Securities Lending as Wholesale Funding: Evidence from the U.S. Life Insurance Industry

Finance and Economics Discussion Series Divisions of Research & Statistics and Monetary Affairs Federal Reserve Board, Washington, D.C. Securities Lending as Wholesale Funding: Evidence from the U.S. Life Insurance Industry Nathan Foley-Fisher, Borghan Narajabad, and Stephane Verani 2016-050 Please cite this paper as: Foley-Fisher, Nathan, Borghan Narajabad, and Stephane Verani (2016). “Securities Lending as Wholesale Funding: Evidence from the U.S. Life Insurance Industry,” Finance and Economics Discussion Series 2016-050. Washington: Board of Governors of the Federal Reserve System, http://dx.doi.org/10.17016/FEDS.2016.050. 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.

Securities Lending as Wholesale Funding: Evidence from the U.S. Life Insurance Industry∗ Nathan Foley-Fisher Borghan Narajabad Stéphane Verani † May 2016 Abstract The existing literature implicitly or explicitly assumes that securities lenders primarily respond to demand from borrowers and reinvest their cash collateral through short-term markets. Using a new dataset that matches every U.S. life insurer’s bond portfolio, as well astheirlendingandreinvestmentdecisions,totheuniverseofsecuritieslendingtransactions, weoffercompellingevidenceforanalternativestrategy,inwhichsecuritieslendingprograms are used to finance a portfolio of long-dated assets. We discuss how the liquidity and maturitymismatchassociatedwithusingsecuritieslendingasasourceofwholesalefunding could potentially impair the functioning of the securities market. JEL Codes: G11, G22, G23 Keywords: securities lending, wholesale funding, life insurers, market liquidity ∗Forprovidingvaluablecomments,wewouldliketothank,withoutimplication,TobiasAdrian;JackBao;Celso Brunetti; Stefan Gissler; Michael Gordy; Diana Hancock; Yesol Huh; Sebastian Infante; Anastasia Kartasheva; FrankKeane;BethKiser;MichaelPalumbo;AndreasUthemann;andseminarparticipantsintheFederalReserve Bank of Cleveland and Philadelphia, LSE Systemic Risk Centre, Bank for International Settlements, Swiss NationalBank,SwissGraduateInstitute,andFederalReserveBoard. AllauthorsareintheDivisionofResearch andStatisticsoftheBoardofGovernorsoftheFederalReserveSystem. WearegratefultoDellaCummingsand Melissa O’Brien for exceptional research assistance. The views in this paper are solely the authors’ and should not be interpreted as reflecting the views of the Board of Governors of the Federal Reserve System or of any other person associated with the Federal Reserve System. †nathan.c.foley-fisher@frb.gov(correspondingauthor),202-452-2350,20th &CStreet,NW,Washington,D.C. 20551; borghan.narajabad@frb.gov; stephane.h.verani@frb.gov.

Introduction Securities lending is widely understood to play an important role in the functioning of securities markets. By agreeing to purchase an asset and return it at a later date, securities borrowers— typically, dealer firms acting for themselves or on behalf of clients such as hedge funds— temporarily gain economic ownership of the asset in exchange for collateral in the form of either cash or another security. In addition to facilitating short positions (Duffie 1996, Keane 2013), financial institutions borrow securities to avoid delivery fails (Musto, Nini & Schwarz 2011) and use the borrowed security as collateral in other transactions (Dive, Hodge, Jones & Purchase 2011). In the absence of securities lending, a large volume of securities would be tied up in institutions that naturally hold big portfolios of assets—pension funds, mutual funds, and life insurance companies—for asset-liability management or regulatory reasons. In this paper, we study the securities lending market from the perspective of the lenders. To understand the decision processes that we analyze, it is helpful to consider the transactions stylistically represented in Figure 1.1 The cloud represents the general functioning of securities markets, illustrated with the example of hedge funds taking long and short positions. Securities market participants typically borrow both cash funding and securities using broker-dealers as intermediaries. Broker-dealers obtain cash–for example, through short-term funding market– from money market mutual funds (MMFs) and securities lenders. Securities lenders decide whether to lend assets from their portfolio and whether to invest the cash collateral they receive back into short-term markets or into long-term markets. In the latter case, they may invest, for example, in long-term corporate bonds or asset-backed securities. The existing literature implicitly or explicitly assumes that securities lenders’ strategy is primarily to respond to demand from borrowers and reinvest their cash collateral through short-term markets.2 We label this lending strategy “passive.” Duffie, Gârleanu & Pedersen (2002) study the effect of search and bargaining in the securities lending market on pricing in the securities market, abstracting from the reinvestment decisions of securities lenders (Figure 2a). Brunnermeier & Pedersen (2009) and Gorton & Metrick (2012) consider securities markettransactionsfundedthroughmarginaccountsandbilateralrepurchaseagreements(repo), abstractingfromthesourceofsecurities(Figure2b). Krishnamurthy,Nagel&Orlov(2014)focus on the cash provided to broker-dealers from MMFs and securities lenders through short-term funding markets, taking as given the lending and reinvestment decisions of securities lenders (Figure 2c). 1 Althoughagentlendersareofteninvolvedinthesecuritieslendingprocess,aswedescribeinSection1,their role is incidental to our analysis. 2 Lenders may reinvest in short-term markets directly through tri-party repo or indirectly through MMFs. 2

Using a new dataset that matches every U.S. life insurer’s bond portfolio as well as their lending and reinvestment decisions, to the universe of securities lending transactions, we offer compelling evidence for an alternative “active” strategy in which securities lending programs are akin to using securities lending as wholesale funding. By raising and maintaining a level of cash collateral, lenders finance a portfolio of longer-duration, higher-yielding assets. The maturity and liquidity transformations associated with this active strategy allow the lender to increase the return on its portfolio of assets. We describe in Section 1 how the degree of passive/active management is likely to be a function of the lender’s business model, regulatory environment, and investment opportunities. Using securities lending as a source of wholesale funding suggests important new channels through which securities market functioning could be impaired. As is widely known, liquidity and maturity mismatches are associated with vulnerabilities to runs (Diamond & Dybvig 1983, Goldstein & Pauzner 2005) and roll-over risk (He & Xiong 2012).3 However, it is important to fully understand the mechanisms through which those vulnerabilities may impact the broader financial system when, for example, designing financial regulatory policy. To be sure, securities lendingmayplayanimportantandefficientrolewithinthefinancialsystem. But,asthe2008-09 financialcrisisamplydemonstrated,itbehoovesregulatorstobefullyawareofanyvulnerabilities associated with financial market activity. Thecollapseduringthe2008-09financialcrisisoftri-partyrepofundingprovidedbysecurities lendersillustrateshowsecuritiesmarketfunctioningispotentiallyassociatedwithanactivecash reinvestment strategy. Figure 3a shows the total cash collateral held by securities lenders (red dashed line) compared to the total assets held by MMFs (blue solid line). Amid widespread concerns about the quality of cash reinvestment portfolios, securities borrowers ran on lenders’ cash collateral.4 By the first quarter of 2009, cash collateral had fallen almost $1 trillion while MMF assets had only begun to decline from pre-crisis levels. Contagion to the broader financial systemoccurredwhen,tomeetthedemandtoreturncash,securitieslendersdrewontheirliquid short-term investments, that is, the portion of cash collateral that was reinvested by securities lenders in short-term funding markets. The effect of the run on securities lenders’ cash collateral onmarketfundingliquiditycanbeseeninFigure3b,showingthetri-partyrepofundingprovided 3 Ataninstitutionallevel, thesevulnerabilitiesweremanifestatAIG’s$80billionsecuritieslendingprogram in 2008, which had retained only about 20 percent of its cash collateral in liquid assets, while 65 percent was reinvested in long-term RMBS and other ABS (Peirce 2014, McDonald & Paulson 2015). Increasing concerns among investors about the value of this reinvestment portfolio drove demands for greater collateral reductions. The cumulative and consequential losses ultimately required AIG to request a series of government bailouts. 4Theexperienceofsecuritieslenderswasrepeatedthroughoutthefinancialsystem,withrunsonrepomarkets (Gorton&Metrick2010a,b,2012),asset-backedcommercialpaper(Covitz,Liang&Suarez2013,Schroth,Suarez & Taylor 2014), MMFs (Schmidt, Timmerman & Wermers forthcoming), and life insurance companies (Foley- Fisher, Narajabad & Verani 2015). 3

by securities lenders (red dashed line) and MMFs (blue solid line). By the first quarter of 2009, repo funding from securities lenders had collapsed by almost $300 billion while MMF funding remained relatively more available. To make progress toward identifying the active channel of securities lending, we collect new regulatorydatabeginningin2010ontheuniverseofU.S.lifeinsurers’securitieslendingprograms at the security level. These regulatory data specify which bonds in an insurer’s portfolio are on loan at the time of filing as well as the composition of the insurer’s cash reinvestment portfolio at the security level. We combine this information with microdata on loan transactions from the most comprehensive source on the securities lending market, which includes information on the amount of individual bonds borrowed and available for loan on any given day as well as the terms of these transactions. By matching these data at the security level, we obtain a detailed view of the securities that life insurers lend, given market conditions. Moreover, because life insurers typically pursue a “buy and hold strategy” on their asset holdings over several years, this dataset permits controlling for individual bond characteristics as well as insurer-specific and time-specific variation. The final dataset provides an unprecedented opportunity to study carefully the lending decisions of an important set of securities lenders. We begin by analyzing the decision to loan bonds as a function of observable security, market, and insurer characteristics. We find a statistically and economically significant positive association between an insurer’s decision to lend a particular security and its market share in that security, which is robust to controls for bond, insurer, and market characteristics. This findingisconsistentwiththeexistingempiricalliteraturethatstudiesthepassivelendingchannel (Kolasinski, Reed & Ringgenberg 2013, Kaplan, Moskowitz & Sensoy 2013) and those theories emphasizing search frictions (Duffie et al. 2002, Vayanos & Weill 2008), which imply that the probability of loaning a bond is higher when a potential borrower is matched with an insurer with a high market share in that bond. However, the finding is also consistent with the active lending channel. In such a case, a life insurer would be more likely to lend bonds for which it has a high market share, as it facilitates financing a portfolio of long-dated assets, and earning a spread between the cost of cash borrowing and the return on these assets. This initial result partly illustrates the severe empirical challenge in identifying the active lending channel through the lending decisions of a single lender. The complication is the need to control for all the factors affecting the borrowing decision that may be influenced by the lenders themselves. For example, life insurers may be able to control their cost of cash borrowing by choosing to lend securities that borrowers cannot easily find. Supporting this view, we show that the cost of borrowing a security is a function of the concentration of holdings 4

