liquidity in a market for unique assets: specified pool and · pdf filespeciﬁed pool and...

THE JOURNAL OF FINANCE • VOL. LXXII, NO. 3 • JUNE 2017

Liquidity in a Market for Unique Assets:Specified Pool and To-Be-Announced Trading

in the Mortgage-Backed Securities Market

PENGJIE GAO, PAUL SCHULTZ, and ZHAOGANG SONG∗

ABSTRACT

Agency mortgage-backed securities (MBS) trade simultaneously in a market for spec-ified pools (SPs) and in the to-be-announced (TBA) forward market. TBA tradingcreates liquidity by allowing thousands of different MBS to be traded in a handful ofTBA contracts. SPs that are eligible to be traded as TBAs have significantly lowertrading costs than other SPs. We present evidence that TBA eligibility, in addition tocharacteristics of TBA-eligible SPs, lowers trading costs. We show that dealers hedgeSP inventory with TBA trades, and they are more likely to prearrange trades in SPsthat are difficult to hedge.

THE MARKET FOR AGENCY MORTGAGE-BACKED securities (MBS) is among thelargest, most active, and most liquid of all securities markets. At first glance,the market’s liquidity is surprising because each MBS is unique and is com-posed of specific mortgages with their own prepayment characteristics. Inthis paper, we study the institutional feature of this market that allows itto work so well, namely, parallel trading in a to-be-announced (TBA) forwardmarket in MBS and a specified pool (SP) market in which specific MBS aretraded.

TBA trading makes the agency MBS market liquid in three ways. First,the TBA market takes thin markets for thousands of different MBS withdifferent prepayment characteristics and trades them through a handful ofthickly traded cheapest-to-deliver contracts. In this way, liquidity is increasedfor the MBS that are traded there instead of as SPs. Second, eligibility for TBA

∗Pengjie Gao and Paul Schultz are with the Mendoza College of Business, University of NotreDame. Zhaogang Song is with Johns Hopkins University. This paper was written while ZhaogangSong was at the Board of Governors of the Federal Reserve. We thank the Editor (Bruno Biais) andreferees for suggestions that significantly improved this paper. We thank Shane Corwin, MartijnCremers, Bill Maxwell, Kumar Venkataraman, and seminar participants at the University ofNebraska, the University of Notre Dame, SMU, and the Vienna Graduate School of Finance forhelpful comments. In addition, we would like to thank John Kandrac, Bernd Schlusche, and ReneeSchultz for their help on MBS characteristics. Finally, we are grateful to Alie Diagne and OlaPersson of FINRA and Christine Black from the Board of Governors of the Federal Reserve, whohelped with data clearance. FINRA screened the paper to ensure that confidential dealer identitieswere not revealed. The authors read the Journal of Finance’s disclosure policy and have no conflictsof interest to disclose.

DOI: 10.1111/jofi.12496

1119

1120 The Journal of Finance R©

trading makes SPs more liquid by giving dealers an option to lay off SP in-ventory through the more liquid TBA market. Third, parallel TBA tradingincreases liquidity for SP trades. There are several possible reasons why TBAtrading may increase SP liquidity and we provide evidence for one: dealers useTBA trades to hedge SP positions.

We are not the first to show that TBA trading is cheaper than SP trading.Bessembinder, Maxwell, and Venkataraman (2013) examine trading costs forstructured credit products that include but are not limited to MBS. They ob-serve that TBA trades are much cheaper than other MBS trades. Friewald,Jankowitsch, and Subrahmanyam (2014) study the tradeoff between accuracyin measuring liquidity and the disclosure of information to market partici-pants. As part of their study, they find that trading costs are much lower forTBA trades than SP trades. Atanasov and Merrick (2012) examine the integra-tion of TBA and SP markets, and show that for large trades the two marketsare well integrated. Like the others, they find that SP trading costs are muchhigher than TBA costs.

Unlike these previous studies, our paper distinguishes between SPs thatare eligible to be traded in the TBA market and those that are ineligible tobe traded there. We show that TBA-eligible SPs are much cheaper to tradethan TBA-ineligible SPs. TBA eligibility itself can increase liquidity by givingdealers and potential buyers of an SP an option to deliver the SP in a TBAtrade if market conditions change. We provide evidence, in two separate tests,that TBA eligibility itself increases liquidity. In the first, we use loan-to-value(LTV) levels and a dummy variable for LTVs greater than 1.05 to see whetherthere is an abrupt change in trading costs at the LTV cutoff for TBA eligibility.We find that trading costs in general decline with LTV ratios, but increasesharply at the 1.05 cutoff. Our second test is a variation on propensity scorematching. In the first stage we estimate the probability that an SP is TBAeligible using characteristics that include minimum and maximum loan values,LTV ratios, and average Fair Isaac Company (henceforth FICO) scores. In thesecond stage, we group SPs by the estimated probability that the SP is TBAeligible and test whether actual eligibility affects trading costs. After adjustingfor the probability that an SP is TBA eligible, we find that TBA eligibility itselfsignificantly decreases trading costs.

We also provide the first evidence that parallel TBA trading makes the SPmarket more liquid. We identify an exogenous factor that directly affects TBAtrading but not SP trading: TBA settlement dates. There is one settlement dateeach month for all TBA trades of MBS with a given maturity and issuer. Thesedates are set by the Securities Industry and Financial Markets Association(SIFMA) well in advance of the settlement month. Traders who do not wishto take or deliver MBS roll over their positions before the settlement date,resulting in TBA trading volume that is three to four times as large in the daysprior to settlement dates as it is during the rest of the month. SP trades canbe settled at any time during the month. Nevertheless, trading costs for SPs,like TBA trading costs, are much lower prior to TBA settlement dates whenthe predictable volume of TBA trading is high.

Liquidity in a Market for Unique Assets 1121

There are several plausible explanations for why parallel TBA trading re-duces SP trading costs. Our data allow us to explore one of them. This paperis the first to show that dealers typically hedge SP inventory changes withoffsetting TBA trades. For individual dealers we regress daily changes in TBAinventory on changes in the inventory of TBA-eligible SPs and TBA-ineligibleSPs with the same maturity and coupon. The coefficients are negative, im-plying that the median dealer hedges SP inventories with TBA trades. Thecoefficients on changes in inventory of SPs that are not TBA eligible are closerto zero than the coefficients on TBA-eligible inventory changes, suggesting thatSPs that are not TBA eligible are less likely to be hedged, all else equal. Thisis not surprising. TBA-ineligible SPs have different characteristics from TBA-eligible ones, and we find that TBA trades are not as effective at hedging pricechanges for TBA-ineligible SPs as they are at hedging changes in TBA-eligibleSP prices.

Difficulty in hedging can reduce liquidity. We present evidence that dealersare reluctant to take hard-to-hedge SPs into inventory. They are more likelyto prearrange a sale of the SP to a second customer before purchasing theSP from the first customer. This means that investors have to wait to sellunwanted MBS while the dealer searches for a buyer. This is a cost that wecannot measure.

Regulators have recently expressed concern about the liquidity of over-the-counter markets for corporate and municipal bonds and have suggested thatmore transparency is needed. Our findings suggest another way to increaseliquidity. Forward market trading of MBS in the TBA market appears to lowertrading costs both for those MBS traded in the TBA market and for the MBStraded in the parallel SP market. Some legal obstacles would need to be over-come, but it may make sense to have similar forward markets in municipaland corporate bonds. Bonds included in such a forward market would needto be sufficiently homogeneous with respect to default risk and other charac-teristics. But there may be, for example, enough relatively homogeneous, 5%20-year, AA-rated industrial bonds to create a liquid cheapest-to-deliver for-ward market.1 The existence of a liquid forward market would be likely tocreate incentives for issuers to issue standardized bonds that could be tradedin the forward market.

The rest of the paper is organized as follows. Section I discusses how thesecondary market for MBS operates. Section II describes the data used here.Section III provides estimates of trading costs in the TBA and SP markets. InSection IV we examine the impact of TBA eligibility on SP liquidity. Section Vpresents evidence that eligibility for TBA trading lowers SP trading costs. InSection VI, we show that dealers use the TBA market to hedge SP positions.Section VII concludes.

1 Spatt (2004) makes the same point. He observes that many corporate and municipal bondsare close substitutes, but that these bonds, unlike MBS, “are not traded interchangeably in themarketplace.”


I. How the Market for Agency MBS Works

Tens of thousands of unique agency MBS have been issued by Fannie Mae,Freddie Mac, or Ginnie Mae. These are pass-through securities in which inter-est and principal payments on the underlying mortgages are passed through tothe MBS investors in proportion to their holdings. If a mortgage holder prepaysa mortgage early, the principal payment is passed on to investors at that time.Agency MBS are backed by the issuing agency and are effectively default-free.

Agency MBS can be traded as SPs where buyer and seller agree to exchangea particular MBS. They are also traded in a forward market known as theTBA market where the seller has an option to deliver any MBS that meetsagreed-upon criteria. Most trading takes place in the TBA market. Mortgageoriginators use TBA trades to sell mortgages forward and to hedge their pro-duction. Investors also use the TBA market to buy and sell already-issuedMBS.

All agency MBS are default-free, but each is unique in its prepayment char-acteristics. From the standpoint of investors, an MBS has more desirable pre-payment characteristics if the mortgages in the MBS are less likely to be paidoff early if interest rates fall, and more likely to be paid off early if interest ratesrise. MBS with the most desirable prepayment characteristics are traded in theSP market where sellers can realize the full value of their MBS rather thanthe cheapest-to-deliver price. Buyers in the SP market know the MBS they aregetting and can be expected to closely examine the prepayment characteristicsof the MBS. Trades in the SP market can be settled at any time rather thanon one day during a month. In contrast to TBA trades, SP market transactionsgenerally result in delivery of the MBS.

For TBA trades, the buyer and seller agree to six parameters: coupon, ma-turity, issuer, settlement date, face value of the MBS, and price. Sellers willattempt to deliver the cheapest MBS that meets the trade requirements, andbuyers expect that this is what they will receive. TBA trading works becausethe MBS exchanged in that market are relatively homogeneous. All TBA tradesof MBS with a specific maturity and issuer settle on the same date each month.Most TBA trades settle within the next month, but TBA trades with settlementdates two or three months out are also common.

Forty-eight hours before the settlement date, on the notification date, theseller tells the buyer which specific MBS will be delivered. In most cases though,TBA buyers do not take delivery and TBA sellers do not deliver MBS; tradersinstead take offsetting positions before the notification date. The ability toeasily close out positions makes TBA trading a useful way to hedge risk frommortgage rate changes. A major source of TBA trading is mortgage originatorswho use the TBA market to sell mortgages forward.

TBA market investors can observe real-time indicative TBA quotes throughTradeweb, an electronic trading platform. For each TBA contract, Tradewebprovides one bid and one ask price after using a proprietary algorithm to filterout meaningless dealer quotes. These indicative quotes are updated continu-ously as dealers update their quotes. Atanasov and Merrick (2012, p. 1) observe


that the TBA market “ . . . has excellent pre-trade transparency via electronicplatforms.” Vickery and Wright (2013) observe that internal Federal Reserveanalysis shows that quotes generally track prices of completed transactionsclosely.

TBA trading converts a market with thousands of MBS into a thick marketthat trades a handful of contracts. In June 2011, the first full month of data inour sample, 24,528 different SPs traded. During the last month of our sample,May 2013, 27,433 SPs traded. In contrast, across all combinations of maturity,coupon, issuer, and settlement date, only 510 different TBA contracts tradedduring June 2011, and only 475 traded during May 2013. This, however, un-derstates the degree to which TBA trading is concentrated in a few contracts.TBA trading takes place in MBS with maturities of 5, 7, 10, 15, 20, 30, and40 years and with coupon yields ending in even percents, in half percents (e.g.,3.50%), and in quarter and three-quarter percents (e.g., 3.25% or 3.75%). Overour entire sample period, 12 maturity-coupon combinations account for 96%of trades: 15 years with 2.5%, 3%, 3.5%, and 4%, and 30 years with 2.5%, 3%,3.5%, 4%, 4.5%, 5%, 5.5%, and 6%. With so much trading volume channeledinto so few TBA contracts, it is easy for dealers to find counterparties and layoff inventory. It is more difficult for dealers to eliminate inventory risk by lay-ing off positions in one of the many thousands of SPs. As we will show, dealersinstead hedge their SP inventory with TBA trades.

The market for agency MBS is almost entirely an institutional market. As of2011, 25% of agency MBS were held by U.S. banks, 9% by insurance companiesand pension funds, 11% by mutual funds, and 14% by foreign investors.2 As aresult of its asset purchase programs, the Federal Reserve held 20% of agencyMBS. Other investors in agency MBS include Fannie Mae, Freddie Mac, theU.S. Treasury, savings institutions, and real estate investment trusts (REITS).These institutions tend to buy and hold MBS for long periods of time. Whenthey trade, they usually trade large quantities of MBS.3

II. Data

The Financial Industry Regulatory Authority (FINRA) began requiring mem-bers to report all trades of MBS through their Trade Reporting and ComplianceEngine (henceforth TRACE) system in May 2011. In this paper, we examineMBS trading using all TBA and SP trades by all dealers who were FINRAmembers over May 16, 2011 through April 2013. This includes virtually all, ifnot all, MBS trades for this period. Data for each trade include the maturity,coupon, and issuer of the MBS; the price, par value, trade date, trade time,and settlement date for the trade; an indicator for whether the trade was a

2 Written statement of Richard Dorfman before the House Committee on Financial Services,Subcommittee on International Monetary Policy and Trade, October 13, 2011.

3 Vickery and Wright (2013) provide a wealth of institutional details about TBA trading and theMBS market. They report that TBA eligibility lowers mortgage interest rates, but are cautious inthe interpretation of the evidence because their data do not allow them to separate differences inliquidity from differences in prepayment risk.


purchase from a customer, a sale to a customer, or an interdealer trade; andidentifying numbers for dealers in the trade. For SP trades, there is a variablethat indicates whether the traded MBS is TBA eligible. The data include bothinterdealer trades and trades between dealers and customers.

Table I provides summary statistics for MBS trading. Panel A reports thenumber of trades of various types, and the volume from these trades. As isalso noted by Vickery and Wright (2013), the great majority of MBS volume isin the TBA market. For our sample period, the volume of dealers’ TBA tradeswith customers totals $64.4 trillion. The volume of interdealer TBA trades is$58.5 trillion. The total value of SP trades with customers is “only” $7.2 trillion.The total dollar volume of interdealer SP trades is $1.9 trillion. It is interestingthat interdealer trades account for almost half the volume in the TBA market,but a much smaller proportion of SP volume. Interdealer trading is more com-mon in the TBA market because dealers lay off TBA inventory by trading withother dealers, and, as we will show, also hedge SP inventory with interdealerTBA trades.

Because the volume of trading in SPs is so much less than TBA volume, it istempting to conclude that the SP market is unimportant. That is not true. Eventhough trading volume is lower in the SP market than in the TBA market, itis still in the trillions of dollars during our sample period. In addition, it isdifficult to compare the dollar volumes directly. SP trades can be expected toresult in delivery of the MBS, but most TBA trades do not result in delivery.4

Finally, without SP trading, MBS traded in the TBA market would be lesshomogeneous, and it is likely that the TBA market would therefore be lessliquid.

Panel A of Table I also provides information on the volume and numbers ofdifferent types of TBA trades. Over the May 2011 through April 2013 period,there are more than 3.3 million interdealer and dealer-to-customer TBA trades.Outright trades make up the majority of TBA trades. Dollar rolls are the sec-ond most common type of trade. Dollar rolls are spread trades that are oftencompared to repos. The seller of a dollar roll sells the front month TBA contractand simultaneously buys a future month contract with the same characteris-tics. Dollar rolls differ from repos in that the securities that are purchased fordelivery in the later month are “substantially similar” to those sold in the frontmonth rather than the same securities. In addition, in a dollar roll, the buyerof the front month contract receives coupon and principal payments over themonth. Dollar rolls tend to be very large trades, and account for most of thebuy, sell, and interdealer volume.

Stipulated trades are TBA trades in which the buyer requires the seller todeliver pools with additional stipulated characteristics. The buyer could specify,for example, that no more than a certain percentage of mortgages in a poolare on California homes. Stipulated dollar rolls are dollar rolls that stipulateadditional characteristics of pools to be delivered. They are less common, andaccount for less than 30,000 trades.

4 See Vickery and Wright (2013, p. 9).


Tab

leI

Su

mm

ary

Sta

tist

ics

for

MB

ST

rad

ing

inth

eT

BA

and

SP

Mar

ket

sT

he

sam

ple

con

sist

sof

alls

econ

dary

mar

ket

MB

Str

ades

from

May

16,2

011

thro

ugh

Apr

il,2

013.

Vol

um

eis

in$1

,000

,000

’sof

face

valu

e.

Pan

elA

:Tot

alTr

adin

gby

Trad

eT

ype

Nu

mbe

rof

Trad

esw

ith

Cu

stom

ers

Vol

um

efr

omTr

ades

wit

hC

ust

omer

sN

um

ber

Inte

rdea

ler

Trad

esIn

terd

eale

rV

olu

me

TB

ATr

ades

Ou

trig

ht

Trad

es83

7,01

127

,363

,239

1,53

1,91

925

,807

,536

Dol

lar

Rol

ls28

7,09

833

,436

,663

544,

153

32,5

25,0

31S

tipu

late

dTr

ades

74,3

962,

126,

427

14,6

2417

0,24

4S

tip.

Dol

lar

Rol

ls26

,121

1,45

6,63

71,

711

28,3

10To

talT

BA

trad

ing

1,22

4,68

164

,382

,977

2,09

2,40

758

,531

,121

Spe

cifi

edP

oolT

rade

s

TB

AE

ligi

ble

949,

378

6,28

1,28

047

2,57

41,

394,

371

Non

elig

ible

150,

299

870,

925

89,6

2547

4,35

5To

talS

peci

fied

Poo

l1,

099,

677

7,15

2,20

556

2,19

91,

868,

726

Pan

elB

:Tra

deS

izes

byTr

ade

Typ

e

Inte

rdea

ler

Trad

esTr

ades

wit

hC

ust

omer

s

Nu

mbe

rA

vg.T

rade

Siz

e($

mil

lion

s)P

erce

nt

>$1

0m

illi

onN

um

ber

Avg

.Tra

deS

ize

($m

illi

ons)

Per

cen

t>

$10

mil

lion

Spe

cifi

edP

ools

562,

067

$3.3

26.

