Why Does the Yield Curve Predict GDP Growth? The Role of Banks

Finance and Economics Discussion Series Federal Reserve Board, Washington, D.C. ISSN 1936-2854 (Print) ISSN 2767-3898 (Online) Why Does the Yield Curve Predict GDP Growth? The Role of Banks Camelia Minoiu, Andres Schneider, Min Wei 2023-049 Please cite this paper as: Minoiu, Camelia, Andres Schneider, and Min Wei (2023). “Why Does the Yield Curve Predict GDP Growth? The Role of Banks,” Finance and Economics Discussion Series 2023-049. Washington: Board of Governors of the Federal Reserve System, https://doi.org/10.17016/FEDS.2023.049. 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.

Why Does the Yield Curve Predict GDP Growth? The Role of Banks* Camelia Minoiu Andres Schneider Min Wei First Draft: October 21, 2022 This Draft: July 10, 2023 Abstract We provide evidence on the effect of the slope of the yield curve on economic activitythroughbanklending. Usingdetaileddataonbanks’lendingactivitiescoupled withtermpremiumshocksidentifiedusinghigh-frequencyeventstudyorinstrumental variables, we show that a steeper yield curve associated with higher term premiums (rather than higher expected short rates) boosts bank profits and the supply of bank loans. Intuitively,ahighertermpremiumrepresentsgreaterexpectedprofitsonmaturity transformation,whichisatthecoreofbanks’businessmodel,andthereforeincentivizes banklending. Thiseffectisstrongerforex-antemoreleveragedbanks. Werationalize ourfindingsinaportfoliomodelforbanks. Keywords: predictive power of the yield curve; term spread; term premium; bank lending; bank profitability;eventstudy;instrumentalvariable. JELclassifications: E44,E52,E58. *MinoiuisaffiliatedwiththeFederalReserveBankofAtlanta. SchneiderandWeiareaffiliatedwiththeFederal Reserve Board. Email addresses: camelia.minoiu@atl.frb.org, andres.schneider@frb.gov, and min.wei@frb.gov. The authors thank Valentina Bruno, Emi Nakamura, Bill Nelson, Pascal Paul, Enrico Sette (discussant), Hiroatsu Tanaka,SkanderVandenHeuvel,AnnetteVissing-Jorgensen,JonathanWright,YuXu,andconferenceandseminar participants at the Swiss Winter Conference on Financial Intermediation and the Federal Reserve Board for helpful commentsandsuggestions. WethankDonKimandMarcelPriebschforsharingtheirtermpremiumseries. Wethank Jack Spira, Jacob Bochner, Alejandro Guillot, and Stephanie Sezen for research assistance at different stages of the project. TheviewsexpressedinthispaperarethoseoftheauthorsanddonotnecessarilyreflectthoseoftheFederal ReserveBoardortheFederalReserveSystem.

1 Introduction The slope of the yield curve has proved to be an enduring predictor of economic activity, but it is not entirely clear where its predictive power comes from.1 One plausible explanation is that long-term interest rates aggregate investors’ predictions about the future state of the economy: if investors foresaw a slowdown in the economy, they would expect the central bank to respond by lowering the short-term interest rate in the future. The expectation of lower short rates in the future would reduce, all else equal, long-term rates, resulting in a smaller slope of the curve today than it would be otherwise. Under this explanation, the slope of the yield curve would reflect, but not cause, future recessions. Another plausible explanation is that the slope contains information beyond investors’ expectations about the future state of the economy. This alternative explanation opens up the possibility for the slope to affect future economic activity through different channels thathavenotyetbeenfullyexploredintheliterature.2 Inthispaper,wearguethattheslopeoftheyieldcurveaffectsbanks’lendingdecisionsthrough an expected profitability channel, which works as follows. Banks’ main business model consists of maturity transformation, whereby they take on short-term liabilities, such as bank deposits and wholesale borrowing, to fund longer-term assets, such as securities and loans. The difference betweenthelongrateandtheaveragefutureshortrates—thetermpremium—representstheexcess returns banks can expect to earn for bearing a given amount of duration risks associated with the maturity mismatch. All else equal, a larger term premium implies higher expected profits to banks’ lending activities funded by shorter-term borrowing and hence incentivizes them to hold more longer-term assets, including by making more loans. A higher supply of bank credit can be expected,inturn,toboosteconomicgrowththroughthewell-documented“creditchannel”through whichbank-dependentfirmscaninvestandgrowrelativelymore(BernankeandGertler,1995). We provide extensive empirical evidence consistent with the expected profitability channel, which we motivate with a banking model along the lines of Gertler and Kiyotaki (2015). We start by showing with aggregate time-series data that a higher term premium is associated with higher subsequent bank profits and loan growth after controlling for short rate changes and expected future economic growth. Motivated by this evidence, we study a dynamic portfolio model in 1Earlystudiesdocumentingthepredictivepoweroftheyieldcurveinclude, amongothers, Harvey(1989,1993), EstrellaandHardouvelis(1991),StockandWatson(1989),HamiltonandKim(2002a),andAngetal.(2006). 2Previous studies present evidence that the slope of the yield curve contains more information about future economicactivitythanexpectationoffutureshortrates(HamiltonandKim(2002a),Wright(2006),andJardetetal. (2013))ormediansurveyforecasts(RudebuschandWilliams(2009)),butdonotelaborateonthecasualimplications ofsuchinformation. 1

which banks are subject to financing constraints, charge a spread on deposits, and are subject to exogenous interest rate and term premium shocks. In equilibrium, banks choose a leveraged exposuretolong-termloansfundedwithshort-termdeposits,withbankssubjecttomorestringent financingconstraintschoosinghigherleverage(lowercapitalratio),allelseequal. Apositiveterm premiumshock,byincreasingtheexpectedreturntomaturitytransformationand,hence,expected bank profitability, can be expected to relax the financing constraint and boost bank lending in the periods after the shock. Furthermore, given that banks’ expected return on wealth (profits) is proportional to their leverage multiplied by the term premium, more leveraged banks have higher exposuretothetermpremiumandshowastrongerresponsetotermpremiumfluctuations. Aswetakethemodel’stestablehypothesestothedatawefaceacriticalidentificationchallenge, namelythatlong-termyieldsareforwardlookingandcontaininformationaboutthefutureeconomic outlook which may confound the effects of the term premium. In particular, a stronger economic outlookmayinducebothasteeperyieldcurvetodayandhigherbanklendinginthefuture,andmay cause an omitted variable bias if the economic outlook is correlated with the term premium. To address this issue, we employ three empirical identification strategies. First, we exploit a measure of high-frequency term premium shocks computed from changes in the Kim and Wright (2005) termpremiumonFOMCeventdays. Weemploythismeasureacrossmostofourspecificationsat thebankandloanlevel. Inaddition,wecontrolforexpectationsoffutureshortrates,computedas thedifferencebetweenthetermspreadandthetermpremiumshocks(oralternativelysurvey-based growthforecasts)andforshort-rateshocksbasedonhigh-frequencyidentification(Kuttner,2001). We show our estimates using identified term premium shocks are similar to, but economically larger,thanthosebasedonsimplelevelsoftermpremiumestimates. Second,weemployaninstrumentalvariablestrategythatisolatesforeigndemand-drivenshocks totermpremiumsthatareunlikelydrivenbystrongereconomicoutlook. Ourinstrumentalvariable istheforeignofficialholdingsofU.S.Treasurysecurities(normalizedbyU.S.GDP).Theidentifying assumption is that changes in foreign official holdings of Treasuries are driven by foreign central banks’ reserve management and FX intervention needs rather than the U.S. economic outlook or other factors driving the supply of U.S. government debt. We show this instrumental variable is a strongpredictorofthetermpremium,aftercontrollingfortheshortrateandothermacrovariables. Our results are invariant to excluding China’s holdings of U.S. Treasuries from the instrumental variable and to controlling for survey-based growth forecasts and the excess bond premium, a measureofriskappetite(GilchristandZakrajs˘ek,2012). 2

Third, we identify a period when the term premium increased sharply and unexpectedly, and examinedifferentiallendingresponsestotheriseinthetermpremiumdependingonbanks’ex-ante leverage. In particular, we study the “taper tantrum” episode following the May 22, 2013 speech by Chair Ben Bernanke regarding the Federal Reserve’s intention to start tapering asset purchases underitsquantitativeeasingprogramatsomefuturedate. Thisspeechwasfollowedbyasteepand persistent rise in the term premium. Even though market commentary generally did not attribute theannouncementtoanimprovingeconomicoutlook,and,ifanything,marketparticipantssawthe Fed’s reaction function as less accommodative than expected, it is still possible that the decision toreducethepaceofassetpurchasesreflectedexpectationsofabrighteroutlook. Therefore,when weanalyzethelendingdecisionsofbankswithdifferentdegreesofleverage,wecontrolforbanks’ ownforecastsofone-yearaheadGDPgrowthbothbeforeandafterthestartofthetapertantrum. The empirical analysis brings together two main data sources: (a) a long-run panel dataset on bank balance sheets and (b) a shorter-span bank-firm loan-level dataset matched with bank- and firm-level balance sheet information. First, we examine the effect of term premiums on bank loan growth and bank profitability (net interest margins and return on equity) using quarterly balance sheet data for individual banks from the U.S. Call Reports over 1994–2019. Second, we study banks’ lending decisions in response to term premium shocks using a loan-level dataset from the Federal Reserve (FR) Y-14Q data collection effort. This dataset contains information about bank-firmloan-levelexposuresoflargebank-holdingcompaniessubjecttosupervisorystresstests, whichaccountforthemajorityofloancommitmentsintheU.S.bankingsector. Weusethesedata for two analyses: First, we use loans originated during a period of several quarters around the taper tantrum event to study the lending decisions of banks and the ex-post performance of their borrowers. Second,wezoomoutfromthetapertantrumepisodeandusetheentireperiodbetween 2013Q1and2019Q4tostudytheeffectoftermpremiumshocksonbanklendingandprofitability andtoestablishmoregeneralpatterns. Our findings can be summarized as follows. Consistent with the implications of model, we show that (i) banks have higher loan growth and higher profits, on average, following an increase in the term premium; and (ii) banks with higher leverage (lower capital ratios) increase lending relatively more following an increase in the term premium.3 Focusing on the taper tantrum episode, we show the rise in the term premium is associated with a lending boost on the extensive and intensive margins: lower-capital banks are more likely to grant new loans, and for approved 3Our main analysis uses the term premium series from the Kim and Wright (2005) term structure model but are invarianttotermpremiumestimatesfromalternativemodelssuchasKimandPriebsch(2020)andAdrianetal.(2013). 3

loans they increase loan volumes and reduce spreads more than other banks. These credit effects affecttherealeconomy: Infirm-levelregressions,weshowthatnonfinancialfirmsborrowingfrom ex-ante more leveraged banks have higher investment rates subsequently. We also show that the creditsupplyeffectsaroundthetapertantrumperiodgeneralizetothelongertimeperiodspanning 2013Q1–2019Q4. Additionally, we document that more leveraged banks respond to a positive term premium shock by increasing the maturity of their lending portfolios (in loan-level data) and byincreasingtheirmaturitygap(inbank-leveldata). Our estimates of the level and heterogeneous effects of the term premium on banks’ lending decisionsandfirms’investmentratesarebothstatisticallysignificantandeconomicallymeaningful. Focusingoninstrumentalvariableestimates,aonestandarddeviationincreaseinthetermpremium, representing 50 basis points (bps), is associated with bank loan growth, net interest margins (NIMs), and return on equity (ROE) that are higher by 1.1 percentage points (ppts), 4 bps, and 32bps,respectively;theseeffectsaresizablewhencomparedwiththeaverageloangrowth,NIMs and ROE of 3.8%, 1.0%, and 2.3% over the sample period. Furthermore, in the four quarters after the start of the taper tantrum, a bank with higher leverage (with capital ratio at the 25th percentile of the distribution) originated loans that were larger by 11% and less expensive by 7 bps compared to a bank with lower leverage (with capital ratio at the 25th percentile). Over the twoyearsfollowingthetapertantrum,afirmborrowingfrommoreleveragedbanks(withaverage exposure to its banks’ capital at the 25th percentile) had an investment rate higher by close to 2 pptsthanafirmborrowingfromlessleveragedbanks(withexposuretothe75thpercentile). Our work contributes to two strands of literature. First, we contribute to the literature on the predictive power of the yield curve for future economic growth (Harvey, 1988; Estrella and Hardouvelis, 1991; Hamilton and Kim, 2002b; Favero et al., 2005; Ang et al., 2006; Rudebusch etal.,2006;Jardetetal.,2013). Ourcontributionistoemphasizetheroleoffinancialintermediaries andtodocumentonespecificchannelthroughwhichahighertermpremiummayboosttheeconomy— anexpected profits channelthatincentivizesbanks toengagein morematuritytransformationand increaseloansupply. Adrianetal.(2019)exploreasimilarchannelthroughwhichmonetarypolicy tightening at the short end flattens the yield curve and reduces credit supply by compressing net interest margins. Our paper differs in two major ways. First, we focus on the role of the term premiumasthecomponentofthetermspreadthatgeneratestheexpectedprofitschannel. Second, webringtogethergranulardataandseveralidentificationstrategiestomitigateendogeneityproblems associated with the forward-looking nature of long-term yields and to isolate a causal mechanism 4

fortheforecastingpowerofthetermspread. Second, this paper contributes to the literature on banks’ exposure to interest rate risks and the implications for monetary policy transmission (Begenau et al., 2015; Di Tella, 2020; Haddad and Sraer, 2020; Drechsler et al., 2021; Gomez et al., 2021). Most studies in this literature focus on the level of the short rate. Alessandri and Nelson (2015) and Paul (2023) examine the link betweentheslopeoftheyieldcurveandbankprofits,butdonotassesstheeffectsonbanklending and the real economy. English et al. (2018) show that bank equity prices fall after increases in the level of interest rates or a steepening of the yield curve. We complement these papers by exploring how changes in the slope of the yield curve affect not only bank profitability but also the supply of bank credit and the investment decisions of bank-dependent firms. Our work further emphasizes the importance of decomposing yields into expectations and term premium componentsforunderstandingthechannelsthroughwhichfinancialintermediariesreacttochanges inlongtermrates. Ourfindingssuggestthatbanksareanimportantchannelthroughwhichtheentireyieldcurve, not just the short end, affects the economy, with implications for policy. Asset purchases have became a standard tool that global central banks employ to provide monetary accommodation whentheshortrateisnearitseffectivelowerbound,andtheyaretypicallythoughtofasoperating by reducing term premiums and long-term yields (Krishnamurthy and Vissing-Jorgensen, 2011). In turn, lower term premiums boost the values of security holdings that are marked-to-market on bankbalancesheets,raisingbanknetworkandsupportingtheabilitytolend—aphenomenonthat Brunnermeier and Sannikov (2014) called stealth recapitalization and was documented, among others,byChakrabortyetal.(2020),Acharyaetal.(2019)andRodnyanskyandDarmouni(2017). Our results suggest that when calibrating those purchases, central banks may want to consider the potential negative effects on bank profits and bank lending implied by the expected bank profitabilitychanneldocumentedhere. 2 A first look at aggregate time-series data We take a first look at the aggregated data and examine how term spreads, and in particular the term premium component, are related to future economic growth, bank profitability, and lending. 5