in that security. Intuitively, lenders of hard-to-find securities can get a better deal on their transaction—as in Duffie (1996)—and the term of the transaction may affect their decision to lendparticularsecurities. Whenspecifyingthelendingdecision,omittingtocontrolfortheeffect of any variable that is correlated with the borrowing decision will invalidate coefficient estimates because the variable will likely be correlated with equilibrium market variables. While we can and do control for the obvious candidates in our basic analysis, the list of possible confounding factors is practically endless. Our strategy for identifying the active channel of securities lending is to exploit the ability to observe the same securities at the same time across different life insurers’ portfolios. Our specification includes security–time fixed effects to control for potentially confounding factors in a reduced form that encompasses the cost of cash borrowing as well as the availability and distribution of holdings associated with each security. Our proxy for the extent of active management is the fraction of assets in an insurer’s cash reinvestment portfolio that have a residual maturity of more than one year. We use this proxy as the main explanatory variable in a specification that includes security–time and insurer fixed effects. Under a passive lending strategy, the variable would be uncorrelated with the decision to lend. By contrast, in an active lending strategy, the proxy would be associated with a greater likelihood that a security will be loaned. Thisapproachapproximatesatheoreticalidealexperimenttoestablishanactivelending channel: testing whether two life insurers with identical bond portfolios would lend differently if one engaged in greater maturity mismatch. We then develop our analysis beyond the baseline reduced form to address a further concern: Our proxy for active management could be correlated with unobservable time-varying insurer heterogeneity that is potentially determined by the decision to lend securities. For example, life insurers may have bundles of securities that facilitate lending in some years. We adopt an instrumental variable (IV) approach. Our instrument for the degree of active cash reinvestment by an insurer is the annual change in unrealized gains/losses as a fraction of that insurer’s total assets. Intuitively, an active securities lender will attempt to compensate for unrealized losses on its underlying portfolio by increasing the return on its cash reinvestment portfolio. The unrealized losses are themselves plausibly beyond the control of the manager of an insurer’s securities lending program, and, therefore, unrelated to the unobserved heterogeneity that is the source of the endogeneity. Our findings from the reduced form and the IV analyses consistently reject the hypothesis thatsecuritieslenders’reinvestmentstrategiesarepassive. OurbaselineIVspecificationsuggests that a one standard deviation increase in the fraction of the cash reinvestment portfolio that has 5

a residual maturity of more than one year, on average, increases the likelihood that a lender will lend a security by 0.7 standard deviation (11 percentage points). Importantly, this finding is based on data available since 2010, indicating that the active management of securities lending programs was not eliminated by regulatory responses to the 2008-09 financial crisis. The remainder of the paper proceeds as follows. In Section 1, we provide an overview of the market for lending securities. Section 2 presents our data and summary statistics. Sections 3, 4, and 5 present our main empirical results, including our IV estimates and robustness tests. We conclude in Section 6 with some remarks on broader implications and further study. 1 Securities lending and life insurance companies Inthissection,webrieflyoutlinethetypicalstructureofasecuritieslendingtransaction,together with the motivations of each party to the deal. We then provide an overview of the securities lending market and describe the activity of insurance companies within this market. Lastly, we explain how the cash collateral reinvestment strategies of different securities lenders depend on their business model and regulatory environment. 1.1 Securities lending transactions In a prototypical loan, the lender transfers full legal and economic ownership of the security to the borrower.5 In exchange, the borrower gives the lender collateral in the form of cash or anothersecurity. Thetermoftheloanisusuallyopen-ended, witheitherpartyabletoterminate thedealatanytimebyreturningthesecurity/collateral.6 Thesecuritieslenderisfreetoreinvest the cash and, in some cases, rehypothecate the securities used as collateral. In the case of non cash collateral, the securities lender earns a fee from the borrower. In the case of cash collateral, the securities lender pays a percentage of the reinvestment income to the securities borrower, called the “rebate.” Both the rebate and fee are equilibrium prices negotiated at the outset of the deal that reflect the scarcity of the security on loan: A hard-to-find “special” security commands a high fee and a low or negative rebate. Typically, the loan is marked to market daily and is “overcollateralized,” with borrowers providing, for example, $102 in cash for every $100innotionalvalueofasecurity. Thepercentageofovercollateralizationiscalledthe“margin,” which serves to insure the securities lender against the cost of replacing the lent security if the borrower defaults. In addition to the loss of collateral, the security borrower is dissuaded from 5 Althoughtheyaretransferredbydefault,thepartiesmayagreethatthesecuritiesborrowerwillreturnany dividend/interest payments and/or voting rights. 6 Flexibility is often preserved, even in term loans, by allowing either party to break the terms early and incurring a fee. 6

defaulting on the loan by reputational effects: lender–borrower relationships are formed through repeated transactions. Overall, the structure of cash-collateralized securities lending is closely related to a sale and repo transaction, in which the securities borrower is entering a reverse-repo arrangement (Duffie 1996, Garbade 2006). A securities lending transaction usually involves three or four parties. The ultimate owner of the security is typically an institutional investor such as a pension fund, insurance company, mutual fund, or sovereign wealth fund. Owners of large portfolios will often conduct their own lending programs, while smaller owners execute their programs through agent lenders, such as custodian banks or asset managers, that act as large warehouses for securities made available for lending. The end users of the borrowed securities are typically dealers and hedge funds. These securitymarketparticipantsgenerallyuselargefinancialinstitutions,forexample,broker-dealers and investment banks, as intermediaries that regularly search for securities and have established relationships with lenders. The ultimate owners decide which securities in their portfolios will be made available to lend and how the cash collateral proceeds of their lending programs will be reinvested. When they choose to employ agent lenders, the owners typically provide guidelines or specific instructions for the type of lending transactions (for example, minimum fee criteria or hard-to-find securities only) and for the reinvestment of cash collateral. In some cases, as we will discuss below, these reinvestment strategies are subject to regulatory limits. If agent lenders are involved, they execute owners’ instructions to lend particular securities and reinvest cash collateral. Because agent lenders often receive the same securities from many ultimate owners, they typically allocate borrowing requests to securities using an algorithm that ensures no owner receives preferential treatment. The agents earn a share of the profits associated with lending securities, including fees and/or reinvestment income after rebate. In exchange, agents will customarily provide indemnification against the risk that the non-cash collateral is insufficient to replace the lent securities if the borrower defaults. To be clear, this indemnification does not protect the owner against the risk of losses associated with reinvesting cash collateral. The borrowing intermediary generally performs three functions as it matches end-user requests for securities with lenders’ availability. First, the intermediary helps to assuage securities lenders’ potential concerns about the credit quality of end users, which may be small and weakly regulated. Second, by establishing relationships with lenders and borrowers, they can lower search costs. In the case of broker-dealers, their securities lending intermediation is oftencombinedwithprimebrokeragetolowercostsfurther. Third,theintermediarymayassume 7

some liquidity risk by establishing open-ended loans with lenders, giving them the freedom to recall the securities as needed, and extending term loans to end users so they can be sure their short positions are covered. In exchange for these services, the borrowing intermediary receives a payment from the end user.7 The end users have a variety of reasons for borrowing a particular security. The most common motivations are to take a short position or to cover a naked short position (Duffie 1996, Keane 2013); to avoid a settlement/delivery failure (Musto et al. 2011), possibly as part of market making activity; to combine one security with other securities as part of an arbitrage trading strategy; to obtain collateral for use in other transactions (Dive et al. 2011); and to take advantage of tax or regulatory arbitrage (Faulkner 2006). The details of these strategies are often complex and we refer the reader to the reference list for further explanation. 1.2 The securities lending market Securities lending is a global market totaling more than one trillion U.S. dollars in outstanding contracts, with U.S. loans accounting for about half of the worldwide market. Figure 4 shows U.S. entities’ securities lending broken down by the type of lender. Retirement and pension funds account for more than 60 percent of securities lending by U.S. institutions, followed by mutual and investment funds, which together cover about 30 percent of the market. Insurance companies are the third largest group of U.S. securities lenders. Because life insurers typically invest in fixed income securities rather than equities, their lending is heavily biased toward the corporate bond market. Indeed, U.S. life insurers were the principal lenders of corporate bonds in the pre-crisis period and remain key participants in the market even though their programs shrank during the 2008-09 financial crisis (Figure 5). The majority of securities lent in the U.S. are against cash collateral. Although lending against non-cash collateral increased in the aftermath of the 2008-09 financial crisis, lending against cash collateral currently accounts for about 80 percent of the U.S. securities lending market. Importantly,theproportionoflendingagainstcashcollateralisgreaterforlifeinsurance companies than for mutual funds or pension funds. Because most of the securities lending, particularly by life insurers, is against cash collateral, reinvestment of cash collateral is an essential component of the securities lending market. When lending against cash collateral, the main source of income is the return from reinvestment. 7 Huh & Infante (2016) show how securities lending allows broker-dealers to separate their own portfolio positions from their ability to fulfill client orders. 8

