GSEs, Mortgage Rates, and Secondary Market Activities

Finance and Economics Discussion Series Divisions of Research & Statistics and Monetary Affairs Federal Reserve Board, Washington, D.C. GSEs, Mortgage Rates, and Secondary Market Activities Andreas Lehnert, Wayne Passmore, and Shane M. Sherlund 2005-07 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.

GSEs, Mortgage Rates, and (cid:3) Secondary Market Activities Andreas Lehnert Wayne Passmore Board ofGovernorsofthe Board ofGovernorsofthe Federal Reserve System Federal ReserveSystem Washington,DC20551 Washington,DC20551 (202)452-3325 (202)452-6432 Andreas.Lehnert@frb.gov Wayne.Passmore@frb.gov Shane M. Sherlund Board ofGovernorsofthe Federal ReserveSystem Washington,DC20551 (202)452-3589 Shane.M.Sherlund@frb.gov First Version: August2004 ThisVersion: January 12, 2005 Preliminary draft. We welcome all comments; please contact us directly for the latestversion. (cid:3)CathyGessertprovidedexcellentresearchassistance. WethankBenBernanke,DarrelCohen, KarenDynan,KieranFallon,MichaelFratantoni,MikeGibson,PaulKupiec,EllenMerry,Steve Oliner,BobPribble,DavidSkidmore,andJonathanWrightforhelpfulcommentsandsuggestions. The opinions, analysis, and conclusionsof this paper are solely those of the authorsand do not necessarilyreflectthoseoftheBoardofGovernorsoftheFederalReserveSystem.

GSEs, Mortgage Rates, and Secondary Market Activities Abstract Fannie Mae and Freddie Mac are government-sponsored enterprises (GSEs) that purchasemortgagesandissuemortgage-backedsecurities(MBS).Inaddition,the GSEs are active participants in the primary and secondary mortgage markets on behalfoftheirownportfoliosofMBS.Becausetheseportfolioshavegrownquite large, portfolio purchases as well as MBS issuance are likely to be important forces in the mortgage market. This paper examines the statistical evidence of a connection between GSE actions and the interest rates paid by mortgage borrowers. We find that both portfolio purchases and MBS issuance have negligible effects on mortgage rate spreads and that purchases are not any more effective than securitization at reducing mortgage interest rate spreads. We also examine the 1998 liquidity crisis and find that GSE portfolio purchases did little to affect interest rates paid by borrowers. These results are robust to alternative assumptionsabout causalityand tomodelspecification. Journalof EconomicLiteratureclassificationnumbers:H81,G18,G21 Keywords: Mortgagefinance,Government-SponsoredEnterprises,Financialstability

PreliminaryDraft 1 Introduction and Summary The housing-related government-sponsored enterprises (GSEs) Fannie Mae and Freddie Mac buy mortgages from originators and use them to issue mortgagebacked securities (MBS). These GSEs also keep many mortgages in their own portfolios, either as whole loans or as MBS. By 2003, these portfolios amounted to almost $1.5 trillion of home mortgages, or more than 22 percent of the entire homemortgagemarket. Earnings from mortgages held in portfolio clearly benefit GSE shareholders. TheGSEs’portfolioholdingsmayalsobenefit mortgageoriginatorsand,tosome degree, homeownerswith conforming mortgages.1 In particular, unusually heavy and sustained portfolio purchases by GSEs might bid up the price of new mortgages, allowing originators to profit more or giving originators greater scope to lowermortgageinterestrates paidby newborrowers. GSEactions,however,maynotbespecial. Manyloansotherthanconforming mortgagesaresecuritized;inthesemarketsavarietyofprimarymarketoriginators sell loans to secondary market entities for securitization. These other markets do notfeature GSEs; instead,participantsare purelyprivateentities. Nonetheless,in acompetitiveequilibrium,totalsecondarymarketpurchasesandmortgagemarket 1One oft-cited measure of the benefit the GSEs pass through to mortgage borrowers is the so-calledjumbo-conformingspread,thatis,thedifferenceininterestratespaidbyborrowerswith conformingmortgagesandthosewithmortgagesthataretoolargetobeboughtbytheGSEs(jumbos). Estimatesofthisspreadrangefrom16to27basispoints;seeMcKenzie(2002);Ambrose, LaCour-Little,andSanders(2004);andPassmore,Sherlund,andBurgess(2004).Passmore,Sherlund,andBurgess(2004)findthatofa16basispointjumbo-conformingspread,only7basispoints areattributabletotheGSEfundingadvantage. 1

PreliminaryDraft spreads are determined simultaneouslythrough theforces of supply and demand. For example, if the risk-adjusted spread between the interest rate on a loan and a benchmark funding rate were unusually wide, secondary market participants will buymoreloans. As aresult,iftheprimarymarket iscompetitive,primarymarket participantsmightextend moreloans; iftheydoso, theequilibriumprimary market interest rate mightdecline and thespread would then return to a morenormal level. However, the secondary market in conforming mortgages is far from a textbook competitive market. On one side of the market, the GSEs are almost the exclusivepurchasers of conforming mortgages. Both their size and their government charters raise the possibility that they are not like other secondary market purchasers. On the other sideof the market, a handful of large mortgageoriginators are the largest sellers of conforming mortgages. Thus, most GSE mortgage purchases are likely the outcome of a complicated dynamic strategic interaction amongahandfuloflargeentities. The negotiatednature ofGSE mortgagepurchases suggeststhat theremay be longandvariablelagsbetweenGSEactionsandanypotentialimpactsonprimary mortgagerates. Inaddition,theinstitutionalstructureoftheconformingmortgage marketsuggestsotherreasonsforsuchlags. First,averagepricingprevailsinconforming mortgage markets because credit risks are small relative to information, servicing,andmarketingcostsand becauseborrowersarerelativelyinsensitiveto smallchangesinmortgagerates.2 Second,mortgageinterestratestendtobequot- 2Foradiscussionofaveragepricinginthemortgagemarket,includingestimatesofborrower 2

PreliminaryDraft edineighthsof1percentagepointincrements,sothatsmallchangesinthecostof funds may not be enough to warrant a change in mortgage rates. Finally, lenders might find it costly to adjust primary market rates to large swings in secondary marketpricing thattheyviewas transient. If secondary market prices are not transmitted automatically into the primary mortgage market, GSE secondary market interventions may not be an effective policy tool for influencing mortgage rates. But even if the social benefit of the GSEportfoliosisnotevidentduringtimeswhenmarketsarefunctioningnormally, purchases could havea particularly importanteffect duringfinancial crises, when the GSEs may act affirmatively to calm market crises with larger-than-normal portfolio purchases. For example, during periods of financial market stress, investors demand larger than normal compensation for holding all risks, including the risks inherent in conforming mortgages. In such periods, GSEs might decide to buffer mortgage originators from financial market shocks, and thus limit the impactofsuchshockson mortgageborrowers. Again, the GSEs may not be special in performing this function. A purely privateinvestorwhobelievedmortgagespreadsto betoolarge wouldbehavelike the GSEs and buy mortgages. What is unique about the GSEs is not that they buy assets when they expect a large return on equity, but that, unlike a purely privateinvestor,GSEscanissuedebtthatotherinvestorstreatasimplicitlyinsured by the government. Clearly, during a time of crisis, such debt might be better received in the markets than purely private debt and, ironically, financial crises sensitivitytoratechanges,seeHancock,Lehnert,Passmore,andSherlund(forthcoming). 3