Inparticular,werunthefollowingpredictiveregressions: (cid:16) (cid:17) Z = α+β ∆ y1+β y20−y1 +γX +ε (1) t,t+4 1 t 2 t t t t,t+4 (cid:16) (cid:17) Z = α+β ∆ y1+β y 20,eh −y1 +β y 20,tp +γX +ε (2) t,t+4 1 t 2 t t 3 t t t,t+4 where yn denotes the n-quarter yields, y 20,eh ≜ E ∑19 y1 represents the average expected t t t i=0 t+i short rates over the next five years, and y 20,tp ≜ y20 − y 20,eh is the 5-year term premium. We t t t measure the term spread as the difference between the 5-year Treasury yield and the 3-month Tbill yield, where the 5-year yield is chosen to match the average maturity of banks’ assets. In 20,tp the second specification, we decompose the term spread into a term premium component (y ), t whichrepresentscompensationtoinvestorsforbearingtheinterestraterisksoverthishorizon;and theexpectationscomponentofthespread(y 20,eh−y1),whichreflectsexpectedchangesintheshort t t rateoverthenextfiveyearsandiscomputedasthedifferencebetweenthetermspreadandtheterm premium estimates. Term premium estimates come from the no-arbitrage term structure model in KimandWright(2005)andisourbaselinetermpremiumseriesinthepaper. Tocontrolforcurrent and expected future macro conditions, we include in X the median forecasts of one-year ahead t realGDPgrowthfromtheSurveyofProfessionalForecasters(SPF)andtheexcessbondpremium (Gilchrist and Zakrajs˘ek, 2012). The dependent variables Z are the four-quarter ahead real t,t+4 GDPgrowthandrealbankloangrowth,andtwomeasuresofbankprofitability(NIMsandROE). Table 2 presents the regression estimates, based on Ordinary Least Squares (OLS), and offers two takeaways. First, the term spread consistently predicts higher future GDP growth, bank loan growth,andbankprofitability,aftercontrollingforshortratechangesandconcurrentmacroeconomic factors. Second,acrossspecifications,thetermpremiumcomponentofthetermspreadisstatistically significant (at conventional levels) in predicting these outcome variables. These results suggest a possible channel through which term spreads predict future growth that goes beyond mere expectations about the future state of the economy. We hypothesize this is an expected bank profitability channel by which a rise in the term premium incentivizes banks to invest in longerterm assets so as to take advantage of higher expected excess returns. In the following section, we present a portfolio model for banks to develop the intuition for this channel and describe the identificationstrategiestoempiricallydocumentthechannel. 6

3 Model 3.1 Setup We present a dynamic partial equilibrium banking model with the objective of understanding how fluctuations in term premium affect banks’ lending decisions. To this end, we use a simple setup witharepresentativebankertakingpricesasgivenandmaximizesthevalueofthebank,subjectto financingconstraints. State of the economy. Time is continuous and denoted by t > 0. There is a pricing kernel, m > 0,thatcapturesthestateoftheeconomy, t dm t = −r dt−κ dW −gdW , (3) t t r,t κ,t m t with √ dr = λ (r−r )dt+σ r dW , t r t r t r,t dκ = λ (κ−κ )dt+σ dW , t κ t κ κ,t where W and W are aggregate Brownian motions representing interest rate (r ) shocks and r,t κ,t t termpremium(κ )shocks,respectively,withaninstantaneouscorrelationof φ .4 t rκ Prices. We use the pricing kernel (3) to price long-term loans. To simplify the analysis and avoid keeping track of the entire maturity structure of loans when solving the bank’s problem, we assume there is a single loan paying coupons τe −τt each instant, corresponding to a duration of 1/τ. Additionally, we assume loans cannot be defaulted on.5 We denote the loan price by P (τ) , t whichisgivenbythediscountedpresentvalueofitsdividends: P (τ) = E (cid:90) ∞ m s τe −τ(s−t) ds. (4) t t m t t 4TheinterestratemodelweuseissimilartothatofCoxetal.(1985). 5Assumingaconstantandnon-zerodefaultprobabilitydoesnotchangetheanalysis. Themodelcouldbeextended toaddatime-varyingdefaultprobabilitytoloans. 7

and is a function of the state variables, r and κ. Using Feymann-Kac we solve the conditional expectationasapartialdifferentialequation: (cid:34) (cid:35) (cid:18) (cid:19) (τ) (τ) (τ) (τ) (τ) τ P 1P P 1P P −τ−r dt+E r dr+ rr dr2+ κ dκ+ κκ dκ2+ κr dκdr = P(τ) t P(τ) 2P(τ) P(τ) 2P(τ) P(τ) (cid:32) (cid:33) dmdP(τ) −cov , t m P(τ) wherethetermpremiumisgivenby (cid:32) (cid:33) (cid:32) (cid:32) (cid:33) (cid:33) dmdP(τ) P (τ) √ P (τ) P (τ) √ P (τ) −cov = κ r σ r+g κ σ + r σ rg+ κ σ κ φ dt. (5) t m P(τ) t P(τ) r P(τ) κ P(τ) r P(τ) κ t rκ Banks. Banks take prices as given and can trade 3 instruments: long-term loans, deposits, and fedfunds. Thebalancesheetisgivenby n +(cid:101)b = x (τ) P (τ) +b , (6) t t t t t (τ) (τ) where n is the wealth of the bank, x is the number of loans at price P , while b and (cid:101)b are t t t t t thevalueofthefedfundsanddepositaccounts,respectively. Theonlydifferencebetweendeposits andfedfundsisthatbankspayalowerrateondepositsthanthefederalfundsrate. Thatis,thefed fundsaccountfollows db = r b dt, t t t andthedepositaccountfollows d(cid:101)b = ϕ(r )(cid:101)b dt, t t t with ϕ(r ) ≤ r representing the fact that banks have market power in the deposit market and pay t t a rate lower than the federal funds rate (Drechsler et al., 2017, 2021).6 The evolution of bank’s wealthisthengivenby dn = x (τ) dP (τ) +db +d(cid:101)b , t t t t t (cid:32) (τ) (cid:33) (cid:16) (cid:17) dP = r n +(ϕ(r )−r )(cid:101)b dt+P (τ) x (τ) t −r dt . t t t t t t t (τ) t P t 6Wespecifythefunctionϕ(r )below. t 8

Banks’ optimization problem. We follow the basic banking structure proposed in Gertler and Kiyotaki(2015)(henceforthGK15). BankspaydividendsexogenouslywithaPoissonprobability λ. As argued in GK15, the purpose of this simple dividend policy is to avoid banks growing out oftheirincentiveconstraint. Weassumeanewgroupofbankersusethenetworthasinitialcapital to restart operations.7 The bank’s problem is to maximize the expected discounted value of the dividendsusing,asinGK15,theaggregatestochasticdiscountfactor (cid:90) ∞ m V = max E s λe −λ(s−t) n ds, t t s (cid:110) x t (τ) ,(cid:101)bt (cid:111) t m t subjectto (cid:32) (τ) (cid:33) (cid:16) (cid:17) dP dn = r n +(ϕ(r )−r )(cid:101)b dt+P (τ) x (τ) t −r dt , (7) t t t t t t t t (τ) t P t V ≥ ρP (τ) x (τ) , (8) t t t (cid:101)b ≤ −δn . (9) t t Constraint (8), as in GK15, is an incentive constraint motivated by a moral hazard problem and implies that the value of the bank V has to be greater than or equal to a fraction ρ of the bank’s t total assets, P (τ) x (τ) . Because banks can earn a positive spread on deposits, r −ϕ(r ), they will t t t t have incentive to issue as many deposits as possible to buy reserves. To avoid this outcome, we imposealeverageconstraintondeposits(9). Recursiveformulation. Wewritetheproblemrecursively (cid:104) (cid:16) (cid:17)(cid:105) 0 = max m λe −λtn dt+E d m e −λtV , (10) t t t t t (cid:110) (τ) (cid:111) x t ,(cid:101)bt subject to (7), (8), and (9). Because the objective function and the constraints are linear in net worth, the solution takes the form of V = ψ n . The variable ψ (κ ,r ) represents the bank’s t t t t t t marginal value of wealth or “Tobin’s Q” (see GK15). Then, the problem can be written as the 7In general equilibrium models, banks pay an aggregate dividend and receive a different amount of resources as startupcapitalinothertoobtainaninvariantdistributionofwealthintheeconomy. Inourpartialequilibriumsetup, however,thewealthdistributionisnotdeterminedandhenceweassumethatalldividendpaymentsareusedasstartup capital,withoutlossofgenerality. 9

followingpartialdifferentialequationfor ψ (κ ,r ): t t t (cid:20) (cid:21) λ−λψ dm dn dψ dψdn dψdm dmdn 0 = max t dt+E + + + + + , t (cid:110) x t (τ) ,(cid:101)bt (cid:111) ψ t m n ψ ψ n ψ m m n subjectto(7),(8),and(9). 3.2 Model calibration and solution We solve for a numerical specification in which the incentive and deposit leverage constraints are always binding. This means V = ρP (τ) x (τ) and (cid:101)b = −δn . As discussed in GK15, the t t t t t incentive constraint is always binding as long as the risky asset yields a positive excess return in equilibrium. Additionally, the leverage constraint on deposits is always binding because the depositspread r−ϕ(r) isalwayspositive. Calibration. We calibrate the processes r and κ using simulated method of moments to match t t the statistical properties of the short interest rate and term premium that we use in the empirical part of the paper. More precisely, we set the parameters {r,κ,σ ,σ ,λ ,λ } to match the mean, r κ r κ the standard deviation, and persistence of the time series of the short-term interest rate and term premiumasdescribedintheempiricalsection. Forsimplicity,wesetthecorrelationbetweenterm premiumandinterestrateshockstobezero. For banks, we calibrate the parameters primarily following the literature. In particular, we set the values of λ and ρ to be the same as GK15, and we set c to match the same average Tobin’s Q in GK15. We use the deposit spread from Drechsler et al. (2021), who show that an increase in the short rate by 100 bps translates into an increase in the average deposit rate by 35 bps. Finally, we set δ to 2.85, which is the average short-term deposits-to-total equity capital ratio in the Y9-C database. Numericalresults: Policyfunctions. Figure1showsthemodel’ssolution. Allthepanelsinthe figure have the state variable κ , which drives term premium fluctuations, in the horizontal axis. t Eachpaneldisplaysasolidblueline,representingthesolutionwhentheshortrater isatitsmean, aswellasadashedyellowlineandadottedredline,representingsolutionswhenr istwostandard deviationsaboveorbelowitsmean,respectively. The upper-left panel shows the term premium, given in equation (5). A more negative κ 10

corresponds to a higher term premium. Intuitively, this is because as the diffusion component of the stochastic discount factor (κ) becomes larger in magnitude, an interest rate shock affects valuations relatively more. Hence, a more negative κ translates into higher term premium and lower marked-to-market loan prices. In addition, as shown by the difference between the yellow dashed and the red dotted line, the term premium is also affected by the level of the short rate because r becomes more volatile as the level of rate increases, thus increasing the quantity of t interest-raterisk.8 The upper-right panel shows banks’ marginal value of wealth (or “Tobin’s Q”), ψ = V/n. A higher ψ means banks value an extra unit of wealth relatively more. Notice that a higher ψ correspondstostatesinwhichthetermpremiumishigherandthemarked-to-marketvalueofloan pricesarelower. Themiddle-leftpanelshowstheexpectedreturnonwealth,givenby (cid:20) (cid:21) dn µ ≡ E t /dt = r [1+(1−ϕ)δ]+α TP (κ ,r ), (11) n,t t t t t t n t where TP (κ ,r ) is the term premium. The expected return on wealth is increasing in the level t t t of term premium as well as in the level of interest rates. This is because a higher term premium translatesintoahigherexpectedexcessreturnonlending(andhencefutureprofits)whileahigher level of rates translates into higher profits from deposit making. Additionally, as shown in the middle-right panel, banks’ leverage on loans, α = x (τ) P (τ) /n , is also increasing in the term t t t t premium. Thisisbecauseαispinneddownbytheincentiveconstraint,andthereforeisproportional to ψ, which is increasing in term premiums as discussed earlier. Together, higher leverage and highertermpremiumimpliesahigherexpectedexcessreturnonwealthwhen κ islow. Finally, the bottom two panels show the solutions for bank lending. The level of lending is pinneddownbytheincentiveconstraint. Thatis,thetotalvalueoftheloanportfolioisgivenby L = P (τ) x (τ) = V t = 1 ψ n . t t t ρ ρ t t Then,applyingIto’slemmato L ,wehavethatlendinggrowthis9 t dL dψ dn dψ dn t = t + t + t t . (12) L ψ n ψ n t t t t t 8Thefactthatthelevelofratesaffectthevolatilityofratesandhencethequantityofriskisamechanicalimplication ofthesquarerootmodel. 9Westudyloangrowthbecausethelevelofloansisnon-stationary. 11

Thedriftof(12)representstheexpectedloangrowth, (cid:20) (cid:21) (cid:20) (cid:21) (cid:20) (cid:21) dψ dn dψ dn µ ≡ E t +E t +E t t , (13) L,t t t t ψ n ψ n t t t t and is shown in the lower-left panel of Figure (1). The second term on the right hand side of equation (13) is the expected return on wealth and, as discussed above, is increasing in term (cid:104) (cid:105) premium. However, the first term, the expected change of Tobin’s Q, E dψt , is decreasing in t ψt term premium. This is because ψ itself is increasing in term premium but stationary; therefore, whentermpremiumincreases,ψalsorisesbutisexpectedtomeanreverttoitsoriginallowerlevel in the future, hence displays a negative expected change when term premium increases. In our calibration,theeffectfromanincreaseinexpectedreturnonwealthmorethanoffsetsthatfroman expected decrease in ψ. Thus, the expected loan growth rises with term premium.10 Intuitively, thismeansthateventhoughahighertermpremiumreducesloanpricesandcausesbankstobecome relativelymore constrained(i.e., banks’marginal valuation ofa unitof wealth, ψ, increases),both of which would dampen lending growth, a higher expected future excess return on loans would dominateandincentivizebankstoincreasetheirlendinginthenearfuture. Onimpact,though,anegativeshocktoκ (i.e.,anincreaseinthetermpremium)causesaslight t decrease in the amount of loans. This is because the diffusion component associated to κ shocks, t σ ,displayedinthelower-middlepanel,isslightlypositive. Thediffusion σ isgivenby Lκ,t Lκ,t (cid:32) (cid:33) (τ) ψ P σ = κ +α κ σ . (14) Lκ,t ψ t P(τ) κ In expression (14), the derivative ψ is negative (i.e., higher κ means a lower term premium κ t and,hence, a lower ψ) while the derivative P is positive (i.e., higher κ, lower term premium, and, κ hence,higherloanprices). Inthebaselinecalibration,themarked-to-marketlossesinloansdueto anincreasein κ ,thatis P > 0,dominatesand σ ispositive. t κ Lκ,t Numericalresults: Impulseresponses. Figure2showsthemodel’simpulse-responsesfunctions toanegativeshocktoκthatcausesanapproximately100basispointsincreaseinthetermpremium. We study the responses of the model in the baseline calibration (shown in blue) as well as for differentvaluesofρ—theparametercapturingthestrengthoftheincentiveconstraint(8)—withthe objectiveofstudyingthemodel’spredictionsforbankswithdifferentcharacteristics. Inparticular, (cid:104) (cid:105) 10Thethirdtermofequation(13),E dψtdnt ,isrelativelysmallanddoesnotmateriallyaffecttheresults. t ψt nt 12

banks with a lower ρ display a higher equilibrium Tobin’s Q, which translates into a higher level of leverage. Banks that have a lower ρ, and thus higher leverage and higher Tobin’s Q, have less capitaland,consequently,valueamarginalunitofwealthmorethanbankswithhigh ρ.11 As elaborated in the discussions above about the policy functions, an increase in the term premium increases banks’ return on wealth (middle-right panel) and, as a consequence, increases loan growth (middle-center panel). On impact, though, loan growth and returns on wealth are slightly negative because of the marked-to-market losses in the loans’ portfolio as loan prices decline when the shock hit. However, because expected return on wealth increases through higher leverage, α (upper-right panel), coupled with higher term premium, loan growth increases in the near horizon after the term premium shock. Tobin’s Q, ψ (middle-left panel), also increases following the term premium shock, which means banks value an extra unit of wealth relatively moreastheexpectedreturnonmaturitytransformationincreases. Importantly,asshownbythered-dottedlines,themodelpredictsthatbankswithalowerρ(i.e., banksthathavehigherleverage,arelesswellcapitalized,andthusvalueamarginalunitofwealth more)displayarelativelystrongerreactiontotermpremiumshocksthaninthebaselinecalibration. The main force driving this result is that low-ρ banks display higher leverage in equilibrium and, therefore,theincreaseintermpremiumhasastrongereffectovertheirreturnonwealth,asshown in (11). The stronger increase in the return on wealth translates into stronger lending growth, as shown in (13). Notice that, although the level of ψ changes with ρ, the expected changes in ψ (cid:104) (cid:105) (E dψt )donotchangemuchforbankswithdifferent ρ.12 t ψt Themodeldeliversthefollowingtestableimplications: Testableimplication1 Banksrespondtohigherexpectedexcessreturnsonmaturitytransformation byincreasingloansupply. Testableimplication2 More leveraged banks increase loan supply relatively more because they experience a stronger increase in expected profitability following a rise in expected excess returns onmaturitytransformationthanotherbanks. 11Appendix Figure B-1 shows the relationship between leverage, Tobin’s Q , and ρ. Intuitively, in the context of themoralhazardproblempresentedinGK15,bankswithahigherρcanobtainarelativelylargerbenefitofdiverting assetsfromthebank. Asaresult,thosebanksmustfundtheirassetholdingswitharelativelylargerfractionoftheir ownequity,henceexhibitingalowerleverageandalowervaluationforanextraunitofwealththanbankswithalower ρ. 12Thisisbecauseψdonotdisplaystrongnonlinearitiesacrossthestatespace. Theexpectedchangeinψisdriven bythesemi-elasticitiesψ /ψ,ψ /ψ,ψ /ψ,ψ /ψandψ /ψ,whichareconstantsifψlinearfunctionofthestate r rr κ κκ rκ variables. 13