1.3 Securities lenders’ cash collateral reinvestment strategies Regardlessoftheirbusinessmodelandregulatoryenvironment,afinancialinstitutioncanalways choose to pursue a passive cash collateral reinvestment strategy. In such cases, the securities lender responds to borrowing requests and reinvests the cash collateral in safe assets of short duration,forexampleingeneralcollateralrepo. Theabsenceofmaturityandliquiditymismatch allows these programs to scale up or down with minimal concern for default by either party. Alternatively, some financial institutions may choose an active cash collateral reinvestment strategy. Such lenders aim to maintain a level of cash from securities borrowers to finance a portfolio of longer-duration, higher-yielding assets, for example, corporate bonds, mortgages, and mortgage-backed securities. Conditional on the size of their lending program, securities lenders have the opportunity to increase the return from their lending programs by reinvesting the cash collateral in riskier and/or longer-term assets. Because the term of the loan is usually open-ended, this strategy creates a liquidity and/or maturity mismatch between the terms of lending and reinvestment. The degree to which a lending institution adopts an active reinvestment strategy depends on its business model and regulatory environment. For example, mutual funds have essentially no capital to absorb losses and their liabilities are mostly on demand. Their business model encourages them to reinvest their cash collateral from securities lending in liquid asset, so that they can return the cash collateral and receive back their lent securities upon demand. Moreover, to protect mutual fund investors, the Securities and Exchange Commission (SEC) subjectsmutualfunds’reinvestmentofcashcollateraltothestringentSECRule2a-7. According to this rule, the weighted average maturity of cash reinvestment must be less than 60 days.8 Enforcementofthisruleonmutualfunds,togetherwithotherregulatoryresponsestothe2008-09 financial crisis, significantly shortened the maturity of reinvestment of cash collateral (Figure 6). In contrast to mutual funds, other securities lenders traditionally have longer-term liabilities and are subject to different regulations on the use of securities lending. For example, life insurance companies and pension funds typically hold long-term contingent liabilities in the form of insurance and pension contracts that are not used by consumers as protection against illiquidity.9 Under their traditional business model, life insurers and pension funds match the terms of their liabilities to those of safe and liquid assets. This model more easily allows life insurersandpensionfundstoenhancethereturnontheirassetportfoliosbecausetheirliabilities 8 In addition, Rule 2a-7 limits exposure to any counterparty to 5 percent, and requires mutual funds to hold atleast10percentoftheircashreinvestmentindailyliquidassetsandatleast30percentinweeklyliquidassets. 9 Healthinsurancecompaniesandpropertyandcasualtylifeinsurersarefarlessengagedinsecuritieslending because the terms of their liabilities are much shorter than life insurers or pension funds. 9

arenotondemand. Inaddition,theregulatoryrulesontheuseofsecuritieslendingaregenerally weakerforlifeinsurersandpensionfunds. Thisweakerregulatoryenvironmentexistsdespitethe NAICadoptingin2010morerestrictiveguidelinesforstateregulationspertainingtolifeinsurers’ securities lending. The new guidelines stem from a review of securities lending practices at AIG that contributed to its collapse during the 2008-09 financial crisis. In particular, the guidelines specify that borrowers should post cash in the amount of at least 102 percent of domestic securities borrowed (and at least 105 percent if the securities are foreign), that individual loans should not be more than 5 percent of admitted assets, that cash reinvestment should be “prudent,” and that all cash reinvestment securities (on- and off-balance sheet) are reported in the NAIC Quarterly and Annual Statutory Filing Schedule DL.10 The differences in institutional business models and regulatory environments help determine the scope of the cash reinvestment strategies of securities lenders. An institution with less liquid liabilities and less stringent regulation on the use of securities lending may more easily enhance the return on its portfolio by allowing greater maturity mismatch between the tenor of its loans and the reinvestment of the cash collateral it receives.11 Currently, U.S. life insurers have significantly greater maturity mismatch between their securities lending and their reinvestment of cash collateral than the median securities lender (Figure 8). The mean U.S. life insurer maturity mismatch (not shown) is about 1,000 days, compared with about 100 days for the mean securities lender, shown by the blue line in Figure 6. Indeed, life insurers’ reinvestment of cash collateral currently has an average maturity more than 15 times larger than the regulatory maximum of mutual funds’ cash reinvestment. That said, to better understand the securities lending strategies of pension funds and life insurance companies requires detailed data on individual loans and cash reinvestment decisions. Inthecaseofpensionfunds, thesedataarenotavailable. Althoughsomefundsreportaggregate level information on their portfolio holdings and lending programs, crucial data are missing on their individual security loan decisions and their cash reinvestment strategies. By contrast, state insurance regulators’ adoption of the NAIC guidelines for enhanced reporting on securities lending programs presents a golden opportunity to observe new and detailed information about securities lending and cash reinvestment activities by U.S. life insurers. We can observe the individual securities that are lent by life insurers, the maturity 10 Inaddition,eachassetfinancedwithcashcollateralrecordedintheNAICQuarterlyandAnnualStatutory Filing Schedule D attracts a risk-based capital charge consistent with its NAIC designation code. http://www.dfs.ny.gov/insurance/circltr/2010/cl2010_16.htm http://www.naic.org/capital_markets_archive/110708.htm 11 Regulatory arbitrage as a source of reach for yield has already been documented in the U.S. life insurance industry through the deliberate portfolio selection of more risky corporate bonds within a rating class (Becker &Ivashina2015),theuseofcaptivereinsurerstolowerregulatorycapitalcharges(Koijen&Yogoforthcoming), and the issuance of institutional funding agreements (Foley-Fisher et al. 2015). 10

of the collateral they received, and their cash reinvestment portfolios. When combined with security-level data on the broader securities lending market, we can deepen our understanding of the strategic use of securities lending by U.S. life insurers to raise the return on their portfolio of assets. To be sure, we do not have data on the maturity structure of life insurers’ reinvestment of cash collateral before 2010 and, thus, cannot gauge how their reinvestment strategy changed following the 2008-09 financial crisis and in response to the new regulatory environment. Nevertheless, these new data offer the best possible insight into the ongoing strategiespursuedbyinstitutionswhosebusinessmodelsandregulatoryenvironmentaresimilar to life insurers’, including, for example, pension funds. In the next section, we describe how we put the data together. 2 Data We combine several data sources to obtain the dataset we use in our analysis. The data on insurance company holdings and securities lending activity come from the NAIC Quarterly and Annual Statutory Filings.12 Within these filings, Schedule DL is a relatively new report of individualinvestmentsmadebylifeinsurersusingcashcollateralreceivedfromsecuritieslending, bothon-andoff-balancesheets. TheScheduleDLwasintroducedin2010asoneofmanychanges tothereportingandstatutoryaccountingofsecuritieslendingtransactionsadoptedasaresponse to the 2008-09 financial crisis. In general, the new data allow us to better track the securities lending transactions entered into by an insurer and to observe detailed information about the lifeinsurers’useofthecollateralreceived. Forexample, from2010, ifthecollateralreceivedfrom securities lending could “be sold or pledged by custom or contract by the reporting entity or its agent,” then the reinvested collateral should be recorded on the balance sheet.13 We hand-coded data about the maturity of the collateral received in the securities lending transactions from the regulatoryNotestotheFinancialStatements. Figure7showsthatthetotalamountofsecurities lending by U.S. life insurers reached over $55 billion at the end of 2013. Because we rely on the detailed information contained in Schedule DL as part of the new reporting requirements, our sample by necessity begins in 2010. The NAIC Quarterly and Annual Statutory Filings also contain Schedule D, reporting life insurers’ individual fixed income holdings at year-end, together with cross-sectional information about each security, including the CUSIP indentifier, the NAIC designation code, and whether 12 Historical NAIC Quarterly and Annual Statutory Filings are contained in the NAIC Financial Data Repository, a centralized warehouse of financial data used primarily by state and federal regulators. 13 Amendments to SSAP No. 91-R, Accounting for Transfers and Servicing of Financial Assets and Extinguishments of Liabilities. 11

the bond was on loan as part of the insurer’s securities lending program. We drew information about the total size and performance of the life insurer’s investment portfolio from the summary balance sheet contained in the NAIC Quarterly and Annual Statutory Filings. We focus on all insurance companies that had a securities lending program at any point during our sample period. Our baseline dataset includes information on 107 life insurers, with holdings data on over half a million bonds. The first four columns of Table 1 report descriptive statistics for the baseline sample. The average bond holding is about $9 million with a standard deviation of $27.6 million. The dummy variable for securities lending indicates that about 3 percent of the bond holdings were lent at some point during the period. We obtain bond characteristics from Mergent FISD and add them to our baseline dataset. FISD provides a wide range of security-level information for mostly U.S. corporate, agency, and government bonds that can be merged with our holdings data using the CUSIP identifier. Our focus on life insurers’ holdings and lending of corporate bonds is consistent with their strong bias toward this sector of the securities lending market, as described in Section 1. Including this additional information reduces our bond holdings data to about 250,000 bond holdings across thesamesetof107lifeinsurers. Columns5through8ofTable1reporttheadditionaldescriptive statistics, including amount issued, offering yield, credit rating, and residual maturity. In this merged sample, the average offering amount of the bonds held is $10 million (with a standard deviation of $22.8 million), with a yield of about 6 percent. The average residual maturity across all year-end bond holdings is 11.2 years (with a standard deviation of 9.6 years). Our numerical rating measure indicates that the average is about 20, equivalent to a Standard & Poor’sbondratingofBBB.14 Lastly, theaveragetotalamountoutstandingacrossallbondsheld by life insurers is $900 million. Markit Securities Finance provides information on individual securities lending transactions. Thisdatasetoffersthemostcomprehensiveviewofthesecuritieslendingmarket, coveringabout 85 percent of the global market and more than 90 percent of the U.S. securities lending market. The daily transaction level data include identifiers of the lent security, such as CUSIP and ISIN; value,quantityandtenureoftheloan;therelevantrates,suchassecuritieslendingfeeandrebate rate; and the type of collateral (cash or non-cash). For each lent security, the total value and quantity of the inventory available to lend is also reported. We cannot observe counterparties to individual loans, nor information on lenders’ reinvestment of cash collateral. We construct weighted averages of the available variables, for each security, across all transactions conducted 14 We collect data on ratings from Moody’s, Fitch, and Standard & Poor’s and combine them into a single rating using the lowest rating when only two are available and the median rating when all three are available. To average the ratings, we set AAA, or equivalent = 28, AA+ = 26, AA = 25, AA- = 24 ... CCC- = 9, CC = 7, and C = 4. 12