PreliminaryDraft may allow GSE shareholders extraordinary profit opportunities. The key policy questions,however,arewhetherGSEactionsactuallyinfluenceprimarymortgage ratespreads, andifso, howlargetheactual benefits tomortgageborrowersare.3 In addition, one can also question the necessity of the GSEs using their portfolios to influence mortgage rates. Mortgage originators can easily swap whole loansforGSE-guaranteed mortgage-backedsecurities.4 TheseMBScarry acreditguarantee, and investorsconsiderthem tobe safeand liquidinvestments. Thus, GSEmortgagesecuritizationmightbeasimportant,orevenmoreimportant,than portfoliopurchasesininfluencingmortgagerates. Indeed,GSEportfoliopurchasesarelikelytocreateasocialbenefitbeyondthatprovidedbyMBSissuanceonly if GSE debt actually tapped a net new source of demand for mortgage assets, thereby lowering the GSEs’ cost of funds. However, since the characteristics of GSEdebtarealreadyavailableinotherdebtinstruments(orcouldbecreatedfrom existingdebtinstruments),itseems difficulttoidentifythesenewinvestors.5 Based on the arguments presented above, GSE portfolio purchases and securitizationactivityarelesslikelytoconferasocialbenefitifGSEssimplyfollowa profit-maximizingstrategyofbuyingmortgageswhenspreads are unusuallyhigh unless mortgage rate spreads react rapidly to GSE purchases. Conversely, the GSEs could confer a social benefit if they actively managed the risk spreads paid 3Notethatinthispaper,wefocusonlyonthegrossbenefitsprovidedbytheGSEs;wedonot attempttonetouttheestimatedcostsassociatedwiththesebenefits. 4Attheendof2003,approximately$3.3trillioninGSE-issuedMBSwereoutstanding;ofthis, about$1.3trillionwereheldinGSEportfolios. 5Indeed,manyofthe“newer”buyersofGSEdebt—Asianbuyers,inparticular—appeartobe simplysubstitutingexplicitlyinsuredTreasurydebtwithimplicitlyinsuredGSEdebt. 4

PreliminaryDraft by conforming mortgage borrowers. In this view, the GSEs might be targeting mortgage rate spreads and heavily intervening in secondary mortgages whenever actual spreadsdeviatefrom atarget, inan effort topromotesocial goals. We analyze the effect of GSE secondary market activities (defined as purchases and securitization) on mortgage interest rate spreads (in both primary and secondary markets). If GSE secondary market activitieshave a social benefit, we wouldexpectGSEactivitiestosignificantlyaffectprimarymarketspreads. Inaddition,iftheGSEswereaffirmativelymanagingmortgageratespreads,wewould expectthemto react quicklytonews likelyto affect mortgagespreads. Our statistical approach directly estimates the effect of MBS issuance on primary and secondary mortgage rate spreads. In addition, we capture the effect of GSE debt issuance via our measures of portfolio purchases, which are almost exclusivelyfinanced viadebt issuance.6 Our main finding is that GSE actions (whether portfolio purchases or gross MBSissuance)havenegligibleeffectsonprimaryorsecondarymortgagespreads. Thatis,asuddenincreaseinGSEportfoliopurchasesorMBSissuancehasessentially no long- or short-run effects on mortgage spreads. This finding is robust to alternativespecifications,variabledefinitions,and identifyingassumptions. Inaddition,wecharacterizethetime-seriescausalityamongmortgagespreads and GSE actions. Intuitively,this procedure measures whether GSE actions react 6NotethatsomeGSEdebtissuanceisusedtofundnon-mortgageassets; however,debtused forthese alternativepurposeswill likely havelittle impacton themortgagemarket. As a result, debt issuance may be a poor proxy to measure the GSEs’ impact on the mortgage market. We thereforeusegrossportfoliopurchasestomeasuretheGSEs’impactonthemortgagemarket. 5

PreliminaryDraft to unexpected movements in spreads, whether spreads move in reaction to unexpected actions by the GSEs, or whether both happen together. Determining the causal relationship among time series is notoriously difficult. Nonetheless, the evidencesupportstheviewthatGSEpurchasesfollowspreads. Thatis,unusually large portfolio purchases are not followed by unusual drops in mortgage interest rate spreads. Instead, the time-series evidence suggests that unusually large increases inspreads arefollowedby fasterthannormalportfoliogrowth. While many studies have examined the general effect of GSEs on mortgage rates, only a few have attempted to estimate the specific effects of GSE portfolio and securitization activity on mortgage rates. A cointegration study by Naranjo and Toevs (2002) uses data covering 1986–98 and concludes that GSE securitizationactivityandportfoliopurchases areassociated withlowerspreads between conforming mortgageinterest rates and comparable Treasury interest rates.7 Further,theyconcludethattotalmortgagepurchases are30percent moreproductive, dollar for dollar, at lowering the conforming-Treasury spread than is securitization activity. Similarly, Gonzalez-Rivera (2001) uses cointegration analysis, but her study of data from 1995–99 shows that larger portfolio purchases are associated with wider spreads. Further, she shows that about 84 percent of movements inthesecondary marketspread are passedthroughtotheprimarymarket spread. ThereareseveralimportantdifferencesbetweenourstudyandthoseofNaranjo and Toevs and Gonzalez-Rivera. First, we use more recent data (1994–2003). 7They also find that portfolio purchases and securitization are associated with lower jumbo- Treasuryspreadsandhigherjumbo-conformingspreads. 6

PreliminaryDraft Second,weusegrossGSEportfoliopurchases,definedaswholeloans,ownMBS, and other MBS purchased, as a measure of the GSEs’ portfolio activity, not just Fannie Mae’s total mortgage purchases or net GSE portfolio purchases. Third, weestimateafull-informationsystemofequationstoavoidconfoundingtheindividual effects of portfolio purchases and securitization activity, and primary and secondarymarketspreads toswapyields. Finally,werecognizethatanylong-run cointegrating relationship does not necessarily specify a unique causal relationship. Wefollowtheearlierstudiesbyusingdataonmortgageratespreads tomodel theGSEportfoliomanagementdecision. However,thisbusinessdecisionislikely primarilyinfluencedbytheoption-adjustedspread(OAS)onmortgagesandGSE debt yields. The GSEs will find it more profitable to invest in mortgages when the OAS is unusually high or GSE debt yields are unusually low. Unfortunately, data on GSE data yields that correct for maturity are only available from 2000 on. In addition, we would ideally use high frequency data, with daily or weekly observations on spreads and GSE activities. However, data on GSE actions are available only at a monthly frequency. Thus, as is standard in this literature, we use monthly averages of raw mortgage rate spreads and monthly data on GSE actionsand includecontrolsforcredit andprepayment risksinouranalysis. 7

PreliminaryDraft 2 Theory Mortgageinterest rates (and theirspreads)are affected by investors’expectations about mortgage risks (mainly credit and prepayment risks), financial market liquidity, and investors’ expectations about the actions of all other participants, including the GSEs. At the same time, the GSEs are buying mortgages for their portfolios for many of these same reasons, and also because of the current level and expectedfuturetrajectory ofmortgageratespreads. 2.1 Structural Model Considerasimplemodelofmortgageinterestratespreads(R )andGSEactivities t suchas portfoliopurchases (P ): t (1) R = a P +a Z +A P +A R +"R; t P t Z t P t−1 R t−1 t (2) P = b R +b Z +B P +B R +"P: t R t Z t P t−1 R t−1 t Z represents (exogenous) variables that affect mortgage pricing, such as proxies t forprepaymentand credit risk.8 Theinnovations(orshocks)"R and"P canbeinducedbyunexpectedchanges t t to liquidity, investor risk aversion, or uncertainty. These shocks are the primi- 8In our empirical work, instead of this stylized two-equation system, we use a richer fourequationsystem. ForR t,weusebothprimaryandsecondarymarketspreads;forP t,weuseboth grossportfoliopurchasesandgrossMBSissuance;andforZ t,weuseproxiesforcreditriskand prepaymentrisk. Theinteractionofthesevariablesisnecessarilycomplex;forsimplicitywewrite thesysteminthisform. 8

PreliminaryDraft tiveexogenousforceswhich“buffetthesystemandcauseoscillations”[Bernanke (1986)]. In our empirical results we discuss the dynamic effects on R and P of onestandarddeviationmovementsintheseshocks. Equations (1)–(2) provide a simple statistical description of the equilibrium interaction of GSE actions, prices, and other observed and unobserved market forces. The theoretical connection among these variables could be quite complicated,inpartbecausetheequilibriumprobablydependsonhowasmallnumberof entities expect the others to behave. Equation (1) captures the stabilizing effects (if any) of GSE actions on spreads. Equation (2) captures the business decision of the GSEs; in particular, how their portfolio managers react to movements in spreads. In this representation, a is the contemporaneous effect of this period’s pur- P chasesonthisperiod’sspreadsandA isthedirecteffectoflastperiod’spurchas- P es on this period’s spreads.9 In the same way, b captures the contemporaneous R effect ofspreads onpurchases andB captures thelaggeddirect effect. P Any model in which GSE purchases pushed down mortgage spreads in the sameperiodwouldimplythata < 0. Inthesameway,anymodelinwhichwider P spreads increased GSE purchases in the same period would imply that b > 0. R These are only the very short-run effects, though. The long-run effect of GSE purchasesonspreadsisthesumofthecontemporaneousdirecteffect(a )andthe P lagged direct effect (A ); in addition, GSE purchases will affect spreads through P 9Forclarity,wehavehereexcludedmorethanonelaggedvaluefromourequations,although intheempiricalworkweallowforanarbitrarynumberoflags. Aswediscussinsection4.1,we usetwolagsoftheendogenousvariablesandonelagoftheexogenousvariables. 9