4 Empirical Strategy To test the main implications of the model, we need to isolate the effects of changes in term premiums on bank lending and profitability. Our analysis faces the econometric challenge that variablesthataredifficulttoobserveandmeasure,suchasexpectationsabouttheeconomicoutlook or uncertainty, might simultaneously affect term premiums and bank behaviors. Such variables could cause an a omitted variable bias and lead to spurious results. We take two concrete steps to mitigate this concern: First, we develop an instrumental variable (IV) approach that isolates variationsintermpremiumsthatareduetofactorsarguablyunrelatedtotheoutlookofthedomestic economy. Second, we exploit plausibly exogenous variation in term premiums borrowing from the literature on high-frequency event study identification of monetary policy shocks. Next we describethetwoapproachesindetail. 4.1 Instrumental variable To identify the effect of changes in term premiums on bank outcomes, we need an instrument thataffectstermpremiumsforreasonsunrelatedtothefutureeconomicoutlook. Theinstrumental variablewefocusonistheforeignofficial(i.e.,centralbanksandforeignexchangereservemanagers) holdingsofTreasurysecurities,normalizedbyU.S.GDP.WearguethatforeignofficialTreasuries holdings are unlikely to be correlated with U.S. economic conditions based on evidence from a large literature on the effects of foreign investor demand for U.S. Treasury securities on Treasury yields (see, e.g., Bernanke et al. (2004), Warnock and Warnock (2009), Beltran et al. (2013), KaminskaandZinna(2020),andAhmedandRebucci(2022)). Thesestudiespostulatethatdemand byforeigninvestors,especiallyforeignreservemanagersandotherofficialaccounts,areprimarily driven by their foreign reserve management and foreign exchange intervention needs rather than profit motives.13 Tabova and Warnock (2021) use annual confidential surveys on security-level foreign holdings of U.S. Treasuries to compare returns to different types of investors and find evidence that foreign officials are less price sensitive than domestic and foreign private investors. Therefore,variationsinforeignofficialholdingscanbeviewedas“shocks”tothedemandforU.S. TreasuriesthatwouldaffectTreasuryyieldsandtermpremiums.14 13See also the literature on capital inflows into the U.S. from foreigners seeking U.S. assets to store value, e.g., CaballeroandKrishnamurthy(2009)andCaballeroetal.(2008)andexplanationsrelatedtotheglobalsavingsglut, summarizedinBernanke(2005). 14Our approach is similar to Krishnamurthy and Vissing-Jorgensen (2015), who use the rapid increase in foreign officialholdingsofTreasuriessincetheearly1970sasashocktothesupplyofTreasuriesavailabletoprivateinvestors 14

Figure 3 shows that our instrument is negatively correlated with the term premium over the sampleperiodbetween1994and2019. Thisnegativecorrelationisespeciallynotableduring2004– 2006 when long-term rates hardly moved despite rising short rates, which former Federal Reserve ChairsAlanGreenspanfamouslycalledaconundrum(Greenspan,2005)andBenBernankeattributed totheglobalsavingsglut(Bernanke,2005). Moreformally,weshowthattheinstrumentisastrong predictoroftermpremiumsinTableOA-1,whereweregressthe5-yearKim-Wrighttermpremium ontheforeignofficialholdingsmeasure,togetherwiththeshortrateeitheraloneorwithadditional macro variables. The estimates in columns 1–2 indicate that higher demand by foreign official investors for Treasury securities is associated with statistically significant lower term premiums, after controlling for current macro conditions. As seen in columns 3–6, the same pattern holds if we replace the Kim-Wright term premium (our baseline measure) with term premium series from alternative term structure models of Adrian et al. (2013) and Kim and Priebsch (2020), which we describeinmoredetailbelow. 4.2 Term premium shocks A second method we use to isolate plausibly exogenous variations in term premiums is based on high-frequency interest rate changes around important monetary policy events. Changes over narrowwindowsaroundsucheventscanbethoughtofasreflectingpredominantlysurprisesassociated withthemonetarypolicyannouncements,ratherthanothermacroeconomicnews. Theeventstudy approachfortheidentificationofmonetarypolicyshockswaspioneeredbyKuttner(2001)andhas becomethestandardapproachinempiricalmonetarypolicystudies. We proceed in three steps. First, we calculate one-day changes in the 3-month Tbill yield and in the expectations and term premiums components (based on the Kim-Wright model) of the 3-month/5-year term spread on days when FOMC statements and minutes are released. Second, we follow Miranda-Agrippino and Ricco (2021) in regressing those event-day changes on past Greenbook/Tealbookforecastsandkeepingtheresidualsasameasureof“trueshocks”thatarefree frompotentialFedinformationeffect(thatis,theydonotreflectrevisionsininvestorbeliefsabout thestateoftheeconomy,see,e.g.,RomerandRomer(2000)andNakamuraandSteinsson(2018)). Finally, we convert the cleaned event-day shocks into quarterly series following the procedure in Gertler and Karadi (2015), by accumulating event-day changes over time and calculating the quarterly averages. This procedure gives more weight to events that occur early in the quarter giventhatforeignofficialholdingsareunlikelytobecorrelatedwithU.S.economicconditions. 15

and allows events that occur later in the quarter to affect the shocks series in the current and the following quarter. The quarterly term premium shocks thus calculated are shown in Figure 4, togetherwiththeKim-Wrighttermpremiumestimates. 5 Data U.S. Call Report data. For the bank-level analysis, we use quarterly balance sheet data from the U.S. Call Reports that reflect banks’ domestic operations. The bank-level panel starts in 1994Q1 and ends in 2019Q4. The baseline sample includes about 11,500 commercial banks with headquarters in 409 MSAs, with the number of banks declining between the start and the end of the sample from about 8,700 banks to some 4,350 banks. Our main regression variables are total bank loan growth (including and excluding off-balance sheet credit lines) and bank profitability metrics (NIMs and ROE) as dependent variables and bank size, capital ratio, the share of core deposits in total liabilities, the share of securities in total assets, and the share of reserves in total assets as controls. In additional tests, we use a measure of the degree to which banks are engaged inmaturitytransformation,namelythematuritygapfromEnglishetal.(2018). Y-14QLoan-leveldata. Ourloan-leveldatacomefromthesupervisoryFRY-14QH.1schedule “Wholesalecreditrisk”collectedbytheFederalReservefrombanksthataresubjecttostresstests. These banks have at least $50 billion in total assets during the period of analysis and together account for three-quarters of total U.S. C&I loans (Crosignani et al., 2023; Favara et al., 2021). The data represent individual C&I loan exposures (of at least $1 million) between each reporting bank and individual borrowers. We limit the sample to loans to nonfinancial U.S.-domiciled firms and to 15 reporting banks for which we also have information on GDP growth forecasts from the Blue Chip Economic and Financial Indicators, a crucial control variable. Caglio et al. (2021) combine the Y-14Q data with the U.S. Flow of Funds and show that the Y-14Q borrowers make up 60% of nonfinancial business debt liabilities. For each BHC’s main commercial bank we have balancesheetdatafromtheCallReport. The Y-14Q data contain detailed information about each loan contract, including the amount (in US$), loan type (credit lines, term loans), interest rate, and maturity. A unique feature of the data is that it includes borrower-level industry, location, and financial information as reported by the lenders (mostly on a yearly basis). For each firm we observe balance sheet variables such as totalassets,fixedassets,debt,cashholdings,tangibility,interestcoverageratios,salesgrowth,and 16

selectincomestatementitemssuchascapitalexpenditure. Ourextensivemarginregressionsample for the period 2013Q1–2019Q4 contains loans extended to more than 25,000 nonfinancial firms. We use the Y-14Q data in two ways: First, in the loan-level dataset we examine banks’ lending decisions in response to changes in the term premium. Second, we extract the firm-level balance sheet information and set it up as a firm-year panel to examine the effects of changes in the term premiumonfirms’investmentdecisions.15 Macroeconomic data. We use a collection of macroeconomic variables in both time-series and bank-levelregressions. Forthetermspread,weusetheFama3-monthyieldandFama-Bliss5-year zero coupon Treasury yields from Center for Research in Security Prices (CRSP). Term premium estimates are downloaded from the website of the Federal Reserve Board for Kim and Wright (2005), from the website of the FRB New York for Adrian et al. (2013), and were provided by the authors for Kim and Priebsch (2020). Real GDP growth and GDP deflator inflation come from the Federal Reserve Economic Data (FRED) at the FRB St. Louis. The excess bond premium (Gilchrist and Zakrajs˘ek, 2012) is sourced from the Federal Reserve Board (see website). We obtain the one-year ahead real GDP forecast and GDP deflator inflation from the SPF at the FRB Philadelphia (see website). We download total foreign official holdings, as well as Chinese holdings,ofTreasurycouponsecuritiesconstructedbyBertautandJudson(2022)usingdatafrom theTreasuryInternationalCapital(TIC)reportingsystem(seewebsite).16 Wedonothavedataon Chineseofficialholdings. However,asnotedinDepartmentoftheTreasury,FederalReserveBank of New York, and Board of Governors of the Federal Reserve System (2023), emerging country holdingsareheavilyconcentratedintheofficialsector. Summarystatistics. AppendixTableC-1reportssummarystatisticsfordatausedinbank-,loanand firm-level analyses. In the bank-level panel, average loan growth during 1994Q1-2019Q4 is 3.8%, average NIMs are 1% and average ROE is 2.3%. In the time series, the Kim-Wright term premium series has an average value of 0.3% and a standard deviation of 0.44% (in the banklevel regression sample, these values are 0.1% and 0.5%, respectively). The Kim-Wright term premiumshockseriesiscenteredonzeroandhasastandarddeviationof0.005%. Forcalculations ofeconomicmagnitudes,weusechangesinonestandarddeviationofthetermpremiumseriesand shockrepresenting50bpsand5bps,respectively. 15ReportingformsanddatadictionariesareavailableontheFederalReservewebsite. 16AlsoseeBertautandJudson(2014)andBertautandTryon(2007). 17

6 Results 6.1 Bank-level analysis over 1994–2019 The goal is to test the model’s implications relating bank lending outcomes to changes in the term premium. For testable implication 1, we use the following specification estimated in bank-quarter paneldata: (cid:16) (cid:17) Loan growthi = α +β ∆ y1+β y 20,eh −y1 +β y 20,tp +τX +γZ +εi (15) t,t+4 i 1 t 2 t t 3 t t it t,t+4 where the dependent variable Loan growthi is the four-quarter ahead loan growth at bank i t,t+4 (excluding and including off-balance sheet credit lines). For the term premium, we employ three estimationstrategies: OLSwiththeKim-Wrighttermpremiumestimates,theIVstrategyoutlined inSection4.1,andOLSwiththeKim-WrighttermpremiumshocksdescribedinSection4.2. The vector of macro controls X includes short-rate changes, excess bond premium, realized real GDP t growth and GDP deflator inflation, and survey forecasts of one-year ahead real GDP growth and GDP deflator inflation. We include bank fixed effects (α ) and bank MSA fixed effects (where the i MSAreferstothebank’sheadquarterslocation)tocontrolforunobservedlocalshockstoallbanks. Furthermore, we include a vector of standard bank-level determinants of lending Z comprising it banksize(logshareofthebank’sassetsintotalbankingsectorassets),capitalratio(ratioofequity tototalassets),coredepositsasashareoftotalliabilities,andsecurities-to-assetratio.17 Theratio ofsecuritiestoassetsaimstocapturepotentialvaluationeffectsfromchangesintheyieldcurve— forinstance,thatadeclineinlong-termrateswouldincreasethevalueofassetsthataremarkedto market in a bank’s balance sheet and boost earnings, hence capital and lending capacity. Testable implication1indicatesthatthecoefficientofinterest β shouldbepositive. 3 Bank lending. In Table 3 we report our main results based on the longest time periods afforded by the data availability. Across specifications, the estimation results indicate that a higher term premium is positively associated with subsequent loan growth. OLS estimates for the simple term premiumseriesarereportedforreferenceincolumns1and3andhavepositivesigns. Focusingon wellidentifiedeffects,theIVestimatesincolumns2and5suggestthatariseinthetermpremium inducedbylowerforeignofficialdemandforTreasurysecuritiesisfollowedbysignificantlyhigher 17Inthebaselineanalysisthecapitalratioisgivenbytheshareofequityoftotalassetsbuttheresultsareinvariant tousingtheregulatoryTier1ratio(Tier1capitaldividedbyrisk-weightedassets). 18