duringthe14daysaroundyear-end. Ourfinaldatasetcontainsdatafor107lifeinsurersandover 187,000 bond holdings.15 Columns 9 through 12 of Table 1 report the descriptive statistics for the dataset following this merge. The information from the securities lending market suggests thattheweightedaveragerebateonthebondsisaboutzero. Onaverage,lifeinsurersholdabout 3 percent of each security’s total lendable amount (with a standard deviation of 16 percent), with a concentration of holdings index (HHI) value equal to 0.37. Lastly, our measure of each security’s market tightness, defined as the ratio of the total amount lent to the total amount that is lendable, indicates that, on average, about 14 percent of the available amount of each security is actually lent. The remaining entries in these columns show that the other observable characteristics of the bond holdings do not vary significantly between the baseline and merged datasets. Takentogether,thesedataofferanewlensthroughwhichwecanobtainacomprehensiveand detailed view of all aspects of the life insurers’ securities lending programs, from the individual bonds lent, through the collateral received, to the reinvestment portfolios. As a simple analysis of the kinds of bonds that life insurers are lending, Table 2 compares descriptive statistics of securities used in lending transactions compared with the securities not lent.16 In general, the securitiesthatlifeinsurerstendtolendhaveaslightlylargerparvalueandhavealongerresidual maturity in comparison with the rest of their portfolio. Life insurers also tend to lend securities with a lower rebate (higher fee) and in which there is a greater concentration of holding and market tightness. Of course, these pairwise comparisons of mean characteristics are indicative. In the following three sections, we provide a more careful empirical analysis of securities lending by U.S. life insurers. First, we explore the bond and market characteristics that determine an insurer’s decision to lend individual securities. Second, we show how fundamental endogeneity concerns preclude finding compelling evidence for active securities lending within an insurer. And third, we present our methodology for showing that active management of the portfolio plays a significant role in the overall strategy for the lending program. 3 Determinants of securities lending by U.S. life insurers Ourfirstsetofempiricaltestsinvestigatestherelationshipbetweenalifeinsurer’sdecisiontolend a particular bond and its market share in that bond. We assume that life insurers are endowed with a portfolio of securities over which they can make lending decisions. This assumption is 15 There are 1,029 out of 14,149 securities reported by life insurers as on loan, but for which no securities lending information is available in Markit; we exclude these observations from our analysis. 16 The table counts each security holding as a separate observation, effectively weighting the descriptive statistics by the number of times that a security is lent or not. 13

in keeping with industry practice, where the agents who manage the securities lending portfolio are not those who decide the broad investment strategy or specific asset purchases. We denote by Loan a binary variable that takes a value of 1 if insurer j is loaning bond i at year t, and ijt 0 otherwise. The main explanatory variable (Market share ) is the year-t holding by insurer j ijt in bond i as a share of the total amount of bond i that is made available to securities borrowers by all lenders. We specify a linear model: Loan = α1+α2+α3+βMarket share +Z γ +(cid:15) , (1) ijt i j t ijt it ijt where Z is a matrix of bond-specific time-varying characteristics. The set of α fixed effects it captures unobservable heterogeneity across securities, life insurers, and time. Table 3 summarizes the results. Column 1 shows the baseline model, controlling for insurer, year, and bond issuer fixed effects. The association between Loan and Market share ijt ijt is positive and statistically significant at less than the 1 percent level. The coefficient on Market share suggests that, on average, a one standard deviation (13.4 percent) increase ijt in an insurer’s market share in a particular bond is associated with a 0.6 percent increase in the probability that the bond is loaned. The remaining columns of Table 3 investigate the robustness of this association to additional security-andmarket-levelcontrolsthatarelikelytobecorrelatedwiththedemandforindividual bonds. Column 2 of Table 3 controls for bond characteristics, including the residual maturity, offering yield, offering amount, amount outstanding, and credit rating. Including these finer controls reduces the original sample by about 60 percent. Nevertheless, the association between the probability that an insurer loans a bond and the market share remains largely unchanged. Furthermore, the coefficient estimate on the offering yield and rating are negative, and the coefficient estimate on residual maturity is positive. These findings are broadly consistent with previous empirical studies, for example, Asquith, Au, Covert & Pathak (2013). Column 3 of Table 3 controls for loan market tightness (Market tightness ), defined as the it fractionofabondthatisonloanrelativetothetotalamountofthatbondthatismadeavailable for loan by all lenders, and the bond rebate (Rebate ). The inclusion of Market tightness it it and Rebate has little effect on the magnitude and significance of the coefficient estimate on it Market share . Inaddition, consistentwithKolasinskietal.(2013), theestimatessuggestthat, ijt when controlling for overall market tightness and an insurer’s market share in that bond, there is a significant positive association between a more special bond—i.e. a more negative rebate— and the probability that the bond will be on loan. Column 4 of Table 3 adds a measure of concentrationforindividualsecuritiesinthelifeinsuranceindustry(HHI )usingtheHerfindahl– it 14

Hirschman Index computed at the bond-year level using only life insurers’ market shares in each individual bond.17 This new finding that life insurers are more likely to lend securities in which they have a greater market share is consistent with the existing empirical literature that studies the passive lendingchannel(Kolasinskietal.2013,Kaplanetal.2013)andthosetheoriesemphasizingsearch frictions (Miller 1977, Duffie et al. 2002, Vayanos & Weill 2008). Intuitively, the probability of loaning a bond is higher if a securities borrower is matched with a lender with a high market share in that bond. Knowing that they face a greater opportunity cost in looking elsewhere for the bond, the match is more likely to be consummated. However, the finding is also consistent with the active lending channel. In such a case, a lenderwouldbemorelikelytolendbondsinwhichithasahighmarketshare. Thelenderknows it can more easily finance a portfolio of long-dated assets and earn a spread between the cost of cash borrowing and the return on these assets.18 4 Endogeneity of the cost of cash from securities lending We build on the analysis of the previous section to understand the major endogeneity challenge in identifying the active lending channel through the lending decisions of a single lender. First, lendersmaybeabletocontrolfactorsthataffecttheborrowingdecision, suchasthetermsofthe loan. Second, those factors and/or the lender’s ability to control them may be correlated with the lender’s strategy. Consider, for example, the previous regression specification representing the loan decision in conjunction with a specification that represents the equilibrium rebate: Loan = α1+α2+α3+βMarket share +δRebate +Z γ +(cid:15) (2) ijt i j t ijt it ijt ijt Rebate = α˜1+α˜3+Z˜ γ˜+(cid:15)˜ . (3) it i t ijt ijt As is widely known, if any variable (Z˜ ) in equation 3 that determines the equilibrium rebate ijt is omitted from the loan decision specification, all the coefficient estimates from equation 2 will be inconsistent (Greene 2012). Wecandemonstratethatconfoundingfactorsarelikelytobeaseriousconcernbyestimating 17Thecalculationislimitedbyourabilitytoobserveonlytheholdingsinourdata–i.e.,bynecessity,itassumes atomistic holdings by other institutions. 18 The statistically significant negative coefficient on HHI is also consistent with both an active and passive it securities lending strategy. With passive lenders, an increase in bond concentration reduces the probability that a borrower seeking a bond is matched with an insurer holding that bond, which decreases the probability of the loan. Withactivelenders,anincreaseinconcentrationimpliesthatthevalueoftherebateonthatbondrises–i.e., thebondbecomeslessspecial–asmoreofthebondissuppliedtosecuritiesborrowers. Asthebondbecomesless scarce and the rebate increases, the cost of borrowing cash by lending that bond also increases, which reduces the likelihood of loaning the bond. 15

equation 3 using known variables from equation 2. Because known variables are correlated with a bond’s rebate, it is likely that other (potentially unobservable) determinants of the lending decision are also correlated with the bond’s rebate. We show that the rebate is a function of the concentration of holdings in that bond. Intuitively, lenders of hard-to-find securities can get a better deal on their transaction—as in Duffie (1996)—and this may affect their decision to lend particular securities. We first show that the cost of raising cash by lending a particular security is a function of life insurers’ holdings of that security. Adding weight to this likely source of endogeneity, we also show that life insurers may affect the shadow value of their entire lending program. We aggregate the lending decisions over segments of life insurers’ bond portfolios and show there is a positive association between the decision to lend bonds within that segment and the holdings distribution of that segment. Taken together, the results suggest that a life insurer may be able to control the cost of cash that it can raise by choosing which bonds to lend, and this confounds the identification of active securities lending. The implication is that we cannot use correlates of the loan decisions within a life insurer to establish the presence of an active securities lending channel. 4.1 Determinants of the rebate on a bond Table 4 summarizes the results from investigating the relationship between a bond’s rebate and the concentration of that bond’s holdings across life insurers. As described in Section 1, a hardto-find security attracts a lower and sometimes negative rebate. To be clear, in the latter case, the security borrower pays a premium to the lender in excess of the return from reinvesting the cash collateral. Our specifications are based on equation 3, with the unit of observation being a bond i held by insurer j at the end of year t. We restrict the sample to bonds that have at least three holders in any given year to obtain meaningful variation in the distribution of holdings. The coefficient on HHI suggests that those bonds that are the most concentrated among it life insurers tend to have a lower rebate (that is, tend to be more special). Column 1 of Table 4 reports the result of a regression of Rebate on HHI . This baseline specification includes the it it loan decision, life insurer, year, and bond issuer fixed effects. Column 2 adds market share, market tightness, and other bond-level controls to the specification.19 Columns 3 and 4 separate the sample into those bonds that are on loan and those that are not loaned by any life insurer, respectively. In all cases, the relationship between Rebate and HHI is statistically significant it it at less than the 1 percent level. The baseline coefficient estimate suggests that, on average, 19 In general, we find results that are qualitatively similar to those of Nashikkar & Pedersen (2007). 16