PreliminaryDraft the indirect effects of equation (2). Because these direct and indirect effects can quickly become complicated, the standard method of describing them is via impulse response functions. This is just the dynamic response of each of the variables in the system to a standard shock to either "P (to capture the effect of an t unexpected increase in purchases) or "R (to capture the effect of an unexpected t increaseinspreads). Granger causality tests and other statistical techniques examine the effects of lagged variables on current variables, that is, whether A = 0 or B = 0. Based P R on these tests and our sensitivityanalysis, as we discuss below, we conclude that the statistical evidence supports the view that GSE actions react to interest rate movementsratherthan theotherway around. 2.2 Reduced-Form Estimates Withoutadditional assumptions,we cannot uniquely identifyboth the contemporaneous effect of purchases on spreads, a , and the contemporaneous effect of P spreads on purchases, b . To estimateb we mustrestrict a , either by forcing it R R P to be zero or by setting it to some value suggested by other data or theory. In the sameway, toestimatea wemustrestrictb . P R In this simplified system, the identification problem can be seen by rewriting the system of equations (1)–(2) with the contemporaneous terms on the left-hand 10

PreliminaryDraft side: R −a P = a Z +A P +A R +"R; t P t Z t P t−1 R t−1 t P −b R = b Z +B P +B R +"P: t R t Z t P t−1 R t−1 t Usingmatrixnotation,thiscan bewrittenas: 2 30 1 2 30 1 2 3 0 1 1 −a R A A R A "R 6 P 7B t C 6 R P 7B t−1 C 6 Z 7 B t C 4 5@ A = 4 5@ A+4 5Z +@ A: t −b 1 P B B P B "P R t R P t−1 Z t Define X as the vector of variables (R ;P )0 and " as the vector of structural t t t t innovations("R;"P)0. Then wecan rewritethesystemas: t t (3) (cid:8) X = (cid:8) X +(cid:8) Z +" : 0 t 1 t−1 Z t t Theinnovationsare assumedto have avariance-covariancematrix of(cid:3). Because the shocks are defined as pure innovations to the structural system, it is standard practicetoassumethattheyareuncorrelated. Thus,weassumethat(cid:3)isadiagonal matrix,i.e.,itsoff-diagonalelementsare allequal tozero. Wecannotestimatetheparametersofthestructuralmodeldirectly. Instead,we estimatetheparametersfromanunrestrictedreduced-formmodeland,underaset of identifying assumptions, compute the structural parameters from the reduced- 11

PreliminaryDraft formparameters. Moreprecisely,weestimatethecoefficientsfrom themodel: (4) X = (cid:8)−1(cid:8) X +(cid:8)−1(cid:8) Z +(cid:8)−1" t 0 1 t−1 0 Z t 0 t (cid:17) ΓX +Γ Z +u : t−1 Z t t Thus, Γ is a mixture of the structural coefficients in (cid:8) and (cid:8) . Further, the 0 1 reduced-form errors, u , are a linear combination of the structural shocks, " . t t Thus, even though the variance-covariance matrix of" is diagonal, the variancet covariance matrix of u will be, in general, non-diagonal. In fact, the variancet covariancematrixofthereduced-form errors is: (5) E[uu0] = (cid:8)−1(cid:3)(cid:8)−10 (cid:17) (cid:6): 0 0 Identifyingrestrictionsare often used to computethe elementsof(cid:8) , (cid:8) , and 0 1 (cid:3)fromestimatesofΓand(cid:6). Iftherearenendogenousvariablesinthesystem,Γ containsn2 uniqueentriesand(cid:6),whichissymmetric,containsn(n+1)=2unique entries. Thus, we can identify n2 + n(n+1)=2 structural parameters. However, (cid:8) containsn(n−1) uniqueentries(itsdiagonalelementsareones),(cid:8) contains 0 1 n2 uniqueentries,and(cid:3),whichisdiagonal,containsnuniqueentries—foratotal of 2n2 structural parameters. Thus, identification requires n(n−1)=2 additional restrictions on the structural parameters. The most common type of identifying restriction is that (cid:8) be triangular (that is, all the entries above [or below] the 0 diagonalareall equal tozero). 12

PreliminaryDraft Assuming that (cid:8) is triangular is tantamount to assuming either that b = 0 0 R or that a = 0. Under the assumption that a = 0, contemporaneous shocks to P P purchases, "P, do not affect spreads, R , in period t. However, contemporaneous t t shocks to spreads, "R , may affect purchases, P , in period t.10 In other words, t t under this assumption, GSE actions may respond to all information in a given period,butspreads respondwitha slightlag. Under the alternative assumption that b = 0, contemporaneous shocks to R spreads, "R , do not affect purchases, P , in period t. However, contemporaneous t t shockstopurchases,"P,mayaffectspreads,R ,inperiodt. Inotherwords,under t t thisassumption,spreadsmayreacttoallcontemporaneousinformationwhileGSE activitiesreact witha slightlag. Thelatteridentifyingassumptionisconsistentwiththenotionthatspreadsare determined continuously over time in financial markets, while GSE activities are theresultofslower-movingbusinessdecisions. However,wealsoconsiderseveral different identifying assumptions as robustness checks on our preferred, baseline specification. 3 Data We obtained consistent data on GSE portfolio purchases, securitization volume, and mortgageinterest rate spreads at a monthlyfrequency for the ten-year period 1994–2003,for120totalobservations. Inaddition,ourdatasetcontainscovariates 10Becausenorestrictionsareplacedon(cid:8) 1,shockstoeitherpurchasesorspreadsinperiodtcan affectbothpurchasesandspreadsinperiodt+1. 13

PreliminaryDraft designedtocontrolforcredit and prepaymentrisk. Our measure of GSE portfolio purchases is the sum of Fannie Mae and Freddie Mac’s gross retained portfolio purchases of mortgageassets, including whole loans, own MBS, and other MBS.11 Our measure of securitization volume is the sum of Fannie Mae and Freddie Mac’s gross issuance of MBS. These data are availableontheGSEs’monthlysummaryreports. Aspartofourrobustnessanalysis,wenormalizeportfoliopurchasesandgross MBSissuancebymeasuresofthesizeoftheresidentialmortgagemarket. Wefollow other studies in using the monthlytotal volumeof new residential mortgages originated (both purchase and refinance) as our measure of total market size. We estimate this measure with the time series of mortgage originations as reported undertheHomeMortgageDisclosureAct (HMDA). Weusebothprimaryandsecondarymarketinterestratestocomputeourmeasures of mortgage rate spreads. The primary market mortgage rate is defined as themonthlyaverageinterestrateonnew30-yearfixed-ratemortgages,fromFreddieMac’sprimarymortgagemarketsurvey. Secondary marketmortgageratesare defined as the monthlyaverage current coupons on Fannie Mae and Freddie Mac 30-yearMBS.12 Wedonotusethelevelsoftheseprimaryandsecondarymarketmortgagerates in our analysis; instead, as is common practice, we use the spread to a relevant 11Measuringthe effectsofGSE portfoliopurchasesis difficultbecausetherecanbelonglags betweenwhenaGSEcommitstoamortgagepurchaseandwhenthepurchaseisbroughtontothe GSEs’ books. However, the GSEs do not release enough data publicly to attempt to adjust for theselags. 12Theuseofmonth-enddatadidnotmateriallyalterourresults. 14

