deposit ratio is considerably higher among banks with high NBFI shares (78.69%) compared to banks with lowerNBFIexposure(40.63%). Incontrast,bankswithhighNBFIsharesholdsignificantlyfewer insured retail deposits (17.66%) compared to banks with lower NBFI exposure (53.84%). These differences highlight substantial variation in banks’ deposit structures linked to NBFI exposure, underscoring their distinct risk profiles and likely divergent liquidity management strategies. Figure 2 further illustrates these differences by showing the evolution of uninsured NBFI deposits for banks with high and low exposure. Before QE, both groups exhibited relatively 10

Table 1: Descriptive statistics: Aggregate volumes Panel A: Deposit categories Feb-20 Mar-20 PreQE QE QT Mean Mean Mean Std. Dev Mean Std. Dev Mean Std. Dev UninsuredNBFI 746.6 953.6 699.7 42.2 978.7 72.6 1051.1 93.2 InsuredNBFI 19.3 19.4 23.3 2.3 17.6 4.5 49.3 21.2 UninsuredRetail 1,383.9 1,449.6 1,240.7 56.9 1,750.5 217.5 2,007.1 100.5 InsuredRetail 3,573.0 3,738.5 3,281.2 119.5 4,162.5 214.8 4,575.1 159.3 TotalDeposits 9,287.5 9,987.6 8,466.6 353.3 11,362.6 771.7 12,480.8 495.5 Panel B: Bank exposure to insured and uninsured Deposits UninsuredNBFIRatio InsuredRetailRatio TotalUninsuredRatio Mean Std. Dev Mean Std. Dev Mean Std. Dev BankswithlowNBFIShare 2.12% 1.79% 53.84% 16.39% 40.63% 15.65% BankswithhighNBFIShare 22.84% 20.14% 17.66% 15.40% 78.69% 17.59% Panel C: Statistics on loan-level data Period TotalCommitments On-balancesheet UndrawnCredit UtilizedCredit TermLoans Commitments Lines Lines 2019q4 1,729 702 1,027 437 267 2020q1 1,762 866 896 580 286 Pre-QE 1,368 550 818 348 210 QE 1,764 658 1,095 397 255 QT 1,965 753 1,211 447 300 Note: Panel A reports the distribution of deposits by counterparty type across key monetary policy periods ($billion),sourcedfromtheFR2052a(Complex Institution Liquidity Monitoring Report Supervisory). Panel B presents deposit ratios for banks with high and low NBFI deposit shares as of February 2020. Panel C summarizesloan-leveldata($billion)fromtheFRY-14Q(CapitalAssessmentsandStressTesting). Variable definitionsanddatasourcesareprovidedinAppendixC. stable trends in uninsured NBFI deposits. However, following QE, banks with high exposure experienced a sharp and persistent increase in uninsured NBFI deposits, whereas those with lowerexposuresawonlymodestchanges. ThisdivergencepersiststhroughQT,underscoringthe structural differences in banks’ reliance on NBFI funding. These patterns suggest that pre-QE exposure may have played a key role in shaping banks’ funding responses to monetary policy interventions, something that we will explore in our empirical methodology. PanelCinTable1providessummarystatisticsonbanks’lendingactivities. On-balancesheet commitments (defined as the sum of utilized credit lines and term loans) rose significantly in 2020Q1,ariseof$164billionrelativeto2019Q4. Utilizedcreditlinesincreasedfrom$437billion to $580 trillion between 2019Q4 and 2020Q1. This development reflects the heavy utilisation 11

Figure 2: Total uninsured deposits from NBFIs Note: ThisfigurepresentsthetotaluninsureddepositsfromNBFIs,expressedin$billion. of credit lines by firms in the last three weeks of March 2020. We also note that total loan commitmentsincreasedfrom$1.37trillionpre-QEto$1.76trillionduringQE,primarilydrivenby theriseinundrawncreditlinesfrom$818billionto$1.10trillion. Thisgrowthinundrawncredit linessuggestsincreasedprovisionofcontingentliquiditytofirms,thuspotentiallyelevatingbanks’ exposuretoliquidityriskassociatedwithfuturecredit-linedraw-downs. Incontrast,theincrease inutilizedcreditlinesandtermloanswascomparativelymodest. Theselendingpatternsindicate thatbanksadjustedtheircreditprovisionprimarilythroughcontingentliquidity,highlightingthe importanceofliquiditymanagementconsiderationsarisingfrommonetarypolicy-drivenshiftsin banks’depositstructures. Atthesametime,wewillshowthattheseareaggregatedevelopments and the overall change in undrawn credit lines are heterogeneously distributed across banks. 2.4 Empirical methodology As discussed above, QE operations can lead to a surge in NBFI deposits at banks. These deposits are highly sensitive to market conditions and prone to rapid withdrawals.12 Although 12See,forexample,Franceschi,Grodzicki,Kagerer,Kaufmann,Lenoci,Mingarelli,Pancaro,andSenner(2023). 12

banks receive reserves at the same time when crediting NBFI deposit accounts, their overall funding fragility can still increase. To illustrate this, consider a bank that aims to maintain a level of high-quality liquid assets that create a buffer over the estimated deposit outflows over a given period. When uninsured NBFI deposits increase, the expected outflow rate rises—not only because total deposits grow but also because these deposits are more volatile. Hence, the additional liquid assets required to preserve the same liquidity buffer exceed the proportional increase in deposits. In other words, despite the mechanical rise in reserves, a bank’s funding stability may deteriorate. Using the deposit dataset, we construct a measure for the ex ante exposure of each bank to such QE-induced fragility. In particularly, we calculate the shares of NBFI uninsured deposits relative to total deposits as of February 2020, prior to the onset of pandemic QE. This measure proxies for the degree that a bank interacts with NBFIs prior to the pandemic. Intuitively, a higher share of NBFI funding suggests that a bank is having more relationships and doing more business with NBFIs, providing a crucial gauge of the bank’s exposure to the creation of NBFI deposits from the QE operations. Prior to QE, the cross-sectional variation in this measure remained stable over time, suggesting that differences in banks’ exposure to uninsured NBFI deposits were persistent rather than driven by transitory factors (see Figure 2). The stability in the aforementioned pattern indicatesthat, ex-ante, banks’exposuretouninsuredNBFIdepositswasnotexpectedtobesystematically affected by the COVID-19 shock, reinforcing the exogeneity of the pandemic to this funding source. Moreover, the banks in our sample were well capitalized and in strong financial condition at the outbreak of pandemic, mitigating concerns that NBFIs would shift their relationships to other banks. As a result, the pre-QE uninsured NBFI share serves as a meaningful and persistent indicator of banks’ exposure to the QE-induced fragility, rather than reflecting a short-term adjustment to pandemic-related disruptions. Ourempiricalstrategyexploitsthecross-sectionalvariationinbanks’pre-QEsharesofNBFI funding and implements a continuous treatment approach to analyze how banks with different NBFIsharesrespondedtotheQE-inducedfragility. Specifically,weexaminetwokeydimensions: (i)banks’adjustmentsindepositcomposition,particularlyshiftsbetweeninsuredanduninsured deposits, and (ii) changes in credit allocation, focusing on loan commitments. 13

Table 2: Identifying exposed banks: Balancing test Low NBFI Exposure High NBFI Exposure Difference Mean SD Mean SD Std. Diff. NBFI Deposits 4.9 8.3 36.8 45.1 -0.98 Uninsured Deposits 101 169.1 183.8 250.8 -0.39 Total Deposits 242 341.5 279.4 400.4 -0.10 Total Assets 522.9 657.3 540.9 757.3 -0.03 Tier 1 Capital Ratio 0.14 0.04 0.15 0.05 -0.03 C&I Loans 43.5 58.1 43.4 49.4 0.003 Treasury + Agency Securities 54.1 103.4 56.7 79.7 -0.03 Note: Thistablereportsstandardizeddifferencesinbankcharacteristicsbetweenbankswith lowandhighNBFIexposurein2019Q4. LowNBFI-exposedbankshavebelow-medianuninsuredNBFIsharesintheirtotaldeposits,whilehighNBFI-exposedbankshaveabove-median shares. FollowingImbensandWooldridge(2009),astandardizeddifferenceabove0.25inabsolutevalueindicatesasubstantialimbalancebetweenthetwogroups. Allvaluesarein$billion, except for the capital ratio. Appendix C provides variable definitions and data sources. Although our identification strategy does not strictly require banks with low and high exposuretouninsuredNBFIdepositstobeidentical,ensuringcomparabilitystrengthenstheinternal validity of our estimates and mitigates concerns about potential omitted variable bias (Roberts and Whited, 2013). Following Imbens and Wooldridge (2009), we assess the balance across key bank characteristics in 2019Q4 using standardized differences, where absolute values below 0.25 indicate sufficient comparability. Table 2 reports mean and standard deviation values for key balance sheet characteristics across banks with high and low exposure. We compare banks along several dimensions, including (1) NBFI Deposits, (2) Uninsured Deposits, (3) Total Deposits, (4) Total Assets, (5) Tier 1 Capital Ratio, (6) Total C&I Loans, and (7) Treasury and Agency Securities. All standardized differences remain far below the 0.25 threshold, except for NBFI Deposits and Uninsured Deposits, which naturally differ between the two groups by construction. The similarity across other balance sheet fundamentals suggests that differences in outcomes are unlikely to be driven by pre-existing structural differences between the two groups. 14

3 NBFI deposits and QE-induced fragility This section establishes how uninsured NBFI deposits evolved based on banks’ ex-ante heterogeneous exposure to the QE-induced fragility. To do so, we estimate a panel regression over the period from January 2016 to February 2023: log(Un.NBFI )=λ·(QE ·Shares )+β·Controls +a +a +ε . (1) i,t t i i,t i t i,t Thedependentvariable,log(Un.NBFI ),representsthelogarithmofuninsuredNBFIdeposits i,t held by bank i in month t. The variables QE is dummy variable equal to one during the QE t period (March 2020–March 2022). Shares measures the share of uninsured NBFI deposits i relative to total deposits in February 2020, providing a measure of banks’ pre-pandemic reliance on NBFI funding. We include bank fixed effects (a ) to account for time-invariant heterogeneity i across banks and month (time) fixed effects (a ) to control for common macroeconomic shocks. t Controls isavectorofcontrolstoaccountforothertime-varyingconfoundingfactors. Finally, i,t we cluster standard errors at the month level to address potential autocorrelation in residuals. Table 3 reports the results from estimating equation (1). In Column 1, we begin with a parsimonious specification that includes the QE interaction term (QE·Shares), as well as bank and month fixed effects. The coefficient on the interaction term is positive and statistically significant, suggesting that banks with a higher share of uninsured NBFI deposits prior to the pandemic saw these deposits increase substantially during QE. Economically, the estimate in Column 1 indicates that a one standard deviation increase in exposure to NBFI deposits is associated with a 3.2% increase in uninsured NBFI deposits during QE. InColumns2,3and4,wesequentiallyaddbanksize,theQTinteractionterm(QT·Shares), and an interaction term that account for differences in deposits between GSIBs and other banks during QE. These extensions allow us to account for characteristics beyond those captured by bank fixed effects. Bank size, measured as the logarithm of total assets, accounts for timevarying banks’ ability to absorb inflows from NBFIs. The QT interaction term allows us to distinguishtheeffectsofQTfromthoseofQE.Moreover,QEinteractedwiththeGSIBindicator captures the distinct role of systemically important banks in deposit flows. The coefficient on 15

Table 3: NBFI uninsured deposits 1 2 3 4 5 6 Dependent variable: Log(Uninsured NBFI deposits) QE * Shares 0.286*** 0.288*** 0.272*** 0.268*** 0.263*** (5.258) (5.436) (4.655) (4.813) (4.469) QT * Shares -0.099 (-1.196) Bank size 0.353*** 0.347*** 0.342*** 0.354*** 0.738*** (3.503) (3.365) (3.446) (3.507) (7.509) QE * GSIBS 0.047** (2.340) QE (SLR rel.)* Shares 0.209*** (3.137) QE (SLR act.)* Shares 0.362*** (7.279) NBFI credit 0.046*** (3.008) Month FE Y Y Y Y Y Y Bank FE Y Y Y Y Y Y Observations 2,079 2,077 2,077 2,077 2,077 2,066 R-squared 0.968 0.968 0.969 0.968 0.968 0.970 Note: The table reports coefficients and t-statistics (in parentheses) for the following regression equation: log(Un. NBFI ) = λ·(QE ·Shares )+β ·Controls +a +a +ε , where i,t t i i,t i t i,t log(Un. NBFI )isthelogarithmofuninsuredNBFIdepositsheldbybankiinmontht. QE is i,t t adummyequaltoonefromMarch2020toMarch2022,andQT isadummyequaltoonefrom t June 2022 onwards. Shares indicates the share of uninsured NBFI deposits in total deposits i forbankiasofFebruary2020. Bank Size isthelogarithmoftotalassets. GSIBS isadummy i,t equaltooneforGlobalSystemicallyImportantBanks. QE (SLR rel.) referstoSLRrelaxation and the exclusion of securities and reserves from SLR calculations between April 1, 2020, and April 1, 2021, while QE (SLR act.) marks the re-activation of SLR criteria. NBFI Credit is the logarithm of total outstanding credit, including credit lines and term loans, that NBFIs received. The terms a anda represent bankandmonthfixed effects, respectively. Observations i t aremonthly,exceptfortotalassets,whicharereportedquarterly. Standarderrorsareclustered atthemonthlevel. VariabledefinitionsanddatasourcesareprovidedinAppendixC.Thesymbols ***, **, and * denote statistical significance at the 1%, 5%, and 10% levels, respectively. 16

