The Real Effects of Secondary Market Trading Structure: Evidence from the Mortgage Market

Abstract

By allowing different agency mortgage-backed securities (MBS) to be traded based on limited characteristics, the to-be-announced (TBA) market generates liquidity and benefits the MBS market broadly. We quantify effects of the TBA structure on mortgage borrowers. Exploiting discontinuities in TBA eligibility, we estimate that TBA eligibility reduces mortgage rates by 7 to 28 basis points. The TBA eligibility benefit is larger for mortgages with higher expected prepayments. We also find that TBA eligibility affects refinancing, which has implications for monetary policy transmission. Our finding is relevant for housing policies, such as housing finance reforms and uniform MBS.

Authors have furnished an Internet Appendix, which is available on the Oxford University Press Web site next to the link to the final published paper online.

The market for agency mortgage-backed securities (MBS) is one of the most active and most liquid fixed-income markets. Its distinct feature is the to-be-announced (TBA) forward contract, through which 90|$\%$| of agency MBS trading is done. In a typical TBA trade, parties do not specify the CUSIP but agree only on limited security characteristics. This TBA trading structure generates liquidity by concentrating trading of MBS with heterogeneous prepayment speeds into a handful of thickly traded TBA contracts although it leads to cheapest-to-deliver pricing (Li and Song, 2020).

A natural policy-relevant question is how much mortgage borrowers benefit from the TBA market liquidity. If investors benefit from a liquid TBA market, agency MBS prices should increase, which may reduce mortgage rates. As a result, the TBA market liquidity may have significant effects on the real economy because a majority of purchase and refinance mortgages are securitized into agency MBS. In the ongoing discussion about how to reform the U.S. housing market, it is often argued that the TBA market structure should be preserved through an explicit federal backstop. Those in favor of this measure have cited the TBA market’s benefit for mortgage borrowers.¹ However, to the best of our knowledge, no studies have provided an estimate of the effect of TBA trading for mortgage borrowers. Most existing studies on the TBA market, for example, Gao, Schultz, and Song (2017), Schultz and Song (2019), and Song and Zhu (2018), focus on secondary market activities. More broadly, we study how financial market liquidity benefits borrowers, and very few studies have investigated how individuals benefit from financial market liquidity.

In this paper, we quantify the impact of the TBA trading structure on mortgage rates and demand for mortgages and study how the impact varies with prepayment characteristics. We use the facts that not all agency MBS are eligible to be delivered for TBA settlement and that TBA-ineligible MBS must be traded in the much less liquid specified-pool (SP) market, where the individual CUSIP is specified at the time of trade. We find that TBA eligibility reduces mortgage rates by 7 to 28 basis points (bps), with the effect being larger for mortgages with faster expected prepayments. Our findings suggest that the option value to trade MBS in the more-liquid TBA market is passed down to mortgage borrowers. Further, the option value is less valuable for loans with low prepayments because TBA-eligible MBS with low prepayments typically trade in the SP market to avoid the cheapest-to-deliver TBA price (Huh and Kim, 2021). Lastly, we show that TBA eligibility also affects mortgage refinancing and has implications for monetary policy transmission.

Our identification strategy exploits two cutoffs that determine the probability that a loan is included in a TBA-eligible MBS. The first cutoff is the national conforming loan limit (CLL). If an agency MBS contains more than 10|$\%$| of its pool value in high-balance loans—Fannie Mae or Freddie Mac loans with sizes above the national CLL—the MBS is ineligible for TBA delivery but still guaranteed by the government-sponsored enterprises (GSEs). The second cutoff is a loan-to-value ratio (LTV) of 105|$\%$| for refinance loans originated under the Home Affordable Refinance Program (HARP). A TBA-eligible MBS is not allowed to include any loans with an LTV greater than 105|$\%$|⁠. We estimate the impact of TBA eligibility for loans around the two cutoffs separately. Moreover, by comparing the estimates from the two cutoffs, we can infer how the TBA benefit varies with prepayment speeds, because loans around the national CLL have higher prepayment than most GSE loans, and loans with LTVs around 105|$\%$| have lower prepayment than most GSE loans.