a one standard deviation increase in a bond concentration is associated with a 0.4 percentage point (1.2 standard deviation) decrease in the rebate rate for that security.20 Taken together, the results reported in Table 4 show the likely endogenity of bond rebates and loan decisions. Although a bond’s rebate is likely to be a function of the size of life insurers’ holdings, life insurers’ decisions to lend the bond are likely to be influenced by its specialness. More generally, the findings support a concern that life insurers may be able to use their market power over a large number of bonds to exert significant control over their cost of cash borrowing via securities lending. This means that, unless the regression that specifies the loan decision accounts for all the co-determinants of a bond’s rebate, the coefficient estimates from equation 2 will be inconsistent. 4.2 Determinants of an insurer’s overall cost of cash collateral We now study the potential endogeneity of life insurers’ overall cost of raising cash collateral via securitieslending. Theprevioussectionprovidedsomeevidencethatindividualbondspecialness is potentially determined by the distribution of bond holdings and by life insurers’ decisions to lend. The findings, however, are only suggestive of an insurer’s ability to exert control over its overall cost of funding. For example, life insurers may have market power for only a small fraction of their portfolio. Although an insurer might exert control over the rebate on some bonds, its control over the rebate on other bonds could be more limited. To assess the determinants of life insurers’ overall cost of funding requires summarizing the content of their portfolios. A crude approach is to run the same test as in Table 4 using averages of the previously defined variables calculated over each insurer’s portfolio. However, aggregating bond-level information to the insurer level is likely to produce a test without much power. We adopt an alternative approach by creating subsets of life insurers’ bond portfolios. Specifically, we rank each life insurer’s bond portfolio according to the rebate of each bond and divide the portfolios according to rebate quantiles, which yields equal-sized groups of bonds that differ by the average rebate of the bonds in each group. Then, for each equal-sized bond group within two rebate quantiles, we compute a weighted average analog of the variables discussed in the previous sections. Throughout the analysis, the weights are the shares of a bond in terms of the fair value with respect to the total fair value of the bonds in the group. The unit of observation for this analysis is a rebate group q for bonds held by a life insurer j at the end of year t. With 20 The positive association between Rebate and Loan , conditional on HHI , is consistent with both a it ijt it passive and active securities lending program. When life insurers are passive, a bond that is more sought after shouldattractahigherrebateinordertosecuretheloan,resultinginapositiveassociationbetweenthelending decisionandtherebate. Conversely,activelifeinsurersmaychoosetolendabondforwhichtheyreceiveabetter deal—amorenegativerebate—tominimizetheiroverallcostofcashborrowing,resultinginthesamecorrelation. 17

variables defined as before, the regression specification is: Loan = α1+α2+α3+βMarket share +Z γ +(cid:15) . (4) qjt q j t qjt qjt qjt Columns 1 and 4 of Table 5 summarize the regression results for 32 and 64 rebate groups, respectively. Thesespecificationscontrolforweightedaveragebondcharacteristicsineachgroup, as well as insurer and year fixed effects. The coefficient estimates suggest there is a positive and robust association between the fraction of bonds on loan and the weighted average market share in those bonds, conditional on the weighted average rebate and concentration. Columns 2 and 3 perform the same test on the subsets of 32 rebate groups above and below the median rebate, respectively. The coefficient on Market share is broadly similar, with marginally weaker qjt significance for rebate groups below the median rebate. Columns 5 and 6 repeat this test on the subsets of 64 rebate groups above and below the median rebate, respectively. Here, the coefficient on Market share is only significant for the sample of groups above the median, qjt suggesting that the positive association between the loan and market share is likely to be more pronounced in the portion of the portfolio that has a higher average rebate. However, since a higher rebate tends to be associated with a higher probability of loaning a bond, increasing the number of quantiles mechanically reduces the fraction of loaned bonds in the lower group. As a result, any correlation between average loan and market share is likely to be weaker. Taken together, these results suggest that the association between the fraction of bonds loaned and the weighted average market share is significant in a large portion of life insurers’ bond portfolios–at least half of the bonds–and especially so in the portion of the portfolio above the median rebate rate. We interpret this result as suggesting that life insurers are likely to have significant control over their overall cost of funding via securities lending. The analysis presented in this section suggests that an analysis of the loan decisions of a single insurer is unlikely to provide compelling evidence for an active lending channel. We find support for the argument that there are factors affecting the securities borrowing decisions that may be influenced by lenders themselves. We can and do control for the obvious candidates in our analysis reported in Section 3, but the list of potential confounding factors is practically endless. 5 Securities lending as wholesale funding To build a compelling case that there is an active securities lending channel, we adopt a twofold strategy. First, we exploit the ability to observe in our data the same bonds at the 18

same time across different life insurers’ portfolios. We introduce security–time fixed effects to control for potentially confounding factors in a reduced form that encompasses the cost of cash borrowing, as well as the availability and distribution of holdings, associated with each security. Second, we propose an instrumental variable to deal with possibly unobservable timevarying insurer heterogeneity. Our two-step approach is intended to approximate the theoretical ideal experiment to establish an active lending channel: testing whether two life insurers with identical bond portfolios would make differentl lending decisions if one exogenously engaged in greater maturity mismatch. 5.1 Controlling for bond-specific factors Our proxy for the extent of active management (Active ) is the fraction of assets in the cash jt reinvestment portfolio of a life insurer j at year t that has a residual maturity of more than one year. This threshold captures investment in assets that are ineligible for purchase by MMFs under Rule 2a-7 and are therefore likely to offer a higher return than cash instruments.21 As is clear from Figure 8, there is variation in this fraction across life insurers and over time. We include Active as the main explanatory variable in a linear specification that includes security– jt time and life insurer fixed effects: Loan = α1+α2+α3+α1×α3+βActive +(cid:15) . (5) ijt i j t i t jt ijt Under a passive lending strategy, Active would be uncorrelated with Lend , that is, β = 0. jt ijt By contrast, in the presence of an active lending strategy, the proxy would be associated with a greater likelihood that a bond will be loaned. Columns 1 and 2 of Table 6 report the results of estimating equation 5 with only life insurer fixed effects and then including security–time fixed effects, respectively.22 The coefficient on Active suggests that, on average, a one standard deviation increase in the fraction of a life jt insurer’s cash reinvestment portfolio invested in assets with maturities greater than 1 year is associated with a 2 percent increase in the probability that the bond is loaned. This association is significant at less than the 1 percent level. 21 By way of further comparison, this threshold is six times the regulatory limit on mutual fund cash reinvestment. 22Thenumberofobservationsismuchlargerthaninsections3and4becausethesampleisnolongerrestricted by the availability of rebate information. 19

5.2 Controlling for insurer-specific unobservable heterogeneity The variable Active may itself be endogenous. The residual error term (cid:15) in equation 5 jt ijt may contain unobservable insurer-specific time-varying factors that are correlated with Active . jt Although we can control for bond-specific, time-varying unobservable factors through the interacted fixed effects, we cannot control for all time-varying insurer heterogeneity that affects its ability to lend bonds. For example, life insurers may have bundles of bonds that facilitate lending only in some years. To obtain exogenous variation in the degree of active cash collateral reinvestment by an insurer (Active ), we introduce, as an instrumental variable, the annual change in the amount jt of unrealized gains/losses on the life insurer’s bond portfolio as a fraction of that insurer’s total assets (∆Unrealized gain ). Under statutory accounting rules, an insurance company jt must report the current market value of the securities in its portfolios. Life insurers calculate the unrealized gain/loss as the difference between the cost of its portfolio and the current liquidation value. The unrealized gain/loss on a portfolio is incorporated into ratings agencies’ evaluationsofthefinancialstrengthofaninsurer,becausetheycanpredictactualeconomiclosses (Standard and Poor’s 2009). Our strategy is predicated on the idea that an active securities lender will attempt to compensate for the change in unrealized losses on its underlying portfolio by increasing the return on its cash reinvestment portfolio. Unrealized losses are themselves plausibly beyond the control of the manager of an insurer’s securities lending program and, therefore, from the unobserved heterogeneity that is the source of the endogeneity concern. Columns 3 and 4 of Table 6 report the first- and second-stage results of a regression of Loan on Active , instrumenting Active with ∆Unrealized gain . As with the reduced form ijt jt jt jt inColumn2ofTable6,theIVregressionincludesCUSIP,CUSIP–year,andinsurerfixedeffects. Column 3 of Table 6 shows that there is a positive first-stage association between Active and jt ∆Unrealized gain . ThecoefficientonActive inColumn4ofTable6suggeststhat,onaverage, jt jt a one standard deviation increase in the fraction of an insurer’s cash reinvestment portfolio invested in assets with maturities greater than 1 year is associated with an 11% increase in the probability that the bond is loaned.23 The parameter estimates imply there would be only a small reduction in lending if Rule 2a-7 limits were imposed on the securities lending programs of U.S. life insurance companies. Table 1 shows that life insurers in our post-crisis sample have on average about 17 percent of their cash collateral portfolio reinvested in bonds with a residual maturity of at least one year. 23 We suspectthat our IVcoefficient is larger than our OLScoefficient because our empiricalapproach places greater weight on life insurers that have more variation in active reinvestment over time (Angrist, Graddy & Imbens 2000). 20