PreliminaryDraft risk-free rate. We have experimented with a variety of different measures of the risk-freerate. However,inouranalysisherewereportresultsusingthespreadsto a simpleaverage of the5-year and 10-year Treasury rates. Our results are almost completely unchanged by the choice of interest rate index. As shown in table 1, changes in GSE debt yields, Treasury yields, and swap yields are very highly correlated.13 Theprimaryriskspricedintomortgagerates(butnotrisk-freerates)arecredit risk(theriskofdefault)and prepaymentrisk(theriskofrefinancing). Asaproxy for credit risk, we use the realized serious delinquency rate on conforming mortgages owned by the GSEs.14 We proxy prepayment risk with a measure of the incentivetorefinance. Inparticular,weusethemortgagecoupongap—thespread between the weighted average coupon on existing securitized 30-year mortgages and the30-yearfixed-rate mortgagerate. Descriptive statistics for key series are provided in table 2. Figure 1 plots thetimeseries ofGSE portfoliopurchases, GSE securitizationvolume,mortgage market spreads, and mortgage delinquencies and the coupon gap. Note that the financial market crisis of October 1998 was associated with a sharp widening of spreads; as shown in the figure, primary market spreads rose about 70 basis points. Serious mortgage delinquencies have a fairly narrow range, between 45 and 62 basis points; the coupon gap has varied substantially from −102 to +168 basispoints. 13We alsouseda simpleaverageof5-yearand10-yearswapsandfoundnosignificantdifferencesinourresults. 14“Seriousdelinquencies”aredefinedasmortgages90ormoredayspastdueorinforeclosure. 15

PreliminaryDraft 4 VAR Results In this section we examinethe relationshipbetween GSE activitiesand mortgage market spreads using VAR techniques. We compute the dynamic responses of mortgage rate spreads to an unexpected shock to GSE activities under our baseline identifying assumption. Our primary specification is in first-differences; we alsoreportafullsetofresultsunderanalternativespecificationinnormalizedlevels. As a robustness check on these results, we examine the cumulative impulse responses under a variety of alternative identifying assumptions. We also show that GSE actions during the liquidity crisis of 1998 were not extraordinary; further, had GSEs done nothing during this period, primary and secondary market spreadswouldhaveevolvedin aboutthesameway. 4.1 Estimation We estimate theparameters of an unrestricted reduced-form VAR as described in section 2.2. Rather than use the levels of the variables, we estimated the model using first-differences of the variables. (We use the logs of portfolio purchases andsecuritizationactivity.) Asshownintable3wefindevidencethatsomeofthe variables might have unit roots in their levels; however, each series is stationary infirst differences.15 Wedeterminedtheoptimalnumberoflagstoincludeinourspecificationusing 15We have also estimated this VAR under a wide variety of alternative variable definitions, including a specification with the level of spreads and the growth of GSE activity. Our results wereinlinewiththosereportedhere:GSEactions,especiallyportfoliopurchases,haveverylittle effectonspreads. 16

PreliminaryDraft the Akaike Information Criterion (AIC). We found that 2 lags of the endogenous variables(GSEactionsandspreads)and1lagoftheexogenousvariables(prepaymentand creditrisk proxies)minimizedthecriterion. We are interested in testing the ability of GSEs to stabilize mortgagemarkets aswellastolowermortgagespreadsinthelongrun. However,thesetwogoalsare notnecessarilyequivalent. TheGSEsmightbeabletodramaticallyaffectspreads in the short run, but then see these effects undone over time, leaving spreads unchanged in the long run. Conversely, the GSEs might be unable to affect spreads verymuchinagivenmonth,butmightbeabletocumulatetheireffectsovertime, producingasignificantlong-runeffect withoutlarge effects inanygivenperiod. For our baseline model, we report a full set of impulse response functions. For the alternative specifications and identifying assumptions, we summarize the short- and long-run effects of GSEs on rates with two statistics: The cumulative impulseresponseafter24monthsandthelargestcumulativeeffect(andwhenthat effect occurs). 4.2 Results Under Baseline Identification We impose a triangular (Wold) representation on our system of equations to examine the impulse response functions. (See section 2 for a discussion of this procedure.) This requires assuming that certain variables in our system do not respondcontemporaneouslytoshocksinotherequations. However,wedonotrestrictanyofthelaggedeffects,soallendogenousvariablesmayreacttoanyshock in the previous period. (As we discussed, we do not estimate reaction functions 17

PreliminaryDraft for credit and prepayment risk, but rather take these two variables as exogenous tooursystem.) Inourbaselineidentification,weassumethatshockstoGSEactivitieshaveno contemporaneouseffect onmortgagemarket spreads. That is,weassumethat the ordering of innovations in the Cholesky decomposition is: [" rs−r , " rp−r , " s , " p ]. Later, we compute impulse response functions under several alternatives to this baselineassumption. Intuitively,ourorderingofcontemporaneousresponsesassumesthattheGSEs observe all available information in a period before reacting. Our ordering thus perforce assumes that shocks to secondary market spreads occur during a period andarenotaffected byGSE actionsduringtherestoftheperiod. Primary market spreads react to secondary market spreads, but not to GSE actions. Of the GSE actions, gross MBS issuance reacts to spreads, and portfolio purchase volume reacts toall contemporaneousvariables. We consider several alternatives to our baseline ordering, including several in which spreads react to GSE actions. Under these alternatives, though, GSE actionshaveless ofalong-runeffect on mortgageratespreads. With an assumption about the contemporaneous effects of shocks in our system, we can estimate the effect of orthogonalized one standard deviation shocks of each innovation on each of our endogenous variables. In other words, we can estimate,forexample,thedynamicresponseofprimarymortgagemarketspreads to a one-time standardized shock to the portfolio purchase equation. A complete setoftheseimpulseresponsefunctionsisshowninfigure2. Eachpanelshowsthe 18

PreliminaryDraft effects ofonestandard deviationshocksofeach innovationonagivenvariable. Effect ofShock to Secondary Market Spreads (" rs−r ) Cumulatingtheeffects across all months,a onestandard deviationshock to thesecondary market spread innovation (about 8 basis points) ultimately leads to a 9.9 basis point increase in the secondary market spread, an 11 basis point increase in the primary market spread,a7.3percentincreaseinsecuritization,anda14.1percentincreaseinpurchases. The largest cumulativeeffects have occurred within three or four months oftheshock. Effect of Shock to Primary Market Spreads (" rp−r ) In the same way, a one standard deviation shock to the primary market spread innovation (about 4 basis points)ultimatelyleadstoa2.5basispointincreaseintheprimarymarketspread, a 1.1 basis point decrease in the secondary market spread, a 3.9 percent increase in securitization,and a 0.2 percent increasein purchases. Here, thelargest cumulativeeffects takeupto fivemonthsto occur. EffectofShocktoGSEActions(" and" ) Turningtotheeffectsofshocksto s p the securitizationand portfolio innovations,we find that a one standard deviation shocktothesecuritizationinnovation(about16percent)ultimatelyleadstoa14.6 percentincreaseinsecuritization,an8.7percentincreaseinpurchases,a2.1basis point decline in the secondary market spread, and a 1.7 basis point decline in the primary market spread—with the largest cumulativeeffects occurring within five monthsoftheshock. 19

PreliminaryDraft A one standard deviation shock to the portfolio purchase innovation (about 26 percent) ultimately leads to a 17 percent increase in purchases, a 3.8 percent increaseinsecuritization,a0.5basispointdeclineinthesecondarymarketspread, and a 0.2 basis point decline in the primary market spread.16 Here, the largest cumulativeeffects havetaken placewithinfourmonthsoftheshock. Thus, based on our impulse response analysis, we estimate that if the GSEs unexpectedly increase their portfolio purchases by $10 billion (or 12.7 percent of the 2003 average), the secondary market spread would decline about 0.3 basis points and the primary market spread would decline about 0.1 basis points over the long-run. But if the GSEs instead unexpectedly increased their securitization activity by $10 billion (or 6.3 percent of the 2003 average), we estimate that the secondary market spread would decline about 0.8 basis points and the primary market spread would decline about 0.7 basis points. These results suggest that GSE portfolio purchases haveeconomicallyand statisticallynegligibleeffects on mortgagemarketspreads,eveninthelongrun. Further,portfoliopurchasesarenot moreeffectiveat reducing mortgagemarket spreads thansecuritizationactivities. Variance Decomposition Figure 3 shows variance decomposition proportions forthekeydataseries. Variancedecompositionsindicatehowmuchoftheforecast error in a series is due to a shock to its own innovation and how much is due to shockstootherinnovations. Asshown,GSEactivitiesaccountforverylittleofthe 16The average (absolute) month-to-month change in GSE portfolio purchases is 24 percent. Moreover,suchashockamountstolessthan0.6percentoftheGSEs’combinedretainedportfolios. 20