QE · Shares remains positive and statistically significant under all controls, confirming that banks with a higher pre-pandemic share of uninsured NBFI deposits experienced a substantial increase in these deposits during QE. Interestingly, the coefficient of QT ·Shares is statistically insignificant, suggesting that QT did not have a persistent effect on uninsured NBFI deposits afteraccountingforbank-specificattributesandtimeeffects. ThesefindingsunderscorethatQEinduced inflows of uninsured NBFI deposits were a dominant driver of bank funding dynamics, while any potential reversal during QT appears to be more muted.13 In Column 5 of Table 3, we explore the impact of the Supplementary Leverage Ratio (SLR) adjustments during the pandemic-related QE.14 Shortly after the initiation of QE, the Federal ReserveBoard(FRB)announcedthatthecalculationoftheSLRwouldbetemporarilymodified. ThemodificationexcludedU.S.Treasurysecuritiesandbankreservesfromthecalculationofthe SLR for bank holding companies. This adjustment aimed to alleviate balance sheet constraints and encourage liquidity provision.15 The FRB specified that the change would be in effect until March31,2021(FederalReserveBoard,2021). TheresultsinColumn5subdividetheQEperiod into the SLR relaxation and re-activation phases and interact them with the NBFI shares. The QE sub-periods are defined using dummy variables corresponding to each SLR phase. Results highlightthattheriseinuninsuredNBFIdepositsofexposedbanksduringQEspansbothphases of the SLR change and is not driven solely by the relaxation phase. In Column 6, we account for the mechanical link between bank lending and deposits by controllingforasset-sideexposurestoNBFIs. Specifically,weincludethelogoftotaloutstanding credit that NBFIs received, recognizing that loans can mechanically influence deposit balances. The results show that our main findings remain unchanged. To gain further insight into the role ofdifferenttypesofNBFIdeposits,inAppendixC,TableOA4focusesexclusivelyondemandable uninsured NBFI deposits, which can be withdrawn without prior notice and even more prone to 13In Appendix C, Table OA3 further explores the heterogeneity in uninsured NBFI deposit dynamics by differentiating between supervised and non-supervised NBFIs. The results show that the effect of QE·Shares on uninsuredNBFIdepositsisprimarilydrivenbysupervisedNBFIs. 14TheSLRwasestablishedin2014aspartoftheBaselIIIregulatoryframework. TheSLRappliesonlytolarge, complexfinancialinstitutionswith$250billionormoreintotalconsolidatedassetsor$10billioninon-balancesheet foreign exposures. Banks must report their SLR since 2015 and must comply with the SLR requirement since January 1, 2018 (binding period). Bank holding companies generally must maintain an SLR of at least 3 percent, and GSIB holding companies in the U.S. must maintain an enhanced SLR (eSLR) of 5 percent. The SLRiscalculatedastheratioofTier1capital(essentiallycommonequitypluspreferredstock)tototalleverage exposure(assetspluscertainoff-balance-sheetitems,suchasOTCderivatives). 15SeeDuffie(2020)andFavara,Infante,andRezende(2024). 17

runs. The results indicate that the observed effects are particularly pronounced for NBFIs with demandabledeposits,reinforcingtheviewthatinflowsofuninsuredNBFIdepositsareassociated with heightened liquidity risk. Finally, a potential alternative driver for the observed increase in uninsured NBFI deposits could be the fiscal transfers to households and firms during the pandemic. Several pieces of evidence suggest that these transfers do not explain our findings. Our empirical design mitigates this concern by including month fixed effects, which absorb aggregate liquidity injections, including fiscal transfers, ensuring that our estimates isolate the effect of banks’ pre-QE exposure to uninsured NBFI deposits from broader system-wide liquidity dynamics. Additionally, in Appendix C, Table OA5, we estimate equation (1) using as the dependent variable the total retail insured deposits (Columns 1–2) and corporate insured deposits (Columns 3–4). Across all specifications, the interaction term QE·Shares is statistically insignificant, indicating that bankswithhigherpre-QEexposuretouninsuredNBFIdepositsdidnotexperienceadifferential increaseininsureddepositsduringQErelativetolessexposedbanks. Thisfindingsuggeststhat fiscal transfers, which primarily flowed into insured retail and corporate deposits, do not explain theobservedpatternsinuninsuredNBFIdeposits. Moreover, fiscaltransfersbeganinmid-April 2020, after $1.3 trillion in QE operations had already been conducted, accounting for approximately40%ofthetotalcumulativeincreaseinreservesduringthePandemic-QEperiod. Finally, in unreported results, we find that the effect of QE is strongest in March 2020, reinforcing the view that monetary policy, not fiscal transfers, drove deposit inflows at exposed banks. Parallel trends. A key assumption underlying our identification strategy is that, in the absence of QE, banks with different levels of exposure to uninsured NBFI deposits would have followed similar trends in deposit accumulation. While this parallel trends assumption cannot be directly tested, we assess its plausibility by examining pre-QE deposit trends. Figure 3A presents the daily growth rates of uninsured NBFI deposits for high- and low-exposure banks fromJanuarytoMarch2020. Thex-axisisnormalizedto100onMarch9forcomparability. The figure shows that in the months leading up to QE, both groups followed largely similar trends, with no systematic differences in growth rates. However, starting in March 2020, a sharp divergence emerges, with banks that had higher pre-QE NBFI exposure experiencing a significantly 18

Figure 3: Security holdings and NBFI deposits (A)DailygrowthinuninsuredNBFIdeposits (B)Paralleltrends Note: Panel (A) shows the daily growth of NBFI deposits from January to March 2020. Panel (B) presents the estimated monthly coefficients for the pre- and post-QE period: Log(Un. NBFIi,t) = (cid:80)T t=1 λt(Montht· Sharesi)+β·BankSizei,t+ai+at+εi,t. Yi,t denotesuninsuredNBFIdepositsatbankiinmontht;Montht are month dummies; Sharesi is the share of uninsured NBFI deposits in total deposits as of February 2020; Bank Sizei,t isthelogarithmoftotalassets; ai,at arebankandmonthfixedeffects. Observationsaremonthly, exceptfortotalassets,whicharequarterly. Standarderrorsclusteredatthemonthlevel. larger increase in uninsured NBFI deposit inflows. This pattern supports our empirical design, suggesting that the differential post-QE response is not driven by pre-existing differences but rather by monetary policy-induced liquidity shocks. To formally assess this assumption, we estimate the following dynamic specification: T (cid:88) log(Un.NBFI )= λ (Month ·Shares )+β·Bank Size +α +α +ε i,t t t i i,t i t i,t t=1 where log(Un. NBFI ) represents uninsured NBFI deposits for bank i at time t, Month i,t t denotes month dummies, Shares captures a bank’s exposure to uninsured NBFI deposits as i of February 2020, Bank Size controls for bank size, and α and α are bank and time fixed i,t i t effects. Standard errors are clustered at the month level. Figure 3B presents the estimated λ t coefficients for the months preceding and following QE. The results indicate that pre-QE trends in uninsured NBFI deposits were statistically indistinguishable between high- and low-exposure 19

banks, with estimated coefficients ranging around zero. However, following QE (March 2020), we observe a sharp divergence, as banks with greater exposure experience a disproportionate increase in uninsured NBFI deposits relative to less exposed banks. This provides empirical support for our identification assumption, validating the credibility of our research design. ThenextsectionsassesshowbanksrespondedtotheQE-inducedfragility,firstbyexamining adjustments in deposit composition and subsequently by analyzing changes in credit allocation. 4 Liquidity risk management: Deposit liabilities We begin by analyzing banks’ responses on the liability side, focusing on how exposed banks adjusted their total deposits and the composition of deposit categories. In Column 1 of Table 4, weexaminetotaldeposits. ThecoefficientonQE·Sharesisstatisticallyinsignificant,indicating that, on average, banks with higher pre-pandemic exposure to NBFI uninsured deposits did not experience a significant change in their total deposits during QE. This suggests that any fundingadjustmentsoccurredprimarilythroughshiftsindepositcompositionratherthanoverall deposit growth. To explore these shifts, we turn to Columns 2 through 5. Column 2 reports the increase in NBFI uninsured deposits for more exposed banks and is included here to maintain comparabilityandthecompletesetofdepositcategories. Banksmanagetheirliabilitystructures by reducing their exposure to total uninsured deposits, as evidenced by the negative coefficient during the QE period (Column 3). This adjustment is even more pronounced when excluding NBFIdeposits(Column4),suggestingastrategiccontractioninuninsuredliabilitieswhereNBFI exposures are not a factor. At the same time, banks with higher NBFI exposure increased their insured deposit holdings, as shown by the positive and highly significant coefficient in Column 5. This shift reflects a deliberate effort to strengthen liquidity buffers and reduce reliance on volatile funding sources during QE. This dual strategy highlights how banks not only respond to immediate financial stresses but also proactively adjust their balance sheets in anticipation of potential liquidity needs.16 The banks in our analysis are subject to Liquidity Coverage Ratio (LCR) regulation, which 16In unreported results (available upon request), we re-estimate Table 4 while controlling for interactions betweenQEwiththeGSIBindicatortoaccountforthedistinctroleofsystemicallyimportantbanks. Theresults remainunchangedwhenwecontrolfortheroleofGSIBs. 20

requires them to hold high-quality liquid assets sufficient to cover estimated net cash outflows over a 30-day stress period. Most banks in our sample maintain an LCR above one, meaning theyholdliquiditybuffersexceedingtheregulatoryminimum(seeTableOA2intheAppendix).17 Importantly, the LCR assigns a 100% run-off factor to NBFI deposits, recognizing their flighty nature and the elevated risk of withdrawal under stress. The influx of uninsured NBFI deposits therefore increases expected cash outflows, tightening banks’ LCRs and bringing them closer to theregulatorythreshold. Therefore,weneedtoaddressthepotentialconcernthatmoreexposed banks to QE-induced fragility may also have had smaller liquidity buffers relative to expected funding outflows, as measured by their LCR. Hence, the relative shift from uninsured to insured depositsformoreexposedbankscouldstemfrombroaderprecautionarymotivestopreservetheir liquiditybuffersratherthanbeingadirectresponsetotheinfluxofuninsuredNBFIdeposits. To address this concern, we re-estimate Table 4 while controlling for interactions between QE with the bank’s LCR in 2019Q4, i.e., before the pandemic. Table OA6 in the Appendix shows that our key results remain robust. More exposed banks continue to actively manage the liquidity risk of their deposit liabilities in response to QE-induced funding fragility.18 Deposit rates. The findings suggest that banks with higher pre-pandemic exposure to NBFI deposits not only experienced a surge in uninsured NBFI deposits during QE, but also actively reshaped their deposit mix in response. This underscores the role of liquidity risk management in mitigating funding instability. A key mechanism through which banks manage liquidity risk is deposit pricing. By adjusting deposit rates, banks influence the volume and composition of their funding sources, either attracting or disincentivizing certain types of deposits. In Table 5, we explore this mechanism by examining how exposed banks adjust deposit rates between insured (Columns 1-3) and uninsured (Columns 4-6) deposits. The coefficient on QE ·Shares is positive and statistically significant for insured deposits, indicating that banks with greater NBFI exposure raise interest rates oninsured depositsduring QE.Economically, a onestandard deviation increase in NBFI exposure is associated with an increase in insured deposit rates by approximately 5.8-7.6 basis points. This suggests that these banks actively sought to attract 17Somebanks,however,aresubjecttoreducedLCRrequirements. 18SeealsoKiernan,Yankov,andZikes(2021)whoshowthatthelargeliquiditybuffersthatthelargestbanks accumulated after the Global Financial Crisis would enable them to provide liquidity to firms even in the most extremedraw-downscenarioswithoutviolatingtheirLCR. 21

Table 4: Other deposit categories 1 2 3 4 5 Dependentvariable: Log(Totaldeposits) Log(Uninsured Log(Totaluninsured Log(Totaluninsured Log(Totalinsured NBFIdeposits) deposits) depositsexc. NBFI) deposits) QE*Shares -0.049 0.272*** -0.253*** -0.398*** 1.711*** (-1.374) (4.655) (-6.469) (-8.444) (14.754) Bankcontrol Y Y Y Y Y QTcontrol Y Y Y Y Y MonthFE Y Y Y Y Y BankFE Y Y Y Y Y Observations 2,145 2,077 2,145 2,145 2,000 R-squared 0.988 0.969 0.981 0.980 0.957 Note: Thetablereportscoefficientsandt-statistics(inparentheses)forthefollowingregressionequation: log(Yi,t)=λ·(QEt· Sharesi)+β·Controlsi,t+ai+at+εi,t,whereYi,tisthedependentvariablelabelledineachcolumnforbankiinmontht. QEtis adummyequaltoonefromMarch2020toMarch2022. SharesiindicatestheshareofuninsuredNBFIdepositsintotaldeposits forbankiasofFebruary2020. TheBankcontrol indicateswhetherwecontrolfortime-varyingbanksize(logarithmoftotalassets),andtheQTcontrolindicateswhethertheinteractiontermQTt·Sharesiisincluded. Thetermsaiandatrepresentbankand monthfixedeffects,respectively. Observationsaremonthly,exceptfortotalassets,whicharereportedquarterly. Standarderrors areclusteredatthemonthlevel. VariabledefinitionsanddatasourcesareprovidedinAppendixC.Thesymbols***,**,and* denotestatisticalsignificanceatthe1%,5%,and10%levels,respectively. more stable funding sources in response to the influx of NBFI uninsured deposits. Incontrast,thecoefficientonQE·Sharesisnegativeandhighlysignificantforuninsureddeposits,indicatingthatthemoreexposedbanksreducedratesonthesedepositsduringQE.Aone standard deviation increase in NBFI exposure is associated with a 5.5-6.6 basis point decline in uninsured deposit rates, further reinforcing the shift toward more stable funding sources. Taken together,thesefindingshighlightthatbanksproactivelymanageliquidityriskbyreshapingtheir liability structures in response to the influx of fragile NBFI deposits. This behavior is consistent with the fact that exposed banks typically have less (more) insured (uninsured) deposits, as previously discussed, hence the incentive to protect liquidity is even stronger. 5 Liquidity risk management: Lending effects Intheprevioussection,weestablishedhowmoreexposedbanksadjusttheirdepositcomposition byshiftingtowardsmoreinsureddepositswhilekeepingthetotaldepositbasethesame, relative to less exposed banks. We argued that this response signals a desire to reduce fragility on the liabilities resulting from the QE-induced increase in NBFI uninsured deposits. In this section, we examine how the QE-induced fragility translates into adjustments on the asset side of banks’ 22