loan growth over the following four quarters. Furthermore, the first-stage F-test for instrument relevance is above 100, suggesting a strong instrument (Lee et al., 2022). The OLS estimates in columns 3 and 6 similarly suggest that a positive term premium shock is associated with higher loangrowth,consistentwithmodelprediction1. Acrossspecifications,thecoefficientestimatesarestatisticallysignificantatconventionallevels and deliver similar economic magnitudes. The estimates in columns 2 and 3 indicate that a one standard deviation increase in the term premium (of 50 and 5 bps, respectively) is associated with an increase in loan growth by 1.1–1.2 ppts, which is an economically sizable effect given that averageloangrowthoverthesampleperiodis3.8%anditsstandarddeviationis1.5%. Bycontrast, the expectations component is less important in predicting bank loan growth, but indicates that expectations of higher short rates are generally associated with lower bank lending. The fact that the term premium and expectations components of the term spread have opposite effects on bank loan growth highlights the importance of this decomposition for understanding the channels throughwhichbanksreacttochangesinlongtermrates. Role of bank leverage. We now examine the model prediction 2 that the response of bank lending to a term premium shock is more pronounced for more leveraged banks, as they value an extraunitofwealthandthusreacttoanincreaseinexpectedreturnsfrommaturitytransformation relatively more. We measure bank leverage with the share of equity capital in total assets and estimate the baseline specifications with an interaction of the term premium with the capital ratio. Given that we focus on estimating differential effects in bank responses to fluctuations in the term premium,weareabletoincludequarterlyfixedeffectsandhenceabsorbtheleveleffectsofmacro shocksonbankbehaviors(includingthatofthetermpremiumitself). Inaddition,tomakesurethat the interaction term of the capital ratio with the term premium is not confounded by other bankspecificormacroeconomicfactors,wesaturatethespecificationwithinteractiontermsbetweenall bankcharacteristicsandthetermpremium,expectations,andshort-ratechanges. The estimates are reported in Table 4 and consistently show that an increase in identified term premium(eitherinducedbyhigherforeignofficialdemandforTreasuriesorhigh-frequencyevent study-based monetary policy shocks) is associated with greater loan growth at banks with lower capital. All estimates on the interaction of the capital ratio and the term premium are statistically significant. In terms of economic magnitude, estimates in columns 2 and 3 indicate that a bank at the 25th percentile of the capital ratio distribution (8%) exhibits loan growth higher by 0.9–1.3 ppts (IV) or 0.6–0.8 ppts higher (term premium shock) than a bank at the 75th percentile of the 19

capitaldistribution(12%),implyingadifferentialeffectof0.3–0.4pptsdependingontheestimate. (Once again, this economic effect is sizable given that the average and standard deviation of loan growthoverthesampleperiodare3.8%and1.5%.) Robustness. We subject the baseline results to three robustness tests. First, we examine the stability of coefficients to a longer time period and to slight changes in the set of bank controls. Specifically, we replace the share of core deposits with that of total deposits in liabilities and we removetheshareofsecuritiesasacontrolvariable(thesechangesaredemandedbythelengthening of the sample period). The results are shown in Table OA-2 and reveal that the level effects of the term premium on bank loan growth (with or without off balance sheet credit lines) are robust to thesechanges,andthatthefirststageF-statisticsinIVestimationsremainabove100. In a second test, we focus on the instrumental variable and remove China’s holdings of U.S. Treasuries from the construction of the total foreign official holdings of U.S. Treasuries. This change is motivated by the potential concern that Chinese official demand for Treasuries may be correlated with the U.S. economic outlook given the strong economic linkages between the two economies. InTableOA-3wereporttheresultsandnoticethatthenewdefinitionoftheIVleaves our conclusions unchanged: across specifications, a higher term premium due to lower demand fromforeignofficialinvestorsforTreasurysecurities(excludingChina)significantlyboostsgrowth inbanks’totalloanbooks.18 Third, we verify that the main results are not driven by the choice of term structure model. The motivation for this exercise is that term premium estimates are subject to some degree of uncertainty regarding the underlying econometric model and its parameters. We focus on two alternativetermpremiumestimatesfromAdrianetal.(2013)(ACM)andKimandPriebsch(2020), respectively. The ACM differs from the Kim-Wright model in that it does not incorporate survey measuresofshortrateexpectationsinthedatasetusedtoestimatethemodel(foracomparisonof thetwomodels,seeLietal.(2017)). TheKim-Priebschmodelhastheadvantageofformallytaking intoaccountthe“effectivelowerbound”periodfrom2008to2015whichoverlapswithoursample period. We repeat the analysis using these alternative term premium series and find, first, that our instrumentalvariable,foreignofficialTreasurysecurityholdings,hasstrongexplanatorypowerfor thealternativetermpremiumestimatesaboveandbeyondtheshortrateandothermacroeconomic factors (Table OA-1, columns 3-6). Second, the estimates for our baseline specifications, as 18MarketcommentariessuggestedthatChinalikelyheldasignificantamountofU.S.TreasuriesoffshoreinBelgium around2013–2015. OurresultsareinvarianttoremovingholdingsbyBelgiumfromtheinstrumentalvariable. 20

shown in Table OA-4, reveal our conclusions are robust to using the ACM and Kim-Priebsch termpremiumestimates. 6.2 Loan-level analysis around the Taper Tantrum Inthissection,wetestthemodel’sprediction2regardingheterogeneousresponsesofbanklending to term premiums depending on ex-ante bank leverage using granular data on bank-firm lending relationships from the Y-14Q dataset. A key advantage of the loan-level data is that it allows us to examine bank lending decisions following a change in the term premium while holding borrower-level loan demand constant each quarter. From the Y-14Q data we extract information on outstanding loans and on new loan originations to nonfinancial domestic firms in a symmetric windowofuptofivequartersaround2013Q2,whenthetapertantrumstarted. For identification, we exploit the unexpected nature of the “taper tantrum” that followed the May22,2013remarksbyChairBenBernankeintheQ&Asessionafterhissemiannualcongressional testimony, indicating that the Federal Reserve might “step down” the pace of its quantitative easing program “in the next few meetings.”19 Following the speech, Treasury bond yields and term premiums surged unexpectedly (Chari et al., 2021), as depicted in Figure 5, inducing a large monetary policy shock (Bernanke, 2015). We take advantage of this unique episode to analyze bank lending behaviors following an unexpected and dramatic rise in the term premium. Identificationofthetermpremium’seffectonbanklendingrequiresthatthetaperingannouncement wasnotdrivenbyexpectationsaboutthefuturestateoftheeconomy,inparticular,byanimproving economic outlook. To mitigate this possibility, we limit the sample to the banks that participate in the Blue Chip Survey and control for their quarterly one-year ahead GDP growth forecasts before and after the taper tantrum. It is also noteworthy that market commentators were skeptical about the economic outlook around that time and believed that, if anything, the Fed’s reaction function waslessaccommodativethanexpected.20 Wespecifythefollowingregressionmodelinadifference-in-differenceframework: Loan outcome = β Capital ratio ×Post +γ ′ X +γ ′ X ×Post + bj,t+k 1 b,2012 t 1 bt 2 bt t (16) +γ ′ Z +δ +ε 3 bj jt bjt whereLoanoutcome referstoanindicatorfornewlyoriginatedloansinthestockofoutstanding bj,t+k 19MinutesoftheFOMC’sApril30-May1meetingwerereleasedlaterthatday. 20See Sinha and Smolyansky (2022) for a detailed discussion of market and policymakers’ narratives around the tapertantrum. 21

loansfrombankbtofirm jatt+kduringaperiodofk = 3,4,5quartersbeforeandafter2013Q2. We also consider lending outcomes for newly originated loans, where Loan outcome is the bj,t+k loan amount (log) or the loan spread. Post is a dummy variable that takes value one starting in t+k 2013Q3 for up to k = 5 quarters subsequently (and zero in a symmetric period before 2013Q2). We drop the loans reported in 2013Q2 to clearly separate the periods before and after the May 2013event. Asinthebank-levelregressions,wecontrolforstandarddeterminantsofbanklending (in vector X ) such as bank size (log-assets), the share of core deposits in total liabilities, and bt securities-to-asset ratio, both in levels and in interactions with the Post dummy. The capital ratio t ismeasuredatend-2012soitispredeterminedwithrespecttotheriseinthetermpremium. Critically, these specifications include the bank-level one-year ahead GDP growth forecasts from the Blue Chip surveys, in levels and interactions with Post . These forecasts account for t the baseline level of banks’ expectations of the economic outlook, an important driver of lending decisions, as well as any revision to those expectations following the tapering announcement. We add a pair-level variable representing the duration of the banking relationship as of end-2012 (Z ) to capture the effect of relationship lending on loan outcomes and to mitigate potential bj concernsofassortativematchingbetweenbankandfirms.21 Thespecificationalsoaddsinteracted borrower×quarter fixed effects (δ ) to control for loan demand changes and other unobserved jt factors at the firm level (Khwaja and Mian (2008); Jiménez et al. (2020)). The estimates are obtainedusingOLSwithstandarderrorsthatareclusteredatthebank-firmlevel. Modelprediction 2indicatesthatthecoefficientofinterestβ shouldbenegativetoreflectthatmoreleveraged(lower 1 capital)banksrespondtoanincreaseintheriskpremiumbyincreasingloansupplyrelativelymore. Bank lending. The regression estimates are reported in Table 5. Across dependent variables (panels A-C) and time periods between three and five quarters around 2013Q2 (columns 1-3), the estimated coefficients on Capital ratio ×Post are statistically significant at conventional levels b t andindicatearobustrelationbetweenbankleverageandthesupplyofbankloansafterthestartof thetapertantrum. InpanelsA-B,theestimatessuggestthat,intheperiodaftertheMay2013event, moreleveragedbankshaveahigherlikelihoodofgrantingnewloansand,conditionalonapproved loans,theygrantlargerloansthanotherbanks. TheresultsinpanelCstrengthentheinterpretation of these estimates as supply-side effects as they show that more leveraged banks are equally more likelytoreduceloanspreadsonnewloansthanotherbanks. 21Ideally,wewouldprefertoincludebank×firmfixedeffectsbutoursampleperiodsofseveralquartersaroundthe tapertantrumaretooshorttoobservemultipleneworiginationsfromagivenbanktoagivenborrower. 22

The estimated relationship between bank leverage and loan supply during the Post period is t economically sizable. Comparing banks with capital ratios at the 25th and 75th percentiles of the distribution (10% vs. 14%), the estimates in column 2 of panels B-C corresponding to lending outcomes in a period of four quarters before and after 2013Q2 indicate that lower capital banks grantednewloansthatwerelargerby11%andlessexpensiveby7bps. Consistentwithourbanklevel evidence from Table 4, these findings support the model’s implication 2 that term premiums haveheterogeneouseffectsonbanklendingdependingonbankleverage. Robustness. We check the robustness of our results in two ways: (a) a placebo test and (b) a specification that accounts for banks’ large reserve accumulation before the the taper tantrum. First, we check the validity of the identifying assumption of “parallel trends” in lending by banks withdifferentlevelsofleverage. Forthispurpose,weidentifyaperiodwhenthetermpremiumwas verystableandexaminechangesinbanks’lendingdecisionsbeforeandafteracutoffpointduring this period. Given that the Y-14Q data start in 2012Q1, we settle on the period between 2012Q2 and 2013Q1 (see Figure OA-1). We run regressions in two sample periods: 2012Q4 vs 2013Q1 and 2012Q2-2013Q3 vs 2012Q4-2013Q1. The specifications are similar to those in Table 5 but the capital ratio is measured in 2012Q1 so it is predetermined relative to lending outcomes in the two placebo periods. The regression estimates are reported in Table OA-5 and show no evidence of pre-existing trends across specifications, as ex-ante bank leverage has statistically insignificant effectontheextensiveandintensivemarginsoflending. Second, we address the concern that our results may pick up lending effects of bank reserves (depositsattheFederalReserve)accumulatedduringtheFed’squantitativeeasingprogramsbefore andduringthetapertantrum. QEprogramswereimplementedstartinginNovember2008,August 2010,andSeptember2012,andresultedinanunprecedentedinjectionofreservesintothebanking system. Omitting reserves from the specifications may induce a bias on the estimated coefficient on Capital ratio × Post whose sign depends on the true effect of reserves on lending and its b t correlation with bank leverage.22 In Table OA-6 we estimate the main specifications in Table 5 where we add the share of reserves in total assets in level and interacted with Post . The results t showthatourmainestimatesarerobusttotheinclusionofthisadditionalvariable. 22Reserves may influence bank lending through different channels, for instance, by affecting relative asset prices and inducing portfolio reallocations, through balance sheet costs, or by influencing aggregate demand. Empirically, Kandrac and Schlusche (2021) document a positive effect of bank-level reserve accumulation on lending and risktaking,whileDiamondetal.(2020)documentacrowding-outeffect. 23

Real effects. Next, we examine if the relative lending boost at lower capital banks during the taper tantrum translates into better firm-level outcomes. The credit effects that we documented above would affect firm performance if bank-dependent firms found it costly to switch across lenders as some lenders contract loan supply relative to others. To examine the effects of the term premium on firm investment, we turn to the firm-year dataset extracted from the Y-14Q data. Specifically, we examine the investment rate of firms that were ex-ante differentially exposed to the taper tantrum through their lenders’ leverage. We use the following specification that relates theinvestmentrateafterthestartofthetapertantrumtothefirms’ex-anteexposuretohigh-capital banksinafirm-yearpanel: Investment Rate = β Post ×Exposure to Bank Capital + jt 1 t j (17) +γ ′ Z +γ ′ Z ×Post +δ +ε 1 jt 2 jt t j˜t jt where the dependent variable is the investment rate, computed as the capital expenditure of firm j in year t divided by capital stock (fixed assets) at t−1 and Post is a dummy variable taking t value one in 2014 or 2014-2015 (and zero in 2013). The key variable “Exposure to bank capital ” j is the firms’ exposure to the rise in term premiums after the start of the taper tantrum through their relationships with lenders differing in degree of leverage. This exposure is computed as the average capital ratio of a firm’s lenders weighted by the share of borrowing from each of those lenders at end-2013. The vector of firm controls Z includes firm size (log-assets), firm jt leverage (total debt/assets), cash holdings (% assets), tangible assets (% assets), interest coverage ratio (EBITDA/interest expense), sales growth (a proxy for future growth opportunities and hence loan demand), and a dummy variable taking value one for listed firms. Firm controls enter the specificationsinlevelsandinteractedwithPost . Weincludeyear,state,andindustryfixedeffects, t or interacted state×industry×year fixed effects (δ ) to control for time-varying geographic and j˜t sectoralshocksaffectingfirmsinagivenlocationandindustry. The coefficient of interest is β which we expect to be negative, supporting the prediction 1 that follows from model’s implication 2, that firms borrowing from lower capital banks exhibit higher investment rates than other firms as the term premium rises and they receive more credit from those banks. The results, reported in Table 6, provide support for this prediction. The coefficient estimates on the interaction of Post with “Firm exposure to bank capital” are negative t and statistically significant across the time horizons examined, suggesting that firms borrowing from more leveraged banks before the taper tantrum had relatively higher investment rates in later years. These estimates are also economically significant. Looking at the coefficients in columns 2 24

and4andcomparingafirmwithmoreleveragedlenders(atthe25thpercentileofthecapitalratio distribution) with a firm with less leveraged lenders (at the 75th percentile of the distribution), the former firm had an investment ratio that was higher by between 1.7 and 2.1 ppts (relative to the meaninvestmentrateof24%)comparedtothelatterfirm. 6.3 Loan-level analysis over 2013–2019 In the previous section we showed that the steep and sustained rise in the term premium after the start of the taper tantrum induces a boost in bank lending on the extensive and intensive margins. Next we investigate if the taper tantrum results generalize to the full time period over which we haveloan-leveldata(2013Q1-2019Q4).23 Tothisend,wespecifythefollowinglendingregression indataonoutstandingloansatthebank-firmloan-level: Loan outcome = β Capital ratio ×Term premium shock + bj,t+1 1 bt t (18) +γ ′ X +γ ′ X ×Term premium shock +γ ′ Z +δ +ε 1 bt 2 bt t 3 bjt j-t bjt where Loanoutcome referstoanindicatorfornewlyoriginatedloansfrombankbtofirm jin bj,t+1 quarter t+1,theloanamount(log),andspread.24 Similartopreviousregressions,wecontrolfor the standard bank characteristics (size, core deposits, securities) and for bank-level GDP growth forecasts in level and interactions with the term premium shock, expectations, and the short rate (X ). We also include a quarterly pair-level relationship duration variable (Z ) and interacted bt bjt borrower×quarter fixed effects (δ ). The estimates are obtained using OLS with standard errors jt clustered at the bank-firm level. According to testable implication 2, the coefficient estimate on β should be negative to reflect that more leveraged banks increase loan supply more than less 1 leveragedbanksfollowingariseintheriskpremium. Banklending. TheresultsarereportedinTable7. Acrossspecifications,theestimatedcoefficient on Capital ratio × Term premium shock are statistically significant and indicate a positive b t link between bank leverage and the supply of bank loans during 2013Q1-2019Q4. The estimates suggest that, following a positive term premium shock, lower capital banks are more likely to grant new loans; in addition, they grant larger and cheaper loans than other banks. In terms of 23The first reporting quarter in the Y-14Q is 2012Q1. However, consistency of data submissions from reporting banksstabilizeswithinafewquarters,thereforewestarttheregressionsamplein2013Q1. Theresultsarefullyrobust toincludingdatafrom2012Q1–2012Q4inthesample. 24A key difference between this and the taper-tantrum specification is that banks’ capital ratio here is lagged one quarter,whileinthetapertantrumanalysisitwasfixedatend-2012(beforetheevent). 25