PreliminaryDraft forecast errors of mortgage market spreads. Secondary market spreads, however, account for a sizable proportion of the forecast errors of primary market spreads and portfoliopurchasevolume. 4.3 The 1998 Liquidity Crisis In a dramatic example of the dynamics of the system, from August 1998 to October 1998, secondary market spreads widened about 69 basis points as a result of a liquidity shock. During the same time, primary market spreads widened 78 basis points, securitization volume decreased 5 percent, and portfolio purchases jumped76percent. Obviously, during this period, spreads and purchases moved much more than during normal periods. As a test of our model’s ability to explain GSE behavior and the evolution of spreads, we initialized the model with the secondary market spread changes in September and October and allowed the system to evolve endogenouslywithoutanyotherinformation. Asshownbythefirstpaneloffigure4, we forced secondary market spreads to increase about 30 basis pointsin Septemberand 39 basis pointsin October. We then allowedsecondary market spreads to evolveaspredictedbythemodel. Asshown,themodelpredictedthatthechanges in spreads wouldgradually declineto zero. In reality,spreads jumped around our model’s prediction in reaction to incoming data (such as prepayment and credit risk information). Primary market spreads, the next panel, show about the same dynamics. Turning to GSE actions, the next two panels show that our model is success- 21

PreliminaryDraft ful in predicting purchase volume in the peak crisis period, September through December 1998. Our model cannot explain the big drop in purchases in January 1999,ortheoffsettingriseinFebruary1999. However,inthelongruntheseforecast errors roughlycancel out, leavingthepredicted net changein purchases over the entire period close to the actual change. Our model also does a decent job of predicting securitization volumethrough this episode. Actual MBS issuance was much more volatile than predicted by our model, although, again, the forecast errors net outtoabout zero. Thus, the GSEs’ portfolio purchases during this period of financial market stress can be explained almost completely by their historical pattern of buying mortgages when spreads are wide. That is, there was nothing special about the GSEs’actionsduringthisperiodoffinancial market stress. However, even though GSE actions during the crisis were roughly consistent withtheirbehavioratothertimes,itcouldbethatthemagnitudesinthisparticular episode were large enough to mitigate the effects on spreads. To test this theory, we computed the changes in spreads that would have occurred had the GSEs not changed their portfolio behavior during the crisis. In other words, we force portfoliopurchases to beconstantat $30.5billionthroughtheepisode. In figure 4, the difference between this counterfactual experiment (the dotted line) and the model’s prediction (the dashed line) is our estimate of the effect of GSE portfolio purchases on the mortgage market during the crisis period. As shown, the GSEs’ portfolio purchases appear to have had little effect on either primaryorsecondary mortgagemarket spreads. However,therewas asubstantial 22

PreliminaryDraft effect onsecuritizationactivity. 4.4 Results Under Alternative Identifying Assumptions With four endogenous variables in our system there are potentially four factorial, or 24, different triangular representations of the system. The previous section reported results in detail under our baseline identification assumptions. In this section,wesummarizeresultsunderseveralalternativerepresentations. Wefocus ontwoclassesofalternatives: first,reasonablereorderingsofourbaselinespecification(VAR)and,second,orderingssuggestedbytheefficientmarketshypothesis (EMH). For each alternative ordering, we summarize the long-run response of variables to shocks with the cumulativeimpulseresponse functionsshownin table4. Wealsosummarizetheshort-runresponseofvariablestoshocksby reportingthe largest cumulative response in table 5. The column labeled VAR 1 in the table reports the results under our baseline specification, which we have already discussed. From a policy perspective, the most interesting results are the effects of GSE actions on mortgage market spreads. As shown, these effects are negligible and donotvary significantlyunderthedifferentidentifyingassumptions. 4.4.1 Reorderings ofBaselineSpecification Itcouldbearguedthatspreadsreacttoallinformationavailableinagivenmonth, while GSE actions are somewhat constrained by prior agreements with originators. In other words, mortgage spreads react continuously to information flow, 23

PreliminaryDraft whileportfoliopurchasesand MBSissuancehavemoreinertia. The alternative representations, labeled VAR 2 and VAR 3, assume some degree of inertia in GSE actions. In VAR 2, the primary market spread reacts contemporaneouslyto shocksto GSE actionsand thesecondary market spread. GSE actions, in turn, react contemporaneously only to shocks to the secondary market spread. InVAR3,primaryandsecondarymarketspreadsreactcontemporaneouslytoshockstoGSEactions. GSEactions,inturn,donotreactcontemporaneously toshocksto spreads. Thefirstthreecolumnsoftables4–5summarizetheresultsunderthebaseline and alternative orderings. As shown, the results are not very different under the alternativesthanunderthebaselineordering. Inparticular,shockstoGSEactions continuetohavesmalleffectson mortgagemarketspreads. 4.4.2 Efficient Markets Hypothesis The efficient markets hypothesis (EMH) maintains that an asset’s price ought to reflect all relevant available information about that asset. The EMH can have strong implications for identification in VARs (see Sarno and Thornton, 2004). Our system contains two endogenous asset price variables: the secondary market spread and the primary market spread. Under a triangular representation, these two variables cannot react to the same set of shocks, so we must order them by their relative inertia. Primary mortgage market interest rates (and thus their spreads) are relatively sticky for the reasons discussed earlier. In addition, our measure of the primary market rate is Freddie Mac’s survey of lenders, which is 24

PreliminaryDraft conductedonlyweekly. In table 4, the column labeled EMH 1 reports results under a strong version of the EMH: GSE actions do not respond to shocks to any asset price equations. ThecolumnslabeledEMH2andEMH3reportresultsunderslightlyrelaxedversions of the EMH. In EMH 2, GSE actions respond to primary market shocks. In EMH 3, GSE actions react to both primary and secondary market shocks. As showninthetables,thesealternativeorderingsdonotsignificantlyalterourbaselineresults. 4.5 Results Under a Levels Specification The results reported in section 4.2 were based on a model estimated using the first-difference of mortgagespreads and the percent changes in GSE activity. We adopted this as our primary specification because GSE activity is clearly nonstationaryand GSE spreads, whileprobably stationary,are highlypersistent. Another approach is to normalize GSE actions by some measure of the size of the market and estimate the model using normalized GSE actions and the level of spreads. Notethat both spreads and normalizedGSE actions are fairly persistent, sotheseresultsmaybecontaminatedbyacommontrendinbothsetsofvariables. Aswediscussedinsection3,wenormalizeGSEportfoliopurchasesandgross MBS issuance by estimates of mortgage originations derived from the HMDA reports. Thenormalizedtimeseries are showninfigure 5. We follow the same estimation procedure as before. For completeness, we report a full set of impulse response functions in figure 6. Note the extreme per- 25

PreliminaryDraft sistence of shocks. In many cases the shocks continue to affect variables five years after the initial period. The top two panels of the figure show the response of spreads to standardized shocks to other variables, including GSE actions. As before, shocks to portfolio purchases have no statistically significant effect on mortgagerate spreads. Thus,ourmainresultsare unchanged. 4.6 Granger Causality Results F-test statistics from bivariate Granger causality tests among all variables in the system are reported in table 6. The null hypothesis in these tests is that the variable in the table row does not Granger-cause the variable in the table column. Forconvenience, if we can reject the nullwe say that the row Granger-causes the column (as opposed to saying that the row does not “not Granger cause” the column). Fromtable6weseethatmortgagemarketspreadsGranger-causeportfolio purchases and securitization volume, but not the converse. Also note that securitization volume and portfolio purchases Granger-cause each other. These results aresummarizedin figure7. 5 Cointegration Analysis Other studies, most notably Naranjo and Toevs (2002), have used cointegration techniquestostudythestatisticalrelationshipbetweenGSEactionsandmortgage market spreads. In effect, these techniques require the assumption that all variables in the system contain a common trend. This assumption is reasonable for 26

PreliminaryDraft variables such as the level of gross MBS issuance and the level of portfolio purchases. However, this assumption is economically less reasonable for variables such as mortgageratespreads and therelativelevelofGSE actions. Nonetheless, weproceed withthecointegrationanalysisforcomparisonwithearlierstudies. In this section we once again examine the impulse responses of variables in thesystemundertheassumptionthatallvariablesarecointegrated. Inplaceofour VAR,weestimatetheparametersfromavectorerror-correctionmodel,orVECM. Asin theVARanalysis,weuseastandardset ofidentifyingrestrictions. There are several economic and statistical reasons to doubt the estimates produced under this specification. First, the estimated cointegrating relationship is primarilybetweensecuritizationandportfoliopurchases;mortgagemarketspreads, thekeyvariablesofpolicyconcern,arestatisticallyinsignificantelementsofcointegratingvector. Second,itisstatisticallydifficulttodistinguishbetweenunitroot and near unit root series, so it is possible that the estimated long-run relationship is spurious.17 Third, we cannot establish that each of our endogenous variables is integrated of the same order. Last, it is generally accepted that interest-rate spreads cannot in principlebe integrated oforder one. That we cannot statistically reject this hypothesis is the result of using cointegration techniques over short time spans. We therefore discount the results in this section and emphasize our VARresultsinsection4. 17If the system is truly not cointegrated,using cointegrationtechniquescan producespurious results. Ifthesystemistrulycointegrated,however,usingadifference-stationaryVARproduces correct,albeitpossiblyinefficient,results. 27