Table 5: Deposit rates 1 2 3 4 5 6 Dependent variable: Rates on Insured Deposits Rates on Uninsured Deposits QE * Shares 0.529*** 0.660*** 0.690*** -0.496*** -0.592*** -0.597*** (6.049) (9.864) (10.257) (-6.074) (-5.715) (-5.733) Bank control Y Y QT control Y Y Y Y Month FE Y Y Y Y Y Y Bank FE Y Y Y Y Y Y Observations 1,224 1,224 1,224 659 659 659 R-squared 0.722 0.724 0.724 0.516 0.519 0.518 Note: The table reports coefficients and t-statistics (in parentheses) for the following regression equation: Deposit rate = λ · (QE · Shares ) + β · Controls + a + a + ε , where i,t t i i,t i t i,t Deposit rate is the deposit rate for bank i in month t. The table analyzes two types of rates: i,t rates on insured deposits (columns 1 to 3) and rates on uninsured deposits (columns 4 to 6). QE is a dummy equal to one from March 2020 to March 2022. Shares indicates the share of t i uninsured NBFI deposits in total deposits for bank i as of February 2020. The Bank control indicates whether we control for time-varying bank size (logarithm of total assets), and the QT control indicates whether the interaction term QT ·Shares is included. The terms a and t i i a represent bank and month fixed effects, respectively. Observations are monthly, except for t total assets, which are reported quarterly. Standard errors are clustered at the month level. Variable definitions and data sources are provided in Appendix C. The symbols ***, **, and * denote statistical significance at the 1%, 5%, and 10% levels, respectively. balance sheets, focusing on lending outcomes. Our results capture the relative effects on lending by more exposed banks compared to their less exposed counterparts.19 In particular, we analyze the effect on loan commitments offered to firms by the same set of 29 banks studied in the previous section. We aggregate loans at the bank-firm-quarter level. The richness of our data enables us to account for time-varying firm characteristics, including a firm’s demand for credit, and unobservable relationships between banks and firms. Section 19Animportantquestioniswhetherbankswithahigherpre-QErelianceonuninsuredNBFIdepositsalready exhibited distinct risk profiles in the pre-QE period. To investigate this, we estimate cross-sectional regressions examining differences in credit risk between high- and low-exposure banks before QE. Specifically, we regress net charge-offs (NCOs) across different loan categories on an indicator for banks with above-median uninsured NBFIdeposits,controllingfortimefixedeffects. Theresults,reportedinAppendixC,TableOA7,suggestthat high-exposurebankshadsignificantlylowercharge-offsoncommercialandindustrial(C&I)loansandconsumer loansbeforeQE.Thisfindingsupportstheviewthatthesebankswerealreadymanagingriskconservativelyand sought to maintain a strong liquidity buffer post-QE rather than responding to a deterioration in asset quality. Consistentwiththis,high-exposurebanksalsoexhibitedhigherLCRspre-QE,reinforcingtheinterpretationthat thesebankswerestructurallymoreconservativeintheirliquidityandriskmanagement. 23

5.1 examines the relative effect on general lending outcomes. Section 5.2 focuses on credit lines, providing additional evidence and intuition for our key finding that more exposed banks reduce the liquidity insurance they extend to firms relative to less exposed banks. 5.1 General lending effects We estimate the following panel regression from 2016Q1 to 2022Q4: log(Y )=λ (QE ·Shares )+λ (QT ·Shares )+βX +α +α +ϵ (2) i,f,t 1 t i 2 t i i,t i,f f,t i,f,t Thedependentvariable, log(Y ), representsthelogarithmofthreelendingmeasures: comi,f,t mitted credit lines, term loans, and total commitments (the sum of the two) for bank i, firm f, and time period t. QE , QT , and Shares are the same variables used in the previous section t t i indicating, respectively, the periods of quantitative easing and tightening, and the level of bank i’s exposure to the QE-induced fragility. X includes a set of time-varying bank controls, ini,t cluding bank size, deposit liabilities, and liquid assets. As shown in the previous section, while total deposits do not differ significantly between more and less exposed banks, there is a notable compositionalshift: moreexposedbanksshiftfromuninsuredtoinsureddepositsrelativetoless exposed banks. To account for this dynamic, we separately control for insured and uninsured deposits in the lending regressions. Additionally, we control for the level of reserves and other high-qualityliquidassets,namelytreasuriesandagencies,toaccountfortheirpotentialimperfect substitutability and the possibility that QE affects them differently. To account for unobserved heterogeneity, we include bank-firm (α ) fixed effects and either i,f industry-location-size-time (ILST) or firm-time (α ) fixed effects. The former allows us to f,t control for persistent bank-firm relationships, while the latter absorbs time-varying firm-level credit demand. Specifically, bank-firm fixed effects control for potential non-random matching between firms and banks, capturing all time-invariant factors that may influence credit within a given bank-firm pair, such as relational banking. Firm-time fixed effects ensure that our estimates capture the supply-side effects of bank lending behavior by absorbing all firm-level demand factors. However, their use results in the exclusion of firms that borrow from only one bank, which is a substantial share of our sample. Given that many smaller firms rely on a single 24

bank, estimatesbasedonfirm-timefixedeffectsmaynotfullyrepresentbroaderfirm-levelcredit dynamics. To address this concern, we also consider an alternative specification that replaces firm-time fixed effects with industry-location-size-time fixed effects, which retains both singleand multi-bank firms, following Degryse, De Jonghe, Jakovljevi´c, Mulier, and Schepens (2019). Finally, we cluster standard errors at the bank-time and firm levels. Table 6 presents the results from estimating equation (2) for credit lines (columns 1-2), term loans (columns 3-4), and total loan commitments (columns 5-6). Across all specifications, we include the full set of time-varying bank controls as discussed above.20 The key difference between columns 1 and 2 (and subsequently between columns 3-4 and 5-6) lies in the choice of fixed effects. Specifically, in columns 1, 3, and 5, we include bank-firm FE alongside ILST FE, while in columns 2, 4, and 6, we replace the ILST FE with firm-time FE. The latter provide a stricter control for firm-level credit demand but reduce the sample size considerably, as firms borrowing from only one bank do not contribute to the estimation. Our coefficient of interest is the interaction term, QE · Shares. With respect to credit lines, this coefficient is negative and significant in all specifications, suggesting that banks with higher exposure to the QE-induced fragility had fewer credit-line commitments to firms after QE, relative to less exposed banks. Economically, the estimates in columns 1-2 indicate that a one percentage point increase in the exposed banks is associated with a 0.08-0.13 percentage point decrease in credit-line commitments during the QE period. Section 5.2 further dissects this result by examining the sub-components of credit-line commitments and the implications for banks’ liquidity management. Note that this result concerns the differential effect on creditline extension between more and less exposed banks. Credit-line commitments kept increasing for both types of banks throughout QE. However, our granular data and the novel identification of QE exposure via the inflow of NBFI deposits, allows to capture this differential effect. With respect to term loans, we find no significant difference between more exposed and less exposed banks after the QE. This result may not be surprising given that total deposits did not evolve differently for more exposed and less exposed banks. Finally, the results in columns 5-6 20Includingbank-levelcontrolshelpstoaccountforpotentialdifferencesinbanks’balancesheetcharacteristics. However, controlling for them may absorb part of the variation through which QE-induced fragility influences lending,potentiallyunderestimatingthefulleffect. Toensurethatourfindingsarenotdrivenbyselectionbias, wepresentresultsbothwithandwithoutthesecontrolsassuggestedbyGormleyandMatsa(2011). Theresults remainunchangedwhenbankcontrolsareexcluded. SeeTableOA8intheAppendix. 25

Table 6: Credit lines, Term loans, and Total Loan Commitments 1 2 3 4 5 6 Dependent variable: Log(Credit lines) Log(Term loans) Log(Total commitments) QE*Shares -0.095** -0.076* 0.065 0.110 -0.142*** -0.153*** (-2.063) (-1.783) (0.752) (1.394) (-3.212) (-3.603) QT*Shares -0.271*** -0.235*** -0.013 -0.028 -0.314*** -0.338*** (-3.761) (-4.045) (-0.086) (-0.211) (-4.278) (-5.124) Bank size 0.047 0.051 0.085 0.086 0.086*** 0.113*** (1.347) (1.408) (1.259) (1.141) (2.659) (3.127) Bank reserves 0.016*** 0.028*** 0.003 0.003 -0.004 0.000 (2.751) (4.273) (0.386) (0.415) (-1.133) (0.052) Bank treasuries & agencies -0.019** -0.009 -0.017 -0.006 -0.025*** -0.019** (-2.390) (-0.981) (-1.434) (-0.426) (-3.509) (-2.087) Bank insured deposits -0.061*** -0.075*** -0.052 -0.067* -0.050*** -0.063*** (-2.833) (-3.723) (-1.378) (-1.757) (-2.715) (-3.429) Bank uninsured deposits 0.094*** 0.112*** -0.073 -0.042 0.028 0.032 (3.625) (4.197) (-1.366) (-0.697) (1.198) (1.279) Bank*Firm FE Y Y Y Y Y Y ILST FE Y Y Y Firm*Time FE Y Y Y Observations 632,635 317,776 236,988 95,199 919,369 391,659 R-squared 0.966 0.944 0.953 0.918 0.962 0.935 Note: The table reports coefficients and t-statistics (in parentheses) for the following regression equation: log(Y )=λ (QE ·Shares )+λ (QT ·Shares )+βX +a +a +ε . Thedependentvariable i,f,t 1 t i 2 t i i,t i f,t i,f,t foreachbanki,firmf,andtimeperiodtisthelogarithmofcreditlines(columns1to2),thelogarithm oftermloans(columns3to4),andthelogarithmoftotal(loan)commitments,whichisthesumofthe two(columns5-6). QE isadummyequaltoonefromMarch2020toMarch2022,andQT isadummy t t equal to one from June 2022 onwards. Shares indicates the share of uninsured NBFI deposits in total i depositsforbankiasofFebruary2020. Bank size isthelogarithmoftotalassets,Bank reservesisthe i,t logarithm of reserves, and Bank treasuries & agencies is the logarithm of the treasuries and agencies a bank holds. The variables Bank insured deposits and Bank uninsured deposits represent the logarithm of insured and uninsured deposits, respectively. In all specifications, we include different levels of fixed effects,asnotedinthelowerpartofthetable. Observationsareatthebank-firm-timelevelandarereportedquarterly. Standarderrorsareclusteredatthebank-quarterandfirmlevels. Variabledefinitions anddatasourcesareprovidedinAppendixC.Thesymbols***,**,and*denotestatisticalsignificance at the 1%, 5%, and 10% levels, respectively. 26

showthattotalloancommitmentsarelowerformoreexposedbanksaftertheQErelativetoless exposed banks.21 In Section 6, we further examine the aggregate implications for firms’ access to credit and real economic outcomes. Taken together, these results indicate that more exposed banks did not expand their credit-line commitments following QE, leading to a relative decline compared to less exposed banks. At the same time, their term loan commitments remained unchanged, resulting in an overall reduction in total loan commitments. 5.2 Credit lines and Liquidity Insurance We now examine in more detail the underlying forces that drive down credit-line commitments for more exposed banks relative to less exposed ones. Credit-line commitments consist of two components: utilized credit lines and undrawn credit lines. When a bank issues a credit line, it doesnotimmediatelyextendaloanonitsbalancesheet. Instead,itprovidesacommitmentthat allows the firm to draw funds when needed. Once a firm withdraws from a committed credit line, the utilized portion (known as draw-down) appears as a credit-line loan on the bank’s balance sheet, while the remaining amount represents undrawn credit, which firms can access in the future. Undrawn credit lines are a key measure of liquidity insurance, as they reflect the funding available to firms for future needs. However, from a bank’s perspective, these undrawn commitments pose liquidity risk, as they represent off-balance-sheet obligations that could be drawn unpredictably (Ippolito et al., 2016; Acharya et al., 2024). Unlike publicly available call reports, which only capture utilized credit lines, our dataset allows us to separately analyze utilized and undrawn credit lines, providing a more precise view of how QE-induced fragility affects banks’ provision of contingent liquidity. Table 7 reports the results from estimating equation (2) for utilized credit lines (columns 1-2)andundrawncreditlines(columns3-4).22 Ourcoefficientofinterestistheinteractionterm, QE ·Shares. Focusing on utilized credit lines, we find no significant difference between more exposed and less exposed banks after the QE. This result is intuitive: the decision to utilize a 21Toensurethatourlendingresultsarenotsensitivetothelogtransformationofloanvariables,were-estimate alllendingregressionsusingPoissonpseudo-maximumlikelihood(PPML).Thismethodiswell-suitedforhandling skeweddataandcaseswheresomeloancommitmentsarezero(Cohn,Liu,andWardlaw,2022). Theresultsremain robustandalignedwithourmainfindings. Unreportedestimatesareavailableuponrequest. 22The results do not change if we exclude bank controls. See Table OA9 in the Appendix. Additionally, in unreported results (available upon request), we control for interactions between QE and QT with the LCRindicatordescribedabovetoaccountforgeneralprecautionaryliquiditymotives. Allresultsremainrobust. 27