economic interpretation, the coefficient estimates in columns 1, 3, and 5 of Table 7 indicate that, if we compare the lending behavior of banks with capital ratios at the 25th and 75th percentiles of the distribution (11% vs. 14%) following a positive term premium shock of 5 bps (approximately 1.5x standard deviations during 2013–2019), then lower capital banks increase loan supply more than other banks in the following quarter—they are more likely to grant new loans by 4.1% (the share of new loans is 11%), to grant loans that are larger by 3% and loans that are less expensive by2bps. Overall,theseresultsareconsistentwithourbank-andloan-levelevidenceandestablish moregeneralpatternsinthemicrodatabeyondthetapertantrumanalysis. Loan maturities. To strengthen our evidence that more leveraged banks increase exposure to maturity transformation in response to changes in the term premium more than other banks, we also examine several loan maturity outcomes. An increase in loan supply, as predicted by model implication 2, would suggest that banks take more maturity risk by lengthening loan maturities when the term premium rises. To test this idea, we aggregate the loan-level data to the bank-firmquarter level and estimate the previous specification with the following outcome variables: the shareofloanswithmaturitygreaterthan3or5yearsandtheaveragevolume-weightedmaturityof the bank’s loan book. The results based on the full Y-14Q sample are reported in Table 8. Across specifications, the estimates are statistically significant and suggest that more leveraged banks are morelikelytolendlongertermthanotherbanksfollowingapositivetermpremiumshock.25 7 Mechanism: Bank profitability Our evidence so far shows that a steeper yield curve associated with a higher term premiums affects banks’ lending decisions. The underlying mechanism is one of expected bank profitability: a higher term premium represents higher expected profits on maturity transformation. In response to an increase in the term premium, banks take more duration risk, in particular by lending more or by lending longer term. In this section we provide direct evidence for this mechanism. For this purpose, we return to the baseline model estimated in bank-level data (over 1994Q1-2019Q4) and 25Twoadditionalsetsofresultsinbank-leveldatasupportthesefindings. First,inTableOA-7weshowthatbanks increasetheirexposuretomaturitytransformationbyrelatingthematuritygap,ameasureofmaturitytransformation at the individual bank level (English et al., 2018), to the term premium in a sample between 1997Q2 and 2019Q4. Second, inTableOA-8welimitthesampletoseveralquartersbefore/afterthetapertantrumandrelatethematurity gaptoabank’sex-anteleverage(inlevelandinteractedwiththe“Post”dummytakingvalueoneafterthestartofthe taper tantrum and zero otherwise). The estimates suggest that on average banks increase their exposure to maturity transformationafterariseinthetermpremium,andtheresultsarerelativelystrongerformoreleveragedbanks. 26

estimate it with dependent variables representing two bank profitability metrics: NIMs and ROE (overafour-quarterhorizon). FortermpremiumsweusetheKim-Wrighttermpremiumseriesand thetwoidentificationstrategiesbasedonIVandtermpremiumshocksdescribedinSection4. The results are reported in Table 9 and reveal a positive and statistically significant relation between term premiums and future bank profitability that holds across estimation strategies. In terms of economic magnitudes, the estimates in columns 3-4 and 5-6 of Table 9 indicate that a one standard deviation increase in the term premium or term premium shock (representing 50 or 5 bps) is associated with an increase in bank NIMs and ROE by 1-4 bps and 6-32 bps over the following four quarters, respectively. These effects are economically sizable given that over the sample period the average NIMs and ROE were 1% and 2.3%. Furthermore, in Table 10 we replicatethebankprofitabilityanalysisinthecontextofthetapertantrumandfindthatbankswith lower capital had higher profitability as the term premium sharply increased after the start of the tapertantrum,withestimatedcoefficientsbecomingstatisticallysignificantinthelaterquarters. 8 Conclusion Thispaperdocumentsanincreaseinbanklendingandprofitabilityfollowinganunexpectedrisein thetermpremium,especiallyformoreleveragedbanks. Wecallthismechanismtheexpectedbank profitabilitychannel. Ourfindingsareconsistentwithpredictionsfromaportfoliochoicemodelfor banksandpointtoonepotentialexplanationforthelong-documentedpredictabilityoftheslopeof the yield curve for future economic growth. Namely, when the yield curve steepens due to higher termpremiums(ratherthanhigherexpectedshortrates),banksincreasetheirexposuretomaturity transformation and provide more credit. The increased supply of bank credit can be expected to boost the overall economy through the usual credit channels. That said, we do not quantify the effectofchangesintheyieldcurveonaggregateborrowing,whichrequiresaccountingforpossible substitutions by borrowers to other sources of credit, or on aggregate economic growth. We leave thoseinterestingquestionstofutureresearch. 27

References Acharya, V. V., Eisert, T., Eufinger, C., and Hirsch, C. (2019). Whatever it takes: The real effects ofunconventionalmonetarypolicy. TheReviewofFinancialStudies,32(9):3366–3411. Adrian, T., Crump, R. K., and Moench, E. (2013). Pricing the term structure with linear regressions. JournalofFinancialEconomics,110(1):110–138. Adrian,T.,Estrella,A.,andShin,H.S.(2019). Risk-takingchannelofmonetarypolicy. Financial Management,48(3):725–738. Ahmed,R.andRebucci,A.(2022). DollarReservesandU.S.Yields: IdentifyingthePriceImpact ofOfficialFlows. NBERWorkingPaperNo.30476. Alessandri,P.andNelson,B.D.(2015). Simplebanking: Profitabilityandtheyieldcurve. Journal ofMoney,CreditandBanking,47(1):143–175. Ang, A., Piazzesi, M., and Wei, M. (2006). What does the yield curve tell us about GDP growth? JournalofEconometrics,131(1-2):359–403. Bansal, R. and Yaron, A. (2004). Risks for the long run: A potential resolution of asset pricing puzzles. TheJournalofFinance,59(4):1481–1509. Begenau,J.,Piazzesi,M.,andSchneider,M.(2015). Banks’riskexposures. NBERWorkingPaper No.21334. Beltran, D. O., Kretchmer, M., Marquez, J., and Thomas, C. P. (2013). Foreign Holdings of U.S. TreasuriesandU.S.TreasuryYields. JournalofInternationalMoneyandFinance,32(1):1120– 1143. Bernanke, B. (2005). The Global Saving Glut and the U.S. Current Account Deficit. Sandridge Lecture,VirginiaAssociationofEconomics(Richmond,VA). Bernanke,B.S.(2015). Thecouragetoact: Amemoirofacrisisanditsaftermath. Bernanke, B. S. and Gertler, M. (1995). Inside the black box: The credit channel of monetary policytransmission. JournalofEconomicPerspectives,9(4):27–48. Bernanke, B. S., Reinhart, V. R., and Sack, B. P. (2004). Monetary policy alternatives at the zero bound: Anempiricalassessment. BrookingsPapersonEconomicActivity,35(2):1–100. Bertaut, C. and Judson, R. (2014). Estimating U.S. Cross-Border Securities Positions: New Data and New Methods. Federal Reserve Board International Finance Discussion Papers No. 2014- 1113. Bertaut, C. and Judson, R. (2022). Estimating U.S. Cross-Border Securities Flows: Ten Years of theTICSLT. FEDSNotes. Bertaut,C.C.andTryon,R.W.(2007). Monthlyestimatesofu.s.cross-bordersecuritiespositions. FederalReserveBoardInternationalFinanceDiscussionPapersNo.2007-910. Brunnermeier, M. K. and Sannikov, Y. (2014). A macroeconomic model with a financial sector. AmericanEconomicReview,104(2):379–421. Caballero, R. J., Farhi, E., and Gourinchas, P.-O. (2008). An equilibrium model of “global imbalances”andlowinterestrates. AmericanEconomicReview,98(1):358–93. Caballero,R.J.andKrishnamurthy,A.(2009).Globalimbalancesandfinancialfragility.American EconomicReview,99(2):584–88. 28

Caglio, C. R., Darst, R. M., and Kalemli-Ozcan, S. (2021). Risk-taking and monetary policy transmission: EvidencefromloanstoSMEsandlargefirms. NBERWorkingPaperNo.28685. Campbell, J. Y., Chan, Y. L., and Viceira, L. M. (2003). A multivariate model of strategic asset allocation. JournalofFinancialEconomics,67(1):41–80. Chakraborty, I., Goldstein, I., and MacKinlay, A. (2020). Monetary stimulus and bank lending. JournalofFinancialEconomics,136(1):189–218. Chari, A., Dilts Stedman, K., and Lundblad, C. (2021). Taper Tantrums: Quantitative easing, Its aftermath, and emerging market capital flows. The Review of Financial Studies, 34(3):1445– 1508. Cox, J. C., Ingersoll, J. E., and Ross, S. A. (1985). A theory of the term structure of interest rates. Econometrica,53(2):385–407. Crosignani, M., Macchiavelli, M., and Silva, A. F. (2023). Pirates without borders: The propagation of cyberattacks through firms’ supply chains. Journal of Financial Economics, 147(2):432–448. Department of the Treasury, Federal Reserve Bank of New York, and Board of Governors of the Federal Reserve System (2023). Foreign portfolio holdings of U.S. securities, as of June 30, 2022. Di Tella, S. (2020). Risk premia and the real effects of money. American Economic Review, 110(7):1995–2040. Diamond, W., Jiang, Z., and Ma, Y. (2020). The reserve supply channel of unconventional monetary policy. Jacobs Levy Equity Management Center for Quantitative Financial Research Paper. Drechsler, I., Savov, A., and Schnabl, P. (2017). The deposits channel of monetary policy*. The QuarterlyJournalofEconomics,132(4):1819–1876. Drechsler, I., Savov, A., and Schnabl, P. (2021). Banking on deposits: Maturity transformation withoutinterestraterisk. TheJournalofFinance,76(3):1091–1143. Duffie,D.andEpstein,L.G.(1992a). Assetpricingwithstochasticdifferentialutility. TheReview ofFinancialStudies,5(3):411–436. Duffie,D.andEpstein,L.G.(1992b). Stochasticdifferentialutility. Econometrica: Journalofthe EconometricSociety,pages353–394. Duffie, D. and Lions, P.-L. (1992). PDE solutions of stochastic differential utility. Journal of MathematicalEconomics,21(6):577–606. English, W. B., den Heuvel, S. J. V., and Zakrajšek, E. (2018). Interest rate risk and bank equity valuations. JournalofMonetaryEconomics,98:80–97. Estrella, A. and Hardouvelis, G. A. (1991). The term structure as a predictor of real economic activity. JournalofFinance,46(2):555–576. Favara,G.,Ivanov,I.,andRezende,M.(2021). GSIBsurchargesandbanklending: Evidencefrom U.S.corporateloandata. JournalofFinancialEconomics,142(3):1426–1443. Favero,C.A.,Kaminska,I.,andSöderström,U.(2005). Thepredictivepoweroftheyieldspread: Furtherevidenceandastructuralinterpretation. AvailableatSSRN743104. Gertler, M. and Karadi, P. (2015). Monetary policy surprises, credit costs, and economic activity. AmericanEconomicJournal: Macroeconomics,7(1):44–76. 29

Gertler, M. and Kiyotaki, N. (2015). Banking, liquidity, and bank runs in an infinite horizon economy. AmericanEconomicReview,105(7):2011–43. Gilchrist, S. and Zakrajs˘ek, E. (2012). Credit spreads and business cycle fluctuations. American EconomicReview,102(4):1692–1720. Gomez, M., Landier, A., Sraer, D., and Thesmar, D. (2021). Banks’ exposure to interest rate risk andthetransmissionofmonetarypolicy. JournalofMonetaryEconomics,117:543–570. Greenspan, A. (2005). Testimony of Alan Greenspan before the Senate Committee on Banking, Housing, and Urban Affairs on Monetary Presenting the Federal Reserve’s Monetary Policy ReporttoCongress. FederalReserveSystem: BoardofGovernors. Haddad, V. and Sraer, D. (2020). The banking view of bond risk premia. The Journal of Finance, 75(5):2465–2502. Hamilton, J. D. and Kim, D. H. (2002a). A re-examination of the predictability of economic activityusingtheyieldspread. JournalofMoney,CreditandBanking,34(2):340–360. Hamilton,J.D.andKim,D.H.(2002b). Are-examinationofthepredictabilityoftheyieldspread forrealeconomicactivity. JournalofMoney,Credit,andBanking,34(2):340–360. Harvey, C. R. (1988). The real term structure and consumption growth. Journal of Financial Economics,22(2):305–333. Harvey, C. R. (1989). Forecasts of economic growth from the bond and stock markets. Financial AnalystsJournal,45(5):38–45. Harvey, C. R. (1993). The term structure forecasts economic growth. Financial Analysts Journal, 49(May/June):6–8. Jardet, C., Monfort, A., and Pegoraro, F. (2013). No-arbitrage Near-Cointegrated VAR(p) term structure models, term premia and GDP growth. Journal of Banking & Finance, 37(2):389 – 402. Jiménez, G., Mian, A., Peydró, J.-L., and Saurina, J. (2020). The real effects of the bank lending channel. JournalofMonetaryEconomics,115:162–179. Kaminska,I.andZinna,G.(2020). Officialdemandforu.s.debt: Implicationsforu.s.realinterest rates. JournalofMoney,CreditandBanking,52(2-3):323–364. Kandrac, J. and Schlusche, B. (2021). Quantitative easing and bank risk-taking: Evidence from lending. JournalofMoney,CreditandBanking,53(4):635–676. Khwaja, A. I. and Mian, A. (2008). Tracing the impact of bank liquidity shocks: Evidence from anemergingmarket. AmericanEconomicReview,98(4):1413–1442. Kim, D. H. and Priebsch, M. (2020). Are Shadow Rate Models of the Treasury Yield Curve StructurallyStable? JournalofFinancialandQuantitativeAnalysis(forthcoming). Kim, D. H. and Wright, J. H. (2005). An arbitrage-free three-factor term structure model and the recent behavior of long-term yields and distant-horizon forward rates. Federal Reserve Board FEDSWorkingPaper2005-33. Krishnamurthy,A.andVissing-Jorgensen,A.(2011). Theeffectsofquantitativeeasingoninterest rates: Channelsandimplicationsforpolicy. BrookingsPapersonEconomicActivity(Fall2011). Krishnamurthy, A. and Vissing-Jorgensen, A. (2015). The impact of Treasury supply on financial sectorlendingandstability. JournalofFinancialEconomics,118(3):571–600. Kuttner, K. N. (2001). Monetary policy surprises and interest rates: Evidence from the fed funds futuresmarket. JournalofMonetaryEconomics,47(3):523–544. 30