Obtaining clean estimates of TBA eligibility benefit is challenging. Importantly, sorting of borrowers into above and below each cutoff may not be random if loans in TBA-eligible MBS have lower rates. We address this issue with instrumental variables. In the test using the CLL cutoff, we analyze variations among mortgages for homes with appraised values around 125|$\%$| of the national CLL, similar to Kaufman (2014), Adelino, Schoar, and Severino (2012), Fuster and Vickery (2014), and DeFusco and Paciorek (2017). This strategy utilizes the fact that GSEs typically require the LTV to be 80|$\%$| or lower without private mortgage insurance. In the test using the LTV cutoff, we analyze variations among loans whose new maximum LTVs allowed under HARP is around 105|$\%$|⁠. This novel strategy utilizes the fact that Freddie Mac limits how much the loan balance could increase under HARP.

We find that TBA eligibility reduces mortgage rates by 28 bps for loans around the national CLL and 7 bps for loans with LTVs around 105|$\%$|⁠.² The large difference in the estimates between the two cutoffs is due to the difference in loan values driven by expected prepayment speeds. Since credit risk is guaranteed for agency MBS, their values are mostly determined by expected prepayment, with MBS with higher expected prepayments having lower values. Because TBA-eligible MBS with higher values (i.e., lower prepayments) typically trade in the SP market to avoid the cheapest-to-deliver TBA price, TBA eligibility would be less valuable for MBS backed by loans with LTVs around 105|$\%$|⁠. In fact, the estimates from the two cutoffs provide estimates for the upper and lower bounds of the value of TBA eligibility, because loans around the national CLL and loans with LTVs around 105|$\%$| are near the upper and lower end of the prepayment distribution, respectively. We also show that the difference in the estimates between the two cutoffs is not driven by differences in other characteristics between the estimation samples used for the two cutoffs. Additional analysis on the effect of TBA eligibility on MBS prices hints that the main mechanism through which TBA eligibility reduces mortgage rates is by providing the option to trade MBS in the liquid TBA market, which raises MBS prices.

Lastly, we investigate the degree to which TBA eligibility affects refinancing decisions of borrowers with remaining loan balances around the national CLL. We find that the monthly probability of rate refinancing increases discontinuously by 1 percentage point (88|$\%$| of the unconditional mean) when the remaining mortgage balance reaches the national CLL from above. Given the abundant evidence for an increased consumption following mortgage refinancing in the literature, our result suggests that the distinct trading structure of the agency MBS market also matters for monetary policy transmission and real economic outcomes.^3,4

In summary, we show that the TBA trading structure lowers mortgage rates, especially for borrowers that typically prepay faster. Because agency mortgages finance a majority of home purchases, our finding implies that the TBA market will eventually affect home-purchase decisions for many households. We find that the TBA market structure also affects refinancing and likely monetary policy transmission. An important policy implication is that preserving a liquid TBA market will benefit mortgage borrowers in the context of the housing finance reform.

Our finding has important policy implications even outside the housing market. Regulators often implement policies based on the view that improvement in the secondary financial market will eventually benefit borrowers in the primary market. For instance, the Securities and Exchange Commission conducted a pilot program that changed the minimum tick sizes of small stocks to bring liquidity to these stocks and ultimately lower the costs of funding for small businesses.⁵ Our paper supports the view that financial markets and their trading structure matter for funding cost even for borrowers who do not directly participate in financial markets.

Our paper makes a number of contributions to the literature. First, we connect secondary market liquidity to the pricing in the primary market. Secondary market liquidity is an important topic because it affects the cost of funding in the primary market, which subsequently affects the real economy. However, only a handful of papers attempt to establish this connection. Both Brugler, Comerton-Forde, and Martin (2021) and Davis, Maslar, and Roseman (2020) show that the introduction of post-trade transparency in the secondary corporate bond market, which decreased trading costs, has decreased cost of capital in the primary market. While the introduction of transparency was done in phases, trading costs do not change sharply on introduction dates, and Asquith, Covert, and Pathak (2019) find that trading volume decreases for high-yield bonds. These effects potentially make it difficult to identify the effect on primary market pricing precisely. Brugler, Comerton-Forde, and Hendershott (2021) study a rule change in NASDAQ that moved the market from a dealer-oriented market toward a more centralized one and find that it decreased the underpricing of seasoned equity offerings. Relative to these papers, our data and identification strategy allow us to cleanly identify the effect of secondary market liquidity on the primary market by comparing otherwise very similar mortgages within narrow bounds around the cutoffs.