Table 7: Utilized & Undrawn Credit Lines 1 2 3 4 Dependent variable: Log(Utilized credit lines) Log(Undrawn credit lines) QE*Shares -0.005 -0.058 -0.291*** -0.182*** (-0.041) (-0.566) (-4.855) (-4.021) QT*Shares -0.160 -0.072 -0.420*** -0.326*** (-0.837) (-0.450) (-4.987) (-5.324) Bank size -0.144* 0.119* 0.119** 0.071** (-1.789) (1.962) (2.577) (2.143) Bank reserves 0.002 0.022** -0.017** -0.005 (0.206) (2.440) (-2.322) (-0.983) Bank treasuries & agencies -0.068*** -0.025 -0.014 -0.019** (-4.057) (-1.348) (-1.125) (-2.166) Bank insured deposits -0.057 -0.064 -0.044* -0.027 (-1.109) (-1.389) (-1.704) (-1.460) Bank uninsured deposits 0.269*** 0.093* 0.005 0.022 (4.477) (1.866) (0.156) (0.883) Bank*Firm FE Y Y Y Y ILST FE Y Y Firm*Time FE Y Y Observations 408,805 184,557 550,076 300,783 R-squared 0.860 0.874 0.897 0.942 Note: The table reports coefficients and t-statistics (in parentheses) for the following regression equation: log(Y ) = λ (QE ·Shares )+λ (QT ·Shares )+βX + i,f,t 1 t i 2 t i i,t a + a + ε . The dependent variable for each bank i, firm f, and time period i f,t i,f,t t is the logarithm of utilized credit lines (Columns 1–2) or the logarithm of undrawn credit lines (Columns 3–4). QE is a dummy equal to one from March 2020 to March t 2022, and QT is a dummy equal to one from June 2022 onwards. Shares indicates t i the share of uninsured NBFI deposits in total deposits for bank i as of February 2020. Bank size is the logarithm of total assets, Bank reserves is the logarithm of reserves, i,t and Bank treasuries & agencies is the logarithm of the treasuries and agencies a bank holds. The variables Bank insured deposits and Bank uninsured deposits represent the logarithm of insured and uninsured deposits, respectively. In all specifications, we include different levels of fixed effects, as noted in the lower part of the table. Observations are at the bank-firm-time level and are reported quarterly. Standard errors are clustered at the bank-quarter and firm levels. Variable definitions and data sources are provided in Appendix C. The symbols ***, **, and * denote statistical significance at the 1%, 5%, and 10% levels, respectively. 28

credit line is primarily driven by firms, and after controlling form firm demand, there is no clear reason why firms would systematically treat more exposed and less exposed banks differently when drawing down their credit lines.23 Turning to the undrawn credit lines, we observe a differentpattern. Thereisasignificantdifferencebetweenmoreexposedandlessexposedbanks, with more exposed banks offering relatively less liquidity insurance to firms after the QE, as measuredbytheundrawncreditlines. Thisreductioninundrawncreditlinesisalsowhatdrives the decline in total loan commitments for more exposed banks relative to less exposed ones, as described above. Liquidity management. These results indicate that the more exposed banks limit or do not topuptheundrawnamountinfirms’creditlinesrelativetolessexposedbanks. Thisadjustment primarily occurs through quantities rather than pricing, as there is no significant difference in interest rate setting on credit lines between more and less exposed banks after QE (see columns 1-4inTableOA10intheAppendix). Furthermore,theeffectisnotdrivenbynewlyissuedcredit lines, where we do not find a significant difference between more and less exposed banks (see columns 5-8 in Table OA10 in the Appendix). Instead, more exposed banks reduce the liquidity insurance they provide to firms, likely as a precautionary measure to mitigate the liquidity risk stemming from QE-induced funding fragility and to lower the probability of a “double run” scenario, in which both depositors withdraw their funds and firms draw down their credit lines (Ippolito et al., 2016). We next examine the underlying mechanism driving this adjustment and investigate the aggregate effects for firms’ access to liquidity in Section 6. But, before that, we presentadditionalevidencethatmoreexposedbanksindeedtrytoreducetheliquidityriskfrom future credit-line draw-downs. We examine whether more exposed banks disproportionately reduced their undrawn creditlineexposurestofirmsmostlyvulnerabletoliquiditystrainsafterthepandemicshock. Toidentify such firms, we focus on industries more affected by Covid-19 and, within those industries, on firms with a higher anticipated need for liquidity. We approximate liquidity needs using the ratio of firms’ sales to account receivables, which serves as a proxy for working capital or bridge liquidity that firms may need. To test this, we extend regression (2) by including a quadruple 23RecallthatoursampleconsistsofthebiggestbankintheUnitedStates, whichwereadequatelycapitalized andservedasasourceofstrengthduringthepandemic. 29

interactiontermQE·Shares·Covid·Liquidity,whereCovidisadummyindicatingwhetherfirm f operatestoamoreCovid-affectedindustryandLiquidity isadummyequaltooneifthefirm’s sales-to-accounts-receivableratioin2019Q4isabovethemedian. TableOA11reportstheresults from our strictest specification that accounts for all variation at the firm-time level. Our key coefficientofinterest,QE·Shares,remainsnegativeandsignificant. Additionally,thecoefficient on the quadruple interaction term is also negative and significant, indicating that more exposed banks not only reduce their undrawn credit lines relative to less exposed banks, but they may do so even more for firms most vulnerable to liquidity strains in the post-QE environment. Economic mechanism. In an influential paper, Kashyap et al. (2002) demonstrate strong complementarities between deposit taking and the provision of liquidity insurance to firms via the extension of credit-line commitments. The underlying idea is straightforward: banks must hold liquid assets to meet deposit withdrawals and credit-line drawdowns, but doing so entails an opportunity cost. When deposit-withdrawals and credit-line-utilization are imperfectly correlated, synergies arise, giving banks a comparative advantage in providing both services. The authors show that credit-line commitments are increasing with deposit-taking. In contrast we showed above that more exposed banks, those receiving larger inflows of uninsured NBFI deposits, reduce their credit-line commitment relative to less exposed banks. ToreconcileourresultswithKashyapetal.(2002),weextendtheirmodeltoexplicitlyincorporate runnable uninsured deposits akin to Diamond and Kashyap (2016) (see the Appendix A forthefullexpositionofthemodel). Weshowthatthissmallmodificationissufficienttoreverse their original result, aligning with our empirical findings. The intuition is straightforward once one accounts for out-of-equilibrium considerations, which are crucial when studying runnable deposits. Managing liquidity in the presence of runnable deposits requires banks to prepare for self-fulfilling withdrawals—not just those expected in equilibrium. In particular, a bank must assess its solvency even in a worst-case scenario where all depositors withdraw and the bank resortstomoreexpensivewholesalefunding. Whentheshareofrunnabledepositsbecomessufficientlylarge, insuringagainstallpotentialout-of-equilibriumwithdrawalsbecomesprohibitively expensive, prompting banks to reduce their lines of credit despite the inflow of runnable deposit funding. Notably, this mechanism does not arise if deposits are insured and therefore not 30

runnable,orifthebankremainssolventunderallpotentialwithdrawalscenarios—anassumption maintained in the numerical example of Kashyap et al. (2002). Figure 4 outlines the mechanism through which runnable deposits may reverse the result of Kashyap et al. (2002), using a calibration similar to theirs (see the Appendix A for analytical results and calibration details).24 Similar to Kashyap et al. (2002), we consider an exogenous increase in deposits, but we analyze two cases: (i) insured and non-runnable deposits, and (ii) uninsured and runnable deposits. If depositors withdraw, then the bank must rely on more expensive funding. The key variable to track is the bank profits under an out-of-equilibrium scenarioinwhichalldepositorswithdraw(topchart). Asdepositsincrease,thebankapproachesa thresholdbeyondwhichitcannolongerremainsolventifalldepositorswithdrawsimultaneously. If all deposits are insured, then the bank can theoretically violate its solvency constraint in outof-equilibrium paths, as insured depositors have no incentives to run in equilibrium. However, whendepositsareuninsuredandrunnable,thesolvencyconstraintbecomesbindingoncedeposits reach a critical level; otherwise, uninsured depositors would decide to run in equilibrium fearing that others may do the same.25 ThemiddlechartofFigure4showstheincrementalcostsofwholesaleborrowingtomeetdepositwithdrawals. Asdepositsincrease,theincrementalcostsriseinequilibriumforbothinsured and uninsured deposits at a similar rate. However, the cost of serving withdrawals in all out-ofequilibrium paths escalates much more rapidly for uninsured, runnable deposits. Consequently, banksinitiallyexpandtheircredit-linecommitmentsasdepositsgrow,regardlessofwhetherthese deposits are insured or uninsured. However, once uninsured deposits exceed a critical threshold, the rising liquidity risk and associated costs become too severe. At this point, banks actively adjust their exposure by reducing credit-line commitments (bottom panel). These dynamics suggest that beyond a certain level of uninsured deposits, the strong complementarity between deposit-takingandtheprovisionofliquidityinsuranceviacredit-linecommitments—emphasized by Kashyap et al. (2002)—breaks down. 24Thekeydifferenceisthatweintroducetheimpatient/patientdepositorsakintoDiamondandDybvig(1983), toexplicitlymodelself-fulfillingrunsandhighlighttheimportanceofrunnabledeposits. 25Forsimplicityandwithoutlossofgenerality,wefollowDiamondandKashyap(2016)inassumingthatbanks aimtoavoidfailureinanyout-of-equilibriumpaths,whichisjustifiedbytheassumtpionthatdepositorsarevery risk averse and only accept riskless deposits. Our argument does not rely strictly on such run-proof contracts andcouldbeextendedtocasesthatrunriskispositiveinequilibriumasinKashyap,Tsomocos,andVardoulakis (2024). Weleavethisextensiontofutureresearch. 31

Figure 4: Model Simulation. Runnable deposits & Credit-line commitments Note: TheFigureplotstheequilibriumoutcomesfromsimulatingthemodeloutlinedintheAppendixfordifferent levelofdeposits. Weconsidertwocases: (i)depositsareinsuredandnotrunnable,and(ii)depositsareuninsured andrunnable. Thetopchartreportsbank’ssolvencyconstraintintheout-of-equilibriumpathsthatalldepositors withdraw. Themiddlechartreportstheincrementalcostofservingdepositswithdrawalin-andout-of-equilibrium. Thebottomchartshowstheequilibriumlevelof(undrawn)credit-linecommitments. 6 Aggregate Lending and Real Effects Intheprevioussection,weshowedthatbanksmoreexposedtotheQE-inducedfragilityactively managedtheirliquidityriskbyadjustingtheirloancommitments,primarilybyreducingundrawn creditlines. Thisresponsemayhaveconstrainedfirms’accesstocontingentliquidity,evenastotal deposits remained stable but shifted toward a composition with more uninsured and runnable deposits. Whiletheseresultssuggestthatexposedbankstookstepstoreducetheirliquidityrisk, a key question remains: Did this shift in exposed banks have broader firm-level consequences? In principle, firms affected by the contraction in credit-line commitments could have offset the impact by switching to less exposed banks. If they were able to do so, the decline in bank-level liquidityprovisionmaynothavenecessarilyledtoacontractioninfirmborrowingorinvestment. To examine this, we aggregate quarterly loan commitments at the firm level and estimate the following panel regression from 2016Q1 to 2022Q4: log(Y )=λ (QE ·Exposure )+λ (QT ·Exposure )+α +α +ϵ (3) f,t 1 t f 2 t f ILST f f,t 32

The dependent variable, log(Y ), represents the logarithm of (i) the following types of lendf,t ing: utilized credit lines, undrawn credit lines, term loans, and total loan commitments, and (ii) firminvestmentasmeasuredbythechangeinfixedassets,foreachfirmf andtimeperiodt. QE t and QT are the same variables used in the previous sections, indicating the periods of quantitat tiveeasingandtightening. Exposure captureshowexposedafirmistotheQE-inducedfragility f through its loan relationships with exposed banks. We construct three measures to capture differentdimensionsoffirmrelianceonmoreexposedbanks. Thefirstmeasure(Unweighted Shares Exposure) is the average share S of uninsured NBFI deposits among the banks with which firm i f has loan relationships as of 2019Q4. The second (Weighted Shares Exposure) is a weighted version of the same measure, where the weights corresponds to the shares of firm f total loan commitments held with each bank i. The third measure (Relationships Dummy) is a dummy equal to one for firms that have more than 50% of their lending relationships with more exposed banks. These measures capture different dimensions of firm exposure, ensuring a broad and consistent pattern across different ways firms interact with exposed banks. To control for firm characteristics and credit demand, we saturate the specification with industry-location-size-time fixed effects and firm fixed effects. Standard errors are clustered at the firm level. Table 8 reports the results for lending (columns 1-4) and real effects (column 5) for the first exposure measure. The findings are consistent for the other two measures, reported in Table OA12 in the Appendix. At the firm level, we confirm that more affected firms experience a decline in liquidity insurance, as reflected in lower undrawn credit lines, but show no significant differences from less affected firms in terms of credit-line utilization or term loans received. Hence, the total impact on credit availability, measured by total loan commitments (column 4), islargelyinsignificant. Theseresultsindicatethatthedeclineinundrawncreditlinesobservedat thebank-firmlevelextendstothefirmlevel,i.e.,affectedfirmsfacereducedaccesstocontingent liquidity. In response, firms reduce investment (column 5), likely as a precautionary measure to preserve liquidity and flexibility amid a diminished ability to hedge future liquidity shocks.26 Overall, the QE-induced fragility can have aggregate effects by reducing liquidity insurance to firms and affecting real economic outcomes. We believe that our analysis is shedding light on 26This aligns with previous studies that highlight the importance of credit lines for firms’ investment, for example,Chang,Chen,andMasulis(2023). 33