Lee, D. S., McCrary, J., Moreira, M. J., and Porter, J. (2022). Valid t-ratio Inference for IV. AmericanEconomicReview,112(10):3260–90. Li, C., Meldrum, A., and Rodriguez, M. (2017). Robustness of long-maturity term premium estimates. FEDSNote. Miranda-Agrippino, S. and Ricco, G. (2021). The transmission of monetary policy shock. AmericanEconomicsJournal: Macroeconomics,13(3):74–107. Nakamura, E. and Steinsson, J. (2018). High-frequency identification of monetary non-neutrality: Theinformationeffect. TheQuarterlyJournalofEconomics,133(3):1283–1330. Paul, P. (2023). Banks, maturity transformation, and monetary policy. Journal of Finanial Intermediation,53:101011. Rodnyansky, A. and Darmouni, O. M. (2017). The effects of quantitative easing on bank lending behavior. TheReviewofFinancialStudies,30(11):3858–3887. Romer, C. D. and Romer, D. H. (2000). Federal reserve information and the behavior of interest rates. AmericanEconomicReview,90(3):429–457. Rudebusch,G.D.,Sack,B.P.,andSwanson,E.T.(2006).Macroeconomicimplicationsofchanges inthetermpremium. FRBofSanFranciscoWorkingPaper2006-46. Rudebusch, G. D. and Williams, J. C. (2009). Forecasting recessions: The puzzle of the enduring poweroftheyieldcurve. JournalofBusinessandEconomicStatistics,27(4):492–503. Sinha, N. R. and Smolyansky, M. (2022). How sensitive is the economy to large interest rate increases? EvidencefromtheTaperTantrum. FEDSWorkingPaper2022–085. Stock, J. H. and Watson, M. W. (1989). New indexes of coincident and leading indicators. In Blanchard, O. and Fischer, S., editors, NBER Macroeconomics Annual 1989. MIT Press, Cambridge,MA. Tabova,A.M.andWarnock,F.E.(2021). ForeignInvestorsandU.S.Treasuries. NBERWorking PaperNo.29313. Warnock, F. E. and Warnock, V. C. (2009). International capital flows and U.S. interest rates. JournalofInternationalMoneyandFinance,28(6):903–919. Wright,J.H.(2006). Theyieldcurveandpredictingrecessions. FEDSWorkingPaper2006–07. 31

Figures and Tables Figure1: Modelsolution Thisfigureshowsthemodel’ssolutionacrosstheκ (termpremium)dimension(horizonalaxis)forthreelevelsofthe t statevariabler (theshortrate). Thesolidlineiswhenr isatitsunconditionalmean,thedashed(dotted)lineiswhen t t r istwostandarddeviationsabove(below)itsmean. t 0.6 0.4 0.2 0 -0.4 -0.2 0 ylretrauq % nI 3 2.5 2 1.5 -0.4 -0.2 0 10 14 12 5 10 0 8 -0.4 -0.2 0 -0.4 -0.2 0 10 3 5 2 0 1 -0.4 -0.2 0 -0.4 -0.2 0 32

Figure2: Impulseresponsestoa κ shockinthemodel This figureshows themodel’s impulse-responses to a κ shockthat causesapproximately a 100bpsincrease interm t premium. Thered-dashedlinesaretheresponsesofthemodelwithtighterleverageconstraints(δ=1.25andρ=0.5). -0.2 -0.3 -0.4 -0.5 0 50 100 level 1 0.5 0 0 50 100 SS morf .p.p 15 10 5 0 0 50 100 SS morf % 15 10 5 0 0 50 100 SS morf % 4 2 0 0 50 100 SS morf p.p 4 2 0 0 50 100 SS morf .p.p 0 -1 -2 -3 0 50 100 SS morf % 0.6 0.4 0.2 0 0 50 100 Quarter SS morf pp 0 -0.2 -0.4 0 50 100 SS morf pp 33

Figure3: Instrumentalvariablevs. Termpremium(Kim-Wright) Thisfigureshows(a)the5-yeartermpremiumseriesfromthetermstructuremodelofKimandWright(2005)and(b) theinstrumentalvariablesrepresntingtheforeignofficialholdingsofU.S.TreasuriesnormalizedbyU.S.GDP. 34

Figure4: Termpremiumvs. Termpremiumshocks(Kim-Wright) Thisfigureshows(a)the5-yeartermpremiumseriesfromthetermstructuremodelofKimandWright(2005)and(b) high-frequencytermpremiumshocksestimatedaschangesintheKim-WrighttermpremiumonFOMCeventdays. 35

Figure5: TreasuryyieldsandthetermpremiumduringTaperTantrum Thisfiguredepictsthesharpandsustainedriseinmarketyieldson5-yearTreasurysecuritiesandinthe5-yearterm premiumduringtheTaperTantrumepisodefollowingformerChairBenBernankespeechonMay222013regarding theFederalReserve’sintentiontostarttaperingassetpurchasesunderitsquantitativeeasingprogram(thedateofthe speechisindicatedbythedashedverticalline).TreasuryyieldsseriesistheMarketYieldonU.S.TreasurySecuritiesat 5-YearConstantMaturity,QuotedonanInvestmentBasis[seriescodeDGS5],retrievedfromFRED,FederalReserve BankofSt. Louis(link). ThetermpremiumseriesistheTermPremiumona5YearZeroCouponBond[seriescode THREEFYTP5],retrievedfromFRED,FederalReserveBankofSt. Louis(link). Source: BoardofGovernorsofthe FederalReserveSystem(US). 36

Table1: ModelCalibration Value Description Source 1. r−process r 0.0115 Meanr SMM λ 0.0241 AC(1)r SMM r σ 0.0071 Volatilityr SMM r 2. κ−process κ -0.2206 Meanκ SMM λ 0.0332 AC(1)κ SMM κ σ 0.0299 Volatilityκ SMM κ 3. Banks λ 0.013 Dividendpayoutintensity GK15 ρ 0.19 Seizurerate GK15 ϕ 0.35 depositspread Drechsleretal.(2021) c 0.003 Fixedcost Avg. Tobin’sQinGK15 δ 2.85 Depositconstraint MatchY-9C Themodelcalibrationisdescribedinthetext. 37

Table2: Afirstlookattheaggregatedata: Time-seriesevidence This table reports OLS regressions in monthly time series. The dependent variables are real GDP growth, banking systemloangrowth,NIMs,andROE.Thesampleperiodis1973Q1–2019Q4incolumns1-4and1984Q4-2019Q4in columns5-8. Specificationsincludeone-monthlaggedmacrocontrols(changeintheshortrategivenbythe3-month Tbill yield), one-year ahead real GDP forecasts, and excess bond premium) and lagged dependent variables. Time seriesonKimandWright(2005)termpremiumestimates,realGDPgrwoth,andaggregatebankingsectorvariables (total loans, NIM, and ROE) come from Federal Reserve Economic Data (FRED). Sources for other variables are discussed in Section 5. Robust and autocorrelation-consistent standard errors in parentheses. Significance: *p<.1; **p<.05;***p<.01. (1) (2) (3) (4) (5) (6) (7) (8) Dependentvariables RealGDPgrowth Bankloangrowth NIM ROE t,t+4 t,t+4 t,t+4 t,t+4 3m5ytermspread 0.64** 0.87** 0.05** 1.33** t (0.26) (0.36) (0.03) (0.53) Termpremium 1.03*** 0.89** 0.13** 1.53** t (0.23) (0.27) (0.02) (0.60) Expectations 0.58** 0.87** 0.02 1.09* t (0.26) (0.39) (0.03) (0.59) ∆Shortrate -0.15 -0.13 0.06 0.06 -0.03 -0.03 0.50 0.49 t (0.23) (0.23) (0.18) (0.18) (0.02) (0.02) (0.36) (0.36) Observations 183 183 183 183 137 137 137 137 R2 0.34 0.37 0.60 0.60 0.91 0.92 0.64 0.64 Macrocontrols Y Y Y Y Y Y Y Y 38

Table3: Termpremiumandbanklending: Bank-levelevidence ThistablereportsOLSandIVestimatesfromaregressionofbankloangrowth(excludingandincludingcreditlines) on the Kim-Wright term premium. In columns 1 and 4 we use the term premium level series, in columns 2 and 5 we instrument for it with foreign official holdings of U.S. Treasuries (normalized by U.S. GDP) and in columns 3 and 6 we use the term Kim-Wright term premium shocks. The sample is 1994Q1-2019Q4. Specifications include the following lagged controls: short-rate changes, expectations (3-month/5-year spread minus the term premium), four-quarterrealGDPgrowth,four-quarterGDPdeflatorinflation,one-yearaheadrealGDPgrowthandGDPdeflator inflationforecastsfromSPF,excessbondpremium,andlaggeddependentvariables. Bankcontrolsincludebanksize (logshareofbankassetsintotalbankingsystemassets),capitalratio,theshareofcoredepositsintotalliabilitiesand the share of securities in total assets. Bank MSA fixed effects are for the MSA of the bank’s headquarters location. Standarderrorsdouble-clusteredbybankandquarterinparentheses. Significance: *p<.1;**p<.05;***p<.01. (1) (2) (3) (4) (5) (6) Dependentvariables Loangrowth Loangrowth t,t+4 (includingcreditlines) t,t+4 OLS IV OLS OLS IV OLS TPshock TPshock Termpremium 1.34*** 2.17*** 24.42** 0.73** 1.46*** 22.04** t (0.35) (0.45) (10.64) (0.31) (0.43) (10.06) Expectations -1.27*** -1.34*** -10.29 -1.14*** -1.20*** -6.80 t (0.20) (0.21) (6.89) (0.18) (0.18) (6.57) ∆Shortrate -1.27*** -1.23*** -11.27** -1.48*** -1.44*** -8.95* t (0.15) (0.17) (5.24) (0.12) (0.14) (4.86) Observations 630,401 630,401 634,153 627,712 627,712 631,463 R2 0.39 0.23 0.22 0.38 0.23 0.22 Macrocontrols Y Y Y Y Y Y Bankcontrols Y Y Y Y Y Y BankFE Y Y Y Y Y Y BankMSAFE Y Y Y Y Y Y First-stageFtest 284.8 284.7 39

Table4: Termpremiumandbanklending: Differentialeffectsbybankleverage ThistablereportsOLSandIVestimatesfromaregressionofbankloangrowth(excludingandincludingcreditlines) ontheKim-Wrighttermpremiumininteractionwithbankcapital. Incolumns1and4weusethetermpremiumlevel series,incolumns2and5weinstrumentforitwithforeignofficialholdingsofU.S.Treasuries(normalizedbyU.S. GDP) andin columns3 and6 we usethe termKim-Wright term premiumshocks. The sample is 1994Q1-2019Q4. Bankcontrolsincludebanksize(logshareofbankassetsintotalbankingsystemassets),theshareofcoredepositsin totalliabilitiesandtheshareofsecuritiesintotalassets. Macrocontrols(fromTable3)arespannedbyquarterfixed effects.BankMSAfixedeffectsarefortheMSAofthebank’sheadquarterslocation.Standarderrorsdouble-clustered bybankandquarterinparentheses. Significance: *p<.1;**p<.05;***p<.01. (1) (2) (3) (4) (5) (6) Dependentvariables Loangrowth Loangrowth t,t+4 (includingcreditlines) t,t+4 OLS IV OLS OLS IV OLS TPshock TPshock Capitalratio ×Termpremium -15.31*** -22.39*** -149.17* -16.07*** -22.74*** -142.79* t t (3.61) (4.89) (86.18) (3.63) (4.91) (85.86) Observations 629,234 629,234 629,234 626,530 626,530 626,530 R2 0.45 0.15 0.45 0.44 0.14 0.44 Bankcontrols Y Y Y Y Y Y Bankcontrols×Termpremium Y Y Y Y Y Y Bankcontrols×Expectations Y Y Y Y Y Y Bankcontrols×Shortrate Y Y Y Y Y Y BankFE Y Y Y Y Y Y QuarterFE Y Y Y Y Y Y BankMSAFE Y Y Y Y Y Y First-stageFtest 55.84 55.33 40

Table5: Termpremiumandbanklending: Loan-levelevidencefromTaperTantrum ThistablereportsOLSestimatesfromaregressionofbanklendingoutcomesontheinteractionofbankcapitaland a“Post”dummythattakesvalueoneafter2013Q2,andzerootherwise. Thedataareatthebank-firmloanleveland cover a period of between three and five quarters before and after 2013Q2. The dependent variables are a dummy variablethattakesvalueonefornewloansandzerootherwiseindataonoutstandingloans(panelA),andlog-volume andspreadsindataonnewloanoriginations(panelsBandC).Specificationscontrolforbank-levelone-yearaheadreal GDPgrowthexpectationsfromtheBlueChipsurveyinlevelsandinteractedwithPost. Relationshipdurationisthe# ofquarterssincethefirstloanisobservedinagivenbank-firmpair(log). Specificationsincludethefollowinglagged bankcontrols:size(log-assets),capitalratio(asofend-2012),andsecurities-to-assetratio,inlevelsandinteractedwith thePostdummy. Standarderrorsclusteredbybank-firminparentheses. Significance: *p<.1;**p<.05;***p<.01. (1) (2) (3) #ofquartersbeforeandaftertapertantrum 3Q 4Q 5Q A.Newloan t+1 Capitalratio ×Post -0.015*** -0.013*** -0.013*** t t (0.002) (0.002) (0.002) Observations 133,262 174,659 293,348 R2 0.651 0.656 0.620 B.Loanvolume(log) t+1 Capitalratio ×Post -0.025*** -0.027*** -0.030*** t t (0.002) (0.006) (0.008) Observations 8,506 11,312 17,489 R2 0.773 0.772 0.882 C.Loanspread t+1 Capitalratio ×Post 2.789*** 1.646* 1.559** t t (0.455) (0.775) (0.639) Observations 2,254 2,915 4,151 R2 0.863 0.851 0.862 Relationshipduration Y Y Y Firm×quarterFE Y Y Y Bankcontrols Y Y Y Bankcontrols×Post Y Y Y Growthexpectations Y Y Y Growthexpectations×Post Y Y Y 41

Table6: Realeffectsoftermpremium: Firm-levelevidencefromtheTaperTantrum ThistablereportsOLSestimatesfromafirm-levelregressionoftheinvestmentrateontheaveragecapitalizationof a firm’s lenders and a “Post” dummy that takes value one in 2014 (columns 1–2) or 2014–2015 (columns 3–4) and zeroin2013. Thedataareatthefirm-yearlevel. Theinvestmentrateiscomputedascapitalexpendituredividedby laggedcapitalstock(fixedassets). “Firmexposuretobankcapital”isdefinedatthefirmlevelastheaveragecapital ratioofafirm’sbanksweightedbytheshareoftotalborrowingfromeachofthosebanksatend-2013. Specifications includethefollowingfirmcontrols: size(log-assets),leverage(totaldebt/assets),cashholdings(%assets),tangibility (tangible assets in % of total assets), interest coverage ratio (EBITDA/interest expense), firm sales growth, and a dummy variable that takes value one for listed firms, in levels and interacted with Post. Industry is given by twodigit NAICS classification. Standard errors double-clustered by firm and year in parentheses. Significance: *p<.1; **p<.05;***p<.01. (1) (2) (3) (4) Investmentrate 2013vs2014 2013vs2014-2015 Firmexposuretobankcapital×Post -0.122** -0.259** -0.095** -0.215** (0.005) (0.006) (0.018) (0.026) Observations 54,181 44,331 80,567 65,800 R2 0.071 0.249 0.065 0.250 Firmcontrols Y Y Y Y Firmcontrols×Post Y Y Y Y YearFE Y Y Y Y StateFE Y Y Y Y IndustryFE Y Y Y Y State×Industry×YearFE Y Y 42