7%1,

099,

260

$6.4

910

.7%

TB

AO

utr

igh

t1,

531,

919

$16.

7137

.1%

837,

011

$32.

6437

.2%

TB

AD

olla

rR

oll

544,

153

$59.

6460

.8%

287,

158

$116

.43

68.1

%T

BA

Sti

pula

ted

14,6

24$1

1.64

12.0

%74

,396

$28.

5836

.4%

TB

AS

tip.

Rol

ls1,

711

$16.

5532

.0%

26,1

21$5

5.77

66.1

%

(Con

tin

ued

)


Tab

leI—

Con

tin

ued

Pan

elC

:Dea

ler

Act

ivit

yL

evel

san

dTr

ade

Typ

e

Dea

ler

Ran

kin

gby

Nu

mbe

rof

Trad

es

Per

cen

tage

ofTr

ades

Th

atA

reS

peci

fied

Poo

ls

Per

cen

tage

ofV

olu

me

from

Spe

cifi

edP

ools

Per

cen

tage

ofTr

ades

Th

atA

reIn

terd

eale

rP

erce

nta

geof

All

Trad

esP

erce

nta

geof

All

Vol

um

e

1–10

23.8

3%13

.55%

56.3

7%54

.9%

64.6

%11

–30

42.8

6%26

.16%

55.0

9%27

.3%

29.3

%31

–50

56.4

4%42

.08%

53.7

3%8.

9%4.

2%51

–100

75.3

5%63

.29%

51.3

7%6.

5%1.

5%10

1–75

891

.36%

87.8

2%44

.73%

2.3%

0.4%


The statistics on dollar rolls and stipulated trades are included to providea complete picture of the MBS market. In most of the rest of the paper, wefocus attention on outright TBA trades. These are most similar to SP trades.For many traders, an SP purchase and an outright TBA purchase are just twodifferent ways to acquire MBS. Similarly, for many traders, an SP sale and anoutright TBA sale are close substitutes.

Panel A of Table I shows that there are about 1.66 million trades of SPs. TBA-eligible SPs make up the great majority of these trades. These SPs could be soldin the TBA market if the seller so desires. The other SPs have characteristicsthat make them ineligible for TBA trading. For example, they could containmortgages with high LTV ratios. Interdealer trades make up a far smallerproportion of SP trades than TBA trades.

Panel B of Table I provides information on trade sizes. The MBS marketis a market for financial institutions, not individual investors, so trade sizesare large. The average size of an outright TBA trade between a dealer anda customer is $32.64 million dollars. The distribution is right-skewed, butnonetheless, over 37% of the TBA trades between dealers and customers arefor more than $10 million. Dollar rolls are especially large. The mean size ofinterdealer dollar roll trades is $59.64 million while the mean size for tradeswith customers is over $100 million. Trade sizes are far smaller for SPs than forTBA trades. Interdealer SP trades have an average size of only $3.32 milliondollars, but the great majority of trade sizes are smaller. Only 6.7% of SPinterdealer trades are for $10 million or more. It is interesting that SP tradeswith customers tend to be larger than interdealer SP trades. The mean tradesize trade with customers is $6.49 million par value, and 10.7% of the tradesare for $10 million or more.

Panel C of Table I reports the proportion of trades of different types for dealerswith different levels of activity. There are over 750 dealers in our sample, butmost trades are handled by a small number of them. Panel C shows that thetop 10 dealers, ranked by number of trades, account for 54.9% of all tradesand 64.6% of all volume. The next 20 dealers account for an additional 27.3%of trades and 29.3% of volume. Active dealers tend to do most of their tradingin the TBA market, while inactive ones trade mainly in SPs. For the 10 mostactive dealers, the average proportion of volume from SPs is 13.55%. For the 20next most active dealers, the proportion of volume from SPs averages 26.16%.For dealers ranked 101–758 by number of trades, the proportion of volume fromSPs reaches 87.82%. As we have seen, TBA trades are usually much larger thanSP trades. To compete effectively as a dealer in the TBA market requires morecapital than it takes to trade SPs—capital that the less active dealers may nothave.

Panel C also reveals that the proportion of trades that are interdealer tradesis higher for more active dealers than for less active ones. Even for the leastactive dealers, however, the average proportion of trades that are interdealertrades is over 44%.

During the sample period, both TBA and SP prices increased. This can beattributed to falling mortgage rates over this time. Figure 1 shows weekly


0.00

1.00

2.00

3.00

4.00

5.00

4/28/2011 10/27/2011 4/26/2012 10/25/2012 4/25/2013

15 Year Rate

30 Year Rate

Figure 1. Weekly 15- and 30-year mortgage rates. Source: Freddie Mac.

national average mortgage rates, from Freddie Mac, for 15- and 30-year mort-gages for the period from April 2011 through April 2013. Over these two years,30-year rates were consistently about 75 basis points higher than 15-yearrates. Rates declined approximately 125 basis points between April 2011 andOctober 2012. The decline in rates led to increased prices of MBS over thesample period, and made prepayment an attractive option for many mortgageholders.

Lower mortgage rates also means that the MBS issued later in the sampleperiod had lower coupon rates than the MBS issued earlier. We obtain fromJP Morgan the gross production, net production, and outstanding balance ofMBS with each coupon and maturity from each issuer. JP Morgan, in turn,obtains these MBS production summary statistics from the Federal NationalMortgage Association (FNMA, or Fannie Mae) Monthly Summary, the FederalHome Loan Mortgage Corporation (FHLMC, or Freddie Mac) Monthly VolumeSummary, and the Government National Mortgage Association (GNMA, orGinnie Mae) Monthly Issuance Report.5 Gross production is the value of newMBS issued and net production is the gross production minus the reductionin value of current MBS from mortgage payments. Figure 2, Panel A showsthe net production in millions of dollars, across all issuers, of 30-year MBSwith 3%, 3.5%, 4%, and 4.5% by month. At the beginning of our sample period,

5 These data are available at the following websites: http://www.fanniemae.com/portal/about-us/investor-relations/monthly-summary.html, http://www.freddiemac.com/investors/volsum/, http://www.ginniemae.gov/data_and_reports/reporting/Pages/monthly_issuance_reports.aspx.

http://www.fanniemae.com/portal/about-us/investor-relations/monthly-summary.html

http://www.fanniemae.com/portal/about-us/investor-relations/monthly-summary.html

http://www.freddiemac.com/investors/volsum/

http://www.ginniemae.gov/data_and_reports/reporting/Pages/monthly_issuance_reports.aspx

http://www.ginniemae.gov/data_and_reports/reporting/Pages/monthly_issuance_reports.aspx


Panel A: Monthly production of 30-year MBS in $millions

Panel B: Monthly production of 15-year MBS in $millions

-$40,000

-$20,000

$0

$20,000

$40,000

$60,000

$80,000

$100,000

May-11 Sep-11 Jan-12 May-12 Sep-12 Jan-13

3.0% 3.5%

4.0% 4.5%

-$5,000

$0

$5,000

$10,000

$15,000

$20,000

$25,000

$30,000

May-11 Sep-11 Jan-12 May-12 Sep-12 Jan-13

2.00% 2.50%

3.00% 3.50%

Figure 2. Monthly production of MBS. All issuers are included. (Color figure can be viewed atwileyonlinelibrary.com)


in May 2011, net production of 30-year 4.5% MBS is positive. With fallingmortgage rates, production of 4.5% 30-year MBS declined quickly however,and turned negative in September 2011. In other words, by September 2011,the dollar value of prepayments of the mortgages in existing 4.5% 30-year MBSexceeded the dollar value of new 4.5% 30-year MBS. Production of 30-year MBSwith coupons of 4% rose from almost nothing in May 2011 to over $20 billionin September 2011. After June 2012, low mortgage rates led to negative netproduction of 30-year 4% MBS. Production of 3.5% MBS began in September2011, and production of 3% 30-year MBS did not begin until 2012.

Figure 2, Panel B depicts net production of 15-year MBS. Patterns of netproduction are similar to those of 30-year MBS. As mortgage rates fell, produc-tion of MBS with high coupon rates declined and turned negative. Productionof MBS with lower coupon rates began. Greater production can be expected toincrease liquidity. One of the major sources of TBA volume is from mortgageoriginators who hedge by selling mortgages forward. When mortgage rates fall,originators shift their hedging demand toward TBA trades with lower coupons.We expect liquidity to be greatest for the TBA trades with demand from orig-inators. Hence, we expect low coupon TBA trades to become more liquid overour sample period.

III. SP and TBA Trading Costs

In this section we extend the MBS portion of the work of Bessembinder,Maxwell, and Venkataraman (2013) by examining the impact of TBA eligibilityand MBS production on trading costs. We have four objectives in this section.First, we want to compare trading costs for TBA-eligible SPs with TBA tradingcosts. Second, we want to examine the impact of TBA eligibility on trading costs.Third, we want to estimate the impact of MBS production and the balance ofoutstanding MBS on MBS trading costs. Finally, because much of the rest ofthis paper revolves around the impact of TBA eligibility and TBA trading onSP trading costs, we establish some trading cost benchmarks in this section.

To date, a handful of studies of the microstructure of MBS markets havebeen conducted. Estimates of MBS trading costs appear in Atanasov andMerrick (2012), Bessembinder, Maxwell, and Venkataraman (2013), andFriewald, Jankowitsch, and Subrahmanyam (2014). The approach of Bessem-binder, Maxwell, and Venkataraman most closely resembles ours. They exam-ine trading of MBS and other structured credit products for the period from May16, 2011 through January 31, 2013. They estimate trading costs by regressingthe price difference between successive trades on a variable for changes from adealer purchase to a dealer sale (+1) or dealer sale to a dealer purchase (−1),along with variables for changes in bond and equity indices over the trade pe-riod. Their estimates of one-way trading costs are 40 basis points for SPs andjust 1 basis point for TBA trades.

Like Bessembinder, Maxwell, and Venkataraman (2013), we employ a regres-sion methodology to estimate trading costs. Each observation is two consecu-tive trades between dealers and customers in an MBS with a specific CUSIP,


but each regression includes observations from all CUSIPs with a particularmaturity. With the change in price between two consecutive trades as the de-pendent variable, we estimate

� Pt = α0 + α1�Qt + α2�Qt ·(

ln(

Sizet

1,000,000

)+ ln

(Sizet−1

1,000,000

))

+ α3�Qt · TBA Eligible + α4�Qt · TBA Eligible ·(

ln(

Sizet

1,000,000

)

+ ln(

Sizet−1

1,000,000

))+ α5�Qt · ln

(MBS Productiont

Avg Production

)

+ α6�Qt · ln(

MBS Balancet

Avg Balance

)+ �βi Reti,t + εt, (1)

where �Pt is the percentage change in prices between trade t and trade t −1, �Qt equals one if the dealer purchases in trade t − 1 and sells in trade tand negative one if the dealer sells in trade t − 1 and purchases in trade t,Size is the par value of the traded securities, TBA Eligible is a dummy variablethat equals one if the SP is eligible to be traded TBA, MBS Production isthe gross amount of new MBS with the same coupon and maturity that wascreated in the previous month, and MBS Balance is the value of the MBS withthe same coupon and maturity outstanding at the end of the previous month.Five return variables are also included to capture changes in MBS valueswhen consecutive trades take place on different days: the percentage changesin (1) Barclay Capital’s U.S. MBS index, (2) Barclay Capital’s 7- to 10-YearU.S. Treasury Bond index, (3) Barclay Capital’s U.S. Corporate Bond Index, (4)Barclay Capital’s U.S. Corporate High-Yield Bond Index, and (5) the S&P 500index. These indices are also used in the study of structured credit products byBessembinder, Maxwell, and Venkataraman (2013). Index values are availabledaily, so if consecutive trades occur on the same day, all of these return valuesare zero. This regression is run separately for SP and TBA trades, but thevariables for TBA eligibility are, of course, omitted in the regressions usingTBA trades.

We include only 30-year MBS with coupon rates of 2.5%, 3.0%, 3.5%, 4.0%,4.5%, 5%, 5.5%, and 6%, and 15-year MBS with coupon rates of 2.5%, 3.0%,3.5%, and 4.0%. Together, these MBS account for 96% of our sample trades.Atanasov and Merrick (2014) show that small lots of MBS are particularlyilliquid because they are not considered suitable for small investors and aredifficult to aggregate into larger lots. Hence, we omit trades of less than$10,000 par value.

Regression estimates are reported in Table II. Panel A reports estimatesfor TBA trades while Panel B reports results for SPs. The first regression inPanel A measures trading costs for 30-year TBA trades. The coefficient on �Qis an estimate of the difference between the prices investors pay for MBS, andthe prices they receive when selling MBS, or, in other words, round-trip trad-ing costs. The coefficients on interactions between �Q and other explanatory


Tab

leII

Est

imat

esof

Tra

din

gC

osts

from

Reg

ress

ion

sof

Pri

ceC

han

ges

onC

han

ges

inT

rad

eT

ype

and

Oth

erV

aria

ble

sW

ere

gres

sth

epe

rcen

tage

chan

gein

pric

esbe

twee

ntw

oco

nse

cuti

vetr

ades

ofth

esa

me

MB

Son

the

chan

gein

trad

ety

pe(�

Q),

onth

ein

tera

ctio

nbe

twee

n�

Qan

dth

esu

mof

the

nat

ura

llo

gsof

the

trad

esi

zes

ofth

etw

oco

nse

cuti

vetr

ades

;on

the

inte

ract

ion

betw

een

�Q

and

adu

mm

yva

riab

lefo

rT

BA

elig

ible

SP

s;on

the

inte

ract

ion

betw

een

�Q

,tra

desi

ze,a

nd

TB

Ael

igib

ilit

y;on

the

inte

ract

ion

betw

een

�Q

and

the

nu

mbe

rof

MB

Sw

ith

the

sam

eco

upo

nan

dm

atu

rity

crea

ted

inth

epr

evio

us

mon

th;o

nth

ein

tera

ctio

nbe

twee

n�

Qan

dth

eou

tsta

ndi

ng

bala

nce

ofM

BS

wit

hth

esa

me

cou

pon

and

mat

uri

tycr

eate

din

the

prev

iou

sm

onth

,an

don

chan

ges

in(1

)aU

.S.A

gen

cyF

ixed

Rat

eM

BS

inde

x,(2

)aU

.S.T

reas

ury

7-to

10-Y

ear

Bon

din

dex,

(3)

aU

.S.I

nve

stm

ent

Gra

deC

orpo

rate

Bon

dIn

dex,

(4)

aU

.S.C

orpo

rate

Hig

h-Y

ield

Bon

din

dex,

and

(5)

the

S&

P50

0in

dex:

�P t

=α

0+

α1�

Qt+

α2�

Qt·(l

n(S

ize t/ 1

,000

,000

)+ln

(Siz

e t−1

/ 1,0

00,0

00))

+α

3�

Qt·T

BA

Eli

gibl

e

+α

4�

Qt·T

BA

Eli

gibl

e·(

ln(S

ize t/ 1

,000

,000

)+ln

(Siz

e t−1

/ 1,0

00,0

00))

+α

5�

Qt·3

0(1

5)Ye

arM

at+

α6�

Qt

30(1

5)Ye

arM

at·(l

nS

ize t

+ln

Siz

e t−1

)+α

7Q

t·M

BS

Pro

duct

ion

+α

8�

Qt·M

BS

Bal

ance

+�

βi

Ret

i,t+

εt

�Q

equ

als

posi

tive

one

wh

enth

ecu

rren

ttr

ade

isa

deal

ersa

lean

dth

epr

evio

us

trad

ew

asa

deal

erpu

rch

ase

and

neg

ativ

eon

ew

hen

the

curr

ent

trad

eis

ade

aler

purc

has

ean

dth

epr

evio

us

trad

ew

asa

deal

ersa

le.

Con

secu

tive

trad

esar

eal

way

sof

the

sam

eM

BS

,bu

ttr

ades

from

all

MB

Sw

ith

the

sam

eco

upo

nan

dm

atu

rity

are

incl

ude

din

the

regr

essi

ons.

Trad

esof

less

than

$10,

000

face

valu

ear

ede

lete

d.R

obu

stt-

stat

isti

csar

ein

pare

nth

eses

.

Pan

elA

:TB

ATr

ades

Mat

uri

ty(Y

ears

)�

Q�

Q×

Trad

eS

ize

�Q

×P

rodu

ctio

n�

Q×

Bal

ance

Ret

urn

Var

iabl

esO

bs.

R2

300.

0357

−0.0

056

Yes

651,

234

0.04

73(2

5.47

)(−

23.3

3)30

0.03

77−0

.005

50.

0003

−0.0

199

Yes

650,

643

0.05

04(2

5.55

)(−

22.9

7)(0

.82)

(−23

.71)

150.

0313

−0.0

052

Yes

144,

531

0.09

91(1

1.29

)(−

10.7

5)15

0.01

40−0

.005

80.

0089

−0.0

145

Yes

144,

531

0.10

03(3

.59)

(−11

.91)

(6.5

5)(−

8.99

)

(Con

tin

ued

)


Tab

leII

—C

onti

nu

ed

Pan

elB

:Spe

cifi

edP

ools

�Q

×

1Tr

ade

Siz

eT

BA

Eli

gibl

eT

BA

Elg

×S

ize

30/1

5Ye

arM

at.

Gro

ssP

rodu

ctio

nM

BS

Bal

ance

Obs

.R

2

30Ye

ars

toM

atu

rity

0.63

24−0

.054

8−0

.395

713

4,11

90.

4594

(46.

05)

(−51

.23)

(−28

.53)

0.78

77−0

.135

1−0

.567

200

30.0

890

134,

119

0.46

37(4

7.51

)(−

34.3

0)(−

32.7

1)(2

1.77

)0.