Second, we add to the literature on the trading structure and liquidity of the agency MBS and TBA market. Vickery and Wright (2013) provide a comprehensive overview of the institutional details of the TBA market. They also provide suggestive evidence that TBA eligibility affects mortgage rates by looking at the difference in mortgage rates between high-balance loans and conforming loans. We advance the understanding of the impact of the TBA market on the primary market in three ways. First, we use discontinuity-based research designs in combination of instrumental variables to tease out the impact of TBA eligibility from unobserved heterogeneity. Second, we also quantify how the benefit of TBA eligibility varies with prepayment characteristics by estimating the benefit for loans with an LTV around 105|$\%$| using a novel identification strategy. Third, we correctly quantify the benefit of TBA eligibility with the data that link each GSE loan to the CUSIP of the agency MBS into which the loan is pooled.

Other studies in this literature focus only on the secondary market. Bessembinder, Maxwell, and Venkataraman (2013) examine trading costs in structured credit products and find that trading costs in TBA trades are very small (1 bp) compared with that of SP trades (40 bps). Gao, Schultz, and Song (2017) find that TBA eligibility also affects trading costs for SP trades. Song and Zhu (2018) study mortgage dollar rolls, and Schultz and Song (2019) study the impact of post-trade transparency in the TBA market. Fusari et al. (2020) study the asset pricing implications of the TBA market. Li and Song (2020) theoretically examine how the TBA market structure can improve liquidity. Huh and Kim (2021) study how the TBA market structure lenders’ MBS pooling and its subsequent impact on market quality. We contribute to this literature by studying the link between the TBA market structure and the primary mortgage market.

Third, we quantify the liquidity benefit of the TBA market separately from other benefits of GSE securitization. A number of papers (Passmore, Sherlund, and Burgess, 2005; Kaufman, 2014; DeFusco and Paciorek, 2017) estimate the spread between jumbo and conforming loans (“jumbo-conforming spread”), which reflects the benefit of securitizing loans through the GSEs. This jumbo-conforming spread includes both the liquidity benefit of the TBA market and other benefits of GSE securitization, such as credit risk guarantees. Disentangling the TBA market benefit from other benefits is useful from a policy perspective because many proposals for housing finance reform desire to keep the TBA market liquid because of its large benefits for mortgage borrowers. Moreover, the aforementioned papers estimate the benefit of GSE securitization only using loans near the CLL. We also contribute to the literature by showing that the TBA liquidity benefit, which is part of the GSE securitization benefit, varies with prepayment characteristics with our analysis of loans near the LTV of 105|$\%$| using a novel identification strategy. This strategy also could be applied to similar settings involving LTV-based cutoffs for refinance loans.

Lastly, a few papers study the pass-through of secondary market MBS yields to primary market mortgage rates. Scharfstein and Sunderam (2017) find that lenders’ market power affects pass-through. Fuster, Lo, and Willen (2017) show that the pass-through is almost perfect at daily frequencies but the price of intermediation depends on capacity constraints at monthly frequencies. Fuster et al. (2013) shows that an increase in lenders’ profit contributed to a rising gap between primary and secondary mortgage rates. We contribute to this literature by studying how the TBA liquidity benefit gets passed through to the primary market.

1. Institutional Details

1.1 TBA eligibility

A majority of mortgages in the United States are securitized and packaged into agency MBS. Agency MBS are backed by mortgages in pools guaranteed by Fannie Mae, Freddie Mac, or Ginnie Mae. In this paper, we focus on MBS guaranteed by the GSEs (Fannie Mae and Freddie Mac).