Table7: Termpremiumandbanklending: Loan-levelevidenceover2013–2019 This table reports OLS estimates from a regression of bank lending terms on bank capital interacted with the term premium shock during 2013Q1-2019Q4. The dependent variables are a dummy variable for new loan originations, loanvolume(log),andloanspread,respectively.Thedataareatthebank-firmloanlevelandrefertooutstandingloans tononfinancialfirmsreportedonaquarterlyfrequencyduring2013Q1-2019Q4.Thetermpremiumshockiscomputed fromchangesintheKim-WrighttermpremiumonFOMCeventdays(seeSection4fordetails);expectationsarebankspecific one-year ahead GDP growth forecasts from the Blue Chip survey. Bank controls include size (log-assets), capital ratio, core deposits (% liabilities) and securities (% assets) in levels and interacted with the term premium shock,expectations,andtheshortrate. Relationshipdurationisdefinedasthenumberofquarterssincethefirstloanis observedforagivenbank-firmpair(log). Standarderrorsclusteredbybank-firminparentheses. Significance:*p<.1; **p<.05;***p<.01. (1) (2) (3) (4) (5) (6) Dependentvariable Newloan Loanamount(log) Loanspread t+1 t+1 t+1 Capitalratio ×Termpremiumshock -0.466*** -0.366*** -0.277*** -0.316*** 0.215** 0.215** t t (0.021) (0.021) (0.064) (0.066) (0.091) (0.089) Observations 616,493 616,493 632,980 632,980 270,422 270,422 R2 0.638 0.639 0.663 0.663 0.822 0.822 Relationshipduration Y Y Y Y Y Y Firm×QuarterFE Y Y Y Y Y Y Bankcontrols Y Y Y Y Y Y Bankcontrols×Termpremiumshock Y Y Y Y Y Y Bankcontrols×Expectations Y Y Y Bankcontrols×Short-rate Y Y Y 43

Table8: Termpremiumandloanmaturities: Loan-levelevidenceover2013–2019 This table reports OLS estimates from a regression of bank lending terms on bank capital interacted with the term premiumshockduring2013Q1-2019Q4. Thedependentvariablesaretheshareofloan-maturityloans(>3yearsor >5 years) and the loan volume-weighted average maturity in data that is aggregated to the bank-firm-quarter level. The data referto outstanding loans tononfinancial firms reported on aquarterly frequency during 2013Q1-2019Q4. The term premium shock is computed from changes in the Kim-Wright term premium on FOMC event days (see Section4fordetails);expectationsarebank-specificone-yearaheadGDPgrowthforecastsfromtheBlueChipsurvey. Bank controls include size (log-assets), capital ratio, core deposits (% liabilities) and securities (% assets) in levels and interacted with the term premium shock, expectations, and the short rate. Relationship duration is defined as thenumberofquarterssincethefirstloanisobservedforagivenbank-firmpair(log). Standarderrorsclusteredby bank-firminparentheses. Significance: *p<.1;**p<.05;***p<.01. (1) (2) (3) (4) (5) (6) Dependentvariable Sharelong-maturity Sharelong-maturity Averagematurity loans(>3years) loans(>5years) (volume-weighted) t+1 t+1 t+1 Capitalratio ×Termpremiumshock -0.267*** -0.242*** -0.734*** -0.712*** -2.045*** -2.203*** t t (0.018) (0.020) (0.028) (0.029) (0.179) (0.180) Observations 175,333 175,333 175,333 175,333 175,333 175,333 R2 0.576 0.577 0.617 0.621 0.542 0.549 Relationshipduration Y Y Y Y Y Y Firm×QuarterFE Y Y Y Y Y Y Bankcontrols Y Y Y Y Y Y Bankcontrols×Termpremiumshock Y Y Y Y Y Y Bankcontrols×Growthexpectations Y Y Y Bankcontrols×Short-rate Y Y Y 44

Table9: Mechanism: Termpremiumandbankprofitability: Bank-levelevidence This table reports OLS and IV (second stage) estimates from a regression of bank profitability (NIMs and ROE) on termpremiumestimates. Incolumns1and4weusetheKim-Wrightestimateofthetermpremium. Incolumns2and 5weuseatermpremiumshockcomputedfromchangesintheKim-WrighttermpremiumonFOMCeventdays(see Section4fordetails). Incolumns3and6theinstrumentalvariablefortheKim-Wrighttermpremiumistheforeign officialholdingsofU.S.Treasuries(normalizedbyU.S.GDP).Thesampleis1994Q1-2019Q4.Specificationsinclude the following lagged controls: short-rate changes, expectations (3-month/5-year spread minus the term premium), four-quarterrealGDPgrowth,four-quarterGDPdeflatorinflation,one-yearaheadrealGDPgrowthandGDPdeflator inflationforecastsfromSPF,excessbondpremium,andlaggeddependentvariables. Bankcontrolsincludebanksize (log share of bank assets in total banking system assets), capital ratio, the share of core deposits in total liabilities, the share of securities in total assets, and the share of trading assets in total assets. Bank MSA fixed effects are for theMSAofthebank’sheadquarterslocation. Standarderrorsclusteredbybankinparentheses. Significance: *p<.1; **p<.05;***p<.01. (1) (2) (3) (4) (5) (6) Dependentvariable NIMs ROE t,t+4 t,t+4 OLS IV OLS OLS IV OLS TPshock TPshock Termpremium 0.055*** 0.087*** 0.193*** 0.403*** 0.630*** 1.198*** t (0.001) (0.001) (0.008) (0.015) (0.020) (0.154) Expectations -0.008*** -0.011*** -0.159*** -0.180*** -0.198*** -1.843*** t (0.001) (0.001) (0.005) (0.009) (0.009) (0.103) ∆Shortrate -0.002*** -0.001** -0.086*** 0.029*** 0.041*** -0.545*** t (0.000) (0.000) (0.004) (0.005) (0.005) (0.086) Observations 640,385 640,385 650240 640,384 640,384 650239 R2 0.754 0.427 0.394 0.399 0.160 0.163 Macrocontrols Y Y Y Y Y Y Bankcontrols Y Y Y Y Y Y BankFE Y Y Y Y Y Y BankMSAFE Y Y Y Y Y Y First-stageFtest 974511 1.529e+06 45

Table10: Mechanism: TermpremiumandbankprofitabilityaroundtheTaperTantrum ThistablereportsOLSestimatesfromaregressionofbankprofitability(NIMsinpanelAandROEinpanelB)onthe interactionofbankcapitalanda“Post”dummythattakesvalueoneafter2013Q2,andzerootherwise. Thedataareat thebank-quarterlevelandcoveraperiodofbetweenthreeandfivequartersbeforeandafter2013Q2. Specifications includethefollowinglaggedbankcontrols: size(log-assets),capitalratio,andsecurities-to-assetratio,inlevelsand interacted with the Post dummy. Standard errors clustered by bank in parentheses. Significance: *p<.1; **p<.05; ***p<.01. (1) (2) (3) #ofquartersbeforeandaftertapertantrum 3Q 4Q 5Q A.NIMs t+1 Capitalratio ×Post -0.057 -0.095* -0.118** t t (0.034) (0.040) (0.040) Observations 22,281 33,469 44,755 R2 0.963 0.951 0.942 B.ROE t+1 Capitalratio ×Post -2.177 -3.146* -4.210** t (1.151) (1.441) (1.529) Observations 22,281 33,469 44,755 R2 0.840 0.800 0.776 Bankcontrols Y Y Y Bankcontrols×Post Y Y Y BankFE Y Y Y QuarterFE Y Y Y 46

Appendix A Stochastic Discount Factor Wepresentasimplerepresentativeagentconsumption-basedmodeltoillustratethetypeofpreferences and endowment process that would deliver a similar stochastic discount factor (SDF) than the one presented in the main text. In brief, the setup is Bansal and Yaron (2004), model 2, but for simplicity we assume there are no shocks to expected consumption growth.26 The setup consists ofarepresentativeagentwithrecursivepreferences,followingDuffieandEpstein(1992b), (cid:90) ∞ U = E f (c ,U )ds, (A-1) t t s s t (cid:40) (cid:41) f (c,U) = 1 ρc 1− ψ 1 [(1−γ)U] ψ 1 1 − − γ γ −ρ(1−γ)U , 1− 1 ψ were γ is the risk aversion parameter, ψ is the elasticity of intertemporal substitution, c is agent’s consumption,andU istheutilitylevel. AsshowninDuffieandEpstein(1992a),theSDF, m ,is t dm df t = c + f dt, U m f t c where f and f is partial derivative of f (c,U) with respect to c and U, respectively. The c U consumptionprocessis dc (cid:113) t = µdt+ exp(v )dW , t 1,t c t dv = λ (v−v )dt+κdW , t v t 2,t where v is the log of variance. It can be shown that the value function depends on two state t variables,consumptionlevelandvolatilitylevel(DuffieandEpstein,1992b;Campbelletal.,2003): (ξ(v)c)1−γ U = , 1−γ 26The analysis would be the same if we assume shocks to expected consumption growth are perfectly correlated withconsumptiongrowth(orlevel)shocks. 47

whereξ(v)isaunknownfunctionthathastobesolvedusingtheintegral(A-1). Then,thepartial derivativescanbewrittenas f = ρc −γξψ 1−γ , c (cid:32) 1 −γ (cid:33) (cid:110) (cid:111) f = ψ ρ ξψ 1−1−1 . U 1− 1 ψ UsingIto’slemma, df dc 1 (cid:18) dc (cid:19)2 c = −γ + γ(γ+1) f c 2 c c (cid:18) 1 (cid:19) dξ 1 (cid:18) 1 (cid:19)(cid:18) 1 (cid:19)(cid:18) dξ (cid:19)2 + −γ + −γ −γ−1 ψ ξ 2 ψ ψ ξ (cid:18) (cid:19) 1 dcdξ −γ −γ ψ c ξ Then,theSDFis dm (cid:113) (cid:18) 1 (cid:19) ξ t = −r dt−γ exp(v )dW − −γ v κdW , (A-2) t t 1,t 2,t m ψ ξ t with (cid:32) γ− 1 (cid:33) (cid:110) (cid:111) r = ψ ρ ξψ 1−1−1 +γµ−γ(γ+1)v t t 1− 1 ψ (cid:18) (cid:19)(cid:20) (cid:21) (cid:18) (cid:19)(cid:18) (cid:19)(cid:18) (cid:19)2 1 ξ 1ξ 1 1 1 ξ + −γ v λ (v−v )+ vv κ2 + −γ −γ−1 v κ v t ψ ξ 2 ξ 2 ψ ψ ξ where ξ is the partial derivative of ξ with respect to v. Notice the SDF (A-2) is similar to the one v proposedintheModelsection, dm t = −r dt−κ dW −κdW . t t r,t κ,t m t Finally, ξ solvesthefollowingordinarydifferentialequation 0 = ρ (cid:110) ξψ 1−1−1 (cid:111) +µ− γ exp(v)+ ξ v λ (v−v )+ 1ξ vv κ2− γ (cid:18) ξ v κ (cid:19)2 , v t 1− 1 2 ξ 2 ξ 2 ξ ψ which can be solved numerically and has a unique solution provided the state variable is strong Markov(DuffieandLions,1992). 48

B Model solution: Tobin’s Q and leverage for different ρ FigureB-1: Relationshipbetween ρ, ψ(r,κ),and α Thisfigureshowsthesolutionforψ(r,κ)andαfordifferentlevelsofρ. 13 2.5 11 2.2 9 1.9 1.6 7 0.16 0.18 0.2 0.22 0.16 0.18 0.2 0.22 49

C Data TableC-1: Descriptivestatistics Thistablereportssummarystatisticsforselectedregressionvariablesinthebank-levelanalysis(Tables3-4,9)(panel A)andthetapertantrumloan-levelanalysis(baselineTables5,10)(panelB).Termpremiumlevelseriesandshocks arebasedontermstructuremodelsfromKimandWright(2005)(baseline),Adrianetal.(2013)andKimandPriebsch (2020)(robustness). InpanelB,one-yearaheadrealGDPgrowthisatthebank-levelandcomesfromtheBlueChip EconomicandFinancialIndicators. N SD Mean P25 P50 P75 A.Variablesinbank-levelregressions Loangrowth 634153 0.14 0.05 -0.02 0.04 0.11 Loangrowth(includingcreditlines) 631463 0.14 0.05 -0.02 0.04 0.11 Netinterestmargins 629324 0.00 0.01 0.01 0.01 0.01 Returnonequity 629323 0.03 0.02 0.02 0.03 0.04 Termpremium(Kim-Wright)x100 634153 0.49 0.50 0.09 0.48 0.84 Termpremiumshock(Kim-Wright)x100 634153 0.05 0.00 -0.03 0.00 0.04 Termpremium(Adrian-Crump-Moench)x100 634153 0.58 0.84 0.34 0.89 1.32 Termpremium(Kim-Priebsch)x100 634153 0.60 0.41 0.05 0.45 0.84 Termpremiumshock(Adrian-Crump-Moench)x100 634153 0.07 0.01 -0.04 0.01 0.04 IV(ForeignofficialholdingsofUST/U.S.GDP) 634153 0.07 0.10 0.05 0.08 0.18 Expectations(Kim-Wright)x100 634153 0.78 0.64 -0.12 0.81 1.38 ∆Shortratex100 634153 1.30 -0.04 -0.56 -0.01 0.83 Shortrate(shock)x100 634153 0.12 -0.01 -0.03 0.00 0.04 SPF:one-yearaheadrealGDPgrowth 634153 0.01 0.03 0.02 0.03 0.03 SPF:one-yearaheadGDPdeflatorinflation 634153 0.00 0.02 0.02 0.02 0.02 RealGDPgrowth 634153 0.02 0.03 0.02 0.03 0.04 GDPdeflatorinflation 634153 0.01 0.02 0.02 0.02 0.02 Excessbondpremium 634153 0.65 0.01 -0.36 -0.22 0.13 Banksize(log-shareofbankassets) 634153 1.11 -10.55 -11.30 -10.62 -9.90 Capitalratio 634153 0.04 0.11 0.08 0.10 0.12 Coredeposits-to-liabilities 634153 0.13 0.80 0.74 0.82 0.88 Securities-to-assets 634153 0.15 0.25 0.14 0.23 0.34 Maturitygap 56453 0.58 0.27 0.38 0.54 0.74 B.VariablesinTaperTantrumloan-levelanalysis Loanamount(US$million) 17489 87.22 29.55 1.55 7.50 26.80 Log(amount) 17489 2.33 15.48 14.25 15.83 17.10 Loanspreadx100(ppts) 4151 1.22 2.35 1.50 2.19 3.00 50