these important unintended consequences of quantitative easing. Table 8: Aggregate effects at firm level 1 2 3 4 5 Dependentvariable: Log(Utilized Log(Undrawn Log(Termloans) Log(Total Log(Investment) Dependentvariable: creditlines) creditlines) commitments) QE*UnweightedSharesExposure 0.201 -0.354* 0.157 -0.063 -2.391** (1.122) (-1.945) (1.063) (-0.959) (-2.638) QT*UnweightedSharesExposure 0.219 -0.595** 0.212 -0.230** -1.718 (0.776) (-2.729) (0.867) (-2.158) (-1.675) Observations 223,976 256,001 122,718 497,200 43,199 R-squared 0.820 0.798 0.929 0.951 0.817 ILSTFE Y Y Y Y Y FirmFE Y Y Y Y Y Note: The table reports coefficients and t-statistics (in parentheses) for the following regression equation: log(Yf,t) = λ1(QEt·Exposuref)+λ2(QTt·Exposuref)+aILST +af +εf,t, where the dependent variable for each firm f in time period t is the sum of credit the firm received. The dependent variables are: the logarithm of utilized credit lines (column 1), the logarithm of undrawn credit lines (column 2), the logarithm of term loans (column 3), the logarithm of total commitments (column 4) and the logarithm of investments (column 5). Exposuref is a measure of how exposed a firm is to the QE-induced fragility via the loan relationships that firm has with exposed banks. Table reports results for Exposuref ={UnweightedSharesExposuref},whichistheaverageshareSi amongthosebankswithwhichafirmf has loanrelationshipsat2019Q4. Theregressionincludesindustry-location-size-time(aILST)andfirm(af)fixedeffects. Standarderrorsareclusteredatthefirmlevel. VariabledefinitionsanddatasourcesareprovidedinAppendixC.Thesymbols ***,**,and*denotestatisticalsignificanceatthe1%,5%,and10%levels,respectively. 7 Conclusions Our findings highlight the critical role of bank liquidity management in response to central bank quantitative policies such as QE and QT. We show that banks with greater exposure to uninsured NBFI deposits during the Pandemic QE adjusted their liability structures by shifting theircompositionofdepositsfrominsuredtouninsured. Ontheassetside, wefindthatexposed bankscutbackoncreditlinecommitmentswhilemaintainingtermloanissuance,therebylimiting firms’ access to contingent liquidity. This active liquidity management on both sides of the balance sheet reflects banks’ efforts to mitigate funding fragility induced by the large influx of flightyNBFIdepositsasanoutcomeoftheQEoperations. ThissuggeststhatwhileQEinjected substantialliquidityintothebankingsystem,italsoledtounintendedconsequencesforcorporate liquidity insurance, reducing firms’ ability to manage future liquidity shocks. Our results carry important implications for monetary policy transmission and financial stability. While QE aims to ease financial conditions and stimulate lending, our findings suggest 34

that its impact is nuanced, as shifts in bank funding composition can lead to constraints on liquidity provision. The persistence of uninsured NBFI deposits post-QE, despite the initiation of QT, underscores the long-term structural changes in bank balance sheets driven by central bank interventions. Future research could explore whether similar patterns emerge in different regulatory environments or during periods of financial stress, shedding further light on the evolving interactions between monetary policy, banks and NBFIs. References Acharya, V. V., N. Cetorelli, and B. Tuckman (2024). Where do banks end and nbfis begin? Technical report, National Bureau of Economic Research. Acharya, V. V., R. S. Chauhan, R. Rajan, and S. Steffen (2023). Liquidity dependence and the waxing and waning of central bank balance sheets. Technical report, National Bureau of Economic Research. Acharya, V. V., R. Engle, M. Jager, and S. Steffen (2024). Why did bank stocks crash during covid-19? The Review of Financial Studies 37(9), 2627–2684. Acharya, V. V. and N. Mora (2015). A crisis of banks as liquidity providers. The Journal of Finance 70(1), 1–43. Acharya, V. V. and R. Rajan (2024). Liquidity, liquidity everywhere, not a drop to use: Why flooding banks with central bank reserves may not expand liquidity. The Journal of Finance 79(5), 2943–2991. Afonso, G., M. Cipriani, and G. La Spada (2022). Banks’ balance-sheet costs, monetary policy, and the on rrp. FRB of New York Staff Report (1041). Bernanke, B. S. (2022). 21st century monetary policy: The Federal Reserve from the great inflation to COVID-19. WW Norton & Company. Chakraborty, I., I. Goldstein, and A. MacKinlay (2020). Monetary stimulus and bank lending. Journal of Financial Economics. Chang, X., Y. Chen, and R. W. Masulis (2023). Bank lines of credit as a source of long-term finance. Journal of Financial and Quantitative Analysis 58(4), 1701–1733. Chen, K., J. Ren, and T. Zha (2018). The nexus of monetary policy and shadow banking in china. American Economic Review 108(12), 3891–3936. Chodorow-Reich,G.(2013,10). Theemploymenteffectsofcreditmarketdisruptions: Firm-level evidence from the 2008–9 financial crisis. The Quarterly Journal of Economics 129(1), 1–59. Cohn, J. B., Z. Liu, and M. I. Wardlaw (2022). Count (and count-like) data in finance. Journal of Financial Economics 146(2), 529–551. Cooperman,H.R.,D.Duffie,S.Luck,Z.Wang,andY.Yang(2023). Bankfundingrisk,reference rates, and credit supply. Technical report, National Bureau of Economic Research. 35

Degryse, H., O. De Jonghe, S. Jakovljevi´c, K. Mulier, and G. Schepens (2019). Identifying credit supply shocks with bank-firm data: Methods and applications. Journal of Financial Intermediation 40, 100813. Di Maggio, M., A. Kermani, and C. J. Palmer (2019, 12). How quantitative easing works: Evidence on the refinancing channel. The Review of Economic Studies 87(3), 1498–1528. Diamond, D. and A. Kashyap (2016). Liquidity requirements, liquidity choice, and financial stability. Volume 2 of Handbook of Macroeconomics, pp. 2263–2303. Elsevier. Diamond, D. W. and P. H. Dybvig (1983). Bank runs, deposit insurance, and liquidity. Journal of political economy 91(3), 401–419. Diamond, W., Z. Jiang, and Y. Ma (2024). The reserve supply channel of unconventional monetary policy. Journal of Financial Economics 159, 103887. Duffie, D. (2020). Still the world’s safe haven? – redesigning the us treasury market after the covid-19 crisis. Hutchins Center Working Paper 62, Brookings Institution. Elliott, D., R. R. Meisenzahl, and J.-L. Peydr´o (2024). Nonbank lenders as global shock absorbers: evidence from us monetary policy spillovers. Journal of International Economics, 103908. Favara, G., S. Infante, and M. Rezende (2024). Leverage regulations and treasury market participation: Evidence from credit line drawdowns. working paper. Federal Reserve Board (2021, March). Press release. https://www.federalreserve.gov/ newsevents/pressreleases/bcreg20210319a.htm. Accessed: 2024-07-10. Franceschi, E., M. Grodzicki, B. Kagerer, C. Kaufmann, F. Lenoci, L. Mingarelli, C. Pancaro, andR.Senner(2023).Keylinkagesbetweenbanksandthenon-bankfinancialsector.Financial Stability Review 1. Gatev, E., T. Schuermann, and P. E. Strahan (2009). Managing bank liquidity risk: How deposit-loan synergies vary with market conditions. The Review of Financial Studies 22(3), 995–1020. Gatev, E. and P. E. Strahan (2006). Banks’ advantage in hedging liquidity risk: Theory and evidence from the commercial paper market. The Journal of Finance 61(2), 867–892. Gopinath, G., S. Kalemli-Ozcan, L. Karabarbounis, and C. Villegas-Sanchez (2017). Capital Allocation and Productivity in South Europe. The Quarterly Journal of Economics 132(4), 1915–1967. Gormley,T.A.andD.A.Matsa(2011). Growingoutoftrouble? corporateresponsestoliability risk. The Review of Financial Studies 24(8), 2781–2821. Hanson, S. G., A. Shleifer, J. C. Stein, and R. W. Vishny (2015). Banks as patient fixed-income investors. Journal of Financial Economics 117(3), 449–469. Holmstr¨om, B. and J. Tirole (1998). Private and public supply of liquidity. Journal of Political Economy 106(1), 1–40. Imbens,G.W.andJ.M.Wooldridge(2009).Recentdevelopmentsintheeconometricsofprogram evaluation. Journal of economic literature 47(1), 5–86. 36

Ippolito, F., J.-L. Peydr´o, A. Polo, and E. Sette (2016). Double bank runs and liquidity risk management. Journal of Financial Economics 122(1), 135–154. Irani, R. M., R. Iyer, R. R. Meisenzahl, and J.-L. Peydro (2021). The rise of shadow banking: Evidence from capital regulation. The Review of Financial Studies 34(5), 2181–2235. Joyce, M., D. Miles, A. Scott, and D. Vayanos (2012). Quantitative easing and unconventional monetary policy–an introduction. The Economic Journal 122(564), F271–F288. Kashyap, A. K., R. Rajan, and J. C. Stein (2002). Banks as liquidity providers: An explanation for the coexistence of lending and deposit-taking. The Journal of Finance 57(1), 33–73. Kashyap, A. K., D. P. Tsomocos, and A. P. Vardoulakis (2024). Optimal bank regulation in the presence of credit and run risk. Journal of Political Economy 132(3), 772–823. Kiernan, K. F., V. Yankov, and F. Zikes (2021). Liquidity provision and co-insurance in bank syndicates. FEDS WP 2021-060. Leonard, D., A. Martin, and S. M. Potter (2017). How the fed changes the size of its balance sheet. Technical report, Federal Reserve Bank of New York. Levine, R., C. Lin, M. Tai, and W. Xie (2021). How did depositors respond to covid-19? The Review of Financial Studies 34(11), 5438–5473. Li, L., P. E. Strahan, and S. Zhang (2020). Banks as lenders of first resort: Evidence from the covid-19 crisis. The Review of Corporate Finance Studies 9(3), 472–500. Roberts, M. R. and T. M. Whited (2013). Endogeneity in empirical corporate finance1. In Handbook of the Economics of Finance, Volume 2, pp. 493–572. Elsevier. Rodnyansky, A. and O. M. Darmouni (2017). The effects of quantitative easing on bank lending behavior. The Review of Financial Studies 30(11), 3858–3887. Vayanos, D. and J.-L. Vila (2021). A preferred-habitat model of the term structure of interest rates. Econometrica 89(1), 77–112. 37

Appendix “QE, Bank Liquidity Management, and Non-Bank Funding: Evidence from Administrative Data” This appendix provides supplementary information and results to support the main paper. The content is organized as follows: Appendix A presents the theoretical model. Appendix B details the filters applied to construct the final dataset. Appendix C provides additional tables that supplement the main results. i

A Theory WeextendKashyapetal.(2002)toexaminehowanincreaseinuninsured,runnable,depositsaffects a bank’s choice to offer credit-line commitments to firms. We maintain the whole structure of their model and only make two modifications. First, we consider that depositors are homogeneous ex ante but receive ex post an idiosyncratic, uninsurable, preference shock to consume early or late. This ex post heterogeneity between impatient and patient depositors allow us to study the incidence of self-fulfilling runs as in Diamond and Dybvig (1983). Second, we consider two types of deposits: insured (non-runnable) deposits and uninsured (runnable) deposits. The former correspond to Kashyap et al. (2002), where only the solvency and liquidity of the bank on the equilibrium path matter. The latter corresponds to the case that runs are possible. To eliminate such runs depositors need to be certain about the solvency and liquidity of the bank not only on the equilibrium path, but also for all out-of-equilibrium paths, or in other words for the worst-case scenario as in Diamond and Kashyap (2016). WebrieflydescribetheKashyapetal.(2002)environmentandreferthereadertotheirpaper foralldetails. Therearethreetimeperiodst=0,1,2andfourtypesofagents: depositors,firms, financiers, and a bank. The per-period net interest rate is set to i > 0. At t = 0 depositors invest all their endowment in deposits and, hence, the amount of deposits, D, in the bank is exogenously determined. Depositors are very risk averse and only accept deposit contracts that carry zero risk. In return, they are willing to accept a zero deposit rate, generating a deposit franchiseforthebank. Depositorsreceiveanuninsurableidiosyncraticpreferenceshockatt=1, urging δ˜of them to withdraw early; δ˜=δ ∈(0,1) and δ˜=0 with equal probability. This is the firstmodificationwemakeonKashyapetal.(2002)thatconsiderδ =1,i.e.,eitheralldepositors or none withdraw. The bank has also access to funding markets both at t = 0 and t = 1 were financiers invest in the bank in the form of wholesale funding or equity injections. Denote by e and e the 0 1 external financing at t = 0 and t = 1, respectively. Financiers demand the market interest rate i but also require an additional premium for period-1 financing equal to α/2e2, with α > 0. 1 This incremental cost of external financing will play an important role in the bank’s liquidity management. Finally, for simplicity, all interest payments accrue at t=2.