These agencies guarantee credit risks of agency MBS; thus, prepayments matter the most in the pricing of MBS in the cross-section. When MBS trades at a premium (i.e., above par), investors dislike prepayment. Given that newly issued MBS generally trade at a premium, MBS with higher expected prepayments are valued lower at issuance.⁶

TBA trade is a forward contract in which two parties agree on a price today for a future delivery of agency MBS. Moreover, instead of agreeing on a specific CUSIP at the onset of the trade, parties only agree on six general parameters: agency (Freddie Mac or Fannie Mae), coupon, maturity, price, par amount, and settlement date. The seller is required to notify the buyer of the specific CUSIP(s) that he will deliver only 48 hours before the delivery date. Given the large number of individual CUSIPs in the agency MBS market, this structure concentrates trading into a handful of TBAs and generates liquidity. According to Vickery and Wright (2013), TBA trades account for 90|$\%$| of trading volume in the agency MBS market.

However, not all MBS are allowed to be delivered for TBA settlement, and TBA-ineligible MBS must be traded in the much less liquid SP market, where the individual CUSIP is specified at the time of trade. An MBS is not eligible for TBA settlement for largely three reasons. First, MBS that include any loans with original LTV greater than 105|$\%$| are not TBA eligible. MBS with very high LTVs were originated under the HARP.⁷ Second, MBS that have more than 10|$\%$| of the pool value at the time of issuance in high-balance loans are not eligible to be delivered for TBA settlement. The GSEs are only allowed to purchase “conforming” loans that are not greater than the CLL. High-balance loans refer to GSE mortgages with loan size greater than the national CLL, but not greater than the county-specific high-cost CLL, which became available in 2008 for certain counties with high home prices.⁸ Lastly, MBS with greater than 15|$\%$| of pool value in loans with other nonstandard features, such as co-op share loans, relocation loans, and loans with significant interest rate buydowns are not eligible for TBA delivery.

Figure 1 shows the evolution of dollar-weighted shares of loans (among 30-year fixed-rate mortgages sold to the GSEs) included in new agency MBS that are not eligible for TBA settlement. We categorize TBA-ineligible MBS into three broad groups: high-balance MBS, high-LTV MBS, and other TBA-ineligible MBS. High-balance MBS consist of high-balance loans only.⁹ High-LTV MBS consist of HARP loans with LTVs greater than 105|$\%$|⁠. Other TBA-ineligible MBS consist of loans with other nonstandard features.

We point out two main takeaways from Figure 1. First, the total share of TBA-ineligible MBS is not negligible during our sample period. In early 2009, the share was close to zero, but it grew substantially in the next few years, reaching close to 20|$\%$| in mid-2012. The increase was in part due to high volume of refinancing into high-balance and high-LTV loans, driven by low mortgage rates. Second, almost all loans included in TBA-ineligible MBS are either high-balance loans or high-LTV loans, and the shares of the two loan types vary over time. Since very few loans are TBA ineligible due to other nonstandard features, we focus on the cutoffs of the LTV of 105|$\%$| and the national CLL in our empirical analysis.

1.2 Specified-pool market

In an SP trade, parties agree and trade on a specific CUSIP, and each CUSIP is thinly traded. As a result, an SP trade usually has a higher trading cost than a TBA trade. Bessembinder, Maxwell, and Venkataraman (2013) find that the trading cost of TBA and SP trades are 1 and 40 bps, respectively. TBA-ineligible MBS must be traded in the SP market.

TBA-eligible MBS may also trade in the SP market; they often do so especially when the value of the MBS is high, that is, when the expected prepayment is low compared with other TBA-eligible MBS (Huh and Kim, 2021). Such selection arises because of cheapest-to-deliver pricing (adverse selection) in TBA trades. Because TBA sellers choose which MBS to deliver, they would usually deliver MBS with low values. Buyers would rationally expect this incentive and thus would not be willing to pay more than the price for the cheapest MBS. As a result, sellers would want to sell high-value TBA-eligible MBS via SP to receive their full prices despite higher trading costs.

TBA eligibility still protects MBS sellers from sudden changes in market environments that make selling high-value MBS in the SP market unprofitable (Gao, Schultz, and Song, 2017), and lenders may pass down this TBA option value to borrowers through lower mortgage rates. Naturally, this option value should be lower for high-value MBS, and the impact of TBA eligibility on mortgage rates also will be smaller for loans with lower expected prepayments.