Rule 2a-7 would remove these bonds entirely from the portfolio.24 Using the coefficient estimate from Table 6 of the effect ofActive on the probability of lending (0.4), elimination of the active jt lendingchannelwouldreducelendingby0.4×0.17,whichisabout7percent. Figure7showsthat life insurers’ lending against cash collateral amounts to about $55 billion, indicating a potential decline in the market of 0.07×$55 billion, or less than $4 billion. This crude calculation suggests that regulation equivalent to Rule 2a-7 on cash collateral reinvestment by life insurers could enhance financial stability with only minor consequences for bond market functioning. 5.3 Robustness Table 7 summarizes tests of the robustness of the association between the decision to lend securities and the degree of active cash collateral reinvestment. One concern is that variation in thesizeofaninsurer’ssecuritieslendingprogramovertimemaybedrivingtheresult. Columns1 and 2 report the instrumental variable results including as a control the size of the insurer’s securities lending program (SL program size ). A further concern is that life insurers may have jt a preference for lending particular securities. Columns 3 and 4 report the results of repeating the analysis including CUSIP–insurer fixed effects. The estimated coefficient of interest in both cases is similar to our previous results. 6 Conclusion This paper offers the first empirical evidence that certain financial insitutions use securities lending as a source of wholesale funding. That is, some firms raise and maintain a level of cash collateral by lending lower-yielding securities and investing the proceeds in longer-term higheryielding securities. Our main empirical exercise shows that U.S. life insurers’ lending decisions are driven in part by the degree of maturity mismatch in their securities lending programs. We acknowledge and address the stringent identification challenge posed by securities lenders’ potentialabilitytoinfluenceunobservablefactorsthataffecttheborrowingdecision. Thefinding stands in contrast to the widespread perception in the existing literature that securities lenders passively respond to borrower demands and reinvest their cash collateral in short-term funding markets. Our evidence based on data from the period after the implementation of sweeping macro- and microprudential reforms that followed the 2008-09 financial crisis demonstrates the vulnerabilities that securities lenders still pose to securities market functioning. A sharp 24 In addition to restricting the weighted average maturity of the reinvestment portfolio to less than 60 days, Rule2a-7specifiesthatmutualfundscanonlyinvestinnon-MMFeligiblesecuritieswithanexceptionfromthe SEC. MMF eligible securities must have a residual maturity of less than 13 months. 21

contraction in repo funding from securities lenders is one way a run on active securities lenders can adversely affect securities markets, as discussed in the introduction and emphasized by Brunnermeier & Pedersen (2009) and Gorton & Metrick (2010a,b). In addition, during times of financial distress, concerns about securities lenders’ cash reinvestment portfolios may require borrowers to post larger margins to compensate broker-dealers for the risk to their cash. In both cases, the underlying vulnerability to runs from liquidity and/or maturity mismatch can be attributed to the active lending channel. One possible reason that life insurers remain active securities lenders is the persistent low interest rate environment. Low returns to institutional investors’ portfolios deriving from a long period of low interest rates increase their incentive to actively manage their securities lending programs. In light of these vulnerabilities, measures such as the degree of maturity mismatch and the extent to which active securities lenders provide tri-party repo funding are important financial stability metrics. Careful monitoring of securities lenders’ programs is especially important for the functioning of corporate bond markets, where financial stability concerns are greatest. AlthoughthesemeasuresareavailabletosomeextentintheregulatoryfilingsofU.S.lifeinsurers, they are not widely available for all securities lenders.25 References Adrian, T., Begalle, B., Copeland, A. & Martin, A. (2012), ‘Repo and Securities Lending’, National Bureau of Economic Research Working Paper No. 18549 . Angrist, J., Graddy, K. & Imbens, G. (2000), ‘The interpretation of instrumental variables estimators in simultaneous equations models with an application to the demand for fish’, Review of Economic Studies 67(3), 499–527. Asquith, P., Au, A. S., Covert, T. & Pathak, P. A. (2013), ‘The market for borrowing corporate bonds’, Journal of Financial Economics 107(1), 155–182. Becker, B. & Ivashina, V. (2015), ‘Reaching for yield in the bond market’, The Journal of Finance 70(5), 1863–1902. Brunnermeier, M. K. & Pedersen, L. H. (2009), ‘Market liquidity and funding liquidity’, Review of Financial Studies 22(6), 2201–2238. 25Measuresofmaturitymismatchandrepofundingassociatedwithsecuritieslendingprogramswouldaugment thedatacollectionproposedbyAdrian,Begalle,Copeland&Martin(2012). Koijen&Yogo(2015)notethatlife insurers’statutoryfilingslackdetailsontheinternationaldimensionsoflifeinsurers’securitieslendingprograms. 22

Covitz, D., Liang, N. & Suarez, G. A. (2013), ‘The evolution of a financial crisis: Collapse of the asset-backed commercial paper market’, The Journal of Finance 68(3), 815–848. Diamond, D. W. & Dybvig, P. H. (1983), ‘Bank Runs, Deposit Insurance, and Liquidity’, The Journal of Political Economy 91(3), 401–419. Dive, M., Hodge, R., Jones, C. & Purchase, J. (2011), ‘Developments in the global securities lending market’, Bank of England Quarterly Bulletin . Duffie, D. (1996), ‘Special repo rates’, The Journal of Finance 51(2), 493–526. Duffie, D., Gârleanu, N. & Pedersen, L. H. (2002), ‘Securities lending, shorting, and pricing’, Journal of Financial Economics 66(2), 307–339. Faulkner, M. (2006), An Introduction to Securities Lending, London: Spitalfields Advisors. Foley-Fisher, N., Narajabad, B. & Verani, S. (2015), ‘Self-fulfilling Runs: Evidence from the U.S. Life Insurance Industry’, Finance and Economics Discussion Paper Series 2015-032 . Garbade, K. (2006), ‘The evolution of repo contracting conventions in the 1980s’, Economic Policy Review 12(1). Goldstein, I. & Pauzner, A. (2005), ‘Demand–deposit contracts and the probability of bank runs’, the Journal of Finance 60(3), 1293–1327. Gorton, G. & Metrick, A. (2010a), ‘Haircuts’, Federal Reserve Bank of St. Louis Review (Nov), 507–520. Gorton, G. & Metrick, A. (2010b), ‘Regulating the shadow banking system’, Brookings Papers on Economic Activity pp. 261–312. Gorton,G.&Metrick,A.(2012),‘Securitizedbankingandtherunonrepo’,Journal of Financial Economics 104(3), 425 – 451. Greene, W. (2012), Econometric Analysis (7th Edition), Pearson Education. He, Z. & Xiong, W. (2012), ‘Dynamic debt runs’, Review of Financial Studies 25(6), 1799–1843. Huh, Y. & Infante, S. (2016), ‘Bond Market Liquidity and the Role of Repo’, mimeo . Kaplan, S. N., Moskowitz, T. J. & Sensoy, B. A. (2013), ‘The effects of stock lending on security prices: An experiment’, The Journal of Finance 68(5), 1891–1936. 23

Keane, F. M. (2013), ‘Securities loans collateralized by cash: Reinvestment risk, run risk, and incentive issues’, Current Issues in Economics and Finance 19(3). Koijen, R. S. & Yogo, M. (2015), ‘Risks of Life Insurers: Recent Trends and Transmission Mechanisms’, mimeo . Koijen, R. S. & Yogo, M. (forthcoming), ‘Shadow Insurance’, Econometrica . Kolasinski, A. C., Reed, A. V. & Ringgenberg, M. C. (2013), ‘A multiple lender approach to understanding supply and search in the equity lending market’, The Journal of Finance 68(2), 559–595. Krishnamurthy, A., Nagel, S. & Orlov, D. (2014), ‘Sizing up repo’, The Journal of Finance 69(6), 2381–2417. McDonald, R. & Paulson, A. (2015), ‘AIG in Hindsight’, Journal of Economic Perspectives 29(2), 81–105. Miller, E. M. (1977), ‘Risk, uncertainty, and divergence of opinion’, The Journal of Finance 32(4), 1151–1168. Musto, D., Nini, G. & Schwarz, K. (2011), ‘Notes on bonds: Liquidity at all costs in the great recession’, mimeo . Nashikkar, A. J. & Pedersen, L. H. (2007), ‘Corporate Bond Specialness’, mimeo . Peirce, H. (2014), ‘Securities Lending and the Untold Story in the Collapse of AIG’, George Mason University Mercatus Center Working Paper No. 14-12 . Schmidt, L., Timmerman, A. & Wermers, R. (forthcoming), ‘Runs on money market mutual funds’, American Economic Review . Schroth, E., Suarez, G. A. & Taylor, L. A. (2014), ‘Dynamic debt runs and financial fragility: Evidence from the 2007 ABCP crisis’, Journal of Financial Economics 112(2), 164–189. Standard and Poor’s (2009), ‘Methodology And Assumptions For Gauging The Impact Of Unrealized Losses On Insurers’ Financial Strength’. Vayanos, D. & Weill, P. (2008), ‘A search-based theory of the on-the-run phenomenon’, The Journal of Finance 63(3), 1361–1398. 24