7668

−0.1

324

−0.5

316

0.08

64−0

.005

2−0

.020

313

3,42

60.

4641

(41.

57)

(−32

.14)

(−26

.52)

(20.

23)

(−1.

93)

(−5.

92)

16–3

0Ye

ars

toM

atu

rity

1.09

53−0

.179

3−0

.419

90.

0513

−0.3

034

−0.0

743

−0.0

318

443,

346

0.17

13(8

9.34

)(−

52.5

5)(−

32.1

9)(1

4.05

)(−

40.2

8)(−

37.9

4)(−

10.2

1)

15Ye

ars

toM

atu

rity

0.61

93−0

.041

5−0

.321

249

,063

0.43

79(1

4.60

)(−

24.5

3)(−

7.55

)0.

7605

−0.1

514

−0.4

652

0.11

1749

,063

0.43

92(1

3.65

)(−

7.65

)(−

8.28

)(5

.63)

0.76

88−0

.149

3−0

.468

20.

1110

−0.0

196

0.01

1149

,063

0.43

96(1

3.68

)(−

7.50

)(−

8.30

)(5

.57)

(−4.

53)

(1.5

6)

0–15

Year

sto

Mat

uri

ty

1.03

51−0

.097

8−0

.653

00.

0609

−0.1

354

−0.0

185

−0.0

064

92,9

730.

2612

(21.

40)

(−5.

65)

(−13

.39)

(3.5

0)(−

14.8

6)(−

4.91

)(−

0.78

)


variables indicate how transaction costs are affected by these other variables. Inthe first regression, the coefficient on �Q is 0.0357 and highly significant. Thedependent variable is the percentage change in the price of the MBS, and thus0.0357 means 3.57 basis points. In estimating this regression, we incorporatethe size of the trade by taking the natural logarithm of the par value of the tradedivided by $1,000,000. Similarly, for our MBS production and outstanding bal-ance variables, we use natural logarithms of the variable divided by its average.Hence, the coefficient estimate of 0.0357 on �Q is an estimate of the round-trip TBA trading costs for $1,000,000 par value trades when monthly produc-tion and the balance of the MBS are at their average. In other regressions inPanel A, the coefficient on �Q reaches as high as 0.0377—still indicating thatthe round-trip TBA trading costs for $1,000,000 par value trades is less than 4basis points.

Trading costs decrease with trade size for every regression inTable II. This is similar to the findings of Bessembinder, Maxwell, andVenkataraman (2013). The coefficient on the interaction between �Q and thelogarithm of trade size is a highly significant −0.0056 in the first regression.The natural logarithm of 2,000,000/1,000,000 is about 0.69, so an increase intrade size from $1,000,000 par value to $2,000,000 would reduce round-tripTBA trading costs by about 0.69 × 0.0056 = 0.39 basis points.

The second regression includes interactions between �Q and the naturallogarithm of the ratio of the gross production to the average gross productionof MBS with the same coupon and maturity and between �Q and the log ofthe ratio of the previous month’s balance to the average balance of outstandingMBS with the same coupon and maturity. Gross production is the dollar valueof new MBS created during the previous month with the same coupon and ma-turity. It varies significantly across coupon rates during a month. Greater grossproduction implies greater demand by originators to hedge new mortgages inthe TBA market, and could affect trading costs in this way. A greater outstand-ing balance of MBS with a given maturity and coupon implies a deeper marketfor TBA trading. It suggests that dealers may know more potential buyers (orsellers) for MBS with the particular coupon and maturity characteristics. Italso suggests higher trading volume as some investors unwind positions.

In the second regression in Panel A, the coefficient on the interaction between�Q and production is insignificant. The coefficient on the interaction between�Q and the balance of outstanding MBS is negative and highly significant. Tosummarize, a $1,000,000 par value round-trip TBA trade costs about 3.5 basispoints, with costs falling for larger trade sizes and during times when there isa large balance of outstanding 30-year MBS with the same coupon.

The remainder of Panel A reports results for regressions using 15-year TBAtrades. Round-trip costs for a $1,000,000 par value trade are about 3.1 basispoints. Trading costs decline with trade size. While trading costs do declinewith the balance of outstanding MBS with the same coupon and maturity, theyappear to anomalously increase with MBS production.

Panel B provides regression estimates of trading costs for SP MBS. In thefirst regression, the percentage change in price for consecutive trades of 30-year


SPs is regressed on �Q and interactions between �Q and the trade size ratioand between �Q and a dummy variable for TBA eligibility. The coefficientof 0.6324 on �Q indicates that the round-trip trading cost for $1,000,000 of30-year SPs that were not TBA eligible was 63.24 basis points—far greaterthan the 3.5 basis points for similar TBA trades.6 For TBA-eligible SPs, theround-trip trading costs were 63.24–39.57 or 23.47 basis points. This is muchless than the trading costs for TBA-ineligible SPs, but much more than TBAtrading costs. The first regression also indicates that trading costs for SPs, likeTBA trading costs, decline with trade size. The second regression includes aninteraction between the TBA eligibility dummy and the trade size ratio. It ispositive and significant, indicating that trading costs do not decline as fast withtrade size for TBA-eligible pools as for TBA-ineligible SPs.

The next regression in Panel B estimates the effects of gross production ofMBS and the balance of outstanding MBS on SP trading costs. A large balanceof outstanding MBS with a particular coupon and maturity means that there isa large supply of these MBS and that dealers probably know which institutionsmay want to buy or sell MBS with these characteristics. For 30-year SPs, bothhigh gross production and a large balance of outstanding MBS are associatedwith lower trading costs. The coefficient on the balance of outstanding MBS ishighly significant with a t-statistic of −5.92, while that on gross production,with a t-statistic of −1.93, is of marginal significance.

The next row reports results for SPs with maturities ranging from 16 through30 years. All of these maturities are eligible for delivery in 30-year TBA trades.Now we include an extra dummy variable that takes a value of one if thematurity is exactly 30 years. SPs with maturities between 16 and 29 yearsare seasoned pools. They can be compared to off-the-run bonds. When theseodd maturities are included in the regressions, trading costs still decline withtrade size, with TBA eligibility, with the previous month’s gross production ofmortgages, and with the balance of MBS at the end of the previous month.The coefficient on the dummy variable for 30-year maturity is negative andhighly significant. SPs with 30 years to maturity are cheaper to trade than theseasoned SPs with maturities from 16 to 29 years.

The remaining rows of the table report regression estimates of trading costsfor SPs with 15 years to maturity. Trading costs are again much higher thanfor similar TBA trades. The first regression for 15-year MBS has a coefficientof 0.6193 on �Q. The round-trip trading cost for a $1,000,000 par value tradeof 15-year SPs is 61.93 basis points if the SP is not TBA eligible, and 61.93– 32.12 = 29.81 basis points if the pool is TBA eligible. For 15-year MBS,like 30-year MBS, trading costs decline with trade size and with greater grossproduction of MBS with the same coupon. The outstanding balance of 15-yearMBS with the same coupon seems to have little impact on trading costs of15-year SPs. Shorter maturities are eligible for delivery as 15-year MBS, sothe last regression includes all SPs with 15 years or less to maturity. The

6 Both the SP and the TBA trading cost estimates are similar to those in Bessembinder, Maxwell,and Venkataraman (2013).


coefficient on the dummy for 15 years to maturity is negative, indicating thatseasoned SPs with less than 15 years to maturity are more expensive to tradethan SPs with 15 years to maturity.

To summarize, we can draw four conclusions about MBS trading costs. First,larger trades have lower trading costs, as a percentage of value, than smallertrades. Second, TBA trades are much cheaper than SP trades of similar size.These findings are similar to those of Bessembinder, Maxwell, and Venkatara-man (2013). In addition, TBA-eligible SPs are cheaper to trade than TBA-ineligible SPs. Finally, trading costs fall with a greater amount of outstandingMBS with the same coupon and maturity.

IV. Does the Option to Trade SPs in the TBA Market Reduce TradingCosts?

A. TBA Eligibility and the Value of the Option to Deliver an SP in a TBA Trade

We have shown that trading costs are significantly lower for TBA-eligibleSPs than for other SPs. One possible reason for this is that characteristicsof the SP that make it TBA eligible also make it more liquid. This would bepossible, for example, if the requirements for TBA eligibility were associatedwith less uncertainty about the likelihood of prepayment. A second possiblereason is that TBA trades are better hedges for TBA-eligible SPs than for otherSPs. Yet another possibility is that TBA prices provide better benchmarks forTBA-eligible SPs than for other SPs. Each of these three potential explanationsfor lower trading costs for TBA-eligible SPs suggests that trading costs declineas SP characteristics get closer to TBA eligibility requirements. There shouldbe little difference in trading costs between SPs that are TBA eligible and thosethat almost TBA eligible.

There is, however, another possible reason why trading costs are lower forTBA-eligible SPs than for other SPs. TBA eligibility gives dealers an option tosell an SP in the much more liquid TBA market. This option may be valuableto dealers, particularly when market conditions change and an SP’s pay-up orpremium over TBA prices disappears. As an article in Mortgage News Dailyobserves

The importance of deliverability stems from the fact that if the pay-ups forthe products disappear, the pools can still be sold into TBAs. This limitsthe risk that changing circumstances will cause investors to be stuck withilliquid and difficult-to-sell securities.7

There is anecdotal evidence that volume does shift between the SP and TBAmarket as conditions change. An article in the July 1, 2012 Asset SecuritizationReport notes that

7 See Bill Berliner, 2012, Trading in MBS pools backed by HARP Loans, Mort-gage News Daily, April 18, 2012. Available at http://www.mortgagenewsdaily.com/channels/secondary markets/04182012-trading-pools-harp-loans.aspx.

http://www.mortgagenewsdaily.com/channels/secondary_markets/04182012-trading-pools-harp-loans.aspx

http://www.mortgagenewsdaily.com/channels/secondary_markets/04182012-trading-pools-harp-loans.aspx


. . . when bond prices move lower, both SP pay-ups and the share of MBStraded as SPs will drop.8

More generally, a particular MBS can become more or less likely to be tradedin the TBA market as mortgage rates change. Sophisticated mortgage holdersmay prepay mortgages to take advantage of lower rates, but there are otherreasons for prepayments. Mortgage holders may prepay if they move for a newjob, or because of a divorce, retirement, or death. A mortgage default can alsolead to prepayment by the government sponsored enterprise (FNMA, FHLMC,or GNMA). A mortgage pool can be characterized as having fast payers or slowpayers. A pool with fast or slow payers can be desirable under some marketconditions but not others. For example, an MBS with slow payers is desirableif mortgage rates have fallen and MBS investors have to reinvest prepaymentsat lower coupon rates. This MBS is likely to be traded as an SP. The same MBScan become undesirable if rates rise. Now investors in the MBS will have towait longer to reinvest at higher coupon rates. This MBS is unlikely to havea pay-up under these circumstances, and will probably be traded in the TBAmarket. Alternatively, an MBS with fast payers is desirable if interest rateshave risen and investors in the MBS want to reinvest in MBS with highercoupons. If mortgage rates later fall sufficiently, the same MBS, with fastpayers, would lose its pay-up and would be delivered in a TBA trade. Note thateven if mortgage rates did not fall, that particular MBS would become lessdesirable over time as the fastest payers among the mortgage holders prepaytheir mortgages. When the remaining borrowers are slow payers, the MBS islikely to trade in the TBA market rather than as an SP.

TBA eligibility gives dealers an option to trade an SP in the TBA market, anoption that may be exercised if market conditions change or as the characteris-tics of the MBS itself change. This option allows the dealer to take advantageof the greater liquidity of the TBA market, and establishes the TBA price as afloor for the value of the MBS.9 If the option to trade an SP in the TBA market isvaluable to dealers, or if it makes it easier to sell the SP to investors, we shouldsee a sharp difference in trading costs between SPs that are TBA eligible andSPs with characteristics that leave them almost TBA eligible.

B. The Value of the Option to Deliver an SP in a TBA Trade: RegressionDiscontinuity Tests

In this section, we explore whether TBA eligibility itself, rather than justMBS characteristics associated with TBA eligibility, leads to greater liquidityfor SPs. Several characteristics of loans make them ineligible for inclusion inTBA-eligible pools. Loans with LTV ratios greater than 1.05 are ineligible for

8 See Bill Berliner, 2012, Specified pool trading and the TBA market, Asset Securitization Report,July 1, 2012. Available at https://www.highbeam.com/doc/1G1-295131066.html.

9 Atanasov and Merrick (2012, p. 6) observe that SP traders should regard the TBA market“ . . . as an extremely liquid backstop, which ensures a price floor equal to the current price of thecorresponding standardized TBA contract.”

https://www.highbeam.com/doc/1G1-295131066.html


inclusion in TBA pools.10 For 30-year TBA pools, maturities must be greaterthan 15 years and less than or equal to 30 years. For 15-year TBA pools,maturities must be less than or equal to 15 years. In addition, mortgages musthave several characteristics for unlimited inclusion in TBA-deliverable pools.These characteristics include a fixed rate, a first lien on the property, levelpayments, full amortization of the loan, and a servicing fee of at least 25 basispoints. The loan should not include a prepayment penalty, should not have anextended buydown provision, should not be a cooperative share loan, shouldnot be a relocation loan, and should not have biweekly payments. Loans thatviolate any of these provisions can be included in TBA-eligible pools only to alimited extent.11

We obtain data from eMBS on characteristics of SPs to see if TBA eligibilityitself, in addition to pool characteristics, creates liquidity. The data consist ofsummary statistics about pool characteristics at the time the MBS is originatedrather than data on individual loans. The data include the average FICO score,the maximum and minimum loan size, the average LTV ratio, the percentageof the loans that are for owner-occupied houses, the percentage that have beenrefinanced, the percentage that are for single-family homes, the state with thelargest percentage of mortgages, and the originator that provided the mostmortgages. These characteristics are widely regarded as important predictorsof prepayment tendencies, but, with the exception of the LTV ratio, are not thecharacteristics that are used to determine TBA eligibility.

For our first tests of whether TBA eligibility affects trading costs, we employa regression discontinuity approach that makes use of the breakpoint for TBAeligibility that occurs for LTV ratios greater than 1.05.12 For SPs with variousranges of LTVs, we run the following regression:

� Pt = α0 + α1�Qt + α2 �Qt ·[ln

(Sizet

1,000,000

)+ ln

(Sizet−1

1,000,000

)]+ α3�Qt · LTV + α4�Qt · DLTV > 105 + α5�Qt · FICO

+�βi Reti,t + εt. (2)

As before, �Pt is the percentage change in price between two successivetrades in the same SP, while �Qt takes a value of one (negative one) if tradet − 1 was a dealer purchase (sale) and trade t was a dealer sale (purchase).We include the interaction between �Q and the LTV ratio to capture anyeffect that the LTV ratio has on trading costs other than helping to determine

10 These unusually high LTV mortgages are “HARP loans.” The HARP refinancing programunder the Making Home Affordable Program offered refinancing to borrowers who were currenton their mortgage payments and had loans guaranteed by Fannie Mae or Freddie Mac. HARPoriginally authorized the agencies to buy fixed-rate loans with LTVs greater than 80% as longas the current LTVs did not exceed 125%. In December 2011, the upper limits on LTVs wereeliminated for loans that closed on or before May 21, 2009.

11 See Securities Industry and Financial Markets Association (2013).12 See Lee and Lemieux (2010) for a survey of regression discontinuity models in economics.


whether a pool is TBA eligible. We also include the interaction between �Qand the average FICO score for loans in the MBS and the fixed income indexreturns used previously. The coefficient of interest is α4, the coefficient on theinteraction between �Q and a dummy variable for LTVs greater than 1.05. IfTBA eligibility affects trading costs, we would expect trading costs to increaseabruptly for LTVs greater than 1.05, and hence for this coefficient to be positiveand significant.13

The first row of Table III reports results for 30-year SPs with mean LTV ratiosof 0.95–1.15.14 We could use a wider range of LTV values and try to capturea nonlinear relation between trading costs and LTV values using interactionsbetween �Q and a polynomial function of LTV. Instead, we restrict the range ofLTV ratios from 0.95 to 1.15 because the impact of LTV ratios on trading costsmay be better approximated with a linear relation over this narrow range ofLTV values than over a wider range. A disadvantage of restricting the range ofLTV ratios from 0.95 to 1.15 is that there are a small number of SPs with LTVratios this high, and hence the number of observations is only 5,450.

In this regression, the estimated coefficient on the interaction between �Qand the LTV ratio is −0.0290 with a t-statistic of −4.04. Higher LTV ratiosmean lower trading costs. This is not surprising. High LTV ratios mean thatmortgage holders are less likely to refinance, and hence there is less uncertaintyabout prepayments and the value of the MBS. Of more interest though is thatthe coefficient on the interaction between �Q and a dummy variable for anLTV ratio above 1.05 is 0.4505 with a t-statistic of 3.87. Trading costs increaseabruptly when the LTV ratio slips above 1.05, the highest LTV level at whichan SP remains eligible for TBA trading. This is what we would expect if TBAeligibility itself, not characteristics that are correlated with TBA eligibility,makes TBA-eligible SPs cheaper to trade. The next row repeats the regressionbut extends the sample by including SPs with LTV ratios from 0.85 to 1.25.This increases the sample size to 33,838, mostly by increasing the number ofobservations with LTVs less than 1.05. The interaction between �Q and theLTV ratio remains negative and statistically significant while the coefficienton the interaction between �Q and a dummy variable for an LTV ratio above1.05 remains positive and statistically significant. Both coefficients, however,are closer to zero than in the original regression.

13 Some mortgage originators have been accused of artificially inflating property values andhence understating the true LTV ratio. See, for example, Superior Court of Washington for KingCounty (2010), Federal Home Loan Bank of Seattle v. Goldman Sachs. If some LTVs are under-stated, any economic relation between the true LTV and trading costs is likely to be blurred. TBAeligibility, on the other hand, depends only on whether the stated LTV is �1.05. Biases in statedLTVs will not affect tests of whether TBA eligibility, as affected by LTV ratios, influences tradingcosts.