2. Data and Summary Statistics

2.1 Data description

We use multiple data sources. First, eMBS provides information on agency MBS and their underlying mortgages. We obtain information on MBS-level characteristics, such as agency and TBA eligibility, and loan-level characteristics. Importantly, the data link each GSE loan with the CUSIP of the agency MBS that the loan is pooled into. This information is crucial in accurately estimating the benefit of TBA eligibility on mortgage rates using the national CLL cutoff. Not all high-balance loans are included in TBA-ineligible MBS because TBA-eligible MBS can include such loans up to 10|$\%$| of its pool value. Thus, the difference in loan rates between high-balance loans and conforming loans is likely much smaller than the benefit of TBA eligibility.

The second is the loan-level Black Knight McDash Data (henceforth, McDash in short). In some of the main analyses, we supplement eMBS with the McDash data by matching the two data sets at the loan level.

The third is loan-level performance data on HARP loans from Fannie Mae and Freddie Mac. They also link each HARP mortgage with its original loan. This linkage is crucial for our empirical test that uses the LTV cutoff.

The fourth data set we use is Equifax Credit Risk Insight Servicing and Black Knight McDash Data (CRISM), which links loan-level mortgage data to each borrower’s credit records from Equifax. We use CRISM to analyze the impact on refinancing in Section 4 and subsequent durable consumption through new auto loan originations in the Internet Appendix.

2.2 Sample selection

Our empirical analyses exploit the fact that TBA eligibility changes discontinuously at the national CLL and the LTV of 105|$\%$|⁠. In all of our analyses of the impact on mortgage rates, we only keep 30-year fixed-rate mortgages (FRMs) originated in or after 2009 that are sold to the GSEs and further restrict the sample to loans originated for owner-occupied single-family houses without second mortgages to keep the sample relatively homogeneous. We also use different subsamples for the two cutoffs to only compare loans near each cutoff as detailed in the following paragraphs.

2.2.1 National CLL

For the analysis using the national CLL cutoff, we match eMBS and McDash in order to supplement eMBS with three additional variables: precise home appraisal value, ZIP code of the property, and prepayments. First, eMBS data only allow us to approximate the appraisal value from the original loan amount and integer-valued LTV. However, this approximation is biased in such a way that the approximated appraisal values bunch at 125|$\%$| of the national CLL although the true appraisal value does not exhibit bunching. Not having the bunching is critical to our identification, which will be discussed in Section 3.1.1. Second, the property ZIP code is available in McDash, whereas eMBS only has state information. Third, we use prepayment information from McDash because eMBS provides prepayment information only for Freddie Mac loans in our sample.

We combine the two data sets by finding a pair of an eMBS loan and a McDash loan that have the following variables in common: agency (Fannie Mae or Freddie Mac), state, origination year-month, original loan amount, mortgage rate, integer-valued LTV, and debt-to-income ratio (DTI). More than 99|$\%$| of eMBS loans that are matched to any McDash loans have a unique match.¹⁰

We impose the following additional sample selection criteria to keep the sample relatively homogeneous. First, we only keep loans for houses in high-cost counties whose CLLs are at least $50,000 above the national CLL ($417,000) to exclude regions without high-balance loans. Second, we only keep purchase loans because our identification strategy, which is discussed later, is most relevant for liquidity-constrained borrowers.¹¹ Third, we exclude loans with mortgage insurance (original LTV greater than 80). Fourth, we exclude loans originated after 2012; thus, the sample period for this analysis is from 2009 to 2012. Since 2013, mortgage lenders increasingly held high-balance loans on their portfolio or privately securitized them instead of selling to the GSEs, which could lead to a selection issue. We discuss this issue further in Section 3.1.5. Lastly, we exclude Fannie Mae loans originated before November 2011 because Fannie Mae loans paid off before January 2012 are not reported in eMBS. We still include loans originated in November or December 2011 because an extremely small fraction of these loans would have been paid off before 2012.