7 Tables fo noitaived dradnats dna ,naidem ,naem ,snoitavresbo fo rebmun eht troper 5 hguorht 2 snmuloC .scitsitats evitpircseD :1 elbaT smargorp gnidnel seitiruceS .CIAN eht yb delipmoc seinapmoc ecnarusni fo sgnilfi yrotutats launna eht morf sisylana ruo ni desu selbairav taht oiloftrop tnemtsevnier eulav riaf eht fo noitcarf eht si tnemtsevnier laretalloc hsac evitcA .tnuocca lareneg eht fo ezis eht yb delacs era snmuloC .tnuocca lareneg eht fo ezis eht yb delacs era sessol/sniag dezilaernu ni egnahc ehT .raey eno tsael ta fo ytirutam laudiser a sah no noitamrofni gnidda ,esabataD seitiruceS emocnI dexiF tnegreM eht htiw sgnilfi yrotutats eht gnigrem retfa scitsitats wohs 9 hguorht 6 ,82 = tnelaviuqe ro ,AAA gnittes ,seicnega niam eerht eht ssorca sgnitar tiderc egareva dna etalsnart yllaciremun eW .scitsiretcarahc dnob gnigrem yltneuqesbus retfa scitsitats edivorp 31 hguorht 01 snmuloC .4 = C dna ,7 = CC ,9 = -CCC ... 42 = -AA ,52 = AA ,62 = +AA .naol eht fo eulav eht gnisu detaluclac era segareva dethgieW .tekram gnidnel seitiruces eht no atad ecnaniF seitiruceS tikraM htiw tikraM htiw egreM DSIF htiw egreM stnemetatS launnA CIAN DS naideM naeM .sbO DS naideM naeM .sbO DS naideM naeM .sbO . . . 701 . . . 701 . . . 701 seinapmoc ecnarusnI 57.34 69.4 13.22 113 57.34 69.4 13.22 113 57.34 69.4 13.22 113 )nb$( eulav tnuocca lareneG 30. 10. 20. 992 30. 10. 20. 992 30. 10. 20. 992 ) ezis margorp LS( gnidnel seitiruceS tj 82. 0 71. 732 82. 0 71. 732 82. 0 71. 732 ) evitcA( tnemtsevnier laretalloc hsac evitcA tj 54. 0 40.- 113 54. 0 40.- 113 54. 0 40.- 113 ) niag dezilaernU∆( )spb( ssol/sniag dezilaernu ∆ tj 22.52 5 82.01 272291 67.22 5 30.01 384942 85.72 4 63.9 396945 )m$( gnidloh dnob fo eulav raP 52. 0 70. 272291 32. 0 50. 384942 61. 0 30. 396945 ) naoL( ytiruces tnel rof elbairav ymmuD tji 18.2 1 41.1 258681 6.4 1 49. 384942 . . . . )nb$( tnuoma gnireffO 84.2 6 95.5 861251 83.2 6 97.5 865402 . . . . )tnecrep( dleiy gnireffO 43.9 8 3.11 889581 95.9 8 91.11 482842 . . . . )sry( ytirutam laudiseR 34.3 1 91.1 91459 20.3 1 39. 441131 . . . . )nb$( gnidnatstuo tnuomA 80.3 02 8.91 680081 40.3 02 78.91 376932 . . . . gnitar tiderC 70. 0 0 272291 . . . . . . . . ) etabeR( etaber gva dethgieW ti 61. 0 30. 272291 . . . . . . . . ) erahs tekraM( dleh elbadnel latot % tji 58.7 0 73. 272291 . . . . . . . . ) IHH( sgnidloh ’srerusni efil fo IHH ti 34. 0 41. 272291 . . . . . . . . ) ssenthgit tekraM( elbadnel latot/tnel latoT ti 25

slenap owt eht ,tesatad degrem lanfi eht gnisU .sdnob tnel-non dna tnel fo scitsitats evitpircseD :2 elbaT .tnel ton era taht esoht htiw seinapmoc ecnarusni yb tnel sdnob fo scitsiretcarahc erapmoc 1 = naoL :seitiruces tneL 0 = naoL :seitiruces tnel-noN tji tji DS naideM naeM .sbO DS naideM naeM .sbO 2.43 37.5 54.31 021,31 24.42 10.5 60.01 251,971 )m$( gnidloh dnob fo eulav raP 9.499,3 3.447 6.284,1 009,21 5.366,2 006 5.160,1 259,371 )m$( tnuoma gnireffO 16.2 6.5 55.5 671,01 54.2 76.5 85.5 299,141 )tnecrep( dleiy gnireffO 26.9 8 40.21 477,21 23.9 8 52.11 412,371 )sry( ytirutam laudiseR 4.672,5 5.786 5.575,1 793,6 4.922,3 006 9.921,1 220,98 )m$( gnidnatstuo tnuomA 55.3 02 16.91 974,21 40.3 02 28.91 706,761 gnitar tiderC 21. 0 20.- 021,31 80. 0 10. 251,971 ) etabeR( etaber dethgiew gvA ti 31. 50. 90. 021,31 41. 40. 90. 251,971 ) erahs tekraM( dleh elbadnel latot % tji 25.8 21. 84. 021,31 97.7 51. 74. 251,971 ) IHH( dleh elbadnel latot fo IHH ti 93. 12. 3. 021,31 63. 31. 42. 251,971 ) ssenthgit tekraM( ssenthgit tekraM ti 26

Table 3: Relationship between bond lending and market share. The dependent variable (Loan ) takes a value of 1 if insurer j is lending bond i at ijt time t and 0 otherwise. The main explanatory variable (Market share ) is the ijt holdingoftheinsurerrelativetotheamountofthebondmadeavailabletosecurities borrowers by all lenders. In addition to the insurer, year, and issuer fixed effects in the baseline specification, reported in Column 1, we incrementally add a variety of bond and market controls, reported in Columns 2 though 4. The sources for the data used in the analysis are described in the notes to Table 1. Huber–White robust standard errors are reported in parentheses. *** p<0.01; ** p<0.05; * p<0.1. Dependent variable: (1) (2) (3) (4) Loan Baseline Security Market Concentration ijt Control Control Index Market share 0.0506*** 0.0501*** 0.0458*** 0.0561*** ijt (0.00490) (0.00781) (0.00778) (0.00831) Rebate -0.229*** -0.231*** it (0.0239) (0.0239) Market tightness 0.0339*** 0.0337*** it (0.00375) (0.00375) HHI -0.0129*** it (0.00344) Residual maturity 0.000696*** 0.000670*** 0.000683*** it (0.000129) (0.000129) (0.000129) Offering yield -0.00340*** -0.00309*** -0.00296*** i (0.00101) (0.000975) (0.000961) Offering amount -0.00384 -0.00510 -0.00532 i (0.00356) (0.00357) (0.00356) Amount outstanding 0.00901*** 0.0110*** 0.0106*** it (0.00318) (0.00320) (0.00319) Rating -0.00752*** -0.00553*** -0.00538*** it (0.00198) (0.00198) (0.00198) Fixed effects: Insurer, issuer, year Y Y Y Y Observations 182,315 69,459 69,459 69,459 R-squared 0.149 0.159 0.163 0.163 27

Table 4: Relationship between bond rebate and holdings distribution. The dependent variable(Rebate )isthepercentageofcashcollateralreinvestmentincomegivenbythesecurities it lender to the borrower. The main explanatory variable (HHI ) is the Herfindahl–Hirschman it Index computed at the bond-year level using life insurers’ holdings. We restrict the sample to bonds that have at least three holders in any given year. Columns 1 and 2 contain results including various bond and market controls, in addition to insurer, year, issuer, and bond type fixed effects. Columns 3 and 4 restrict the sample to bonds on loan and to bonds not loaned by any life insurer, respectively. The sources for the data used in the analysis are described in the notes to Table 1. Huber–White robust standard errors are reported in parentheses. *** p<0.01; ** p<0.05; * p<0.1. Dep. variable: (1) (2) (3) (4) Rebate All Securities All Securities Lent Securities Not Lent Securities it HHI -1.306*** -1.016*** -4.009** -0.537*** it (0.0889) (0.0974) (1.748) (0.109) Loan -1.313*** -1.201*** ijt (0.0827) (0.131) Market share 1.790*** 27.69*** 0.827*** ijt (0.232) (5.189) (0.292) Market tightness 0.464*** 1.440** 0.194 it (0.0746) (0.731) (0.159) Residual maturity 0.0221*** -0.0560 0.0256*** it (0.00272) (0.0370) (0.00329) Offering yield 0.0572*** 1.043*** 0.0117 i (0.0169) (0.373) (0.0113) Offering amount -0.170* -1.221 0.155 i (0.0992) (0.850) (0.152) Amount outst 0.541*** 2.338*** 0.595*** it (0.0971) (0.780) (0.159) Rating 0.463*** 2.297*** 0.0136 it (0.0679) (0.754) (0.0745) Fixed effects: Insurer, issuer, year Y Y Y Y Bond type N Y Y Y Observations 177,397 67,538 1,637 35,427 R-squared 0.526 0.577 0.803 0.679 28

Table 5: Relationship between fraction lent and weighted average market share in the portfolio. We divide life insurers’ portfolios into groups according to rebate quantile. The dependent variable Loan is the fraction of bonds in group q lent by insurer j at time qjt t. The main explanatory variable (Market share ) is the fair-value-weighted average holding qjt of bonds in group q by insurer j relative to the total amount of the bonds made available to securities borrowers by all lenders. Columns 1 and 4 divide the porfolios into 32 and 64 groups, respectively, and include controls for various bond and market characteristics as well as insurer and year fixed effects. Columns 2 and 3 and Columns 5 and 6 divide the groups into those that are above and below the median, respectively. The sources for the data used in the analysis are described in the notes to Table 1. Huber–White robust standard errors are reported in parentheses. *** p<0.01; ** p<0.05; * p<0.1. Dep. variable: (1) (2) (3) (4) (5) (6) Loan All 32 parts 32 parts All 64 parts 64 parts qjt perc. > 50 perc. ≥ 50 perc. perc. > 50 perc. ≥ 50 perc. Market share 7.806** 9.984** 9.081* 5.242** 9.684** 4.736 qjt (3.532) (5.028) (5.013) (2.463) (3.801) (3.301) Rebate -27.13*** -41.13*** -28.23*** -28.10*** -55.50*** -30.28*** qt (3.060) (15.46) (3.439) (2.677) (11.24) (2.926) HHI -5.496*** -5.614*** -4.974*** -3.910*** -3.150** -3.709*** qt (1.202) (1.851) (1.602) (0.813) (1.367) (0.978) Residual mat -0.0562 0.0179 -0.110* -0.0139 0.0807* -0.0934** qt (0.0432) (0.0612) (0.0606) (0.0312) (0.0464) (0.0410) Offering yield -0.217* -0.00620 -0.386** -0.282*** -0.158 -0.368*** qt (0.123) (0.185) (0.164) (0.0836) (0.150) (0.101) Offering amt 0.471*** 0.335*** 1.630*** 0.473*** 0.384*** 1.457*** qt (0.115) (0.117) (0.339) (0.0992) (0.0974) (0.302) Amt outst 0.110 0.234* -0.720 0.179 0.274** -0.252 qt (0.141) (0.136) (0.500) (0.132) (0.134) (0.405) Rating 0.0385 0.0260 0.0444 0.0216 0.0119 0.0113 qt (0.0652) (0.0968) (0.0913) (0.0510) (0.0730) (0.0705) Fixed effects: Insurer, year Y Y Y Y Y Y Observations 9,941 4,926 5,015 16,703 8,208 8,495 R-squared 0.340 0.373 0.359 0.287 0.314 0.304 29