PreliminaryDraft 5.1 VECM Results For the system defined in equation (3) to be cointegrated, a necessary condition is that each component of X be integrated of the same order (for example, each t componentofX hasaunitrootandisstationaryinfirstdifferences). Additionalt ly,theremustbeatleastonelinearcombinationofX thatisstationary. Asshown t in table3, portfoliopurchases mightnot contain a unit root. This is similarto the resultreported byGonzalez-Rivera(2001). Toexaminethesecondnecessaryconditionforcointegration,weuseJohansen’s (1988,1991)maximumlikelihoodprocedurestoestimateanylong-runcointegrating relationships in equation 3. Johansen’s trace and max statistics suggest that our system of equations has multiple cointegrating relationships. The first relationshipcan bewrittenas: (6) "b = 1:1882 + 0:0188(rp − r ) − 0:0050(rs − r ) − 1:8534s + p : t t t t t t t Here, "b denotes the deviations from the estimated long-run relationship. Note t thatwehavenormalizedwithrespect toGSE portfoliopurchases, sop hasa unit t coefficient inthisrepresentation. Figure 8 shows actual purchases, the equilibrium level of purchases (that is, the level consistent with the long-run relation given the actual values for other variables in equation 6), and the deviations from the long-run path. Purchases do track their long-run levels, but the deviations are frequent and exhibit serial correlation for extended periods, reaffirming the potential problems with using 28

PreliminaryDraft cointegrationtechniquesinsuch asetting. Tables 7–8 summarizetheimpulseresponsesfor each endogenous variableto shocks to each structural innovation for three different orderings of innovations. Theresults,onceagain, are similartothosefrom ourbaselineVARspecification: GSE portfoliopurchases havenegligibleeffects onmortgagemarketspreads. 5.2 Comparison With Other Studies Forcomparability,we also estimatetheNaranjo and Toevs (2002)and Gonzalez- Rivera(2001)specificationsusingourdataset. WefindthatunderNaranjoandToevs’sspecification,unanticipatedshockstoGSEportfoliopurchasesagainhaveno meaningfuleffect onmortgagemarket spreads. UnderGonzalez-Rivera’sspecification, unanticipated shocks to GSE portfolio purchases increase secondary market mortgage spreads slightly. However, as we discussed, these results are obtainedunderstrongand unrealisticassumptions. Moreover,NaranjoandToevsnormalizeportfoliopurchasesandsecuritization bythesizeofthemortgagemarkettorelateGSEactivitytotheprimarymortgage spread. WeuseanestimateofmortgageoriginationsderivedfromHMDAdataas thenormalization. Asshownintables7–8,anunanticipatedincreaseintheGSEs’ portfoliopurchaseshareofthemarketofabout5percenthaslittleeffectontheprimary market spread. Conversely,an unanticipatedincrease intheprimary market spread of about 9 basis points leads to an increase in the GSE portfolio purchase shareofthemarketofslightlyover1percent. Similarly,anunanticipated5.5percent increase in the GSEs securitization share of mortgage originations leads to 29

PreliminaryDraft 1.8 basis point increase in the primary market spread, and an unanticipated 9 basispointincreaseintheprimarymarketspreadincreasestheGSEs’securitization shareofmortgageoriginationsnearly2percent. Notethatunderthisspecification, wecan find no evidencethat GSE activitiesdecrease mortgagemarket spreads or thatportfoliopurchasesare inany sensebetterthan securitizationactivity. Gonzalez-Rivera(2001)relatestherawlevelof(net)portfoliopurchasestothe secondarymarketspread. Underthisspecification,anunexpectedincreaseinGSE portfolio purchases of $11 billion decreases the secondary market spread nearly 3 basis points. An unexpected 8 basis point increase in the secondary mortgage marketspread increases GSE portfoliopurchases nearly $8billion. 6 Conclusion This paper examines the statistical evidence of a connection between GSE secondary market actions and the interest rates paid by mortgage borrowers. While GSE portfolio purchases benefit GSE shareholders directly, the purchases must lower the mortgage rate paid by the homeowner in order to have a wider social benefit. We find, however, that portfolio purchases has economically and statistically negligibleeffectsonmortgagerates. Further,portfoliopurchasesarenotanymore effectiveatdecreasingspreadsthansecuritizationvolume. Ourresultswererobust to several alternative identifying assumptions, including those suggested by the efficient marketshypothesis. 30

PreliminaryDraft EarlierstudiesbyNaranjoandToevs(2002)andGonzalez-Rivera(2001)found thatGSE actionsdid significantlyaffect mortgagespreads. Werepeated thecointegration analyses used in those studies with our more recent data set and found, again,thatGSE actionshaveonlynegligibleeffects onmortgagespreads. Further,westudiedthe1998liquiditycrisisandfoundthatGSEactionsgenerallyfollowedthepredictionsofthemodel. HadGSEactionsremainedunchanged throughthisepisode,weestimatethatmortgagespreadspaidbyborrowerswould have been essentially unchanged. Thus, the GSEs do not appear to have played a significantroleinmanagingmortgagemarketrisksthroughthe1998crisis. References Ambrose, B., M. LaCour-Little, and A. Sanders (2004). The effect of conforming loan status on mortgageyield spreads: A loan level analysis.Real EstateEconomics32, 541–69. Bernanke, B. S. (1986). Alternative explanations of the money-income correlation.Carnegie-RochesterConferenceSeries onPublicPolicy25,49–99. Gonzalez-Rivera,G.(2001).Linkagesbetweensecondaryandprimarymarkets formortgages.JournalofFixed Income11,29–36. Hancock, D., A. Lehnert, W. Passmore, and S. Sherlund (Forthcoming). Basel II capitalstandards: Potentialimpactson mortgagerates and securitization markets.Manuscript,Federal ReserveBoard, WashingtonDC. Johansen, S. (1988). Statistical analysis of cointegration vectors. Journal of EconomicDynamicsandControl12,231–54. Johansen,S.(1991).Estimationandhypothesistestingofcointegrationvectors in Gaussianvectorautoregressivemodels.Econometrica59, 1551–80. Johansen, S. and K. Juselius (1992). Testing structural hypotheses in a multivariate cointegration analysis of the PPP and the UIP for the UK. Journal 31

PreliminaryDraft of Econometrics53, 211–44. McKenzie,J.(2002).Areconsiderationofthejumbo/non-jumbomortgagerate differential.JournalofRealEstateFinanceand Economics25,197–214. Naranjo, A. and A. Toevs (2002). The effects of purchases of mortgages and securitization by government sponsored enterprises on mortgageyield spreads and volatility. Journal of Real Estate Finance and Economics 25, 173–96. Passmore, W., S. Sherlund, and G. Burgess (2004). The effect of housing government-sponsoredenterprises onmortgagerates. Forthcoming. Sarno,L.andD.L.Thornton(2004).TheefficientmarkethypothesisandidentificationinstructuralVARs.FederalReserveBankofSt.LouisReview86, 49–60. 32

PreliminaryDraft TABLE 1: CorrelationsAmongInterest Rates 10-year GSE Debt Treasury Swaps GSE Debt 1:0000 Treasury 0:9557 1:0000 Swaps 0:9915 0:9688 1:0000 5-year GSE Debt Treasury Swaps GSE Debt 1:0000 Treasury 0:9720 1:0000 Swaps 0:9960 0:9756 1:0000 NOTE. Table gives correlation coefficients among monthly differences in yields on comparable debt instruments. Data are monthly averages, Jan. 2000 through Dec. 2003. 33