14 We use only SPs with maturities of 15 or 30 years even though SPs with shorter maturitiescan be used to settle TBA trades. Our LTV measure is a snapshot taken at one point in time. AsSPs age, the LTV can change as a result of prepayments or defaults. A small number of SPs withLTVs in our data that are greater than 1.05 become TBA eligible as they age. There are no SPswith 15 or 30 years to maturity with LTVs greater than 1.05 that are TBA eligible.


Tab

leII

IR

egre

ssio

nD

isco

nti

nu

ity

Est

imat

esar

oun

dL

TV

Rat

ios

For

30(1

5)-y

ear

SP

sw

ith

mea

nL

TV

rati

osw

ith

ina

spec

ified

ran

ge,w

ees

tim

ate

the

foll

owin

gre

gres

sion

usi

ng

con

secu

tive

trad

es:

�P t

=α

0+

α1�

Qt+

α2

�Q

t·[ ln

( Siz

e t1,

000,

000

) +ln

( Siz

e t−1

1,00

0,00

0

)] +α

3�

Qt·L

TV

+α

4�

Qt·D

LT

V>

x

+α

6�

Qt·F

ICO

+�

βi

Ret

i,t+

εt,

wh

ere

�P

isth

epe

rcen

tage

chan

gein

the

spec

ified

pool

,�Q

equ

alon

e(n

egat

ive

one)

for

ade

aler

purc

has

e(s

ale)

foll

owed

bya

sale

(pu

rch

ase)

,LT

Vis

the

loan

tova

lue

rati

o,D

isa

dum

my

vari

able

that

equ

als

one

ifth

eL

TV

rati

ois

abov

ea

spec

ific

leve

l.A

nL

TV

rati

ogr

eate

rth

an1.

05ca

nn

otbe

incl

ude

din

TB

A-e

ligi

ble

SP

s. �Q

�Q

×ln

Trad

eS

ize

�Q

×D

LT

V>

x�

Q×

LT

V�

Q×

FIC

OS

core

×e6

Obs

.R

2

30-Y

ears

toM

atu

rity

DL

TV

>1.

05,0

.95

<

LT

V<

1.15

2.80

20−0

.033

90.

4505

−0.0

290

0.70

905,

450

0.07

66(3

.65)

(−4.

05)

(3.8

7)(−

4.04

)(0

.17)

DL

TV

>1.

05,0

.85

<

LT

V<

1.25

2.07

73−0

.081

10.

1666

−0.0

189

1.45

0033

,838

0.06

09(6

.76)

(−25

.73)

(2.4

9)(−

7.11

)(0

.56)

30-Y

ear,

DL

TV

>0.

95−1

.524

6−0

.083

1−0

.453

10.

0174

5.67

0031

,818

0.06

650.

85<

LT

V<

1.05

(−3.

51)

(−25

.94)

(−10

.97)

(4.4

8)(2

.06)

30-Y

ear,

DL

TV

>1.

15−1

.404

3−0

.064

1−0

.125

40.

0152

−2.2

500

2,10

60.

1157

1.05

<L

TV

<1.

25(−

0.40

)(−

4.71

)(−

1.28

)(0

.53)

(−0.

14)

15-Y

ear,

DL

TV

>1.

054.

4364

−0.1

992

0.32

59−0

.020

2−0

.002

727

50.

3274

0.95

<L

TV

<1.

15(1

.80)

(−3.

15)

(1.5

1)(−

1.64

)(−

1.08

)15

-Yea

r,D

LT

V>

1.05

3.33

10−0

.115

10.

3011

−0.0

176

−0.0

018

1,32

80.

1405

0.85

<L

TV

<1.

25(3

.61)

(−6.

03)

(2.1

0)(−

3.49

)(−

1.39

)15

-Yea

r,D

LT

V>

0.95

3.51

80−0

.123

0−0

.066

9−0

.016

3−2

1.80

001,

034

0.15

400.

85<

LT

V<

1.05

(2.9

8)(−

6.21

)(−

0.53

)(−

1.79

)(−

1.61

)15

-Yea

r,D

LT

V>

1.15

−7.7

204

−0.0

071

−0.1

553

0.08

31−0

.002

229

40.

0786

1.05

<L

TV

<1.

25(−

1.08

)(−

0.16

)(−

0.51

)(1

.18)

(−0.

53)


To summarize, the regressions in the first two rows of Table III show thattrading costs generally decline with the LTV ratio, but they increase sharplyat the 1.05 breakpoint between TBA eligibility and ineligibility. This suggeststhat TBA eligibility itself, rather than characteristics correlated with TBAeligibility, is responsible for lower trading costs.

The next two rows of the table provide the results of “placebo” regressions.In the third regression, the range of LTVs is from 0.85 to 1.05 and a dummyvariable for LTVs in excess of 0.95 is included. In the fourth regression, therange of LTVs is from 1.05 to 1.25 and a dummy variable for LTV ratios of1.15 or greater is included. For both of these regressions, the coefficient on theinteraction between �Q and the dummy variable is the opposite sign from theregressions with a 1.05 breakpoint. Thus, it is the breakpoint of 1.05, abovewhich an SP is not TBA eligible, that is important.

The next four rows of the table report results of identical regressions with15-year SPs. In the first two rows, the coefficient on the interaction between�Q and the dummy variable is positive, as expected. There are far fewer ob-servations for 15-year than for 30-year SPs though, so the t-statistic is only1.51 when the range of LTV ratios is from 0.95 to 1.15, and 2.10 when therange is from 0.85 to 1.25. The last two rows of the table present the placeboregressions for the 15-year SPs. There are far fewer observations with 15-yearSPs, so we would expect lower significance levels in these regressions. As withthe 30-year SPs, the coefficient on the interaction between �Q and the dummyvariable is negative, the opposite of what we find with the 1.05 breakpoint. Tosummarize, for the 15-year SPs, like the 30-year SPs, it appears that there is asharp increase in trading costs when LTV ratios exceed 1.05 and the SP is noteligible for TBA trading. TBA eligibility again appears to reduce trading costs.

Researchers who employ regression discontinuities often graph the variableaffected at the discontinuity against the explanatory variable. In our case,this makes little sense as the LTV ratio is just one variable in the tradingcost regression. So, to graph the discontinuity, we calculate abnormal costs bysubtracting the product of estimated regression coefficients from equation (2)and the size, FICO, change in trade types, and index returns for each successivepair of trades. That is,

AbNCostst = �Pt − α1�Qt − α2 �Qt ·[ln

(Sizet

1,000,000

)+ ln

(Sizet−1

1,000,000

)]−α6�Qt · FICO − �βi Reti,t. (3)

We do not subtract the intercept or the coefficients on the LTV ratio orthe dummy variable for an LTV ratio greater than 1.05 from �Pt. Hence, theabnormal costs are a measure of costs after adjusting for everything except theLTV ratio. We then calculate the average abnormal trading costs for each LTVratio from 0.96 through 1.14.

Figure 3, Panel A shows average abnormal trading costs for 30-year SPsfor each LTV ratio. This graph shows a sharp increase in trading costs when


Panel A: Mean abnormal trading costs

Panel B: Mean abnormal trading costs and the standard error confidence intervals

-1.5

-1.0

-0.5

0.0

0.5

1.0

1.5

96 98 100 102 104 106 108 110 112 114

Abn

orm

al C

osts

(%)

LTV Ratio (x100)

-9.0

-6.0

-3.0

0.0

3.0

6.0

9.0

96 98 100 102 104 106 108 110 112 114

Abn

orm

al C

osts

(%)

LTV Ratio (x100)

Figure 3. Abnormal trading costs for SPs with LTV ratios greater than 0.95 and less than 1.15.(Color figure can be viewed at wileyonlinelibrary.com)

LTV ratios pass the 1.05 threshold for TBA eligibility. The limitations of us-ing this procedure with a small sample are revealed in Figure 3, Panel B,which shows mean abnormal trading costs and two standard error confidenceintervals. As the LTV ratio increases over this range, the number of observa-tions drops (there are none at LTV ratios of 1.06 or 1.07) and the standarderrors increase sharply. Differences between abnormal trading costs measured


at a single LTV ratio greater than 1.05 and LTV ratios less than or equal to1.05 are not statistically significant.

C. The Value of the Option to Deliver an SP in a TBA Trade: Propensity ScoreMatching Tests

We employ a type of propensity score matching as an additional test to see ifTBA eligibility affects trading costs after adjusting for characteristics that mayaffect pool value. In a first step, we use a logistic regression and characteristicsof the SPs to predict whether an SP is TBA eligible. The predictive variablesinclude the average FICO score of the home buyers in the pool, the maximumand minimum sizes of loans in the pool, the proportion of loans that are forowner-occupied housing, the percentage of loans that have been refinanced,and the proportion of mortgages that are on single-family properties. Theselogistic regressions are run separately for SPs with maturities of 16–30 yearsand for SPs with maturities of 15 years or less. TBA-eligible pools with 16–30 years to maturity may be traded as 30-year TBAs, while TBA-eligible SPswith 15 or fewer years to maturity may be traded as 15-year TBAs. In discussingthese results, we refer to SPs with 16–30 years to maturity as 30-year SPs andSPs with 15 or fewer years to maturity as 15-year SPs.

Logistic regression estimates are reported in Table IV. The first two columnsreport coefficients and z-statistics, while the last two report marginal effectsand corresponding z-statistics. Panel A reports results for 30-year SPs. Coeffi-cients on minimum and maximum loan size are negative and highly significant.Coefficients on the mean FICO score and the percentage of mortgages that arefor owner-occupied houses are positive and highly significant. The coefficient onthe proportion of mortgages that are for single-family homes and the percent-age of mortgages that are refinanced are negative and highly significant. Thecoefficient on LTV is positive and significant while the coefficient on a dummyvariable for LTV greater than 1.05 is negative and highly significant.15 Thesemortgage characteristics have significant ability to predict TBA eligibility. For30-year SPs, the pseudo-R2 is 0.4247.

Results for 15-year SPs, reported in Panel B, are similar. Coefficients on theminimum loan size and the percent of mortgages that are for single-familyhomes are negative and significant, while coefficients on the mean FICO score,the percentage of mortgages that are for owner-occupied houses, and the per-centage refinanced are positive and statistically significant. As with 30-yearSPs, the coefficient on LTV is positive and significant, while the coefficient onthe dummy variable for LTVs in excess of 1.05 is negative and highly signifi-cant. The logistic regression does a better job of explaining TBA eligibility for15-year SPs than for 30-year SPs. Here the pseudo-R2 is 0.8336.

15 An LTV greater than 1.05 makes an SP ineligible for TBA trading. Our MBS characteristicsare taken from a snapshot at one point in time and LTV ratios are as of that time. LTVs can changeas a result of prepayments and defaults and MBS can become or cease being TBA eligible. Hence,with seasoned MBS, an LTV > 1.05 from a different point in time need not indicate with certaintythat the SP is TBA eligible.


Table IVLogistic Regression Estimates of TBA Eligibility on Specified Pool

CharacteristicsSPs with 16–30 years to maturity that are TBA eligible can be traded as a 30-year TBA, while SPswith 15 years to maturity or less can be traded as a 15-year TBA.

Panel A: SPs with 16–30 Years to Maturity

Coefficient z-Statistic dy/dx z-Statistic

Average FICO Score 0.0122 52.77 0.00037 42.86Maximum Loan Size −1.92e−6 −68.37 −5.77e−8 −60.60Minimum Loan Size −1.54e−5 −174.18 −4.61e−7 −81.98Percent Owner Occupied 2.9463 86.32 0.08837 65.24Percent Refinanced −0.2278 −4.15 −0.00683 −4.25Percent Single Family −1.3409 −14.24 −0.04022 −13.83LTV 0.0034 2.26 0.00010 2.23LTV > 1.05 −12.5178 −17.43 −0.97320 −2,762.12Constant −3.8932 −16.19Observations 401,346Pseudo-R2 0.4247

Panel B: SPs with 15 or Fewer Years to Maturity

Coefficient z-Statistic dy/dx z-Statistic

Average FICO Score 0.0189 8.80 1.13e−5 7.26Maximum Loan Size 5.81e−8 0.19 3.48e−11 0.19Minimum Loan Size −2.62e−5 −57.57 −1.57e−8 −11.39Percent Owner Occupied 4.8898 19.65 0.00598 11.66Percent Refinanced 9.9795 47.28 −0.00527 −4.36Percent Single Family −8.7973 −4.59 0.00293 9.48LTV 0.0648 7.74 3.89e−5 6.08LTV > 1.05 −32.9982 −41.17 −0.9995 < −10,000Constant −13.5466 −6.18Observations 91,954Pseudo-R2 0.8336

We use the probability that an SP is TBA eligible from the first-stage logisticsregressions to see if TBA eligibility itself, rather than variables correlated withTBA eligibility, are associated with lower trading costs. For the second stagewe estimate

� Pt = α0 + α1�Qt + α2�Qt · (ln Sizet + lnSizet−1) + α3�Qt · TBA Eligible

+α4�Qt · Probability TBA Eligible + �βi Reti,t + εt. (4)

Note that we include the interaction between �Q and the probability that anSP is TBA eligible in the regressions. Any remaining impact of TBA eligibilityon trading costs is likely to reflect the impact of TBA eligibility itself, not othercharacteristics associated with TBA eligibility. Regressions are run separatelyfor SPs with probabilities of less than 10% of being TBA eligible, of 10% to


19.99% of being TBA eligible, and so forth. We estimate the regressions sepa-rately for different probabilities of TBA eligibility as a robustness check, andalso as a way to restrict observations in the regressions to similar SPs. We alsoestimate equation (4) using all observations. Results are reported in Table V.

Panel A of Table V reports results for SPs with 16–30 years to maturity. IfTBA eligible, these SPs can be delivered to settle 30-year TBA trades. The first10 rows report regressions with SPs with similar probabilities of being TBAeligible. The number of SPs in the regression that are actually TBA eligible andineligible is reported for each regression. As expected, the proportion of SPsthat are TBA eligible increases with the probability of TBA eligibility estimatedin the first stage. The coefficient on the interaction between �Q and TBA eli-gibility is negative for each of the regressions. For 7 of the 10 regressions, thecoefficient on the interaction between �Q and TBA eligibility is negative andsignificant at the 1% level. The last regression reports results when all obser-vations, regardless of the predicted likelihood of TBA eligibility, are included.As in the other regressions, the coefficient on the interaction between �Q andTBA eligibility is negative and significant at the 1% level. It is interesting thatthe coefficient is −0.4078, suggesting that TBA eligibility decreases tradingcosts by 40.78 basis points. The regression discontinuity estimates using 30-year SPs and a range of LTVs from 0.96 to 1.14 suggest that TBA eligibilitydecreases trading costs by very similar 45.05 basis points, and our regressionsin Table II suggest that TBA eligibility decreases trading costs by 39.6 basispoints. Panel B reports results for the regressions with 15-year SPs. In eachregression, the coefficient on TBA eligibility is negative and statistically sig-nificant. As with 30-year SPs, after adjusting for SP characteristics that areassociated with TBA eligibility, it appears that trading costs for TBA-eligibleSPs are lower than those for TBA-ineligible SPs.

To summarize, we have shown that TBA-eligible and TBA-ineligible SPsdiffer in FICO score, loan size, LTV ratio, percentage of mortgages that are forowner-occupied homes, percentage of mortgages that have been refinanced, andpercentage of mortgages that are for single-family homes. Our propensity scorematching results indicate, however, that, despite predicting TBA eligibilitywell, these characteristics do not explain the relation between TBA eligibilityand liquidity. Like our finding that trading costs increase sharply when theLTV ratio threshold for TBA eligibility is crossed, these results are consistentwith TBA eligibility itself leading to greater liquidity.

We do not expect the effectiveness of TBA hedging to change abruptly be-tween SPs that are very close to TBA eligibility and those that are TBA eligible.We instead expect the effectiveness of TBA hedging to change continuously asthe characteristics of the SP come closer to those of TBA-eligible SPs. Like-wise, we do not expect the usefulness of TBA prices as benchmarks to changeabruptly with TBA eligibility. Rather, we expect TBA prices to become con-tinuously better benchmarks as the characteristics of SPs come closer to thecharacteristics of TBA-eligible pools.

What does change abruptly with TBA eligibility itself is that the holder gainsan option to deliver the SP to fulfill a TBA trade. Like all options, the option to


Tab

leV

Pro

pen

sity

Sco

reM

atch

ing

and

the

Imp

act

ofT

BA

Eli

gib

ilit

yon

Tra

din

gC

osts

Th

epr

obab

ilit

yth

atan

SP

isT

BA

elig

ible

ises

tim

ated

usi

ng

alo

gist

icre

gres

sion

wit

hT

BA

elig

ibil

ity

asth

ede

pen

den

tva

riab

lean

dth

eS

P’s

aver

age

FIC

Osc

ore,

max

imu

mlo

ansi

ze,m

inim

um

loan

size

,per

cen

tow

ner

-occ

upi

ed,p

erce

nt

refi

nan

ced,

and

perc

ent

sin

gle-

fam

ily

asex

plan

ator

yva

riab

les.

Usi

ng

the

esti

mat

edpr

obab

ilit

yof

TB

Ael

igib

ilit

y,w

ees

tim

ate

trad

ing

cost

su

sin

gal

lmat

uri

ty-c

oupo

nco

mbi

nat

ion

san

dth

ere

gres

sion

�P t

=α

0+

α1�

Qt+

α2�

Qt·(l

n(S

ize t/ 1

,000

,000

)+ln

(Siz

e t−1

/ 1,0

00,0

00))

+α

3�

Qt·T

BA

Eli

gibl

e+

α4�

Qt·P

roba

bili

tyT

BA

Eli

gibl

e+

�β

iR

eti,

t+

εt,

wh

ere

�P

isth

epe

rcen

tage

chan

gein

the

spec

ified

pool

pric

eov

erco

nse

cuti

vetr

ades

,�

Qeq

ual

son

e(n

egat

ive

one)

for

ade

aler

purc

has

e(s

ale)

foll

owed

bya

sale

(pu

rch

ase)

,T

BA

Eli

gibl

eis

adu

mm

yva

riab

leth

atta

kes

ava

lue

ofon

efo

rT

BA

-eli

gibl

eS

Ps,

and

TB

AIn

elig

ible

isa

dum

my

vari

able

that

take

sa

valu

eof

one

ifth

eS

Pis

TB

Ain

elig

ible

.Th

ere

gres

sion

isru

nse

para

tely

for

SP

sw

ith

esti

mat

edpr

obab

ilit

ies

ofT

BA

elig

ibil

ity

of0.