2.2.2 LTV of 105|$\%$|

For the analysis using the LTV cutoff, we mainly use the HARP data set. We impose the following sample selection criteria in addition to the common sample selection criteria. First, we only keep HARP loans because a vast majority of loans with LTVs close to 105|$\%$| are HARP loans. Thus, the estimated effect of TBA eligibility on mortgage rates with this sample does not reflect differences in mortgage rates between HARP and non-HARP loans. Second, we only keep loans securitized by Freddie Mac as our identification strategy relies on restrictions only imposed by Freddie Mac. Last, we only keep HARP 2.0 loans (HARP loans originated in the period from 2012 to 2016) since there were fairly few loans with LTV greater than 105|$\%$| in earlier years.¹²

2.3 Summary statistics

2.3.1 Prepayment and loan characteristics

An important goal of this paper is to examine how the benefit of TBA liquidity varies with prepayments. Here, we show that loans around the two cutoffs used for the estimation—the national CLL and the LTV of 105|$\%$|—are near the opposite ends of the prepayment distribution. Eventually, the difference in the estimated impact of TBA liquidity on mortgage rates will inform us about how the TBA liquidity benefit varies with prepayments.

Figure 2 plots the relationship between ex post prepayments and loan amounts (panel A) and between ex post prepayments and LTVs (panel B). Ex post prepayments are measured in terms of whether a loan was paid off completely by 36 months after the loan origination, but the patterns remain qualitatively unchanged when we use other loan ages. To control for the potential interactive effect of mortgage rates and interest rate paths on prepayments, we consider residual prepayments, which are calculated by removing variation accounted for by the origination year-month |$\times$| mortgage rate bin fixed effects and state and lender fixed effects.

Panel A shows that prepayment and loan amounts are positively correlated. Larger loans prepay faster because the dollar gains from refinancing into lower interest rates are greater with larger loans. Panel B shows that prepayment and LTVs are negatively correlated. It usually takes longer for borrowers with higher initial LTVs to build equity in their houses to be able to refinance later. Thus, such borrowers prepay slowly. Overall, given that loans around the national CLL are larger than most GSE loans and that loans with LTVs around 105|$\%$| have higher LTVs than most GSE loans, loans with LTVs around 105|$\%$| are likely around the lower end and loans with sizes around the national CLL are around the upper end in terms of expected prepayments.

The way in which lenders typically pool loans into MBS and trade them can lead to differences in TBA-eligibility benefits for the two loan types. Lenders often pool together loans with LTVs right below 105|$\%$| to create TBA-eligible MBS with low prepayments and typically sell them via SP. In contrast, TBA-eligible MBS that include loans near the national CLL have usually the highest prepayments and are much more likely traded via TBA (Huh and Kim, 2021). Thus, the option value to trade in the TBA market should vary depending on prepayment characteristics, and its impact on mortgage rates should be smaller for loans with LTVs around 105. In Section 3, we quantify the impact on mortgage rates and compare the estimates from the two cutoffs to test how TBA eligibility benefit varies with prepayments.

2.3.2 Share of loans in TBA-eligible MBS around the cutoffs

Figure 3 shows that the fraction of loans included in TBA-eligible MBS changes substantially and discontinuously at each of the two cutoffs. In panel A, all loans with sizes below the national CLL are included in TBA-eligible MBS. However, the fraction decreases to around 0.6 for loans right above the national CLL. This fraction does not decrease to zero because high-balance loans can still be included in a TBA-eligible MBS as long as their share does not exceed 10|$\%$|⁠.

The discontinuity presented in panel A indicates that the 10|$\%$| rule is binding. Given that high-balance loans are generally included in either TBA-ineligible MBS or cheapest-to-deliver TBA-eligible MBS, an alternative way to examine whether the rule is binding is to calculate the share of high-balance loans relative to loans that are typically pooled into cheapest-to-deliver TBA-eligible MBS. In our sample, high-balance loans account for about 16|$\%$| of loans that have both balances greater than $175,000 and LTVs up to 80|$\%$|⁠, which are low-value loans that are typically pooled into cheapest-to-deliver MBS for TBA trading. Thus, the 10|$\%$| rule is binding.¹³