Table 6: Securities lending as wholesale funding. The dependent variable (Loan ) ijt takes a value of 1 if insurer j is lending bond i at time t and 0 otherwise. The main explanatoryvariable(Active )isthefractionofcashcollateralreinvestmentthathasaresidual jt maturity of more than one year. Columns 1 and 2 report reduced form correlations including insurer, CUSIP, year, and CUSIP–year fixed effects. Columns 3 and 4 contain the first and second stages, respectively, of a two-stage least squares estimation. The instrumental variable (∆Unrealized gain ) is the unrealized gain/loss made by insurer j on its bond portfolio. The jt sources for the data used in the analysis are described in Section 5 of the main text. *** p<0.01; ** p<0.05; * p<0.1. Dependent variable: (1) (2) (3) (4) Loan Reduced form Reduced form First stage Second stage ijt Active 0.071*** 0.103*** 0.389*** jt (0.00431) (0.00676) (0.0520) ∆Unrealized gain -2.904*** jt (0.0367) Fixed effects: Insurer Y Y Y Y CUSIP N Y Y Y Year N Y Y Y CUSIP×Year N Y Y Y Cragg–Donald Wald F statistic 3849.06 Stock–Yogo weak ID test critical values 10% maximal IV size 16.38 Observations 462,433 462,433 462,433 462,433 R-squared 0.0364 0.291 0.965 0.284 30

Table 7: Robustness tests. The dependent variable, main explanatory variable, and instrumental variable are the same as in Table 6. Columns 1 and 2 contain the first and second stages, respectively, of a two-stage least square estimation including as a control the size of insurer j’s securities lending program. Columns 3 and 4 contain the first and second stages, respectively, of a two-stage least square estimation including CUSIP– Insurer fixed effects. The sources for the data used in the analysis are described in Section 5 of the main text. *** p<0.01; ** p<0.05; * p<0.1. Dependent variable: (1) (2) (3) (4) Loan Stage 1 Stage 2 Stage 1 Stage 2 ijt Active 0.430*** 0.394*** jt (0.05127) (0.05983) ∆Unrealized gain -3.437*** -3.133*** jt (0.04590) (0.06566) SL program size 0.007*** 0.005*** jt (0.00236) (0.00098) Fixed effects: Insurer, year, CUSIP Y Y Y Y CUSIP×Year Y Y Y Y CUSIP×Insurer N N Y Y Observations 462,433 462,433 462,433 462,433 31

8 Figures Figure 1: A stylized map of securities lenders’ role in the shadow banking system. Broker-dealers obtain cash from money market mutual funds and securities lenders through short-term funding markets. The dealers provide the cash to securities market participants. The cloud represents the general functioning of securities markets, illustrated with the example of hedge funds taking long and short positions. Securities market participants borrow both cash funding and securities. Securities lenders decide whether to lend assets from their portfolios in exchange for collateral in the form of either cash or other securities, from broker-dealers. When they receive cash collateral, securities lenders decide whether to invest back into shortterm markets or into long-term markets. In the latter case, they may invest, for example, in long-term corporate bonds or asset-backed securities. MMF security cash Short-term market Hedge Fund Long Securities Broker/ Securities market Lender Dealer Hedge Fund Short Long-term market 32

Figure 2: Coverage of the stylized map of securities lending by the existing literature. Duffie et al. (2002) (Panel (a)) study the effect of interactions between securities lenders and broker-dealers on pricing in the securities market. Brunnermeier & Pedersen (2009) and Gorton & Metrick (2012) (Panel (b)) consider the effect of haircuts on cash funding and the functioningofsecuritiesmarkets. Krishnamurthyetal.(2014)(Panel(c))describehowsecurities lenders and MMFs participate in short-term funding markets. The overall map is described in the notes to Figure 1. (a) Duffie, Gârleanu & Pedersen (2002) MMF security cash Short-termmarket HedgeFundLong Securities Broker/ Securitiesmarket Lender Dealer HedgeFundShort Long-termmarket (b) Brunnermeier & Pedersen (2009), Gorton & Metrick (2012) MMF security cash Short-termmarket HedgeFundLong Securities Broker/ Securitiesmarket Lender Dealer HedgeFundShort Long-termmarket (c) Krishnamurthy, Nagel & Orlov (2014) MMF security cash Short-termmarket HedgeFundLong Securities Broker/ Securitiesmarket Lender Dealer HedgeFundShort Long-termmarket 33

Figure 3: Securities lenders’ and MMFs’ tri-party repo market funding. Panel (a) shows the total resources potentially available to securities lenders and MMFs for lending. Panel (b) shows the amount of funding that securities lenders and MMFs provided through the tri-party repo market. Source: Krishnamurthy, Nagel & Orlov (2014), based on data from Risk Management Associates. (a) Total resources available to lend 4200 3700 3200 2700 2200 1700 1200 700 nb DSU 4 1 2 3 4 1 2 3 4 1 2 3 4 1 Q Q Q Q Q Q Q Q Q Q Q Q Q Q 6 7 7 7 7 8 8 8 8 9 9 9 9 0 0 0 0 0 0 0 0 0 0 0 0 0 0 1 0 0 0 0 0 0 0 0 0 0 0 0 0 0 2 2 2 2 2 2 2 2 2 2 2 2 2 2 MMF total assets Total cash collateral (b) Tri-party repo funding 650 600 550 500 450 400 350 300 250 200 nb DSU 4 1 2 3 4 1 2 3 4 1 2 3 4 1 Q Q Q Q Q Q Q Q Q Q Q Q Q Q 6 7 7 7 7 8 8 8 8 9 9 9 9 0 0 0 0 0 0 0 0 0 0 0 0 0 0 1 0 0 0 0 0 0 0 0 0 0 0 0 0 0 2 2 2 2 2 2 2 2 2 2 2 2 2 2 Repo from MMFs Total repo from cash collateral 34

Figure 4: Securities lending against cash collateral in the United States. These daily data aggregate the fair value of all securities lent against cash collateral in the United States, including equity, Treasuries, agency securities, and corporate bonds. The category of other lenders includes corporations, endowments, foundations, and government bodies. Source: authors’ calculations based on data from Markit Securities Finance. $SU fo snoillirT 5.1 2.1 9. 6. 3. 0 5.1 2.1 9. 6. 3. 0 $SU fo snoillirT 2007 2008 2009 2010 2011 2012 2013 2014 2015 Retirement & Pension Plans Mutual Funds Insurance Other 35

Figure 5: Corporate bond lending against cash collateral in the United States. These daily data aggregate the fair value of all corporate bonds lent against cash collateral in the United States. The category of other lenders includes corporations, endowments, foundations, and government bodies. Source: authors’ calculations based on data from Markit Securities Finance. $SU fo snoillirT 2. 51. 1. 50. 0 2. 51. 1. 50. 0 $SU fo snoillirT 2007 2008 2009 2010 2011 2012 2013 2014 2015 Retirement & Pension Plans Mutual Funds Insurance Other Figure 6: Securities lending cash collateral reinvestment. The blue shaded region shows thequarterlytotalamountofcashcollateralreceivedbysecuritieslenders. Theblackandyellow lines, respectively, show the mean and median weighted average maturity of the reinvestment portfolio. Source: authors’ calculations based on data from Risk Management Associates, from a survey of about 14 agent lenders. Mean WAM (right axis) Median WAM (right axis) $SU fo snoillirT 8.1 5.1 2.1 9. 6. 3. 0 052 002 051 001 05 0 syaD Total Cash Reinvestment (left axis) 2000Q1 2002Q1 2004Q1 2006Q1 2008Q1 2010Q1 2012Q1 2014Q1 36

Figure 7: Securities lending by U.S. life insurers. The line shows thequarterly amount of cash collateral received by U.S. life insurers through their securities lending programs. Source: authors’ calculations based on data from NAIC Quarterly Statutory Filings. nb DSU 55 05 54 04 2011q1 2011q3 2012q1 2012q3 2013q1 2013q3 Figure 8: Maturity mismatch in the securities lending programs of U.S. life insurers. Each box-and-whisker plot represents the distribution of U.S. life insurance companies, with the solid area covering the 25-th to 75-th percentiles and the line within the solid area showing the median of the distribution. The whiskers reflect the remaining observations, up to 1.5 times the interquartile range, and the dots are outliers. The blue plot is calculated using the weighted average maturity of the cash collateral that life insurers received for lending securities, where the weights are the amounts of cash collateral received. The red plot is calculated using the weighted average maturity of life insurers’ reinvestment portfolio, where the weights are the amounts reinvested. Source: authors’ calculations based on data from NAIC Annual Statutory Filings. syaD 000,1 008 006 004 002 0 2011 2012 2013 WAM of cash collateral received WAM of reinvestment portfolio 37