0–0.

1,0.

1–0.

2,et

c.S

Ps

wit

hm

atu

riti

esof

16–3

0ye

ars

can

beu

sed

tofu

lfill

30-y

ear

TB

Atr

ades

,wh

ile

SP

wit

hm

atu

riti

esof

15ye

ars

orle

ssca

nbe

deli

vere

din

15-y

ear

TB

Atr

ades

.Rob

ust

t-st

atis

tics

are

inpa

ren

thes

es.

Pan

elA

:Spe

cifi

edP

ools

wit

h16

–30

Year

sto

Mat

uri

ty

�Q

�Q

×ln

Trad

eS

ize

�Q

×T

BA

Eli

gibl

e�

Q×

Pro

bT

BA

Elg

Inde

xR

etu

rns

TB

AE

ligi

ble

Obs

erva

tion

sT

BA

Inel

igib

leO

bser

vati

ons

R2

0.0

�P

rob

<0.

10.

6375

−0.1

338

−0.1

519

0.15

93Ye

s76

9,55

40.

4570

(19.

02)

(−21

.74)

(−0.

29)

(9.5

4)0.

1�

Pro

b<

0.2

1.67

25−0

.169

3−0

.773

4−3

.739

9Ye

s63

1,30

10.

3988

(5.4

5)(−

12.6

4)(−

3.54

)(−

2.05

)0.

2�

Pro

b<

0.3

−0.0

116

−0.1

504

−0.2

241

4.37

61Ye

s10

21,

925

0.35

24(−

0.03

)(−

12.8

3)(−

0.85

)(2

.70)

0.3

�P

rob

<0.

41.

1993

−0.1

481

−0.6

930

−0.4

099

Yes

225

1,31

80.

3425

(2.2

9)(−

9.26

)(−

5.66

)(−

0.27

)0.

4�

Pro

b<

0.5

3.05

17−0

.198

5−0

.612

2−3

.544

9Ye

s73

070

90.

2924

(1.8

9)(−

8.23

)(−

3.11

)(−

1.02

)0.

5�

Pro

b<

0.6

2.01

92−0

.193

3−0

.382

7−1

.604

9Ye

s1,

428

509

0.39

84(1

.88)

(−13

.96)

(−3.

03)

(−0.

83)

(Con

tin

ued

)


Tab

leV

—C

onti

nu

ed

Pan

elA

:Spe

cifi

edP

ools

wit

h16

–30

Year

sto

Mat

uri

ty

�Q

�Q

×ln

Trad

eS

ize

�Q

×T

BA

Eli

gibl

e�

Q×

Pro

bT

BA

Elg

Inde

xR

etu

rns

TB

AE

ligi

ble

Obs

erva

tion

sT

BA

Inel

igib

leO

bser

vati

ons

R2

0.6

�P

rob

<0.

74.

6466

−0.1

863

−0.3

510

−5.2

814

Yes

2,94

71,

945

0.28

71(4

.96)

(−16

.89)

(−3.

95)

(−3.

73)

0.7

�P

rob

<0.

81.

3558

−0.1

626

−0.2

831

−0.5

358

Yes

10,1

764,

676

0.24

49(2

.87)

(−29

.23)

(−7.

91)

(−0.

86)

0.8

�P

rob

<0.

9−1

.266

3−0

.183

7−0

.041

12.

6198

Yes

42,7

086,

496

0.29

82(−

3.19

)(−

74.1

3)(−

1.61

)(5

.58)

0.9

�P

rob

<1.

08.

7195

−0.0

997

−0.9

167

−7.5

077

Yes

308,

646

5,81

20.

1199

(30.

30)

(−58

.21)

(−28

.22)

(−25

.56)

0.0

�P

rob

<1.

01.

1177

−0.1

295

−0.4

078

−0.1

740

Yes

367,

101

34,2

450.

1441

(62.

24)

(−95

.50)

(−21

.83)

(−6.

17)

Pan

elB

:Spe

cifi

edP

ools

wit

h15

Year

sor

Les

sto

Mat

uri

ty

�Q

�Q

×ln

Trad

eS

ize

�Q

×T

BA

Eli

gibl

e�

Q×

Pro

bT

BA

Elg

Inde

xR

etu

rns

TB

AE

ligi

ble

Obs

erva

tion

sT

BA

Inel

igib

leO

bser

vati

ons

R2

0.0

�P

rob

<0.

11.

1680

−0.0

929

−1.3

043

4.68

66Ye

s68

784

0.23

73(1

0.55

)(−

3.71

)(−

4.97

)(1

.28)

0.1

�P

rob

<0.

20.

3730

−0.1

330

−1.0

489

4.31

55Ye

s59

175

0.25

36(0

.54)

(−1.

08)

(−2.

37)

(0.7

2)

(Con

tin

ued

)


Tab

leV

—C

onti

nu

ed

Pan

elB

:Spe

cifi

edP

ools

wit

h15

Year

sor

Les

sto

Mat

uri

ty

�Q

�Q

×ln

Trad

eS

ize

�Q

×T

BA

Eli

gibl

e�

Q×

Pro

bT

BA

Elg

Inde

xR

etu

rns

TB

AE

ligi

ble

Obs

erva

tion

sT

BA

Inel

igib

leO

bser

vati

ons

R2

0.2

�P

rob

<0.

31.

0483

−0.0

937

−1.1

062

−0.3

578

Yes

7719

70.

4028

(1.2

5)(−

2.90

)(−

4.31

)(−

0.11

)0.

3�

Pro

b<

0.4

1.19

15−0

.075

8−0

.558

7−1

.610

2Ye

s83

365

0.20

79(1

.37)

(−2.

41)

(−4.

24)

(−0.

63)

0.4

�P

rob

<0.

54.

1117

−0.0

034

−0.3

382

−8.3

389

Yes

8415

50.

2831

(2.4

8)(−

0.11

)(−

2.43

)(−

2.09

)0.

5�

Pro

b<

0.6

4.99

020.

0079

−1.0

707

−7.6

779

Yes

2388

0.53

79(1

.35)

(0.1

9)(−

4.00

)(−

1.12

)0.

6�

Pro

b<

0.7

4.32

840.

0288

−2.1

788

−3.5

208

Yes

513

0.83

02(1

.53)

(1.2

0)(−

15.1

7)(−

0.77

)0.

7�

Pro

b<

0.8

1.82

48−0

.061

7−0

.612

6−0

.946

2Ye

s47

560.

6562

(0.6

1)(−

0.95

)(−

2.07

)(−

0.23

)0.

8�

Pro

b<

0.9

3.35

28−0

.045

4−1

.862

0−1

.407

5Ye

s19

42

0.38

57(1

.33)

(−1.

92)

(−21

.53)

(−0.

49)

0.9

�P

rob

<1.

01.

2844

−0.0

390

−0.9

058

−0.0

559

Yes

88,6

7377

00.

2838

(1.9

2)(−

26.7

6)(−

3.85

)(−

0.08

)0.

0�

Pro

b<

1.0

1.06

1−0

.039

9−0

.664

9−0

.076

5Ye

s89

,359

2,59

50.

2754

(17.

83)

(−27

.23)

(−7.

27)

(−0.

65)


deliver the SP in a TBA trade is valuable. Dealers may exercise this option ifmarket conditions change and a previously desirable SP loses its pay-up, or ifthe difference between the SP value and TBA price is small and time and effortare required to sell the SP. Our results are consistent with dealers valuing thisoption enough to trade TBA-eligible SPs for smaller markups than otherwisesimilar TBA-ineligible SPs.

V. The Impact of TBA Trading on SP Liquidity

In this section, we examine whether the existence of parallel TBA tradingincreases liquidity for MBS while they are traded as SPs. There are severalreasons to expect a liquidity spillover from TBA trading to SP trading. First,TBA prices may provide a benchmark for SP pricing. Price discovery maytake place in the TBA market rather than the SP market. In addition, anactive TBA market may allow dealers to hedge SP positions with minimal basisrisk.

It is not straightforward to test whether TBA trading affects SP liquidity.We would expect many of the factors that affect TBA trading to also directlyaffect SP liquidity. We, however, identify an exogenous factor that directlyaffects trading volume in the TBA market but not the SP market, namely, TBAsettlement dates. TBA contracts for a given maturity and issuer settle on oneday during a month. FNMA and FHLMC 30-year TBA trades settle on thesame Class A schedule. Their settlement dates are typically around the 12th or13th of each month. The Class B schedule is for 15-year TBA trades. Settlementdates for these trades are typically three trading days after Class A settlementdates. The Class C schedule is for GNMA 30-year TBA trades. Settlement datesare about two trading days after Class B dates.

The impact of settlement dates on TBA trading is particularly strong fordollar rolls. Recall that the purchaser of a dollar roll buys a TBA contractfor settlement in the current month and simultaneously sells a TBA contractfor settlement in a future (usually the next) month. Likewise, the seller of adollar roll sells a TBA contract for settlement in the current month and simul-taneously buys a TBA contract for settlement in a future month. Investors whotrade dollar rolls typically either terminate or roll over their positions beforesettlement. To avoid being assigned a delivery, TBA traders must terminatepositions at least 48 hours before the settlement date. This results in a spike intrading volume from seven trading days through two trading days before eachsettlement date.

Figure 4, Panel A shows daily trading volume from dollar rolls of 30-year TBAtrades over our sample period. It is easy to see monthly trading volume spikesduring which daily volume is three to five times that in the rest of the month.These volume spikes are two to five days before the Schedule A settlementdates. It is clear from the figure that timing relative to the settlement date is amajor determinant of daily dollar roll trading volume. Since dollar roll tradingaccounts for most of the dollar volume of TBA trading, we can say that thesettlement date is a major determinant of TBA trading volume in general.


Panel A: Daily 30-year dollar roll trading volume

Panel B: Scatter plot of daily 30-year SP trading volume and dollar roll volume

0

100

200

300

400

500

600

5/16/2011 11/16/2011 5/18/2012 11/20/2012

$ Bi

llion

s

0

5

10

15

20

25

30

0 100 200 300 400 500 600

SP V

olum

me

Dollar Roll Volume $Billions

Figure 4. Daily trading volume of 30-year TBA dollar rolls and 30-year SPs. (Color figure can beviewed at wileyonlinelibrary.com)


SP trades, in contrast, may be settled on any day during the month. Atanasovand Merrick (2012) note that many dealers choose to settle SP trades onTBA settlement dates for convenience,16 but unlike dollar rolls and other TBAtrades, SP trades almost always lead to delivery. Hence, the monthly settlementdates and the corresponding trades to avoid delivery that are so important inthe TBA market are unimportant for SP trades. Nevertheless, if dollar rolltrading makes the market for SPs more liquid, we would expect SP tradingto peak when dollar roll trading is high.17 Figure 4, Panel B is a scatter plotof daily trading volume for 30-year SPs against daily trading volume of TBAdollar rolls. Days with dollar roll volume of more than $200 billion correspondto monthly peaks in trading volume before settlement dates. The figure showsthat SP volume is also typically high on these days. The correlation betweenthe daily 30-year dollar roll and SP volume is 0.45.18

We test whether TBA trading affects SP liquidity by running the followingregression:

� Pt = α0 + α1�Qt + α2�Qt ·(

ln(

Sizet

1,000,000

)+ ln

(Sizet−1

1,000,000

))

+ α3�Qt ·(

ln(

PredDRolltMedianDRoll

)+ ln

(PredDRollt−1

MedianDRoll

))

+ α4�Qt ·(

ln(

PredDRolltMedianDRoll

)+ ln

(PredDRollt−1

MedianDRoll

))

·(

ln(

Sizet

1,000,000

)+ ln

(Sizet−1

1,000,000

))+ α5�Qt D30yr

15yr+ �βi Reti,t + εt. (5)

Here, PredDRoll is the predicted dollar roll volume using only settlementdates and volume from the previous month. For days that are between twoand seven days before a Class A (30-year Fannie Mae and Freddie Mac) TBAsettlement date or between two and seven days before a Class B (15-year) TBAsettlement date, we use the dollar roll volume from the corresponding day in

16 Atanasov and Merrick (2012) find that 71.5% of 30-year SP trades of more than $250,000settle on TBA settlement dates. A smaller proportion of smaller trades and 15-year SP trades aresettled on TBA settlement dates.

17 See Admati and Pfleiderer (1988) for a model in which those liquidity traders with discretionover the timing of their trades may endogenously choose to concentrate their trading in the sameperiod.

18 The volume for TBA trades is 3.30 times as high in the period before the settlement date asit is on other dates. When we separate 30-year SP volume into TBA-eligible and TBA-ineligiblevolume, we find that TBA-eligible volume is 2.03 time as high on days −7 to −2 before a settlementdate, and TBA ineligible volume is 2.00 times as high. The seasonality for TBA-ineligible SP volumeis as pronounced as the seasonality for TBA-eligible volume. This suggests that the seasonality inSP volume is not driven by market participants trading to acquire SPs to deliver into TBA trades.


the previous month as the prediction of dollar roll volume.19 For other days, weuse the average volume from days t − 40 to t − 20, excluding days that weretwo to seven days before a settlement date, as a forecast of volume. Hence, ourpredicted dollar roll volume is based only on volume from the previous monthand the publicly known settlement date.20 Other factors that would simultane-ously affect TBA dollar roll volume and SP trading costs would be most likelyto show up in unexpected dollar roll volume. The variable MedianDRoll is themedian predicted dollar roll volume across all trades.

Results are reported in Table VI. Here, the α1 coefficient represents round-trip trading costs for $1,000,000 trades when the predicted dollar roll volumeis at its median level and the SP does not have 15 or 30 years to maturity.The α1 coefficient plus the α5 coefficient provides round-trip trading costs forSPs with 15 or 30 years to maturity. The α3 coefficient on the interaction be-tween the change in trade type, �Q, and the predicted dollar roll volume,PredDRoll, shows how trading volume affects trading costs. For 30-year out-right TBA trades, 30-year TBA-eligible SPs, and 30-year TBA-ineligible SPs,the coefficients are negative and highly significant. Increases in TBA dollar rollvolume are associated with lower trading costs. The α3 coefficients are similarfor TBA-eligible and TBA-ineligible SPs, so dollar roll volume seems to reducetrading costs by similar amounts for TBA-ineligible and TBA-eligible SPs. Thetrading costs for TBA-ineligible SPs are much higher, however. So, dollar rolltrading reduces trading costs for 30-year TBA-eligible and TBA-ineligible SPsby about the same amount, even though TBA-ineligible SPs are much moreexpensive to trade. Some of the regressions include a further interaction be-tween �Q, predicted dollar roll volume, and trade size. Dollar roll volume hasa smaller impact on trading costs for large trades.

These results are economically significant as well. We multiply the coefficienton the interaction between �Q and the log of predicted dollar roll volume forTBA-eligible SPs with 16–30 years to maturity (−0.0431) on the differencebetween the 90th and 50th deciles of the log of predicted dollar roll volume. Thisyields a difference in round-trip trading costs of 7.4 basis points when predicteddollar roll volume goes from the median to the 90th percentile. As a rough ruleof thumb, this suggests that trading costs are about 7.4 basis points lower onthe day just before the settlement date than they are on a typical day. For theTBA-ineligible SPs the coefficient is larger in absolute value (−0.0570) and thedifference in round-trip trading costs is 9.8 basis points.

The last five rows of the table report results for similar regressions using15-year TBA and SPs with 15 years or less to maturity. Trading costs declinesignificantly with predicted dollar roll trading volume both for TBA trades andfor trades of TBA-eligible SPs. The volume of dollar roll trading seems, however,

19 We find that Class B settlement dates, which apply to 30-year Ginnie Mae trades, have littlepredictive power for volume. Ginnie Mae TBA volume in general is significantly less than FannieMae and Freddie Mac TBA volume.

20 This simple model predicts well. When we regress actual 30-year TBA dollar roll volume onthe predicted volume, the coefficient on the predicted volume is 0.9625 and the adjusted R2 is0.7156. For 15-year TBA dollar roll volume, the coefficient is 1.0023 and the adjusted R2 is 0.6552.

Liquidity in a Market for Unique Assets 1153T

able

VI

Tra

din

gC

osts

and

Pre

dic

ted

Exo

gen

ous

Dol

lar

Rol

lV

olu

me

Dol

lar

roll

volu

me

pred

icti

ons

are

obta

ined

from

the

prev

iou

sm

onth

’svo

lum

ear

oun

dth

ese

ttle

men

tda

te.F

orda

ysfr

omtw

oto

seve

nda

ysbe

fore

the

sett

lem

ent

date

,we

use

the

volu

me

from

doll

arro

lls

wit

hth

esa

me

cou

pon

and

mat

uri

tyfr

omth

esa

me

day

rela

tive

toth

ese

ttle

men

tin

the

prev

iou

sm

onth

asa

pred

icti

onof

curr

ent

mon

thvo

lum

e.F

orot

her

days

,we

use

the

aver

age

dail

yvo

lum

efr

om20

to40

days

befo

re,n

otin

clu

din

gda

ysfr

omtw

oto

seve

nda

ysbe

fore

,th

ese

ttle

men

tda

te.T

oes

tim

ate

trad

ing

cost

s,w

eth

enes

tim

ate

the

foll

owin

gre

gres

sion

usi

ng

pred

icte

ddo

llar

roll

volu

me:

�P t

=α

0+

α1�

Qt+

α2�

Qt·( ln

( Siz

e t1,

000,

000

) +ln

( Siz

e t−1

1,00

0,00

0

)) +α

3�

Qt·( ln

( Pre

dD

Rol

l tM

edia

nD

Rol

l) +ln

( Pre

dD

Rol

l t−1

Med

ian

DR

oll))

+α

4�

Qt·( ln

( Pre

dD

Rol

l tM

edia

nD

Rol

l) +ln

( Pre

dD

Rol

l t−1

Med

ian

DR

oll)) ·( ln

( Siz

e t1,

000,

000

) +ln

( Siz

e t−1

1,00

0,00

0

)) +α

5�

QtD

30yr

15yr

+�

βi

Ret

i,t+

εt.