The bank uses the period-0 deposits and external funds to extend term loans and also invest in liquid assets, denoted by L and S . Term loans mature at t = 2 yielding a net loan rate 0 r. Liquid assets mature after one period and pay net interest i−τ, with τ > 0 to account for the fact that hoarding liquidity is costly for the bank. The bank can also extend credit-line commitments to firms at t = 0, denoted by C. These lines of credit constitute a promise to extendloansuptoC att=1ifthefirmsdecidestodraw-downthelineofcredit. Utilizedcredit lines carry an net interest i, but firms also pay a fee fC, with dfC/dC >0 and d2fC/dC2 <0, to have access to such lines of credit irrespective if they end up using or not. But, credit lines do not take balance-sheet capacity unless drawn and, thus, the balance sheet of the bank at t=0 is L+S =D+e . At t=1 firms receive a shock urging them to utilize z˜portion of the 0 0 credit lines; z˜= 1 and z˜= 0 with equal probability. Importantly, z˜ and δ˜ may be imperfectly correlated with correlation ρ≤1. Next, we examine the balance sheet constraint at t = 1 and bank solvency at t = 2 when λ depositors decide to withdraw. Given that self-fulfilling runs are possible under uninsured deposits, we study the general case the λ ∈ (δ˜,1). This is the second modification we make on Kashyap et al. (2002) that consider only the equilibrium level of (impatient) withdrawals λ=δ˜. The balance sheet constraint at t = 1 is, then, z˜C +λD = S +e if z˜C +λD−S > 0 and 0 1 0 z˜C +λD+S = S otherwise. In other words, the bank resorts to external financing, e , at 1 0 1 t = 1 only if their liquidity outflows cannot be met by the liquidity they carried over from the previous period; otherwise, the bank rolls over any remaining liquidity, S , to t = 2. We focus 1 on the case that S <min(δD,C), i.e., the bank requires external financing at t=1 apart from 0 the cases that δ˜=z˜=0. This is the interesting case in Kashyap et al. (2002) that gives rise to the mechanism they highlight, but we also show the other cases in the numerical solution. Weexaminethesolvencyofthebankatt=2whenitfacesaliquidityshortfallz˜C+λD−S > 0 0 at t = 1 and needs to borrow at more expensive rates. Given a λ, the highest shortfall is for z˜=1, so if a bank can survive that state, it is solvent for z˜=0 as well. The profits for deposit iii

withdrawals λD and credit-line draw-downs for z =1 C are given by: Π(λ)=(1+r)L+fC+S (i−τ)+(1+i)C−(1+2i)e −(1+i)e −α/2e2−(1−λ)D 0 0 1 1 =rL−2iL+fC−τS −α/2(C+λD−S )2+(2−λ)Di, (4) 0 0 using the expressions for e and e from the balance-sheet constraints. Π(λ) is decreasing in λ. 0 1 Thus,aslongasititpositiveforthehighestpossibleλgiventhedepositcontract,thenthebank is always solvent. In the case of insured deposits, only impatient depositors would withdraw early, and hence the highest possible λ is equal to δ. It suffices then that the equilibrium profits Π(δ) for δ˜= δ and z˜= 1 are positive, which is always true from optimality; otherwise deposits wouldberiskyanddepositorswouldnotdepositinthebank. Inthecaseofuninsureddepositors, thestricterconditionΠ(1)≥0isneededtoeliminateallfearsaboutpotentialruns,i.e.,thebank needs to remain solvent in all out-of-equilibrium paths for potential withdrawals, which is true is profits are positive for λ = 1 and z˜ = 1. Note that if this condition is satisfied, then only impatient depositors withdraw in equilibrium, i.e., λ = δ˜. But the bank may need to make adjustments to eliminate out-of-equilibrium fears. In fact, Π(1) ≥ 0 can be regarded as an incentive compatibility constraint for the bank, since the slightest probability of a run would make deposits risky and push the very risk-averse depositors away. 27 Then, the bank chooses L,C,S to maximize 0 ρ/2· (cid:2) rL−2iL+fC−τS −α/2(C+δD−S )2+(2−δ)Di (cid:3) 0 0 +(1−ρ)/2· (cid:2) rL−2iL+fC−τS −α/2(C−S )2]+2iD (cid:3) 0 0 +(1−ρ)/2· (cid:2) rL−2iL+fC−τS −α/2(δD−S )2]+(2−δ)Di (cid:3) 0 0 +ρ/2·[rL−2iL+fC−τS +2iD], (5) 0 subject to Π(1)=rL−2iL+fC−τS −α/2(C+D−S )2+Di≥0 (µ), (6) 0 0 27Our arguments should carry through for at least certain cases with positive run risk in equilibrium akin to Kashyapetal.(2024),whomicrofoundtheprobabilityofarunusingaglobalgameandderivetheoptimalcapital andliquidityregulation.. Weleavethisextensiontofuturework. iv

where µ is the Lagrange multiplier on the out-of-equilibrium solvency constraint (6). The first-order conditions with respect to L, C, and S are: 0 L: (r L+r−2i)(1+µ)=0, (7) L α C : dfC/dC− (ρδD+C−S )+µ(dfC/dC−α(C+D−S ))=0, (8) 2 0 0 α S : −τ + (δD+C−S (2−ρ))+µ(α(C+D−S )−τ)=0, (9) 0 2 0 0 where r <0 is the derivative of the loan rate with respect to loan amount along loan demand. L Recall that the bank is internalizing the loan demand schedule. L is determined by exogenous parameterssimilartoKashyapetal.(2002). C andS dependonwhetherthesolvencyconstraint 0 binds, i.e., on µ. Supposefirstthatµ=0. Then,substituting(9)in(8)andtotallydifferentiatingwithrespect , we obtain the same result as in Kashyap et al. (2002): dC −αδ(1−ρ)2 = 4−ρ >0 for ρ<1. (10) dD d2fC − α(1−ρ) dC2 4−2ρ Hence, an exogenous increase in deposits D results in higher credit lines commitment C, as long as deposit withdrawals and credit line draw-downs are not perfectly correlated, i.e., ρ < 1, and theout-of-equilibriumsolvencyconstraintdoesnotbind,i.e.,µ=0. Inthenumericalexercisein section 5.2 of the paper, we show that this result can revert once the solvency constraint binds as the level of deposit increases.28 Below we provide an analytical proof for this reversal. Proposition 1 shows that Π(1) is decreasing in the level of deposit funding and, hence, the out-of-equilibrium solvency constraint (6) will start binding after a level of deposits. It follows that the complementarity between deposit-fundingandcredit-lineissuanceceasestoexistwhen(6)bindsand,insteadbanksreduce the issuance of credit lines when uninsured deposit funding increases. Proposition 1. For low enough α and i: Π(1) ≥ 0 for D ≤ D¯, Π(1) < 0 for D > D¯, and 28Weemploythefollowingparameterization, whichissimilartoKashyapetal.(2002): Theloanratederived from firm’s loan demand is r = A·γ·Lγ−1, with A = 2, α = 0.09 and γ = 0.9; fC = C−0.0.25C2; i = 0.8, τ =0.45,δ=0.5,andρ=0.5. v

dΠ(1)/dD <0. Then, C(D)<C(D¯) for D >D¯ Proof. Recall that we are interested in the region that S <min(δD,C), where the complemen- 0 taritybetweendeposit-fundingandcredit-lineissuanceexistsinKashyapetal.(2002). Consider such an equilibrium. Then, evaluate Π(1) at a level of deposits D′ →S /δ: 0 α (cid:18) 1−δ (cid:19)2 lim Π(1)=rL−2iL+fC+S (i/δ−τ)− C+S , D′→S0/δ 0 2 0 δ which is strictly positive for rL−2iL+fC+S (i/δ−τ) α<αˆ ≡2 0 >0, (cid:0) C+S 1−δ(cid:1)2 0 δ sincerL−2iL=−r L2 >0andi>τ. Thus, bycontinuity, Π(1)canbeslackinanequilibrium L with D >S /δ in the neighborhood of D′ as along as α is sufficiently low. 0 Next take the derivative of Π(1) with respect to D: (cid:18) (cid:19) dΠ(1) dfC dC dS dC dS = −τ 0 −α(C+D−S ) +1− 0 +i. (11) dD dC dD dD 0 dD dD Totally differentiating (9), for µ=0, we get that dS δ+dC/dD 0 = . (12) dD 2−ρ Substituting (8) and (9) for µ=0, as well as (12) in (11), after some algebra, we get dΠ(1) =− dC α (cid:2) (1−ρ)2δD+(1−ρ)(C+D−S )+(1−ρ)(D−S ) (cid:3) dD dD4−2ρ 0 0 τδ 2−ρ−δ − −α(C+D−S ) +i<0, (13) 2−ρ 0 2−ρ for sufficiently low i since dC/dD > 0 for µ = 0 from 10. Given that for D → ∞ we have that Π(1)→−∞, there exists a level of deposits D¯ that the constraint becomes binding. Finally, totally differentiating Π(1) = 0 requires that dΠ(1)/dD = 0. Evaluating this condition at D = D¯, at which point µ → 0, implies that dC/dD < 0 from (13). Hence, the vi

complementarity between deposit-funding and credit-line issuance breaks down when the outof-equilibrium solvency constraint starts binding. It follows that for all D > D¯, the level of C is below its level before the solvency constraint starts binding; otherwise the bank would need to fund these higher commitments with more expensive funding given that it would need to, inefficiently, hold excess liquid assets to cover all out-of-equilibrium deposit withdrawals. B Data We make use of several confidential and public data sources to reconstruct bank balance sheets and lending terms. This appendix outlines the filtering criteria applied to construct the final dataset used in the analysis. We implement a series of selection rules to ensure data consistency and mitigate potential biases. FR 2052a. The unit of analysis from the FR 2052a is the consolidated Bank Holding Company. Our analysis focuses on Product Instruction O.D which reports bank deposits by type (operational (O.D.4), non-operational (O.D.6), transactional, etc), where each product instruction sub-category reports on the status of deposit insurance (FDIC insured or not), maturity (openordatestomaturity),currency(USD,EUR,etc),andcounterparty(retail,corporate,government, financial institution, etc). In our analysis we consider USD-denominated deposits and aggregate over all deposit types. We mainly differentiate along the deposit insurance status and counterparty-type, focusing in particular on NBFIs. Daily data are then aggregated to monthly averages for each bank-year. There is a reporting transition for FR2052a in April 2022 that expanded the set of NBFI counterpartycategories. BeforethereportingchangetherewerethreebroadcategoriesofNBFIs: SupervisedNon-BankFinancialEntities(SNBFEs),DebtIssuingSpecialPurposeEntities(DIS- PEs), and Other Financial Entities (OFEs). SNBFEs include supervised institutions such as investment advisors, (certain) investment companies, brokers/dealers, and insurance companies. DISPEsissue(orhaveissued)commercialpaperorsecuritiestofinancetheirpurchasesoroperations. OFEscompriseinstitutionssuchas(certain)investmentcompaniesaswellashedgefunds or private equity funds. Our main NBFI deposits series aggregates all these three categories. vii

The change introduced additional granularity in NBFI types which further included Broker Dealers, Non-regulated Funds, Debt Issuing Special Purpose Entities, Pension Funds, Other SupervisedNon-BankFinancialEntities,FinancialMarketUtilities,andInvestmentCompanies. Our main NBFI series aggregates all these new categories after the change. Note that from the threeNBFIcategoriesbeforethechangeonlyDebtIssuingSpecialPurposeEntitiescontinuedto bereportedthesamewayafterthechange,whiletheinformationintheotherwasdisaggregated in way that they cannot be unambiguously reconstructed. Toavoiddiscontinuitiesand/ordoublereportingduringthefirstseveralmonthsofthetransition, we hand-checked, bank-by-bank, the NBFI aggregate series and, separately, each subseries was reported the same way before and after the change. We interpolated the data at the daily level for each series when the reporting transition led to a big discontinuity in reported values and then aggregate our series to the monthly level. Depositrates. WeutilizetwodifferentRatewatchdatasets,onewithretailratesandtheother withbusinessrates. WeleveragedaRatewatchretailratedatabasethatincludedinformationon different deposit products and associate rate information that was aggregated to the BHC level andfilteredoutforY-14reportingbanks. Aftermergingmonthlyrawbusinessratefilestogether with raw institutional detail data, and appending each monthly file together, the business rate data was in a similar state to the cleaned retail data. From here, were able to roll up and subsetthebusinessratedatainasimilarway. Thenwecreatesomedummyvariables,onewhich denotes if a product is for amounts greater than $250k, and another if the rate is retail. From here, we append the retail and business rate data together. Banks balance sheet. Bank balance sheet data are collected from the Y9-C using bank holding company RSSDs, which accounts for any bank mergers. For the second stage of analysis (Stage 2) on credit commitments, we convert the monthly data to quarterly averages for every column in FR2052a to merge with FRY-14Q. Y-14Q. The FR Y-14Q dataset covers bank holding companies (BHCs), savings and loan holding companies (SLHCs), and U.S. intermediate holding companies (IHCs) of foreign banking organizations (FBOs). It includes quarterly loan-level data collected as part of the Federal viii

Reserve’s Comprehensive Capital Analysis and Review (CCAR). Institutions covered have consolidated assets exceeding $50 billion (increased to $100 billion from 2019 onward), capturing more than 85 percent of the U.S. banking sector assets. The population of loans in the FR Y-14Q is reported at the credit facility (loan) level and is restricted to commercial and industrial loans with a committed balance of at least $1 million. Eachfacilityisreportedseparately,evenifaborrowerhasmultiplefacilitieswiththesamebank. Facility-level details include total committed and utilized amounts, pricing and spread information,originationandmaturitydates,andcollateralinformation. Loansarecategorizedprimarily as held-for-investment (HFI), representing approximately 98 percent of total loan amounts. The total committed amount reported on the FR Y-14Q as of 2019Q4 is approximately $3.3 trillion, accounting for around 70 percent of U.S. commercial and industrial lending relative to FR Y-9C reports.29 The FR Y-14Q also provides comprehensive financial information (balance sheets and incomestatements)onborrowingfirms, whichisparticularlyvaluableforprivatelyheldU.S.firms that are typically not covered in other datasets. Borrower identifiers, such as tax identification numbers, CUSIPs, andcompanynamesandaddresses, enablematchingwithexternalsourcesto distinguishborrowertypes(e.g.,publicversusprivatefirms,SMEsversuslargefirms,syndicated versus non-syndicated loans). For public companies we merge FR Y-14 with Compustat by firm EIN to obtain their balance sheet information. Finally, we merge in geographic census data information to get MSAs for each firms in our sample. Data Cleaning and Sample Construction. Thissectiondescribestheintensivedatacleaning process needed to use the FR Y14 data for our purposes. 1. Remove from the raw loan-level data loans issued to “Individuals” and loans to foreign addresses. 2. Remove any loans to financial firms (NAICS 52); real estate REITS (NAICS 513); educationalservies(NAICS611);religious,grantmaking,andcivilandprofessionalorganizations (NAICS 813); and private household (NAICS 814). 29We keep loans identified on the FR Y-9C as C&I loans domiciled in the U.S. (item 4(a)), loans to finance agriculturalproduction(item3),loanssecuredbyowner-occupiedrealestatedomiciledintheU.S.(item1(e)(1)), andotherleases(item10(b)). ix