In panel B, all loans with LTVs of 105|$\%$| or lower are in TBA-eligible MBS. The fraction decreases sharply to zero once the LTV exceeds 105|$\%$|⁠, because any such loans cannot be included in TBA-eligible MBS. Although loans with LTV of 105|$\%$| or slightly lower are valuable and MBS backed by them rarely trade in the TBA market, there is a value to keeping the option to trade via TBA. For example, Gao, Schultz, and Song (2017) find that SP-trading of TBA-eligible MBS incurs lower trading costs than TBA-ineligible MBS. Therefore, lenders do not mix loans with LTVs below and above 105|$\%$| in the same MBS.

Note that Figure 3 is created using only GSE loans. The changes in TBA eligibility at the cutoffs are not because loans above the cutoffs cannot be sold to the GSEs. This is the main difference from papers that estimate the jumbo-conforming spread in the period before high-cost CLLs were introduced in 2008.¹⁴ They estimate how much GSE securitization reduces mortgage rates by comparing loans that are eligible and ineligible for GSE securitization. The effect of GSE eligibility will capture not only the value of having access to the TBA market, which is only available for agency MBS, but also the value of credit guarantees for GSE loans. In contrast, we estimate the effect of TBA eligibility controlling for the effect of the credit guarantees.

3. Effects on Mortgage Rates

In this section, we quantify the benefit of TBA eligibility on the mortgage rate. We present the tests and results using the national CLL cutoff in Section 3.1 and the LTV 105|$\%$| cutoff in Section 3.2. We then compare the estimates from the two cutoffs and discuss how TBA eligibility benefit varies with expected prepayments in Section 3.3.

3.1 National CLL

3.1.1 Identification strategy

As shown by Figure 3, panel A, the probability that a loan is packaged into TBA-eligible MBS decreases discretely at the national CLL. One way to exploit this discontinuity is to estimate the difference in mortgage rates between loans above and below the cutoff. However, Figure 4, panel A, shows that many loans bunch at the national CLL. Although this bunching suggests higher rates for loans above the cutoff, it also makes identification challenging. Borrowers who bunch might have different unobserved characteristics from those who take out loans just above the cutoff. Thus, the rate spread between loans above and below the cutoff could be due to unobservables.

We address this issue with an instrument variable (IV) strategy that utilizes an alternative cutoff based on the home appraisal value.¹⁵ The GSEs usually require a borrower with less than a 20|$\%$| down payment to have mortgage insurance. Therefore, the appraisal value of 125|$\%$| of the national CLL can serve as the alternative cutoff. Most borrowers with the appraisal value up to the new cutoff would take out conventional conforming mortgages (loans not larger than the national CLL). In contrast, some borrowers with the appraisal value greater than the new cutoff would take out high-balance mortgages. If the appraisal value is just above 125|$\%$| of the national CLL, then many borrowers would be able to pay small extra payments to originate conventional conforming loans. As the appraisal value increases from the national CLL, more borrowers will take out high-balance loans, and the probability that the loan is packaged into TBA-ineligible MBS will also increase.

Figure 5, panel A, plots the probability that the loan is packaged into a TBA-ineligible MBS against the home appraisal value. A visible kink at the home appraisal value cutoff shows that the relationship between the two variables changes discretely at the cutoff. Therefore, we study the causal relationship between TBA eligibility and mortgage rates using the regression kink design (RKD). Figure 5, panel B, plots mortgage rates against the home appraisal value, which also exhibits a visible kink at the cutoff. Intuitively, the impact of TBA eligibility on mortgage rates is estimated by dividing the change in the slope at the kink in the mortgage rate by the change in the slope at the kink in TBA eligibility.

The identification assumption is that the kink for the mortgage rate is solely driven by the kink in TBA eligibility. A potential concern about this identification strategy is that borrowers above the cutoff might be different from those below the cutoff because financing costs may be higher for homes above the cutoff. However, we will show later that ex post prepayment patterns and borrower characteristics, such as credit scores, do not change discretely at the cutoff. RKD also requires both the density of the appraisal value and its first derivative to be smooth at the cutoff (Card et al., 2015). Figure 4, panel B, does not show either a jump or bunching at the cutoff. Despite bunching at multiples of $5,000, there is no obvious kink around the cutoff.