Rob

ust

t-st

atis

tics

are

inpa

ren

thes

es.

Year

sto

Mat

uri

tyT

ype

ofM

BS

�Q

�Q

×L

nTr

ade

Siz

e�

Q×

Pre

d.D

.R

ollV

olu

me

�Q

×S

ize

×P

red.

Dol

lar

Rol

lVol

.�

Q×

30-Y

ear

or15

-Yea

rM

atu

rity

Obs

.R

2

30T

BA

0.02

85−0

.005

1−0

.007

961

4,80

50.

0359

(20.

54)

(−21

.10)

(−17

.52)

16–3

0T

BA

Eli

gibl

e0.

4936

−0.0

936

−0.0

431

−0.2

643

331,

533

0.16

23(6

0.61

)(−

89.5

2)(−

19.1

3)(−

29.8

8)16

–30

TB

AE

ligi

ble

0.49

15−0

.068

0−0

.033

40.

0115

−0.3

161

331,

533

0.16

50(6

0.00

)(−

57.6

2)(−

16.3

2)(2

5.72

)(−

37.0

2)16

–30

TB

AIn

elig

ible

1.11

04−0

.178

1−0

.057

0−0

.247

625

,511

0.40

45(3

9.34

)(−

51.9

3)(−

6.33

)(−

6.72

)16

–30

TB

AIn

elig

ible

1.12

36−0

.160

2−0

.043

30.

0112

−0.3

093

25,5

110.

4066

(40.

61)

(−46

.64)

(−5.

30)

(6.2

0)(−

8.68

)15

TB

A0.

0278

−0.0

050

−0.0

071

137,

666

0.07

30(9

.80)

(−9.

99)

(−8.

68)

�15

TB

AE

ligi

ble

0.28

75−0

.035

1−0

.024

0−0

.021

678

,626

0.31

45(2

4.12

)(−

24.4

3)(−

10.8

2)(−

1.84

)�

15T

BA

Eli

gibl

e0.

2909

−0.0

372

−0.0

231

−0.0

010

−0.0

209

78,6

260.

3145

(23.

44)

(−22

.54)

(−9.

34)

(−1.

73)

(−1.

79)

�15

TB

AIn

elig

ible

0.97

03−0

.150

30.

0117

−0.2

356

1,94

10.

3334

(7.4

8)(−

9.64

)(0

.65)

(−1.

68)

�15

TB

AIn

elig

ible

0.98

52−0

.154

80.

0134

−0.0

032

−0.2

435

1,94

10.

3336

(7.3

5)(−

9.10

)(0

.74)

(−0.

83)

(−1.

72)


to have little impact on trading costs for 15-year TBA-ineligible SPs. It is truethat only 1,941 observations are included in the regression with 15-year TBA-ineligible SPs, but the coefficients on �Q and the interaction between �Q andtrade size remain highly significant in the regression.

To summarize, predictable TBA dollar roll volume spikes that occur beforeexogenously determined settlement dates are associated with lower SP tradingcosts. Unlike TBA trades, SP trades can be settled at any time within a month,and SP traders, unlike dollar roll TBA traders, generally take delivery. Hence,the reasons for a spike in TBA volume before settlement dates do not apply toSP trades. But, before TBA settlement dates SP volume is high and SP tradingcosts are low. This suggests that heavy TBA trading volume increases liquidityfor SPs.

VI. TBA Market Hedging and SP Liquidity

Our results suggest that TBA trading creates liquidity not only for the MBSthat are traded in the TBA market, but also for SPs. In this section, we examineone (but not necessarily the only) way in which TBA trading can make the SPmarket more liquid: by providing a way for dealers to hedge SP positions.Because each SP is unique, it may take some time for dealers to sell them. Thelow trading costs in the TBA market allow dealers to hedge the inventory theyintend to sell in the SP market cheaply. They also have the option, in mostcases, to deliver the SP to close out the TBA position. In addition, hedging withTBA trades rather than, say, Treasury securities, minimizes basis risk for thedealer. The likelihood of prepayment changes with interest rates. This can becaptured at least in part with a hedge from a TBA trade, but not by hedgingwith derivatives on Treasuries.

A. Hedging with TBA Trades

We study dealers’ use of the TBA market to hedge the risk of SP inventory byexamining daily changes in TBA and SP inventory for each dealer i each day.For a given maturity-coupon combination (i.e., 30 years, 3.5%) we calculate thechange in dealer i’s TBA inventory each day using all of the dealer’s tradeswith customers and with other dealers. In calculating the changes in dealerinventory, we aggregate across all issuers (e.g., FNMA) and all settlementdates. For each dealer each day, we also calculate changes in inventory of SPswith the same maturity and coupon. For each dealer i, we then regress dailychanges in TBA inventory on same-day changes in TBA-eligible and TBA-ineligible SP inventory with the same coupon and maturity. That is,

� TBA Invi,c,t = α1i + α2i�TBAEligSP Invi,c,t

+ α3i�TBAIneligSP Invi,c,t + εi,t. (6)


Here, the c subscript refers to maturity-coupon combination c and t refers today t.

A simple way to hedge is to offset a long (short) position in SPs by selling(buying) an equivalent amount of MBS with the same maturity and couponin the TBA market. If a dealer follows this strategy, the coefficients in theregression should equal negative one. If the dealer hedges some SP positionsbut not others, we would expect the coefficients to be negative, but betweennegative one and zero. If a dealer hedges more TBA-eligible SP positions thanTBA-ineligible SP positions, we would expect the α2 coefficients to be closer tonegative one than the α3 coefficients.

Some maturity-coupon combinations are not traded actively throughout thesample period. For example, 30-year 5%, 5.5%, and 6% MBS were only tradedin the TBA market in the early part of the sample period. By 2013, tradingin these maturity-coupon combinations had virtually disappeared from theTBA market. Hence, in estimating the inventory regressions, we use only dayswhen the dealer had a change in either TBA-eligible or TBA-ineligible SPinventory.

Table VII provides statistics on the distribution of regression coefficientsacross dealers. We weight the individual dealer regression coefficients by thenumber of observations in the regression to provide a weighted median. Thenumber of observations in each regression is the number of days when thedealer had positive or negative changes in SP inventory. Hence, these weightedmedians can be interpreted as the proportion of positions hedged. As an exam-ple, for 3% 30-year maturity TBA-eligible SPs, the weighted median is −0.9131,so about 91% of the SP positions taken in that maturity-coupon combinationare hedged. Across maturity-coupon combinations, the weighted-median co-efficients for TBA-eligible SP inventory range from −0.9220 to −0.1130. Acoefficient of −1.0 would imply that SP trades were offset one-for-one withTBA trades. So, our results imply partial hedging. Weighted-median coeffi-cients are closest to zero for 30-year SPs with coupons of 5% or more. TheseMBS were not in production during our sample period and there was relativelylittle TBA trading for these maturity-coupon combinations. This suggests thatdealers are more likely to hedge SP inventory when there is an active TBAmarket.

There are large differences in the number of observations in each individ-ual dealer regression and large differences across dealers in the standard er-rors of the regression coefficients. So, in calculating mean coefficients, we usethe Bayesian framework employed by Panayides (2007) and Bessembinder,Panayides, and Venkataraman (2009). This framework assumes that theestimated α coefficient for each dealer i is distributed,

α̃i| αi ∼ i.i.d. N(αi, s2

i

)and

αi ∼ i.i.d.N(α, σ 2) .

1156 The Journal of Finance R©T

able

VII

Th

eIm

pac

tof

Dai

lyC

han

ges

inS

PIn

ven

tory

onS

ame-

Day

Ch

ange

sin

TB

AIn

ven

tory

ofM

BS

wit

hth

eS

ame

Cou

pon

and

Mat

uri

tyF

orea

chde

aler

ian

dm

atu

rity

-cou

pon

com

bin

atio

nc,

day

tch

ange

sin

TB

Ain

ven

tory

are

regr

esse

don

sam

eda

ych

ange

sin

TB

A-e

ligi

ble

SP

san

dT

BA

Inel

igib

leS

Ps.

Spe

cifi

call

y,w

ere

gres

s

�T

BA

Inv

i,c,

t=

α1i

+α

2i�

TB

AE

ligS

PIn

vi,

c,t

+α

3i�

TB

AIn

elig

SP

Inv

i,c,

t+

ε i,t.

Day

sar

eon

lyin

clu

ded

inth

ere

gres

sion

ifth

ere

was

ach

ange

inT

BA

-eli

gibl

eor

TB

A-i

nel

igib

leS

Ps.

SP

mat

uri

ties

of16

–30

(�15

)ye

ars

are

elig

ible

for

incl

usi

onin

30(1

5)-y

ear

TB

As

and

are

thu

sin

clu

ded

inth

ere

gres

sion

sw

ith

30(1

5)-y

ear

TB

Ain

ven

tory

chan

ges

asth

ede

pen

den

tva

riab

le.

Med

ian

san

dpe

rcen

tile

sar

eof

the

dist

ribu

tion

ofin

divi

dual

deal

erco

effi

cien

ts.T

he

aver

age

αes

tim

ate

acro

ssN

deal

ers

isa

wei

ghte

dav

erag

ew

her

ew

eigh

tsar

ede

term

ined

byth

esa

mpl

eva

rian

ceof

the

deal

erco

effi

cien

t(s

i2)

and

the

sam

ple

vari

ance

ofth

em

axim

um

like

lih

ood

esti

mat

eof

the

aver

age

coef

fici

ent

(σ̃2 m.l.

e.),

that

is,

α̃=

∑ N i=

1α̃

i( s2 i

+σ̃2 m.l.

e.

)∑ N i

=1

1( s2 i

+σ̃2 m.l.

e.

).

Wit

hin

depe

nde

nce

acro

ssde

aler

s,th

eva

rian

ceof

the

aggr

egat

ees

tim

ate

is

Var

(̃α)=

1∑ N i

=1

1( s2 i

+σ̃2 m.l.

e.

).

�T

BA

-Eli

gibl

eS

PIn

ven

tory

(α2)

Coe

ffici

ents

�T

BA

-In

elig

ible

SP

Inve

nto

ry(α

3)

Coe

ffici

ents

Dea

lers

Wei

ghte

dM

edia

nM

ean

t-S

tati

stic

25th

Per

cen

tile

75th

Per

cen

tile

Dea

lers

Wei

ghte

dM

edia

nM

ean

t-S

tati

stic

25th

Per

cen

tile

75th

Per

cen

tile

30Y

R2.

5%35

−0.4

001

−0.4

985

−2.1

4−0

.838

3−0

.154

536

−0.0

215

−0.0

785

−0.2

8−0

.062

70.

0046

30Y

R3.

0%65

−0.9

131

−0.6

454

−3.9

9−0

.980

1−0

.360

962

−0.3

980

−0.3

597

−3.1

1−0

.853

0−0

.002

830

YR

3.5%

82−0

.877

6−0

.708

5−1

9.87

−0.9

664

−0.4

615

76−0

.648

4−0

.397

8−4

.53

−0.8

700

0.00

4130

YR

4.0%

88−0

.799

7−0

.546

6−4

.58

−0.8

912

−0.1

528

76−0

.569

8−0

.292

9−2

.51

−0.7

412

−0.0

032

30Y

R4.

5%86

−0.6

404

−0.4

440

−5.3

7−0

.727

9−0

.149

573

−0.5

178

−0.3

610

−3.4

9−0

.815

70.

0092

30Y

R5.

0%76

−0.3

027

−0.2

594

−1.6

9−0

.358

3−0

.085

166

−0.1

643

−0.2

351

−3.2

5−0

.527

00.

0253

30Y

R5.

5%65

−0.1

899

−0.2

122

−1.6

5−0

.252

8−0

.063

156

−0.0

104

−0.0

895

−0.8

6−0

.170

20.

0194

30Y

R6.

0%60

−0.1

130

−0.1

620

−0.7

0−0

.230

0−0

.023

844

−0.0

046

−0.0

777

−0.3

5−0

.121

10.

0196

15Y

R2.

5%59

−0.9

220

−0.7

012

−5.8

1−0

.950

0−0

.391

633

−1.1

087

−0.7

076

−3.4

1−1

.335

5−0

.199

815

YR

3.0%

71−0

.670

9−0

.484

4−2

.71

−0.7

073

−0.2

280

54−0

.820

3−0

.348

9−2

.40

−0.9

109

0.01

0915

YR

3.5%

78−0

.566

7−0

.432

8−3

.51

−0.7

680

−0.2

242

47−0

.303

0N

AN

A−1

.084

50.

0249

15Y

R4.

0%73

−0.4

423

−0.3

869

−2.9

6−0

.590

3−0

.146

852

−0.2

548

−0.1

751

−1.5

0−0

.656

60.

0660


The average α estimate across N dealers is given by

α̃ =∑N

i = 1α̃i

(s2i +σ̃ 2

m.l.e.)∑Ni = 1

1(s2

i +σ̃ 2m.l.e.)

. (7)

With independence across dealers, the variance of the aggregate estimate isgiven by

Var (α̃) = 1∑Ni = 1

1(s2

i +σ̃ 2m.l.e.)

. (8)

Our individual dealer regressions produce sample variances si22 and si3

2 foreach dealer i. We next use maximum likelihood to jointly estimate the meancoefficient α2, its variance σ 2

2m.l.e., as well as the mean coefficient α3 and its

variance σ 32

m.l.e..The cross-sectional mean and t-statistics for the coefficients from the dealer

hedging regressions are reported in the columns just after the weighted me-dian. The 30-year SPs with yields of 5.0%, 5.5%, and 6.0% trade relatively in-frequently. Mean coefficients on TBA-eligible SP inventory range from −0.2594to −0.1620 for these maturity-coupon combinations. They are not significantlydifferent from zero. The mean coefficients for the other maturity-coupon combi-nations are closer to negative one and are significantly different from zero. Onaverage, for the actively traded TBA-eligible SPs, dealers hedge most of theirSP inventory with TBA trades, and the coefficients are significantly less thanzero.

The 25th and 75th percentiles of individual dealer regression coefficients foreach maturity-coupon combination are reported next. For example, for the 3%30-year maturity TBA-eligible SP, the 25th percentile of coefficients is −0.9801and the 75th percentile is −0.3609. The 75th percentile of dealer coefficientsis negative for all maturity-coupon combinations of TBA-eligible SP inventorychanges, implying that more than 75% of dealer coefficients are negative re-gardless of maturity or coupon. The great majority of dealers seem to hedge atleast some of their positions in TBA-eligible SPs.

Another way to look at dealers’ use of TBA trades to offset SP trades isthat the dealers are selling forward the MBS that they have purchased in theSP market. That is, rather than maintaining an offsetting TBA position, theyintend to deliver the SPs to settle their TBA trades but also have an optionto deliver other MBS. It is not clear that there is a useful distinction betweenhedging and selling their SP positions in the forward market. TBA-ineligibleSPs, however, cannot be delivered to fulfill a TBA trade. For TBA-ineligible SPpositions, offsetting TBA trades would instead be hedges that the dealer wouldexpect to maintain until the SP was sold.

The last six columns of the table report the weighted median, the median, andthe 25th and 75th percentiles of individual dealer regression coefficients for eachmaturity-coupon combination for TBA-ineligible SPs along with the number of


dealers for which the coefficients can be estimated. Weighted-median and meancoefficients are negative for TBA-ineligible SPs, suggesting that dealers hedgeTBA-ineligible SPs. They are not selling them forward as they cannot deliverthem to settle the TBA trades. For most maturity-coupon combinations, bothweighted median and mean coefficients are closer to negative one for TBA-eligible trades than for TBA-ineligible trades. This suggests that the mediandealer is less inclined to hedge TBA-ineligible SP inventory than TBA-eligibleSP inventory, or, alternatively, that some dealers with TBA-eligible positionsexpect to deliver the SPs to settle the TBA trade rather than looking at theTBA trade as one to be reversed when the SP position is closed.

B. Hedging with TBA Trades versus Hedging with Derivatives Tiedto Treasuries

Our evidence suggests that dealers hedge SP positions with offsetting TBAtrades. Dealers could alternatively hedge using derivatives on Treasury secu-rities. Agency MBS, like Treasuries, are default-free securities with prices thatvary inversely with interest rates. For MBS, though, changes in interest ratesaffect the likelihood of prepayment as well as the present value of future cashflows. This suggests that hedging with TBA trades will mean lower basis riskas the value of TBA positions, like the value of SPs, is affected by changesin the likelihood of prepayment. On page 4 of the Securitization Module, TheFederal Housing Finance Agency’s (2013) Examination Manual notes that

While this interest-rate risk could also be hedged with other types ofinstruments, TBAs are a superior hedging instrument because they moredirectly offset the risks associated with a mortgage. The price movementsof Treasury futures, for instance, can diverge significantly from those ofMBS for a number of reasons, including different sensitivities to interestrates.

To examine the effectiveness of hedging, we estimate returns on actual dealerpositions and see how well these returns can be replicated by returns from TBAor Treasury trading. We estimate returns on dealer positions by identifyingcases in which dealers bought and sold the exact same par value of the sameSP. We include only purchases from and sales to customers. We miss cases inwhich a dealer split the position into multiple transactions, or in which an SPposition was liquidated by delivering it to fulfill a TBA trade. We omit positionsthat were opened and closed on the same day. Nevertheless, we identify over100,000 round-trip positions and can use these to compare the effectiveness ofTBA hedging versus hedging with Treasuries.