3. Drop all observations for which there is no financial data reported and when total firm assets are missing or equal to 0. 4. Drop all facilities where the total value of commitments is less than $1 million (probable errors given reporting threshold). 5. To consistently identify firms across banks with missing or different tax ids, we first apply a name cleaning algorithm to make a consistent names for firms that are the same based on string matches, zipcode, and city. For example Firm A LLC, 20002 Washington D.C, Firm A Limited Liability Corporation 20002 Washington D.C., and Firm a LLC, 20002 Washington D.C. are all treated as the same firm, etc. 6. Once we have a clean and uniform set of firm names, we can fill in missing tax ids. For observations loans where firm tax id is missing, we fill in missing observations if the bank reports a consistent tax id through any portion of the loan; for multi-bank borrowers for which one bank does not report the tax id, we use a consistent tax id reported by other banks. 7. To ensure that firm income statement and balance sheet variables are reasonable and reportedinconsistentunits,weapplyacleaningalgorithmthatsearchesforlargereporting discrepancies within and across banks over time for the same firm. We set threshold for potential misreported to be a difference in a variable either by the same bank or across differentbanksofeither103, 106, 109sincethesearemostcommonunitdifferencesreported in the data. We also note that when there is miss reporting, all variables appear to be consistently miss reported in the same units, so financial ratios are correct. InternalConsistencyofBalanceSheetInformation. WefollowGopinath,Kalemli-Ozcan, Karabarbounis, and Villegas-Sanchez (2017) to check the sensibility of our cleaning procedure by comparing the sum of variables belonging to some aggregate of their respective category: 1. The sum of tangible fixed assets, intangible fixed assets, and other fixed assets as a ratio of total fixed assets. 2. The sum of fixed assets and current assets as a ratio of total assets x

3. The sum of long-term debt and other non-current liabilities as a ratio of total non-current liabilities 4. The sum of cash and securities, inventory, and accounts receivable as a ratio of current assets 5. The sum of current assets and tangible assets as a ratio of total assets 6. The sum of accounts payable, short-term debt, and current maturity long-term debt as a ratio of current liabilities 7. The sum of current liabilities, long-term debt and minority interest as a ratio of total liabilities 8. The sum of total liabilities, retained earnings, and capital expenditure as a ratio of total assets. xi

C Additional tables Table OA1: Variable definitions and sources Name Description Source Policy Variables: QE AdummyequaltoonefromMarch2020toMarch2022, Own calculations indicating the quantitative easing period. QT A dummy equal to one from June 2022 onwards, indi- Own calculations cating the quantitative tightening period. Deposits and Shares: Shares The share of uninsured NBFI deposits in total deposits FR 2052a as of February 2020. Average shares A dummy equal to one for firms with more than 50% of FR 2052a, FR Y-14Q their lending relationships with exposed banks. Total uninsured deposits Thetotalamountofdepositsthatarenotcoveredbyde- FR 2052a exc. NBFI posit insurance excluding non-bank financial institution (NBFI) deposits. Total uninsured deposits The total amount of deposits that are not covered by FR 2052a deposit insurance. Total insured deposits Thetotalamountofdepositsthatarecoveredbydeposit FR 2052a insurance. Uninsured NBFI The amount of (uninsured) deposits from NBFIs that FR 2052a are not covered by deposit insurance. Insured NBFI The amount of (insured) deposits from NBFIs that are FR 2052a covered by deposit insurance. Uninsured retail The amount of (uninsured) deposits from retail cus- FR 2052a tomers that are not covered by deposit insurance. Insured retail The amount of (insured) deposits from retail customers FR 2052a that are covered by deposit insurance. Total deposits The total amount of total deposits, including both in- FR 2052a sured and uninsured deposits. Rates on insured deposits The interest rate paid on insured deposits. RateWatch Rates on uninsured de- The interest rate paid on uninsured deposits. RateWatch posits NBFI Variables: NBFI credit The total amount of credit extended to NBFIs. FR 2052a, FR Y-14Q Supervised NBFI NBFIdepositsfromsupervisedentities,includinginvest- FR 2052a mentadvisors,insurancecompanies,andbroker-dealers. Non-supervised NBFI NBFI deposits from non-supervised entities, including FR 2052a hedge funds, private equity funds, investment companies, and REITs. Loan-Level Variables: Continued on next page

Table OA1 – continued from previous page Name Description Source Total commitments The total amount committed across all credit lines FR Y-14Q and term loans, including both utilized and undrawn amounts. On-balancesheetcommit- The sum of utilized credit lines and term loans. FR Y-14Q ments Utilized & drawn credit The combined total of drawn credit lines and utilized FR Y-14Q term loans. Undrawn credit Lines Theamountofcreditlinesthathasbeencommittedbut FR Y-14Q not yet drawn. Utilized credit lines The amount drawn and used from the available credit FR Y-14Q line. Term loans The amount of term loans. FR Y-14Q Rate on credit lines The interest rate charged on utilized credit lines. FR Y-14Q Rate on term loans The interest rate charged on term loans. FR Y-14Q Bank Characteristics: Bank size The logarithm of total bank assets. FR Y-9C GSIBS A dummy equal to one for Global Systemically Impor- Own calculations tant Banks. Tier 1 capital ratio The ratio of Tier 1 capital to total assets. FR Y-9C C&I loans The total amount of commercial and industrial loans. FR Y-9C Treasury & agency securi- ThetotalamountofTreasuryandagencysecuritiesheld FR Y-9C ties by the bank. xiii

Table OA2: List of banks in FR 2052a and FR Y-14 samples BankName Totalassets($bn) Totaldeposits($bn) C&I/TA CR LCR JPMORGANCHASE&CO$+ 2688 1563 0.05 0.14 1.16 BANKOFAMERCORP$+ 2434 1435 0.10 0.13 1.16 CITIGROUP$+ 1951 1071 0.04 0.13 1.15 WELLSFARGO&CO$+ 1928 1323 0.09 0.13 1.20 GOLDMANSACHSGROUPTHE$+ 993 190 0.02 0.15 1.27 MORGANSTANLEY$+ 895 190 0.02 0.19 1.34 USBC 495 362 0.16 0.11 1.07 PNCFNCLSVCGROUP 410 289 0.22 0.11 1.07 TDGRPUSHOLDSLLC 409 285 0.08 0.16 1.06 CAPITALONEFC 390 263 0.10 0.14 1.41 BANKOFNYMELLONCORP$+ 382 260 0.00 0.14 1.20 HSBCNAMERHOLDS 249 116 0.11 0.14 1.14 STATESTREETCORP$+ 246 182 0.01 0.15 1.10 ALLYFNCL 181 121 0.22 0.11 1.24 BMOFNCLCORP 173 104 0.23 0.12 1.49 MUFGAMERSHOLDSCORP 171 96 0.10 0.14 1.52∗ FIFTHTHIRDBC 169 127 0.27 0.11 1.15 CITIZENSFNCLGRP 166 126 0.23 0.11 1.15∗ SANTANDERHOLDSUSA 149 67 0.12 0.16 1.44∗ KEYCORP 146 112 0.27 0.11 1.45 RBCUSGRPHOLDSLLC 140 53 0.06 0.17 1.28 UBS 139 56 0.04 0.28 1.34∗ NORTHERNTRCORP 137 109 0.03 0.14 1.10 REGIONSFC 127 98 0.19 0.11 1.10 BNPPARIBAS 125 67 0.11 0.16 1.25∗ M&TBKCORP 120 95 0.16 0.11 1.21 DEUTSCHEBANK 109 19 0.02 0.38 1.75 HUNTINGTONBSHRS 109 82 0.21 0.11 1.49 BBVAUSABSHRS 94 75 0.18 0.13 1.28∗ Note: Thetableliststhebanksinourfinalsample,whichreportbothFR2052aandFRY-14data. $+indicates daily FR2052a filers. Total assets and total deposits are in $ billion in 2019Q4. C&I/TA is the share of C&I loans in total assets in 2019Q4. CR and LCR are the Tier-1 capital ratio and the Liquidity Coverage Ratio in 2019Q3 or 2019Q4. Sources for balance sheet data: FR Y-9C and public disclosures. (*) indicates the global LCR. xiv

Table OA3: NBFI uninsured deposits: Supervised and non-supervised NBFIs 1 2 3 4 Dependent variable: Log(uninsured NBFI deposits) Group Supervised NBFI Non-supervised NBFI QE * Shares 2.615*** 2.735*** 0.064 0.061 (3.556) (4.106) (1.045) (0.977) Bank control Y Y Month FE Y Y Y Y Bank FE Y Y Y Y Observations 1,736 1,734 1,625 1,623 R-squared 0.877 0.877 0.952 0.952 Note: The table reports coefficients and t-statistics (in parentheses) for the following regression equation: log(Un. NBFI ) = λ·(QE · i,t t Shares )+β·Bank Size +a +a +ε , where log(Un. NBFI ) is i i,t i t i,t i,t the logarithm of uninsured NBFI deposits held by bank i in month t. QE is a dummy equal to one from March 2020 to March 2022, and t QT is a dummy equal to one from June 2022 onwards. Shares indit i cates the share of uninsured NBFI deposits in total deposits for bank i as of February 2020. The Bank control indicates whether we control for time-varying bank size (logarithm of total assets). The terms a i and a represent bank and month fixed effects, respectively. Columns t (1)-(2) examine uninsured deposits from Supervised Non-Bank Financial Entities, which include regulated institutions such as investment advisors, brokers/dealers, and insurance companies. Columns (3)-(4) analyze uninsured deposits from Non-Supervized Non-Bank Financial Entities,comprisinginstitutionsregisteredwiththeSECundertheInvestmentCompanyActof1940,aswellashedgefundsandprivateequityfunds. Observationsaremonthly,exceptfortotalassets,whichare reported quarterly. Standard errors are clustered at the month level. The symbols ***, **, and * denote statistical significance at the 1%, 5%, and 10% levels, respectively. xv

Table OA4: Demandable NBFI uninsured deposits 1 2 3 4 5 6 Dependent variable: Log(Demandable uninsured NBFI deposits) QE * Shares 0.423*** 0.410*** 0.350*** 0.435**** 0.344*** (4.153) (4.063) (2.994) (3.824) (2.971) QT * Shares -0.351** (-2.578) Bank size -0.207 -0.227 -0.195 -0.207 -0.142 (-1.283) (-1.411) (-1.227) (-1.285) (-0.516) QE * GSIBS -0.055 (-1.158) QE (SLR rel.)* Shares 0.444*** (4.302) QE (SLR act.)* Shares 0.377*** (3.618) NBFI credit 0.117*** (3.031) Month FE Y Y Y Y Y Y Bank FE Y Y Y Y Y Y Observations 2,028 2,026 2,026 2,026 2,026 2,015 R-squared 0.907 0.906 0.906 0.906 0.906 0.907 Note: The table reports coefficients and t-statistics (in parentheses) for the following regression equation: log(Un. NBFI ) = λ·(QE ·Shares )+β ·Controls +a +a +ε , where i,t t i i,t i t i,t log(Un. NBFI ) is the logarithm of demandable uninsured NBFI deposits held by bank i in i,t month t. QE is a dummy equal to one from March 2020 to March 2022, and QT is a dummy t t equal to one from June 2022 onwards. Shares indicates the share of uninsured NBFI deposits i in total deposits for bank i as of February 2020. Bank Size is the logarithm of total assets. i,t GSIBS isadummyequaltooneforGlobalSystemicallyImportantBanks. QE (SLR rel.) refers to the SLR relaxation and exclusion of securities and reserves from SLR calculations between April 1, 2020, and April 1, 2021, while QE (SLR act.) marks the re-activation of SLR criteria. NBFI Credit is the logarithm of total outstanding credit, including credit lines and term loans, that NBFIs received. The terms a and a represent bank and month fixed effects, respectively. i t Observationsaremonthly, exceptfortotalassets, whicharereportedquarterly. Standarderrors are clustered at the month level. The symbols ***, **, and * denote statistical significance at the 1%, 5%, and 10% levels, respectively. xvi

Table OA5: Fiscal transfers during Covid-19 1 2 3 4 Dependent variable: Log(Total retail insured deposits) Log(Corporate insured deposits) QE * Shares 0.158 0.157 -0.086 -0.093 (1.460) (1.185) (-0.799) (-0.895) Bank control Y Y QT control Y Y Month FE Y Y Y Y Bank FE Y Y Y Y Observations 2,339 2,335 2,212 2,208 R-squared 0.974 0.978 0.958 0.967 Note: Thetablereportscoefficientsandt-statistics(inparentheses)forthefollowingregression equation: log(Y )=λ·(QE ·Shares )+β·Controls +a +a +ε ,whereY )isthedependent i,t t i i,t i t i,t i,t variable, denoting either total retail insured deposits (Columns 1–2) or corporate insured deposits (Columns 3–4) for bank i in month t. QE is a dummy equal to one from March 2020 to t March 2022. Shares indicates the share of uninsured NBFI deposits in total deposits for bank i i as of February 2020. Bank Size is the logarithm of total assets. The Bank control indicates i,t whether we control for time-varying bank size (logarithm of total assets), and the QT control indicates whether the interaction term QT ·Shares is included. The terms a and a represent t i i t bank and month fixed effects, respectively. Observations are monthly, except for total assets, which are reported quarterly. Standard errors are clustered at the month level. The symbols ***, **, and * denote statistical significance at the 1%, 5%, and 10% levels, respectively. xvii

Table OA6: Other deposit categories & Liquidity Coverage Ratios 1 2 3 4 Dependentvariable: Log(Totaldeposits) Log(Totaluninsured Log(Totaluninsured Log(Totalinsured deposits) depositsexc. NBFI) deposits) QE*Shares -0.042 -0.237*** -0.384*** 1.599*** (-1.175) (-6.093) (-8.268) (14.072) QE*LCR 0.001*** 0.002*** 0.002*** -0.012*** (2.856) (6.648) (6.123) (-6.466) Bankcontrol Y Y Y Y QTcontrol Y Y Y Y QT·LCRcontrol Y Y Y Y MonthFE Y Y Y Y BankFE Y Y Y Y Observations 2,145 2,145 2,145 2,000 R-squared 0.988 0.982 0.980 0.959 Note: The table reports coefficients and t-statistics (in parentheses) for the following regression equation: log(Y )=λ·(QE ·Shares)+µ·(QE ·LCR)+β·Controls +a +a +ε ,whereY isthedependentvariable i,t t i t i i,t i t i,t i,t labeled in each column for bank i in month t. QE is a dummy equal to one from March 2020 to March 2022. t Shares indicatestheshareofuninsuredNBFIdepositsintotaldepositsforbankiasofFebruary2020. LCR is i i theliquiditycoverageratioofbankiin2019Q4. TheBankcontrol indicateswhetherwecontrolfortime-varying banksize(logarithmoftotalassets),andtheQTcontrol indicateswhethertheinteractiontermQT ·Shares and t i QT ·LCR isincluded. Thetermsa anda representbankandmonthfixedeffects,respectively. Observations t i i t are monthly, except for total assets, which are reported quarterly. Standard errors are clustered at the month level. VariabledefinitionsanddatasourcesareprovidedinAppendixC.Thesymbols***,**,and*denotestatisticalsignificanceatthe1%,5%,and10%levels,respectively. Table OA7: Net Charge-Offs categories and NBFI Exposure: Pre-QE period 1 2 3 4 Dependent variable: C&I Loans Land Loans Consumer Loans Family Residential Shares −5.31∗∗∗ −0.13 −2.73∗∗ −0.02 (1.21) (1.02) (0.12) (0.35) Time FE Yes Yes Yes Yes Observations 350 350 350 350 R-squared 0.131 0.042 0.019 0.038 Note: This table reports coefficients from cross-sectional regressions of net charge-offs on bank’s shares. All dependent variables are net charge-offs (NCOs) normalized by total assets. The sample includes data from the pre-QE period. Time fixed effects are included in all specifications. Robust standard errors are in parentheses. The symbols ***, **, and * denote statistical significance at the 1%, 5%, and 10% levels, respectively. xviii