PreliminaryDraft TABLE 2: DescriptiveStatistics Variable Symbol Mean Median Std. Dev. Min Max SecuritizationVolume($bn) s 54:70 35:11 47:70 7:47 204:28 t PortfolioPurchase Volume($bn) p 28:44 19:84 26:82 3:45 148:67 t Primary Market Spread (bps) rp −r 181:05 173:69 39:20 122:95 268:52 t t Secondary Market Spread (bps) rs −r 144:81 143:70 28:21 102:68 203:56 t t MortgageCouponGap (bps) pr 45:12 55:92 61:26 −101:92 168:45 t MortgageDelinquencies(bps) cr 53:27 54:00 5:10 45:00 62:00 t NOTE. Statisticsarefor120 monthlyobservationsrunningfromJanuary 1994throughDecember2003. 34

PreliminaryDraft TABLE 3: AugmentedDickey-FullerUnitRoot Tests Levels First Differences Variable Intercept Intercept+Trend Intercept Intercept+Trend SecuritizationVolume s −1:12 −3:21 −4:70?? −4:61?? t PortfolioPurchase Volume p −1:60 −3:52? −7:00?? −6:96?? t Primary MarketSpread rp −r −1:67 −3:87? −5:75?? −5:72?? t t Secondary Market Spread rs −r −2:22 −3:70? −6:37?? −6:35?? t t MortgageCouponGap pr −2:67 −2:78 −5:75?? −5:68?? t MortgageDelinquencies cr −1:74 −1:73 −3:88?? −3:86?? t NOTE. H 0 : Unitroot. ? and ?? denotestatisticalsignificanceatthe5-and 1-percent levels,respectively. 35

PreliminaryDraft TABLE 4: CumulativeLong-RunEffects BaselineVectorAutoRegression EfficientMarkets Hypothesis Shock to Effect on VAR1 VAR2 VAR3 EMH1 EMH2 EMH3 Percent " s 14:6? 14:8? 15:0? 15:0? 14:6? 14:6? s " s 3:8? 3:7? 5:3? 5:3? 3:7? 3:8? p " s 7:3? 7:3? 5:8? −0:7 −0:7 −0:8 rs−r " s 3:9 3:2 3:2 6:5? 8:3? 8:3? rp−r " p 8:7? 8:7? 9:1? 9:1? 8:5? 8:7? s " p 17:0? 17:0? 20:0? 20:0? 17:6? 17:0? p " p 14:1? 14:1? 9:2? 3:4 3:4 5:2 rs−r " p 0:2 0:2 0:2 8:5? 13:1? 13:1? rp−r BasisPoints " rs −r −2:1 −2:2 −1:9 −1:9 −2:3 −2:1 s " rs −r −0:5 −0:5 2:0 2:0 0:0 −0:5 p " rs −r 9:9? 9:9? 9:7? 4:7? 4:7? 4:8? rs−r " rs −r −1:1 −1:0 −1:0 8:6? 8:7? 8:7? rp−r " rp −r −1:7 −1:6 −1:3 −1:3 −1:8 −1:7 s " rp −r −0:2 −0:3 2:4 2:4 −0:0 −0:2 p " rp −r 11:0? 11:0? 10:8? 1:8 1:8 1:9 rs−r " rp −r 2:5 2:6 2:6 10:9? 11:1? 11:1? rp−r NOTE. The tablegivesthe long-run effects of shocks to the indicated innovationson the indicated variables under various assumptionsabout causality. The columns refer to different orderings of innovations,where VAR 1 is our baseline assumption (see the text for details). VAR 1: [" rs−r , " rp−r , " s , " p ]. VAR 2: [" rs−r , " s , " p , " rp−r ]. VAR 3: [" s , " p , " rs−r , " rp−r ]. EMH 1: [" s , " p , " rp−r , " rs−r ]. EMH 2: [" rp−r , " s , " p , " rs−r ]. EMH 3: [" rp−r , " rs−r , " s , " p ]. ?denotesstatisticalsignificanceatthe5-percent level. 36

PreliminaryDraft TABLE 5: MaximalCumulativeEffects BaselineVectorAutoRegression EfficientMarkets Hypothesis Shock to Effect on VAR1 VAR2 VAR3 EMH1 EMH2 EMH3 Percent " s 16:2? 16:2? 16:3? 16:3? 16:2? 16:2? s 0 0 0 0 0 0 " s 5:5? 5:4? 6:6? 6:6? 5:3? 5:5? p 2 2 2 2 2 2 " s 7:8? 7:8? 6:2? −2:0 −2:0 −1:9 rs−r 4 4 4 2 2 2 " s 4:6 3:9 3:9 7:1? 8:8? 8:8? rp−r 2 2 2 4 4 4 " p 11:1? 11:1? 11:4? 11:4? 11:0? 11:1? s 1 1 1 1 1 1 " p 26:1? 26:1? 27:0? 27:0? 26:3? 26:1? p 0 0 0 0 0 0 " p 15:2? 15:2? 9:9? 3:7 3:7 5:4 rs−r 3 3 2 5 5 5 " p 0:9 0:9 0:9 9:4? 14:3? 14:3? rp−r 3 3 3 3 3 3 BasisPoints " rs −r −2:3 −2:3 −2:0 −2:0 −2:4 −2:3 s 5 5 5 5 5 5 " rs −r −0:7 −0:7 2:6? 2:6? 0:5 −0:7 p 4 4 2 2 2 4 " rs −r 10:8? 10:8? 10:5? 4:9? 4:9? 4:9? rs−r 2 2 2 4 4 4 " rs −r −1:1 −1:0 −1:0 9:5? 9:7? 9:7? rp−r 5 5 5 2 2 2 " rp −r −1:9 −1:7 −1:4 −1:4 −1:9 −1:9 s 5 5 5 5 5 5 " rp −r −0:3 0:5 3:3? 3:3? 0:7 0:6 p 4 1 1 1 1 1 " rp −r 11:8? 11:8? 11:5? 1:9 1:9 2:0 rs−r 2 2 3 4 4 4 " rp −r 3:6? 3:6? 3:6? 11:7? 12:0? 12:0? rp−r 0 0 0 2 2 2 NOTE. x j gives the maximal cumulative effects of shocks to the indicated innovations on the indicated variables undervariousassumptionsaboutcausality. jindicatesinwhatj-stepaheadperiodthecumulativeeffectismaximal. The columns refer to different orderings of innovations, where VAR 1 is our baseline assumption (see the text for details). ?denotesstatisticalsignificanceatthe5-percent level. 37

PreliminaryDraft TABLE 6: GrangerCausalityF-Tests (cid:1)s (cid:1)p (cid:1)(rp −r ) (cid:1)(rs −r ) t t t t t t (cid:1)s : 2:27 0:71 1:26 t (cid:1)p 6:32?? : 0:74 0:19 t (cid:1)(rp −r ) 6:42?? 5:54?? : 0:21 t t (cid:1)(rs −r ) 4:51? 4:79? 0:61 : t t (cid:1)pr 8:58?? 3:09? 1:34 1:40 t (cid:1)cr 2:94 1:58 1:36 2:69 t NOTE. H 0 : TherowdoesnotGrangercausethecolumn. ? and ?? denotestatistical significanceatthe5-and 1-percent levels,respectively. 38

PreliminaryDraft TABLE 7: Long-Run Effects(VectorErrorCorrection) VectorErrorCorrection Shock to Effect on VEC1 VEC 2 VEC3 N&T1 N&T2 G-R Percent " s 9:6 9:8 11:5 : 0:3 : s " s 5:3 5:2 6:7 : : : p " s 11:0 11:0 8:2 : : : rs−r " s 3:3 3:0 3:0 : 1:7 : rp−r " p 14:0 14:0 15:6 : : : s " p 13:6 13:6 14:8 0:0 : 2:2 p " p 10:2 10:2 5:2 : : 7:7 rs−r " p 0:4 0:4 0:4 1:3 : : rp−r BasisPoints " rs −r 1:9 1:9 3:0 : : : s " rs −r −2:6 −2:6 −1:4 : : 2:8 p " rs −r 6:7 6:9 6:8 : : 9:9 rs−r " rs −r −0:6 −0:8 −0:8 : : : rp−r " rp −r 2:6 2:7 3:9 : 1:8 : s " rp −r −2:7 −2:9 −1:6 0:1 : : p " rp −r 7:2 7:2 7:1 : : : rs−r " rp −r 2:8 2:5 2:5 12:0 9:4 : rp−r NOTE. The tablegivesthe long-run effects of shocks to the indicated innovationson the indicated variables under various specifications and assumptions about causality. The columns refer to different specifications. VEC 1: [" rs−r , " rp−r , " s , " p ]. VEC 2: [" rs−r , " s , " p , " rp−r ]. VEC 3: [" s , " p , " rs−r , " rp−r ]. N&T 1: [" p , " rp−r ]. N&T 2: [" s , " rp−r ]. G-R: [" p , " rs−r ]. 39