We categorize SP positions along three dimensions: the maturity of the SPs(�15 years versus 16–30 years), whether they are TBA eligible or ineligible,and whether the position was held for 1–5 days, 6–20 days, 21–60 days, or


more than 60 days. For each group of SP positions, we run the following cross-sectional regression:

�Pi = α0 + α1�Qi + α2�Qi · ln(

Sizei

1,000,000

)+ βiRetHedge

i, j + εi, (9)

where �Pi is the percentage price change in the SP position i, and �Qi takes avalue of one if i was a long position and negative one if it was a short position.We include these variables to adjust for trading costs in the return regressionto enable a more accurate reading of how well the hedges could work. Theregressions are run separately for hedges with TBA positions and hedges withTreasuries. To obtain the TBA hedge return, we first find the last interdealerTBA trade each day with the same coupon and maturity as the SP. If there areTBA trades with multiple settlement dates, we choose the earliest settlementdate. The TBA return over the life of the hedge is the ratio of the TBA price onthe day the position is established to the TBA price on the day the position wasclosed. Hedge returns for Treasuries are the total returns of a CRSP Treasuryindex over the life of the position. We use the CRSP 5-, 7-, and 10-year Treasuryindices.

We report results in Table VIII. To compare how well different hedges work,we compare the adjusted R2’s across the otherwise identical regressions. Ahigher R2 indicates that the hedging variable explains a larger portion of theSP return and therefore provides a hedge with less basis risk. We also reportcoefficients and t-statistics for the hedging variables.

Panel A reports the results for SPs with 16–30 years to maturity. If they areTBA eligible, SPs with these maturities can be delivered to settle TBA trades.There are 15,833 30-year TBA-eligible SP positions that are held for one tofive days. With no hedge variable in the regression, the adjusted R2 is 0.1710.This is the proportion of position returns that can be explained by transactioncosts. When the TBA hedge is included, the coefficient on the hedge return ispositive with a t-statistic of 3.56, but the adjusted R2 increases only slightly to0.1716. For this short time period, the adjusted R2’s are slightly higher for five-and seven-year Treasuries at 0.1717, and slightly lower for 10-year Treasuriesat 0.1714. For this short holding period, none of the potential hedges explainmuch of the SP returns.

The next three rows of Panel A look at SP positions of 30-year TBA-eligibleSPs that were held for 6–20 days, 21–60 days, and more than 60 days. Aswe move to longer and longer holding periods, going from no hedge to a TBAhedge means a larger and larger increase in the adjusted R2. When positionsheld more than 60 days are considered, the adjusted R2 goes from 0.0423 to0.2815. This is not surprising. As the holding period gets longer, the varianceof returns on MBS increases, while the variance of returns from microstruc-ture effects remains constant. Hence, a larger proportion of SP holding-periodreturns can be explained by changes in TBA prices. The adjusted R2 from theregressions with the TBA hedge is much higher than the adjusted R2 from 5-,7-, or 10-year Treasuries. Hence, a hedge with an offsetting TBA trade can


Tab

leV

III

Reg

ress

ion

sof

Ret

urn

sof

Dea

ler

Pos

itio

ns

inS

Ps

onR

etu

rns

ofP

oten

tial

Hed

gin

gIn

stru

men

tsW

eid

enti

fya

deal

erpo

siti

onas

apu

rch

ase

bya

deal

erfr

oma

cust

omer

foll

owed

bya

sale

ofth

esa

me

par

valu

eof

the

sam

eS

Pfr

oma

deal

erto

acu

stom

er.

We

also

incl

ude

posi

tion

sth

atar

ein

itia

ted

wit

ha

sale

and

clos

edby

apu

rch

ase.

For

each

posi

tion

that

ish

eld

for

atle

ast

one

day,

we

esti

mat

e

�P i

=α

0+

α1�

Qi+

α2�

Qi·ln

( Siz

e i1,

000,

000

) +β

iRet

Hed

gei,

j+

ε i,

wh

ere

�P

iis

the

perc

enta

gepr

ice

chan

gein

the

SP

posi

tion

i,�

Qi

take

sa

valu

eof

one

ifi

was

alo

ng

posi

tion

and

neg

ativ

eon

eif

itw

asa

shor

tpo

siti

on.

Sep

arat

ere

gres

sion

sin

clu

deh

oldi

ng

peri

odre

turn

sof

fou

rpo

ten

tial

hed

ges:

TB

Atr

ades

wit

hth

esa

me

mat

uri

tyan

dco

upo

n,

five

-yea

rTr

easu

ryn

otes

,sev

en-y

ear

Trea

sury

not

es,a

nd

10-y

ear

Trea

sury

not

es.A

dju

sted

R2’s

from

the

regr

essi

ons

and

coef

fici

ents

onth

eh

edgi

ng

vari

able

sar

ere

port

edin

the

tabl

e.R

obu

stt-

stat

isti

csar

ein

pare

nth

eses

.

Pan

elA

:Pos

itio

ns

ofS

Ps

wit

h16

–30

Year

sto

Mat

uri

ty

No

Hed

geT

BA

Hed

ge5-

Year

Trea

sury

7-Ye

arTr

easu

ry10

-Yea

rTr

easu

ry

Obs

.A

dj.R

2A

dj.R

2C

oef.

Adj

.R2

Coe

f.A

dj.R

2C

oef.

Adj

.R2

Coe

f.

TB

AE

ligi

ble,

�5

Day

s15

,833

0.17

100.

1716

0.19

870.

1717

0.24

080.

1717

0.13

660.

1714

0.08

26(3

.56)

(3.6

8)(3

.70)

(2.9

1)T

BA

Eli

gibl

e,6–

20D

ays

14,8

570.

1895

0.20

920.

5916

0.19

570.

3402

0.19

570.

1995

0.19

650.

1567

(19.

27)

(10.

71)

(10.

71)

(11.

38)

TB

AE

ligi

ble,

21–6

0D

ays

16,9

920.

1420

0.21

260.

7661

0.17

310.

4382

0.17

540.

2643

0.17

240.

1818

(39.

05)

(25.

30)

(26.

26)

(25.

01)

TB

AE

ligi

ble,

>60

Day

s29

,665

0.04

230.

2815

0.87

020.

0928

0.29

130.

1010

0.19

120.

1010

0.14

31(9

9.39

)(4

0.65

)(4

4.03

)(4

4.03

)T

BA

Inel

igib

le,�

5D

ays

1,98

30.

3160

0.31

560.

0098

0.31

590.

1010

0.31

570.

0226

0.31

570.

0207

(0.1

7)(0

.85)

(0.3

4)(0

.40)

TB

AIn

elig

ible

,6–2

0D

ays

1,51

30.

4487

0.45

840.

2638

0.45

840.

3255

0.45

770.

1790

0.45

790.

1327

(5.2

8)(5

.29)

(5.1

0)(5

.16)

TB

AIn

elig

ible

,21–

60D

ays

1,47

90.

4354

0.48

400.

5081

0.48

050.

5403

0.48

000.

3116

0.47

820.

2261

(11.

83)

(11.

37)

(11.

30)

(11.

05)

TB

AIn

elig

ible

,>60

Day

s2,

239

0.26

330.

4656

0.59

510.

4266

0.54

760.

4203

0.33

020.

4105

0.23

83(2

9.12

)(2

5.26

)(2

4.63

)(2

3.66

)

(Con

tin

ued

)


Tab

leV

III—

Con

tin

ued

Pan

elB

:Pos

itio

ns

ofS

Ps

wit

h15

orF

ewer

Year

sto

Mat

uri

ty

No

Hed

geT

BA

Hed

ge5-

Year

Trea

sury

7-Ye

arTr

easu

ry10

-Yea

rTr

easu

ry

Obs

.A

dj.R

2A

dj.R

2C

oef.

Adj

.R2

Coe

f.A

dj.R

2C

oef.

Adj

.R2

Coe

f.

TB

AE

ligi

ble,

�5

Day

s2,

578

0.17

130.

2041

0.29

440.

1796

0.20

660.

1798

0.11

840.

1794

0.08

87(1

0.35

)(5

.21)

(5.2

8)(5

.16)

TB

AE

ligi

ble,

6–20

Day

s2,

727

0.14

150.

4133

0.60

690.

2595

0.40

330.

2560

0.23

200.

2507

0.17

14(3

5.54

)(2

0.86

)(2

0.50

)(1

9.96

)T

BA

Eli

gibl

e,21

–60

Day

s2,

983

0.04

100.

4419

0.61

960.

2544

0.34

390.

2521

0.20

030.

2297

0.13

95(4

6.28

)(2

9.22

)(2

9.02

)(2

7.03

)T

BA

Eli

gibl

e,>

60D

ays

9,91

10.

0159

0.43

890.

6803

0.21

800.

2723

0.23

840.

1756

0.22

680.

1280

(86.

44)

(50.

61)

(53.

82)

(51.

99)

TB

AIn

elig

ible

,�5

Day

s21

30.

0364

0.03

810.

4713

0.04

02−0

.753

70.

0630

−0.8

182

0.04

03−0

.347

1(1

.17)

(−1.

36)

(−2.

64)

(−1.

36)

TB

AIn

elig

ible

,6–2

0D

ays

930.

0260

0.01

74−0

.308

10.

0154

−0.1

007

0.01

51−0

.019

30.

0153

0.03

06(−

0.46

)(−

0.19

)(−

0.06

)(0

.14)

TB

AIn

elig

ible

,21–

60D

ays

910.

2118

0.35

470.

6787

0.25

070.

3204

0.23

860.

1573

0.23

170.

1030

(4.5

3)(2

.36)

(2.0

2)(1

.81)

TB

AIn

elig

ible

,>60

Day

s15

00.

0287

0.04

650.

2559

0.03

900.

1249

0.03

930.

0774

0.03

280.

0458

(1.9

3)(1

.60)

(1.6

2)(1

.27)


reduce uncertainty far more than a hedge with Treasuries. Notice also thatas we go toward longer holding periods, the coefficient on the TBA return ap-proaches one, indicating that returns on a TBA position get closer to offsettingreturns on the SP position one-for-one. That is not true for the coefficients onTreasury returns.

The next four rows of Panel A report results for positions of TBA-ineligibleSPs. Note that positions without hedges have much higher R2’s than compa-rable TBA-eligible positions without hedges. This is because trading costs arelarger and explain a larger proportion of returns for TBA-ineligible positionsthan for TBA-eligible positions. As we have seen in Table V, TBA-ineligibleSPs typically differ from TBA-eligible SPs in FICO score, loan size, the propor-tion of mortgages refinanced, the proportion of homes that are owner-occupied,and other characteristics. Nevertheless, Panel A shows that, for TBA-ineligibleSPs, adjusted R2’s increase when the return on the TBA hedge is included inthe regression, particularly when the holding period is long. This indicatesthat TBA trades can be useful for hedging TBA-ineligible SPs. The increasein R2’s, however, is much smaller for TBA ineligible SPs than for TBA eligibleSPs. So, for example, when the holding period is 21–60 days, the adjusted R2

increases from 0.1420 to 0.2126, or 0.0706 when the TBA returns are addedto the TBA-eligible regression, but only from 0.4354 to 0.4840, or 0.0486 whenthe TBA returns are added to the TBA-ineligible regression. Finally, noticethat, for TBA-ineligible securities, R2’s with Treasuries as the hedging vari-able are similar to R2’s for the TBA hedge. The advantages of hedging withTBA trades appear to be smaller for TBA-ineligible SPs than for TBA-eligibleSPs.

Panel B of Table VIII reports analogous regressions for TBA-eligible andTBA-ineligible SPs with 15 years or less to maturity. At least for 15-year TBA-eligible SPs, we can draw the same conclusions as in Panel A. Adjusted R2’sare greater when TBA returns are included in the regression than when 5-,7-, or 10-year Treasuries are included. Hence, hedging with TBA trades can bemore effective than hedging with Treasuries. Also, as in Panel A, increases inadjusted R2’s from hedging are greater for longer holding periods. For 15-yearTBA-ineligible SPs, sample sizes are very small and estimates are noisy. Forlong holding periods, TBA hedging appears to work better than hedging withTreasuries, but estimates are too noisy to draw strong conclusions.

C. Prearranged Trades as a Way to Reduce Risk When Dealers Cannot Hedge

If a dealer is unwilling to take the risk of holding inventory, he can actas a broker and find a buyer for SPs that a customer wants to sell ratherthan take the securities into inventory before finding a buyer. These brokeredtrades should play an especially important role for SPs that cannot be easilyhedged with TBA trades. To examine this, we look at all purchases of all SPs—regardless of coupon or maturity—by dealers from customers. We then see ifthe dealer that purchased the SP sold the same par value of the same SP


within five minutes.21 We define purchases that were sold within five minutesas brokered or prearranged trades.22 Trades that are not prearranged are likelyto remain in a dealer’s inventory for days or weeks.

We separate these prearranged trades into those in which the dealer soldto another dealer and those in which the dealer sold to a customer. We viewthese as very different transactions. When a dealer sells to a customer in aprearranged trade, the dealer takes no inventory risk. We hypothesize thatprearranged trades with customers should be especially common for SPs thatare difficult to hedge with TBA sales. In contrast, when a dealer purchases froma customer and immediately resells to another dealer, we find that it is veryunusual for the second dealer to have a prearranged trade with a customer.The second dealer almost always takes the SP into inventory and assumesinventory risk. It is likely that the second dealer would prefer to be able tohedge this risk. A dealer sale to another dealer may occur because the seconddealer specializes in a particular type of SP. It is possible that some of the casesin which a dealer sells to another dealer within five minutes are not actuallyprearranged, but reflect dealers’ knowledge of the positions and interests ofother dealers.23

Panel A of Table IX describes the proportion of trades of various types of SPsthat are prearranged with other dealers and with customers. In total, we have699,263 purchases of SPs by dealers from investors. Of these, 3.99% representtrades that are prearranged with customers and 29.28% are trades that areprearranged with other dealers. When we subtract out the trades that areprearranged with other dealers, we are left with 494,513 trades that dealerselected to handle themselves. Of these, 5.65% were prearranged, and dealerstook the other 94.35% into inventory.

The next two rows of the table report the proportion of prearranged tradesthat are larger than the median size and the proportion that are smaller thanthe median size. Almost half of the small trades are prearranged with anotherdealer, but less than 9% of the large trades are immediately resold to anotherdealer. When we look at the trades that dealers handled themselves, we findthat 7.73% of the trades that are smaller than the median are prearranged withcustomers, but only 4.49% of the large trades. Trades that are prearranged withcustomers tend to be small, but the difference between the proportions of small

21 Harris (2015) examines riskless principal trades in corporate bonds and defines them asoffsetting trades within one minute of each other. These are equivalent to prearranged trades.

22 The choice of five minutes is arbitrary. We tried to pick a time interval that would allowdealers enough time to get back to the other leg of a prearranged trade and execute it, but not solong as to include trades that are not prearranged. When we use two minutes instead of five, theproportion of trades that are prearranged with customers falls from 3.99% to 3.21%. Other resultsare qualitatively the same.

23 Zitzewitz (2010) finds that 46% of dealer trades of corporate bonds with customers are followedby an interdealer transaction within 60 seconds if the trade size is less than $100,000. Only 4.5% ofcustomer trades of over $500,000 are matched with an interdealer transaction within 60 seconds.These figures are upper bounds on the proportion of trades that dealers reverse through offsettinginterdealer trades. Zitzewitz uses data without dealer identities though, so in some cases theinterdealer trade following a trade with a customer will involve different dealers.


Tab

leIX

Pre

arra

nge

dT

rad

esO

bser

vati

ons

incl

ude

allp

urc

has

esof

SP

sby

ade

aler

from

acu

stom

er.T

he

purc

has

eis

prea

rran

ged

wit

ha

cust

omer

(oth

erde

aler

)ift

he

purc

has

ing

deal

erse

lls

the

sam

epa

rva

lue

ofth

eS

Pto

acu

stom

er(o

ther

deal

er)

wit

hin

five

min

ute

sof

the

purc

has

e.T

he

SP

has

am

atch

ing

TB

Aif

ith

asa

mat

uri

tyof

15ye

ars

and

aco

upo

nof

2.5%

,3.0

%,3

.5%

,or

4%,o

rif

ith

asa

mat

uri

tyof

30ye

ars

and

aco

upo

nof

2.5%

,3.0

%,3

.5%

,4%

,4.5

%,5

.0%

,5.

5%,o

r6.

0%.T

he

SP

’syi

eld

mat

ches

TB

Ayi

elds

ifit

is2.

5%,3

.0%

,3.5

%,4

%,4

.5%

,5.0

%,5

.5%

,or

6.0%

.In

the

regr

essi

ons

inP

anel

Bth

ede

pen

den

tva

riab

leeq

ual

son

eif

the

trad

eis

apr

earr

ange

dtr

ade

wit

han

oth

erde

aler

.In

Pan

elC

the

depe

nde

nt

vari

able

equ

als

one

ifth

etr

ade

isa

prea

rran

ged

trad

ew

ith

acu

stom

er.I

nal

lre

gres

sion

s,st

anda

rder

rors

are

clu

ster

edby

deal

er.F

orO

LS

regr

essi

ons,

t-st

atis

tics

are

show

nin

pare

nth

eses

,wh

ile

z-st

atis

tics

are

repo

rted

for

logi

stic

regr

essi

ons.

For

logi

stic

regr

essi

ons,

mar

gin

alef

fect

sar

ein

brac

kets

.Rob

ust

t-st

atis

tics

are

inpa

ren

thes

es.

Pan

elA

:Th

eP

ropo

rtio

nof

SP

Trad

esT

hat

Are

Pre

arra

nge

d

All

Obs

erva

tion

sP

rear

ran

ged

Inte

rdea

ler

Om

itte

d

Obs

erva

tion

sP

rear

ran

ged

wit

hO

ther

Dea

ler

Pre

arra

nge

dw

ith

Cu

stom

erO

bser

vati

ons

Pre

arra

nge

dw

ith

Cu

stom

er

All

699,

263

29.2

8%3.

99%

494,

513

5.65

%Tr

ade

Siz

e�

Med

ian

350,

087

49.5

4%3.

90%

176,

646

7.73

%Tr

ade

Siz

e>

Med

ian

349,

176

8.97

%4.

09%

317,

867

4.49

%T

BA

Eli

gibl

e65

7,97

429

.94%

3.56

%46

0,99

35.

08%

TB

AIn

elig

ible

41,2

8918

.82%

10.8

8%33

,520

13.4

1%M

atch

ing

TB

A11

8,34

014

.08%

3.14

%10

1,67

43.