TableC-1: Descriptivestatistics(continued) Thistablereportssummarystatisticsforselectedregressionvariablesinthefirm-levelrealeffectsanalysisaroundthe taper tantrum (Table 6) (panel C) and the full Y-14Q sample (Tables 7-8) (panel D). In Panel D, the data structure isbank-firmloan-levelforallvariablesexceptshareoflong-maturityloansand(volume-weighted)averagematurity whicharereportedintheaggregatedbank-firmquarterpanel. N SD Mean P25 P50 P75 B.VariablesinTaperTantrumloan-levelanalysis(continued) Newloan 293348 0.27 0.08 0.00 0.00 0.00 Netinterestmarginsx100 33469 0.19 0.86 0.74 0.84 0.95 Returnonequityx100 33469 2.79 1.70 1.08 1.92 2.88 Maturitygap/100 33912 0.27 0.58 0.38 0.55 0.75 Capitalratio(x100) 17489 2.65 12.96 10.74 12.17 14.72 Banksize(log-assets) 17489 1.13 6.55 5.68 7.17 7.28 Securities-to-assets 17489 0.07 0.19 0.16 0.18 0.20 Coredeposits-to-liabilities 17489 0.10 0.57 0.49 0.55 0.63 Reserves-to-assets 17489 0.09 0.07 0.02 0.05 0.08 BlueChip: one-yearaheadrealGDPgrowth 17489 0.44 2.34 2.08 2.36 2.60 C.VariablesinTaperTantrumfirm-levelanalysis Investmentrate(Capex/L.Capitalstock) 80567 0.33 0.24 0.00 0.11 0.32 Firmexposuretobankcapital 80567 0.03 0.11 0.11 0.12 0.13 Firmsize(log-assets) 80567 2.51 17.50 16.04 17.16 18.84 Firmleverage(debt/assets) 80567 0.28 0.35 0.14 0.31 0.51 Cashratio(cash/assets) 80567 13.75 10.17 1.16 4.90 13.70 Tangibility(%assets) 80567 19.41 89.30 88.56 99.39 100.00 InterestCoverageRatio 80567 0.86 0.44 0.04 0.11 0.33 Realsalesgrowth 80567 40.65 13.87 -0.26 6.70 17.07 1: Listedfirm 80567 0.23 0.06 0.00 0.00 0.00 D.Variablesinfull-sampleloan-levelanalysis Loanamount(US$million) 632980 64.34 27.27 2.50 8.50 30.00 Loanamount(log) 632980 1.48 16.04 14.73 15.96 17.22 Loanspread(ppts) 318846 1.06 2.12 1.50 1.95 2.50 Newloan 616493 0.32 0.11 0.00 0.00 0.00 Sharelong-maturityloans(>3years) 175333 0.24 0.93 1.00 1.00 1.00 Sharelong-maturityloans(>5years) 175333 0.39 0.77 0.71 1.00 1.00 Averagematurity(volume-weighted) 175333 2.66 5.80 4.96 5.00 6.44 Capitalratio 632980 1.41 12.79 11.89 12.57 13.67 Banksize(log-assets) 632980 1.22 20.41 19.52 21.16 21.38 Securities-to-assets 632980 0.01 0.03 0.02 0.02 0.03 Coredeposits-to-liabilities 632980 13.99 63.15 55.78 67.65 69.83 Relationshipduration(log) 632980 1.05 0.38 0.00 0.00 0.00 Termpremiumshock(Kim-Wright)x100 632980 0.03 0.00 -0.02 0.00 0.03 BlueChip: one-yearaheadrealGDPgrowth 632980 0.34 2.42 2.18 2.40 2.65 Short-ratex100 632980 2.75 0.78 -1.59 1.37 2.49 51

Online Appendix Why Does the Yield Curve Predict GDP Growth? The Role of Banks FigureOA-1: TermpremiumduringTaperTantrumplacebotests This figure depicts the 5-year term premium during the period over which we conduct placebo tests for the Taper Tantrum analysis between 2012Q2 and 2013Q1. The term premium series is the Term Premium on a 5 Year Zero CouponBond[seriescodeTHREEFYTP5],retrievedfromFRED,FederalReserveBankofSt. Louis(link). Source: BoardofGovernorsoftheFederalReserveSystem(US). 52

TableOA-1: Firststage—Instrumentrelevance This table reports OLS estimates of contemporaneous time series regressions of the 5-year term premium estimated using the Kim and Wright (2005), Adrian et al. (2013) and Kim and Priebsch (2020) models, respectively, on the instrumentalvariable(ForeignofficialholdingsofTreasurysecurities,normalizedbyGDP)andtheshortrate(3-month T-billyield).Regressionsincolumns2,4,and6addmacrocontrolsincludingrealGDPgrowth,GDPdeflatorinflation, one-yearaheadrealGDPgrowthandGDPdeflatorinflationforecastsfromSPF,andtheexcessbondpremium. The sample period is 1961Q3-2019Q4 in columns 1, 3 and 5; and 1973Q1-2019Q4 in columns 2, 4, and 6 (due to the availability of the excess bond premium time series). Robust and kernel-based autocorrelation-consistent standard errorsinparentheses. Significance: *p<.1;**p<.05;***p<.01. (1) (2) (3) (4) (5) (6) Dependentvariable Kim-Wright Adrian-Crump-Moench Kim-Priebsch termpremium termpremium termpremium t t t Foreignofficialholdings -0.03*** -0.07*** -0.04*** -0.05*** -0.05*** -0.07*** ofU.S.Treasuries (0.01) (0.01) (0.01) (0.01) (0.01) (0.01) t Short-rate 0.07*** -0.02 0.12*** 0.08*** 0.12*** 0.09*** t (0.02) (0.03) (0.02) (0.02) (0.02) (0.02) Observations 234 187 234 187 232 185 R2 0.19 0.62 0.60 0.82 0.67 0.85 Macrocontrols Y Y Y 53

TableOA-2: Robustnessofbank-levelevidencetolongersampleperiod ThistablereportsOLSandIVestimatesfromaregressionofbankloangrowth(excludingandincludingcreditlines) ontheKim-Wrighttermpremium. Incolumns1and4weusethetermpremiumlevelseries,incolumns2and5we instrumentforitwithforeignofficialholdingsofU.S.Treasuries(normalizedbyU.S.GDP)andincolumns3and6we usethetermKim-Wrighttermpremiumshocks. Thesampleis1973Q4-2019Q4incolumns1-2,1994Q1-2019Q4in columns3and6,and1991Q1-2019Q4incolumns4-5.Specificationsincludethefollowinglaggedcontrols:short-rate changes,expectations(3-month/5-yearspreadminusthetermpremium),four-quarterrealGDPgrowth,four-quarter GDPdeflatorinflation,one-yearaheadrealGDPgrowthandGDPdeflatorinflationforecastsfromSPF,excessbond premium,andlaggeddependentvariables. Bankcontrolsincludebanksize(logshareofbankassetsintotalbanking systemassets),capitalratio,andtheshareofdepositsintotalliabilities. BankMSAfixedeffectsarefortheMSAof thebank’sheadquarterslocation. Standarderrorsdouble-clusteredbybankandquarterinparentheses. Significance: *p<.1;**p<.05;***p<.01. (1) (2) (3) (4) (5) (6) Dependentvariable Loangrowth Loangrowth t,t+4 (includingcreditlines) t,t+4 OLS IV OLS OLS IV OLS MPshock MPshock Termpremium 0.13 2.83*** 0.28** 0.83** 3.40*** 0.25** t (0.40) (0.70) (0.12) (0.40) (0.73) (0.11) Expectations -0.28 0.14 -0.13* -1.23*** -1.39*** -0.09 t (0.32) (0.38) (0.08) (0.26) (0.29) (0.07) ∆Shortrate -0.39** 0.03 -0.12** -0.79*** -0.59*** -0.10* t (0.17) (0.21) (0.06) (0.17) (0.22) (0.05) Observations 1,603,535 1,603,535 638,962 748,003 748,003 631,463 R2 0.27 0.16 0.19 0.34 0.18 0.20 Macrocontrols Y Y Y Y Y Y Bankcontrols Y Y Y Y Y Y BankMSAFE Y Y Y Y Y Y First-stageFtest 110.1 193.4 54

TableOA-3: RobustnessofIVestimatestodroppingChina’sofficialU.S.Treasuryholdings This table reports IV (second stage) estimates from a regression of bank loan growth (excluding and including credit lines) and bank profitability (NIMs and ROE) on term premium estimates. Compared to baseline Table 3, the instrumental variable is constructed by excluding Chinese holdings of U.S. Treasuries. The sample period is 1994Q1-2019Q4. Specification details are as in baseline Table 3. Standard errors are double-clustered by bank and quarter. Significance: *p<.1;**p<.05;***p<.01. (1) (2) (3) (4) Dependentvariable Loangrowth Loangrowth NIM ROE t,t+4 t,t+4 t,t+4 (incl. creditlines) t,t+4 IV IV IV IV Termpremium 1.77*** 1.19*** 0.08*** 0.53*** t (0.41) (0.40) (0.01) (0.12) Expectations -1.31*** -1.19*** -0.01 -0.19*** t (0.20) (0.18) (0.01) (0.07) ∆Shortrate -1.24*** -1.44*** -0.00 0.04 t (0.15) (0.13) (0.00) (0.06) Observations 634,153 631,463 640,385 640,384 R2 0.23 0.23 0.43 0.16 Macrocontrols Y Y Y Y Bankcontrols Y Y Y Y BankMSAFE Y Y Y Y First-stageFtest 357.1 357.2 343.4 363.0 55

TableOA-4: Robustnessofbank-levelestimatestoalternativetermpremiumestimates ThistablereportsOLSestimatesfromaregressionofbankloangrowth(excludingandincludingcreditlines)onterm premiumestimatesfromthetermstructuremodelinAdrianetal.(2013)(columns1-4)andKimandPriebsch(2020) (columns 5-8). The sample period is 1994Q1-2019Q4. Specification details are as in baseline Table 3. TP shock estimatesbasedontheKim-Priebschmodelarenotreportedastheirtermpremiumseriesismonthlyandprecludesa high-frequencyeventstudy. Standarderrorsaredouble-clusteredbybankandquarter. Significance: *p<.1;**p<.05; ***p<.01. (1) (2) (3) (4) (5) (6) (7) (8) Dependentvariable Loangrowth Loangrowth Loangrowth Loangrowth t,t+4 t,t+4 (includingcreditlines) (includingcreditlines) t t Adrianetal.(2013) KimandPriebsch(2020) IV OLS IV OLS OLS IV OLS IV TPshock TPshock Termpremium 2.70*** 3.76 1.75*** 5.48** 0.99*** 1.49*** 0.45 0.99*** t (0.73) (2.99) (0.60) (2.72) (0.29) (0.35) (0.27) (0.35) Expectations -1.80*** 3.64 -1.68*** 4.41** -1.10*** -1.06*** -1.01*** -0.97*** t (0.38) (2.20) (0.31) (2.00) (0.19) (0.19) (0.17) (0.17) ∆Shortrate -0.88*** -1.13 -1.19*** -0.70 -1.18*** -1.13*** -1.41*** -1.35*** t (0.19) (2.50) (0.16) (2.40) (0.14) (0.15) (0.11) (0.12) Observations 634,153 634,153 631,463 631,463 634,153 634,153 631,463 631,463 R2 0.22 0.22 0.23 0.22 0.39 0.23 0.38 0.23 Macrocontrols Y Y Y Y Y Y Y Y Bankcontrols Y Y Y Y Y Y Y Y BankMSAFE Y Y Y Y Y Y Y Y First-stageFtest 51.83 53.25 463.2 462.8 56

TableOA-5: PlacebotestforTaperTantrumanalysis ThistablereportsOLSestimatesfromaregressionofbanklendingoutcomesontheinteractionofbankcapitalanda “Post”dummy. Thedataareatthebank-firmloanlevelandcovertwodistinctperiods: 2012Q4vs2013Q1(labelled 1Qincolumns1and3),and2012Q2-2013Q3vs2012Q4-2013Q1(columns2,4,and5).Incolumns1and3,the“Post” dummy takes value one in 2013Q1 (zero in 2012Q4). In columns 2, 4, and 5 it takes value one in 2012Q4-2013Q1 (and zero in 2012Q2-2013Q3). The number of loans spread observations is insufficient to estimate the regression forthe2012Q4vs2013Q1period. SpecificationdetailsareasinTable5. Standarderrorsclusteredbybank-firmin parentheses. Significance: *p<.1;**p<.05;***p<.01. (1) (2) (3) (4) (5) Placeboperiod 1Q 2Q 1Q 2Q 2Q A.Newloan B.Loanvolume C.Loanspread t+1 t+1 (log) t+1 Capitalratio ×Post 0.005 -0.003 0.034 -0.002 -0.011 t t (0.006) (0.004) (0.010) (0.017) (0.040) Observations 40,009 80,067 2,262 5,200 1,660 R2 0.635 0.656 0.761 0.761 0.800 Relationshipduration Y Y Y Y Y Firm×quarterFE Y Y Y Y Y Bankcontrols Y Y Y Y Y Bankcontrols×Post Y Y Y Y Y Growthexpectations Y Y Y Y Y Growthexpectations×Post Y Y Y Y Y 57

TableOA-6: RobustnessofTaperTantrumanalysistocontrollingforbankreserves ThistablereportsOLSestimatesfromaregressionofbanklendingoutcomesontheinteractionofbankcapitalanda “Post”dummythattakesvalueoneafter2013Q2,andzerootherwise,whereweadditionallycontrolforbankreserves. Thedataareatthebank-firmloanlevelandcoveraperiodofbetweenthreeandfivequartersbeforeandafter2013Q2. SpecificationdetailsareasinTable5exceptwealsoincludethebank-levelshareofreservesintotalassets(measured attheendof2012Q1)inlevelsandinteractedwiththePostdummy(coefficientsnotshown).Standarderrorsclustered bybank-firminparentheses. Significance: *p<.1;**p<.05;***p<.01. (1) (2) (3) #ofquartersbeforeandaftertapertantrum 3Q 4Q 5Q A.Newloan t+1 Capitalratio ×Post -0.013*** -0.010*** -0.010*** t t (0.003) (0.002) (0.002) Observations 133,262 174,659 293,348 R2 0.651 0.656 0.620 B.Loanvolume(log) t+1 Capitalratio ×Post -0.044*** -0.059*** -0.055*** t t (0.009) (0.013) (0.012) Observations 8,506 11,312 17,489 R2 0.774 0.774 0.883 C.Loanspread t+1 Capitalratio ×Post 3.011*** 1.957** 2.097* t t (0.551) (0.709) (0.934) Observations 2,254 2,915 4,151 R2 0.830 0.825 0.837 Relationshipduration Y Y Y Firm×quarterFE Y Y Y Bankcontrols Y Y Y Bankcontrols×Post Y Y Y Growthexpectations Y Y Y Growthexpectations×Post Y Y Y 58

TableOA-7: Termpremiumandbank’smaturitygap ThistablereportsOLSestimatesfromaregressionofbanks’maturitygaponKim-Wrighttermpremiumestimates. Thesampleperiodis1997Q2-2019Q4. SpecificationdetailsareasinbaselineTable3. Standarderrorsareclustered bybank. Significance: *p<.1;**p<.05;***p<.01. (1) (2) (3) Dependentvariable Maturitygap t+1 OLS IV OLS TPshock Termpremium -17.060*** -24.346*** -16.201*** t (0.197) (1.548) (1.005) Expectations 4.743*** 4.998*** 15.825*** t (0.112) (0.785) (0.665) ∆Shortrate 0.436*** 0.037 7.947*** t (0.035) (0.605) (0.452) Observations 573,302 573,272 582,207 R2 0.729 0.260 0.669 Macrocontrols Y Y Y Bankcontrols Y Y Y BankMSAFE Y Y Y First-stageFtest 226.95 59

TableOA-8: Termpremiumandbank’smaturitygaparoundtheTaperTantrum This table reports OLS estimates from a regression of banks’ maturity gap on the interaction of bank capital and a “Post” dummy that takes value one after 2013Q2, and zero otherwise. The data are at the bank-quarter level and coveraperiodofbetweenthreeandfivequartersbeforeandafter2013Q2. Specificationsincludethefollowingbank controls: size(log-assets),capitalratio,coredeposits(%liabilities)andsecurities(%assets),inlevelsandinteracted with the Post dummy. Standard errors double-clustered by bank and quarter in parentheses. Significance: *p<.1; **p<.05;***p<.01. (1) (2) (3) #ofquartersbeforeandaftertapertantrum 3Q 4Q 5Q Dependentvariable Maturitygap t+1 Capitalratio ×Post -0.229*** -0.224*** -0.204*** t t (0.028) (0.021) (0.017) Observations 22,599 33,912 45,229 R2 0.181 0.181 0.181 Bankcontrols Y Y Y BankcontrolsxPost Y Y Y QuarterFE Y Y Y 60