Table OA8: Credit lines, Term loans, and Total Loan Commitments 1 2 3 4 5 6 Dependent variable: Log(Credit lines) Log(Term loans) Log(Total loan commitments) QE*Shares -0.120*** -0.133*** -0.007 0.038 -0.134*** -0.164*** (-2.888) (-3.354) (-0.073) (0.416) (-3.192) (-3.895) QT*Shares -0.260*** -0.270*** -0.090 -0.087 -0.305*** -0.359*** (-3.891) (-4.633) (-0.570) (-0.579) (-4.221) (-5.186) Bank*Firm FE Y Y Y Y Y Y ILST FE Y Y Y Firm*Time FE Y Y Y Observations 655,814 328,905 243,258 95,469 952,707 404,116 R-squared 0.966 0.942 0.953 0.919 0.962 0.935 Note: Thetablereportscoefficientsandt-statistics(inparentheses)forthefollowingregressionequation: log(Y ) = λ (QE ·Shares )+λ (QT ·Shares )+βX +a +a +ε . The dependent i,f,t 1 t i 2 t i i,t i f,t i,f,t variableforeachbanki,firmf,andtimeperiodtisthelogarithmofcreditlines(columns1to2),the logarithm of term loans (columns 3 to 4), and the logarithm of total commitment, which is the sum of the two (columns 5-6). QE is a dummy equal to one from March 2020 to March 2022, and QT is t t a dummy equal to one from June 2022 onwards. Shares indicates the share of uninsured NBFI dei positsintotaldepositsforbankiasofFebruary2020. Inallspecifications,weincludedifferentlevels of fixed effects, as noted in the lower part of the table. Observations are at the bank-firm-time level and are reported quarterly. Standard errors are clustered at the bank-quarter and firm levels. The symbols ***, **, and * denote statistical significance at the 1%, 5%, and 10% levels, respectively. xix

Table OA9: Utilized & Undrawn Credit Lines, and Total On-balance sheet Loan Commitments 1 2 3 4 5 6 Dependentvariable: Log(Utilizedcreditlines) Log(Undrawncreditlines) Log(On-balancesheetcommit.) QE*Shares -0.006 -0.115 -0.191*** -0.142*** -0.090 -0.120 (-0.057) (-1.178) (-3.472) (-3.337) (-1.033) (-1.452) QT*Shares -0.055 -0.119 -0.360*** -0.295*** -0.122 -0.167 (-0.322) (-0.819) (-4.384) (-4.807) (-0.883) (-1.218) Bank*FirmFE Y Y Y Y Y Y ILSTFE Y Y Y Firm*TimeFE Y Y Y Observations 425,895 192,968 569,704 310,230 736,103 296,247 R-squared 0.859 0.870 0.897 0.941 0.869 0.867 Note: The table reports coefficients and t-statistics (in parentheses) for the following regression equation: log(Y ) = λ (QE ·Shares)+λ (QT ·Shares)+βX +a +a +ε . The dependent variable for each i,f,t 1 t i 2 t i i,t i f,t i,f,t bank i, firm f, and time period t is the logarithm of utilized credit lines (columns 1 to 2), the logarithm of undrawncreditlines(columns3to4),andthelogarithmofon-balancesheetloancommitments,whichisthesumof thetwo(columns5-6). QE isadummyequaltoonefromMarch2020toMarch2022,andQT isadummyequal t t toonefromJune2022onwards. Shares indicatestheshareofuninsuredNBFIdepositsintotaldepositsforbank i iasofFebruary2020. Inallspecifications,weincludedifferentlevelsoffixedeffects,asnotedinthelowerpartof thetable. Observationsareatthebank-firm-timelevelandarereportedquarterly. Standarderrorsareclustered atthebank-quarterandfirmlevels. Thesymbols***,**,and*denotestatisticalsignificanceatthe1%,5%,and 10%levels,respectively. xx

Table OA10: Credit lines: Interest rates & new issuance 1 2 3 4 5 6 7 8 Dependentvariable: Log(Interestrateoncreditlines) Log(Newlyissuedcreditlines) QE*Shares 0.003 0.001 0.003 0.002 -0.197 -0.294** -0.143 -0.234 (0.787) (0.354) (0.645) (0.474) (-1.362) (-2.417) (-0.796) (-1.561) QT*Shares 0.021*** 0.021*** 0.025*** 0.026*** -0.119 -0.117 -0.114 -0.103 (3.067) (3.091) (3.275) (3.645) (-0.811) (-0.990) (-0.706) (-0.804) Banksize 0.002 0.004 -0.163 -0.228 (0.618) (1.240) (-0.636) (-1.126) Bankreserves 0.000 0.000 0.029 0.022 (0.519) (0.098) (1.059) (1.050) Banktreasuries&agencies 0.001* 0.002** 0.005 0.013 (1.888) (2.254) (0.096) (0.304) Bankinsureddeposits 0.000 -0.000 0.077 0.093 (0.172) (-0.122) (0.807) (1.159) Bankuninsureddeposits 0.009*** 0.009*** -0.049 -0.028 (4.340) (4.206) (-0.410) (-0.309) Bank*FirmFE Y Y Y Y Y Y Y Y ILSTFE Y Y Y Y Firm*TimeFE Y Y Y Y Observations 617,829 308,407 595,106 297,643 9,991 9,058 9,825 8,899 R-squared 0.775 0.798 0.777 0.799 0.839 0.780 0.843 0.783 Note: Thetablereportscoefficientsandt-statistics(inparentheses)forthefollowingregressionequation: log(Y )= i,f,t λ (QE ·Shares)+λ (QT ·Shares)+βX +a +a +ε . Thedependentvariableforeachbanki,firmf,and 1 t i 2 t i i,t i f,t i,f,t timeperiodtisthelogarithmofratesoncreditlines(columns1and4)andthenewcreditlines(columns5to8). QE t isadummyequaltoonefromMarch2020toMarch2022,andQT isadummyequaltoonefromJune2022onwards. t Shares indicatestheshareofuninsuredNBFIdepositsintotaldepositsforbankiasofFebruary2020. Banksize is i i,t thelogarithmoftotalassets,andBankreservesisthelogarithmofreserves. ThevariablesBankinsureddepositsand Bankuninsureddepositsrepresentthelogarithmofinsuredanduninsureddeposits,respectively. Inallspecifications, weincludedifferentlevelsoffixedeffects,asnotedinthelowerpartofthetable. Observationsareatthebank-firmtimelevelandarereportedquarterly. Standarderrorsareclusteredatthebank-quarterandfirmlevels. Thesymbols ***,**,and*denotestatisticalsignificanceatthe1%,5%,and10%levels,respectively. xxi

Table OA11: Undrawn credit lines for liquidity constrained firms 1 2 Dependentvariable: Log(Undrawncreditlines) QE*Shares -0.150*** -0.181*** (-3.148) (-3.619) QE*Shares*Liquidity 0.060 0.026 (0.688) (0.283) QE*Shares*Covid -0.507 -0.419 (-1.057) (-0.814) QE*Shares*Covid*Liquidity -1.616*** -1.663*** (-2.766) (-2.731) QT*Shares -0.320*** -0.349*** (-4.651) (-5.065) QT*Shares*Liquidity 0.092 0.059 (0.766) (0.485) QT*Shares*Covid -0.029 0.040 (-0.064) (0.084) QT*Shares*Covid*Liquidity 1.202 1.413 (0.633) (0.658) BankSize 0.070** (2.083) Bankreserves -0.005 (-0.976) Banktreasuries&agencies -0.021** (-2.278) Bankinsureddeposits -0.029 (-1.542) Bankuninsureddeposits 0.023 (0.933) Bank*FirmFE Y Y Firm*TimeFE Y Y Observations 293,416 284,562 R-squared 0.941 0.941 Note: The table reports coefficients and t-statistics (in parentheses) for the following regression equation: log(Y )=λ (QE ·Shares)+µ (QE ·Shares·Liquidity )+ i,f,t 1 t i 1 t i f ζ (QE · Shares · Covid ) + ψ (QE · Shares · Covid · 1 t i f 1 t i f Liquidity )+λ (QT ·Shares)+µ (QT ·Shares ·Liquidity )+ f 2 t i 2 t i f ζ (QT · Shares · Covid ) + ψ (QT · Shares · Covid · 2 t i f 2 t i f Liquidity )+βX +a +a +ε . Thedependentvariable f i,t i f,t i,f,t for each bank i, firm f, and time period t is the logarithm ofundrawncreditlines. QE isadummyequaltoonefrom t March2020toMarch2022,andQT isadummyequaltoone t fromJune2022onward. Shares indicatestheshareofunini suredNBFIdepositsintotaldepositsforbankiasofFebruary2020. Covid isadummyindicatingthatfirmf operates f inanindustryheavilyimpactedbytheCOVID-19pandemic. WefollowingNAICSindustriesaredefinedtobeheavilyimpactedbythepandemic: 721110–Hotels(exceptCasinoHotels) and Motels; 722511–Full-service restaurants; 722513– Limited-Service Restaurants; 722514–Cafeterias, Grill Buffets,andBuffets;and722515–SnackandNonalcoholicBeverageBars. Liquidity isadummythattakesthevalueofoneif f theratioofsalestoaccountsreceivableforfirmfat2019Q4is higherthanthemedianforallfirmsat2019Q4. Banksize i,t isthelogarithmoftotalassets,andBankreservesisthelogarithmofreserves. ThevariablesBankinsureddepositsand Bankuninsureddeposits represent the logarithm of insured anduninsureddeposits,respectively. Inallspecifications,we includedifferentlevelsoffixedeffects,asnotedinthelower part of the table. Observations are at the bank-firm-time level and are reported quarterly. Standard errors are clusteredatthebank-quarterandfirmlevels. Thesymbols***, **, and*denotestatisticalsignificanceatthe1%, 5%, and 10%levels,respectively.

Table OA12: Aggregate effects at firm level: Alternative exposure measures 1 2 3 4 5 Dependentvariable: Log(Utilized Log(Undrawn Log(Termloans) Log(Total Log(Investment) Dependentvariable: creditlines) creditlines) commitments) QE*WeightedSharesExposure 0.217 -0.346* 0.124 -0.051 -2.629*** (1.266) (-1.902) (0.866) (-0.763) (-2.931) QT*WeightedSharesExposure 0.165 -0.596** 0.133 -0.220* -1.792 (0.579) (-2.656) (0.570) (-2.022) (-1.554) Observations 223,976 256,001 122,718 497,200 43,199 R-squared 0.820 0.798 0.929 0.951 0.817 QE*RelationshipsDummy -0.014 -0.071*** 0.001 -0.014* -0.354*** (-0.515) (-2.926) (0.072) (-1.800) (-4.130) QT*RelationshipsDummy -0.105** -0.095*** 0.036 -0.029** -0.454*** (-2.729) (-3.507) (1.428) (-2.301) (-3.573) Observations 223,976 256,001 122,718 497,200 43,199 R-squared 0.820 0.798 0.929 0.951 0.817 ILSTFE Y Y Y Y Y FirmFE Y Y Y Y Y Note: The table reports coefficients and t-statistics (in parentheses) for the following regression equation: log(Y ) = f,t λ (QE ·Exposure )+λ (QT ·Exposure )+a +a +ε ,wherethedependentvariableforeachfirmf intimeperiod 1 t f 2 t f ILST f f,t tisthesumofcreditthefirmreceived. Thedependentvariablesare: thelogarithmofutilizedcreditlines(column1),the logarithmofundrawncreditlines(column2),thelogarithmoftermloans(column3),thelogarithmoftotalcommitments (column 4) and the logarithm of investments (column 5). Exposure is a measure of how exposed a firm is to the QEf inducedfragilityviatheloanrelationshipsthatfirmhaswithexposedbanks. Wereportresultsfortwomeasuresofexposure,Exposure ={WeightedSharesExposure ,RelationshipsDummy },allcomputedusinginformationat2019Q4. f f f WeightedSharesExposure istheweightedaverageshareS amongthosebankswithwhichafirmf hasloanrelationf i shipsat2019Q4,wheretheweightsaregivenbytheloancommitmentsfirmf haswithbankiotherthetotalfirm-f commitmentswithallbanks. RelationshipsDummy isadummyequaltooneforfirmswithAverageRelationships >0.5, F f i.e,withmorethan50%oftheirlendingrelationshipsarewithmoreexposedbanks. QE isadummyequaltoonefrom t March2020toMarch2022,andQT isadummyequaltoonefromJune2022onwards. Theregressionincludesindustryt location-size-time(a )andfirm(a )fixedeffects. Standarderrorsareclusteredatthefirmlevel. Variabledefinitions ILST f anddatasourcesareprovidedinAppendixC.Thesymbols***, **, and*denotestatisticalsignificanceatthe1%, 5%, and10%levels,respectively.