65%

No

Mat

chin

gT

BA

580,

923

32.3

8%4.

17%

392,

839

6.16

%C

oupo

n=

TB

A59

1,29

629

.79%

2.93

%41

5,12

84.

17%

Cou

pon

�T

BA

107,

967

26.4

7%9.

83%

79,3

8513

.36%

(Con

tin

ued

)


Tab

leIX

—C

onti

nu

ed

Pan

elB

:Det

erm

inan

tsof

Pre

arra

nge

dTr

ades

wit

hO

ther

Dea

lers

Reg

ress

ion

Typ

eD

eale

rF

EIn

terc

ept

Cou

pon

=T

BA

Cou

p.D

eale

rT

BA

Sh

are

TB

AE

ligi

ble

Log

Trad

eS

ize

Obs

.R

2

OL

SYe

s0.

2976

0.04

420.

0172

−0.0

097

698,

770

0.73

41(1

4.01

)(2

.33)

(0.8

7)(−

2.32

)L

ogis

tic

No

1.19

930.

4812

−50.

6842

0.28

23−0

.395

169

8,77

00.

4033

(2.4

1)(2

.75)

(−4.

45)

(0.9

5)(−

5.24

)[0

.037

3][−

4.28

91]

[0.0

216]

[−0.

0334

]L

ogis

tic

No

1.67

660.

0507

−0.0

979

−0.5

225

698,

770

0.23

50(3

.16)

(0.2

6)(−

0.34

)(−

6.21

)[0

.008

7][−

0.01

73]

[−0.

0904

]

Pan

elC

:Det

erm

inan

tsof

Pre

arra

nge

dTr

ades

wit

hC

ust

omer

s

Reg

ress

ion

Typ

eD

eale

rF

EIn

terc

ept

Cou

pon

=T

BA

Cou

p.D

eale

rT

BA

Sh

are

TB

AE

ligi

ble

Log

Trad

eS

ize

Obs

.R

2

OL

SYe

s0.

1095

−0.0

332

−0.0

478

0.00

2549

4,06

60.

2649

(5.8

8)(−

3.02

)(−

3.66

)(1

.93)

Log

isti

cN

o−1

.044

0−0

.731

0−2

3.45

61−0

.687

10.

0039

494,

066

0.13

18(−

2.43

)(−

4.60

)(−

5.06

)(−

3.68

)(0

.07)

[−0.

0264

][−

0.70

85]

[−0.

0277

][0

.000

1]L

ogis

tic

No

−0.8

375

−1.1

140

−0.9

039

−0.0

783

494,

066

0.05

84(−

1.92

)(−

6.64

)(−

5.17

)(−

1.50

)[−

0.06

58]

[−0.

0588

][−

0.00

35]


and large trades that are prearranged is not as dramatic as it is for trades thatare prearranged with other dealers.

The next two rows show the proportion of prearranged trades for TBA-eligibleand TBA-ineligible SPs. TBA-ineligible pools are more difficult to hedge thanthe eligible pools, so we might expect dealers to be especially likely to prear-range trades for ineligible SPs. This is indeed true for trades that are prear-ranged with customers. Looking just at the trades that dealers chose to handlethemselves, we see that dealers prearrange trades with customers for 5.08%of the TBA-eligible SPs and 13.41% of the TBA-ineligible pools. Trades of SPsthat are difficult to hedge are more likely to be prearranged.

The difference in the proportions of TBA-eligible and TBA-ineligible SPstrades that are prearranged with customers suggests that we have understatedthe differences in liquidity between TBA-eligible and TBA-ineligible SPs. Notonly are trading costs much higher for TBA-ineligible SPs than for TBA-eligibleSPs, investors are more likely to have to wait until a counterparty is foundbefore trading. We have no way to determine whether TBA-ineligible tradershave to wait minutes, hours, or days before a counterparty is located.

When we compare TBA-eligible and TBA-ineligible trades, the contrast be-tween trades that are prearranged with customers and trades that are prear-ranged with other dealers is striking. TBA-eligible SPs trades are much morelikely to be prearranged with other dealers than TBA-ineligible trades. This isthe opposite of what we observe for prearranged trades with customers. Dealersprearrange trades of TBA-ineligible SPs with customers because the dealersdo not want to take the TBA-ineligible SPs into inventory. They are difficult tohedge and cannot be delivered to fulfill a TBA trade. Other dealers do not wantto take them into inventory for the same reasons. So, TBA-ineligible SPs areunlikely to be part of a prearranged trade with another dealer.

Finally, in the last two rows of Panel A we compare the proportion of prear-ranged trades of SPs with coupons that match TBA coupons with the proportionof prearranged trades for SPs that do not match TBA coupons. If the SP hada coupon yield that ended in an even percent or half percent and was between2.5% and 6%, we define it as having a matching yield among TBA trades. SPswith coupons that ended in a quarter or three-quarter percent, like 3.25%, orwith coupons greater than 6% are among those that do not match TBA coupons.Many of these SPs have coupon payments that are higher than coupons used inTBA specifications at the time. Because prepayment is a complicated nonlinearfunction of yields, it is difficult to hedge SPs with TBA trades with differentyields.

Because they are more difficult to hedge, we expect that SPs with couponsthat do not match TBA coupons are more likely to be purchased by dealers incombination with a prearranged trade to a customer than are SPs with couponsthat do match TBA coupons. When we omit prearranged interdealer trades andlook only at the trades that dealers chose to handle themselves, we see that13.36% of SP trades are prearranged with customers if the SP has a coupon thatdoes not match TBA coupons. For the SPs with coupons that do match TBAcoupons, the proportion that is prearranged is only 4.17%. In this case, the


difference between prearranged trades with dealers and prearranged tradeswith customers is again striking. Of the trades of SPs with yields that matchTBA coupons, 29.79% are prearranged with other dealers. For trades of SPswith yields that do not match TBA coupons, only 26.47% are prearranged withother dealers. Prearranged trades with other dealers are again less common ifthe SP is difficult to hedge.

In Panel B of Table IX, we report regressions that test more formally how SPcharacteristics influence the decision to prearrange a trade with another dealer.In all regressions in this panel, the dependent variable takes the value of onewhen a trade is prearranged with another dealer. In each regression, standarderrors are clustered by dealer. The first regression is an OLS regression of thedummy for a prearranged trade on a dummy variable that equals one whenthe coupon of the SP matches a coupon rate used for TBA trades, a dummyvariable that equals one if the SP is TBA eligible, and the natural log of thesize of the trade. Dealer fixed effects are included. The coefficient on the yieldmatch dummy variable is 0.0442 with a t-statistic of 2.33. The coefficient on thenatural logarithm of trade size is −0.0097 with a t-statistic of −2.32. Largertrades are less likely to be prearranged with another dealer. The coefficient onTBA eligibility is 0.0172, with a t-statistic of 0.87. The insignificance may besurprising given the univariate differences in prearranged trades with otherdealers for TBA-eligible and TBA-ineligible SPs shown in Panel A. Much of thedifference in prearranged trades with other dealers can be explained by dealeridentity, however, so clustering by dealer and dealer fixed effects renders TBAeligibility insignificant in determining prearranged trades with dealers.

The next two rows report analogous logistic regressions. Here, z-statistics arereported in parentheses under the coefficients, and marginal effects in bracketsunder the z-statistics. The first logistic regression includes dealer TBA marketshare. Its coefficient is negative and highly significant. A dealer with a largemarket share is less likely to prearrange a trade with another dealer. Thecoefficient on the dummy for a coupon equal to a TBA coupon is positive andsignificant. Trade size is also negative and highly significant, indicating thatdealers are less likely to prearrange large trades with other dealers. The secondlogistic regression drops TBA dealer share. The dummy variable for a couponequal to a TBA coupon is now insignificant. As before, the logistic regressionresults indicate that dealers are less likely to prearrange a large trade withanother dealer. On the whole, Panel B suggests that trades that are easy tohedge are prearranged with other dealers, but this evidence is weak.

Panel C reports similar regressions, but now the dependent variable is adummy variable that equals one if the trade is prearranged with a customer.In each regression, we use only trades that were not prearranged with anotherdealer. The decision to prearrange a trade with a customer differs from thedecision to prearrange a trade with another dealer in important ways, so theresults here are very different from those of Panel B.

The first regression is an OLS regression and the last two are logistic re-gressions. Each has standard errors clustered by dealer. The regression resultsfor prearranged trades with customers in Panel C are very different from the


regression results for prearranged trades with other dealers in Panel B. If anSP has a coupon that matches a TBA coupon, Panel B shows that the trade ismore likely to be prearranged with another dealer while Panel C shows it isless likely to be prearranged with a customer. If an SP has a TBA-matchingcoupon, dealers can hedge it. They are more willing to take the SP into inven-tory and other dealers are also more willing to take it into inventory. If an SP isTBA eligible, Panel C shows that dealers are less likely to prearrange a tradewith a customer. Dealers do not mind taking it into inventory because it canbe hedged more easily and can be delivered through a TBA trade. In contrast,as Panel B shows, TBA eligibility seems to have little impact on determiningwhether a trade is prearranged with another dealer.

The results in Panel C suggest that dealers are more likely to take an SPinto inventory if it can be hedged with a TBA trade. If it cannot be hedged, theinvestor who owns the SP may have to bear risk himself until the dealer canfind a buyer. This suggests a benefit to TBA trading that is not captured bytrading costs. If similar MBS trade in the TBA market, an investor does notneed to hold an unwanted SP while a dealer searches for a buyer.

Our findings indicate that dealers hedge SP positions with TBA trades andare reluctant to hold inventory that they cannot hedge. We do not, however,mean to imply that the only reason TBA trading improves SP liquidity isthat it provides a superior hedging option. We can posit other ways in whichTBA trading could contribute to SP liquidity. For instance, it could providebenchmark prices for SPs.24 MBS trade frequently in the TBA market withlow trading costs and minimal price impact. Less frequently traded SPs can bepriced off of the TBA trades. We hope to explore this possibility in future work.

VII. Conclusions

The secondary market for agency MBS is among the largest and most liquidsecurities markets in the world. In a way, this is surprising because each ofthe tens of thousands of MBS is a claim on the cash flows of a different set ofmortgages and is therefore unique. An important reason why this market is soliquid is the existence of TBA trading, whereby MBS are traded in a forwardmarket on a cheapest-to-deliver basis. TBA trading takes thousands of thinlytraded MBS and combines them into a few thickly traded forward contracts.

We provide evidence that TBA eligibility makes SPs more liquid. SPs can beineligible for TBA trading if they are not fully amortizing, include a prepaymentpenalty, or have an LTV ratios above 1.05. In particular, we present evidencethat TBA eligibility itself, not just SP characteristics associated with eligibility,leads to greater SP liquidity. We show that trading costs in general decreasewith LTV ratios, but increase abruptly when the 1.05 threshold is crossed. Inaddition, we use characteristics of the SPs to predict the likelihood that the

24 Duffie, Dworczak, and Zhu (2014) analyze the use of benchmarks in over-the-counter markets.They demonstrate that benchmarks can lower search costs, increase trading volume, and generatemore efficient trade matching between dealers and investors.


SP is TBA eligible. After adjusting for the estimated probability that the SP isTBA eligible, actual eligibility decreases trading costs significantly.

We also present evidence that the existence of parallel trading in the TBAmarket increases liquidity for the MBS that are traded individually in the SPmarket. For a given issuer and maturity, all TBA trades settle on the sameday of the month. Most traders do not want to take delivery or deliver on theTBA trades they have made, and so will reverse their positions in the few daysjust before the settlement date. This leads to high TBA volume, particularlyfrom dollar rolls, just prior to the settlement date. This high volume is easilypredicted well in advance of the actual trading. SPs can settle any day of themonth. Nevertheless, trading costs for SPs decrease significantly on days ofpredictable high TBA volume. Exogenous increases in TBA trading volumelower SP trading costs.

One way in which TBA trading can increase the liquidity of the SP marketis by providing a way for dealers to hedge their SP inventory positions. Consis-tent with this explanation, we find that when we regress dealers’ daily changesin TBA inventory on the same day, changes in SP inventory are consistentlynegative and usually between −0.5 and −1.0. Dealers offset most of their posi-tions in SPs with TBA trades. Coefficients are smaller for TBA-ineligible SPs,indicating that a smaller proportion of these trades are hedged.

In some cases, dealers act as brokers and find a purchaser for an SP beforebuying it from a customer. This appears in the data as dealer purchases followedby offsetting sales within five minutes. We find that prearranged or brokeredtrades are most common for the SPs that are least likely to be hedged with TBAtrades, that is, TBA-ineligible SPs, or SPs with coupons that are unmatchedby TBA coupons.

Corporate and municipal bonds trade in relatively illiquid over-the-countermarkets. Parallel trading in the securities themselves and a forward contracton a generic security may increase the liquidity of those markets. Specific highlyrated municipal bonds, for example, could trade in parallel with forward con-tracts on, say, 30-year, AA-rated, general obligation New York municipals. Ourresults suggest that a forward contract of this type could lower trading costs forthe municipal bonds themselves by converting thin markets for multiple mu-nicipals to thick markets in a handful of forward contracts. Forward contractswould also allow dealers to hedge inventory. A forward contract on municipalbonds could also lower the risk to underwriters by allowing them to hedgewhile selling a bond issue. It is true that municipals, unlike corporates, havestate clienteles. Separate contracts would be needed for different large statesand forward markets may not be possible for small-state municipals. On theother hand, within states, highly rated municipals are very good substitutes foreach other because they are very unlikely to default.25 The unique structure ofparallel trading in SPs and the forward TBA market looks like it could enhanceliquidity in other markets as well.

25 The one-year default rate for all rated municipals over 1970–2011 was 0.01%. See Moody’sInvestors Services (2012).


Initial submission: December 24, 2014; Accepted: July 25, 2016Editors: Bruno Biais, Michael R. Roberts, and Kenneth J. Singleton

REFERENCES

Admati, Anat, and Paul Pfleiderer, 1988, A theory of intraday patterns: Volume and price variabil-ity, Review of Financial Studies 1, 3–40.

Atanasov, Vladimir, and John Merrick, 2012, Liquidity and value in the deep vs. shallow ends ofmortgage-backed securities pools, Working paper, College of William and Mary.

Atanasov, Vladimir, and John Merrick, 2014, Why do dealers buy high and sell low? An analysisof persistent crossing in extremely segmented markets, Review of Finance (forthcoming).

Bessembinder, Hendrik, William Maxwell, and Kumar Venkataraman, 2013, Trading activity andtransaction costs in structured credit products, Financial Analysts Journal 69, 55–68.

Bessembinder, Hendrik, Marios Panayides, and Kumar Venkataraman, 2009, Hidden liquidity:An analysis of order exposure strategies in electronic stock markets, Journal of FinancialEconomics 94, 361–383.

Duffie, Darrell, Piotr Dworczak, and Haoxiang Zhu, 2014, Benchmarks in search markets, Workingpaper, Stanford University.

Federal Housing Finance Agency, 2013, Securitization Module, Examination Manual. Availableat https://www.fhfa.gov/SupervisionRegulation/Documents/Securitizations Module FinalVersion 1.0 508.pdf.

Friewald, Nils, Ranier Jankowitsch, and Marti Subrahmanyam, 2014, Transparency and liquidityin the structured product market, Working paper, WU Vienna University of Economics andBusiness.

Harris, Larry, 2015, Transaction costs, trade throughs, and riskless principal trading in corporatebonds, Working paper, University of Southern California.

Lee, David, and Thomas Lemieux, 2010, Regression discontinuity designs in economics, Journalof Economic Literature 48, 281–355.

Moody’s Investors Services, 2012, U.S. Municipal Bond Defaults and Recoveries, 1970–2011. Avail-able at http://www.nhhefa.com/documents/moodysMunicipalDefaultStudy1970-2011.pdf.

Panayides, Marios, 2007, Affirmative obligations and market making with inventory, Journal ofFinancial Economics 86, 513–542.

Securities Industry and Financial Markets Association, 2013, Uniform Practices Manual,Chapter 8, Standard Requirements for Delivery on Settlements of Fannie Mae, Fred-die Mac, and Ginnie Mae Securities. Available at http://www.sifma.org/uploadedfiles/services/standard forms and documentation/ch08.pdf?n=42389.

Spatt, Chester, 2004, Frictions in the bonds market, Keynote speech for the Second MTS Conferenceon Financial Markets: The Organization and Performance of Fixed-Income Markets, availableat http://www.sec.gov/news/speech/spch121604cs.htm.

Superior Court of Washington for King County, 2010, Federal Home Loan Bank of Seattle v.Goldman Sachs, Case Number 09-2-46349-2 SEA.

Vickery, James, and Joshua Wright, 2013, TBA trading and liquidity in the agency MBSmarket, FRBNY Economic Policy Review, May, 1–16. Available at https://www.newyorkfed.org/medialibrary/media/research/epr/2013/EPRvol19no1.pdf.

Zitzewitz, Eric, 2010, Paired corporate bond trades, Working paper, Dartmouth College.

http://dx.doi.org/10.1093/rfs/1.1.3

http://dx.doi.org/10.2469/faj.v69.n6.2

http://dx.doi.org/10.1016/j.jfineco.2009.02.001


https://www.fhfa.gov/SupervisionRegulation/Documents/Securitizations_Module_Final_Version_1.0_508.pdf

https://www.fhfa.gov/SupervisionRegulation/Documents/Securitizations_Module_Final_Version_1.0_508.pdf

http://dx.doi.org/10.1257/jel.48.2.281

http://dx.doi.org/10.1257/jel.48.2.281

http://www.nhhefa.com/documents/moodysMunicipalDefaultStudy1970-2011.pdf



http://www.sifma.org/uploadedfiles/services/standard_forms_and_documentation/ch08.pdf?n=42389

http://www.sifma.org/uploadedfiles/services/standard_forms_and_documentation/ch08.pdf?n=42389

http://www.sec.gov/news/speech/spch121604cs.htm

https://www.newyorkfed.org/medialibrary/media/research/epr/2013/EPRvol19no1.pdf

https://www.newyorkfed.org/medialibrary/media/research/epr/2013/EPRvol19no1.pdf