PreliminaryDraft TABLE 8: MaximalEffects(VectorErrorCorrection) VectorErrorCorrection Shock to Effecton VEC1 VEC2 VEC 3 N&T 1 N&T2 G-R Percent " s 14:7 14:7 14:9 : 5:5 : s 0 0 0 0 " s 7:4 7:2 8:3 : : : p 2 2 2 " s 12:4 12:4 9:5 : : : rs−r 4 4 4 " s 3:5 3:4 3:4 : 2:3 : rp−r 4 4 4 5 " p 14:2 14:2 15:8 : : : s 8 8 5 " p 25:4 25:5 25:9 5:1 : 10:7 p 0 0 0 0 0 " p 13:7 13:7 9:0 : : 7:7 rs−r 2 2 2 24 " p 1:0 1:0 1:0 1:4 : : rp−r 2 3 3 3 BasisPoints " rs −r 2:1 2:1 3:1 : : : s 6 6 5 " rs −r −3:0 −2:9 −1:8 : : 2:8 p 5 5 6 24 " rs −r 9:9 9:9 9:8 : : 10:9 rs−r 2 2 2 2 " rs −r −0:7 −0:9 −0:9 : : : rp−r 6 6 6 " rp −r 2:8 2:9 4:0 : 2:4 : s 6 6 6 5 " rp −r −3:1 −3:2 −2:1 −0:8 : : p 6 6 6 2 " rp −r 10:7 10:7 10:3 : : : rs−r 1 1 2 " rp −r 3:6 3:5 3:5 12:3 11:7 : rp−r 0 0 0 2 2 NOTE. x j givesthemaximaleffectsofshockstotheindicatedinnovationsontheindicatedvariablesundervarious assumptions about causality. j indicates in what j-step ahead period the effect is maximal. The columns refer to differentorderingsofinnovations(seethetextfordetails). 40

FIGURE 1: Data (a)GSE purchaseand securitizationvolume(logscale). 320 160 80 40 20 10 5 1994 1995 1996 1997 1998 1999 2000 2001 2002 2003 2004 Date snoilliB GSE Purchases GSE Securitization (b)Primary and secondary mortgagemarketspreads toTreasury rates. 280 260 240 220 200 180 160 140 120 100 1994 1995 1996 1997 1998 1999 2000 2001 2002 2003 2004 Date stnioP sisaB Primary Market Secondary Market (c)Credit and Prepayment Characteristics. 80 60 40 1994 1995 1996 1997 1998 1999 2000 2001 2002 2003 2004 Date stnioP sisaB 200 Delinquency Rate (left scale) Coupon Gap (right scale) 0 −200 1994 1995 1996 1997 1998 1999 2000 2001 2002 2003 2004 41

FIGURE 2: CumulativeImpulseResponseFunctions Primary Market Spread 16 14 12 10 8 6 4 2 0 −2 0 2 4 6 8 10 12 j−months ahead stnioP sisaB Secondary Market Spread 16 Primary Market Spread Secondary Market Spread 14 Portfolio Purchases 12 Securitization Volume 10 8 6 4 2 0 −2 −4 0 2 4 6 8 10 12 j−months ahead stnioP sisaB Primary Market Spread Secondary Market Spread Portfolio Purchases Securitization Volume Portfolio Purchase Volume 30 25 20 15 10 5 0 −5 0 2 4 6 8 10 12 j−months ahead tnecreP Securitization Volume 25 Primary Market Spread Secondary Market Spread Portfolio Purchases 20 Securitization Volume 15 10 5 0 0 2 4 6 8 10 12 j−months ahead tnecreP Primary Market Spread Secondary Market Spread Portfolio Purchases Securitization Volume NOTE. Each panel gives the cumulative effect in month j for a given variable of a shock to the indicated equation. Thus, the dashed thick line in the upper left panel showstheeffect on primary market spreads ofa onestandard deviation shock to secondary market spreads. Note that these impulse response functions wereestimatedunderourbaselineassumptionaboutcausation(seesection4.2for moredetails). 42

FIGURE 3: Variance Decomposition Primary Market Spread 100 80 60 40 20 0 0 5 10 15 20 25 j−months ahead tnecreP Secondary Market Spread 100 80 Primary Market Spread 60 Secondary Market Spread Portfolio Purchases Securitization Volume 40 20 0 0 5 10 15 20 25 j−months ahead tnecreP Primary Market Spread Secondary Market Spread Portfolio Purchases Securitization Volume Portfolio Purchase Volume 100 80 60 40 20 0 0 5 10 15 20 25 j−months ahead tnecreP Securitization Volume 100 80 Primary Market Spread 60 Secondary Market Spread Portfolio Purchases Securitization Volume 40 20 0 0 5 10 15 20 25 j−months ahead tnecreP Primary Market Spread Secondary Market Spread Portfolio Purchases Securitization Volume NOTE. Each panel gives the fraction of the j−month ahead forecast error variance for a given variable that can be explained by innovations to the indicated equations. 43

FIGURE 4: ResponsetoLiquidityCrisisof1998 Secondary Market Spread 40 30 20 10 0 −10 −20 SEP98 OCT98 NOV98 DEC98 JAN99 FEB99 MAR99 stnioP sisaB Primary Market Spread 50 Actual Predicted 40 Counterfactual 30 20 10 0 −10 −20 −30 SEP98 OCT98 NOV98 DEC98 JAN99 FEB99 MAR99 stnioP sisaB Actual Predicted Counterfactual Portfolio Purchase Volume 50 0 −50 −100 SEP98 OCT98 NOV98 DEC98 JAN99 FEB99 MAR99 tnecreP Securitization Volume 40 30 20 10 0 −10 Actual −20 Predicted Counterfactual −30 SEP98 OCT98 NOV98 DEC98 JAN99 FEB99 MAR99 tnecreP Actual Predicted Counterfactual NOTE. Each panel gives the month-by-month effects of the liquidity shock to secondary mortgage market spreads. Thus, the lower left graph shows that the model does remarkably well in tracing out the effects of the liquidity shock on GSEportfoliopurchases. Theseimpulseresponsefunctionswereestimatedunder ourbaselineassumptionaboutcausation(seesection 4.2formoredetails). 44

FIGURE 5: Normalized GSE Actions 70 60 50 40 30 20 10 0 1994 1995 1996 1997 1998 1999 2000 2001 2002 2003 2004 Date snoitanigirO fo tnecreP GSE Securitization GSE Portfolio Purchases NOTE. Figure shows monthly total GSE portfolio purchases and gross MBS issuancedividedbyestimatedmonthlytotalmortgageoriginationsderivedfromthe HMDAreports. 45

FIGURE 6: ImpulseResponseFunctionsUndertheLevelsSpecification Primary Market Spread 12 10 8 6 4 2 0 −2 0 2 4 6 8 10 12 j−months ahead stnioP sisaB Secondary Market Spread 10 Primary Market Spread Secondary Market Spread Portfolio Purchases 8 Securitization Volume 6 4 2 0 −2 0 2 4 6 8 10 12 j−months ahead stnioP sisaB Primary Market Spread Secondary Market Spread Portfolio Purchases Securitization Volume Portfolio Purchase Volume 5 4 3 2 1 0 −1 −2 0 2 4 6 8 10 12 j−months ahead tnecreP Securitization Volume 6 Primary Market Spread Secondary Market Spread 5 Portfolio Purchases Securitization Volume 4 3 2 1 0 −1 −2 0 2 4 6 8 10 12 j−months ahead tnecreP Primary Market Spread Secondary Market Spread Portfolio Purchases Securitization Volume NOTE. Figure shows the impulse response functions of endogenous variables to normalizedshocksunderthelevelspecification. See section4.5 formoredetail. 46

FIGURE 7: GrangerCausalityTest Results Primary Market Spreads PortfolioPurchases Secondary MarketSpreads SecuritizationVolume NOTE. Arrows show the direction of causality from bivariate Granger tests. (Dashed lineindicatesmarginalstatisticalsignificance.) FIGURE 8: CointegratingRelationship 6 5 4 3 2 1 0 −1 −2 1994 1995 1996 1997 1998 1999 2000 2001 2002 2003 2004 Date snoilliB goL Actual Purchases Equilibrium Purchases Error 47