3.1.2 Summary statistics

Table 1 presents summary statistics of the estimation sample. 40|$\%$| of loans in the sample are above the cutoff based on the appraisal value. High-balance loans and loans included in TBA-ineligible MBS account for 18|$\%$| and 8|$\%$|⁠, respectively. Thus, 44|$\%$| (=.8/.18) of high-balance loans are included in TBA-ineligible MBS because almost all loans with sizes up to the national CLL are included in TBA-eligible MBS as shown in Figure 3. 33|$\%$| of loans in the sample were paid off within 3 years of origination, suggesting fast prepayments for these loans.

3.1.3 Regression specification

The dependent variable in Equation (1), |$NoTBA_{i}$|⁠, is a dummy variable that equals one if loan |$i$| is included in a TBA-ineligible MBS. In Equation (2), |$Rate_{i}$| is the mortgage rate.

On the right-hand side, |$z_{i}$| represents the difference between the appraisal value for loan |$i$| and the cutoff (125|$\%$| of the national CLL), in thousands of dollars. Thus, |$1[z_{i}>0]$| is a dummy variable that is equal to one if the appraisal value is greater than the cutoff. In Equation (1), |$\alpha_{11}$| measures the change in the slope at the cutoff. The variable |$z_{i}\times1[z_{i}>0]$|⁠, which only shows up in Equation (1), serves as the instrument. The coefficient |$\beta_{11}$| in Equation (2) measures the treatment effect of being packaged into TBA-ineligible MBS, or equivalently, the value of the option to trade via TBA. Next, |$g^{-}(z_{i})$| and |$g^{+}(z_{i})$| represent polynomials of |$z_{i}$| for values not greater than zero and values greater than zero, respectively.¹⁷ Parameters |$\theta_{0}$|⁠, |$\theta_{1}$|⁠, |$\phi_{0}$|⁠, and |$\phi_{1}$| are the coefficients of the polynomials and will be estimated.

Vector |$K_{i}$| contains other loan characteristics: credit score, loan-to-income ratio, whether a loan is originated by a broker, and whether a loan is originated by a correspondent lender. In addition, we include dummy variables (⁠|$\xi_{LLPA}$| and |$\chi_{LLPA}$|⁠) for each cell in the loan-level price adjustment (LLPA) matrix, which is determined mainly by the credit score and LTV.¹⁸ The GSEs use the LLPA matrix to adjust guarantee fees to account for credit risks. Finally, |$\xi_{zipcode}$| (⁠|$\chi_{zipcode})$|⁠, |$\xi_{lender}$| (⁠|$\chi_{lender})$|⁠, and |$\xi_{t}$| (⁠|$\chi_{t})$| refer to the fixed effects for ZIP code, lender, and origination year-month |$t$|⁠.

3.1.4 Regression results

Table 2 presents the regression estimates with different bandwidths. For each bandwidth, we experiment with different maximum numbers of polynomials for functions |$g^{-}$| and |$g^{+}$|⁠. We report robust standard errors for our estimates.¹⁹

Panel A presents the estimates from the first-stage regression. The estimates of |$\alpha_{11}$| show that the kink at the cutoff is statistically significant. With our preferred specification (column 3), a $10,000 increase in the appraisal value from the cutoff raises the probability to be included in TBA-ineligible MBS by |$0.008 \times 100 \times 10 = 8$| percentage points (pp). Because our sample includes GSE loans only, this change in the slope is only due to TBA eligibility and not GSE guarantees.

Panel B displays the estimates from the second-stage regression. Our estimates of the effects of TBA eligibility on the mortgage rate are 28–47 bps and statistically significant across almost all specifications.²⁰ Our preferred estimate (column 3) shows that the option value to trade MBS in the TBA market reduces the mortgage rate by 28 bps for loans near the national CLL. As discussed earlier, the TBA option value is likely higher for such loans due to higher prepayments.

Note that the estimate does not measure the difference in mortgage rates between conventional conforming loans and high-balance loans. In our sample, about 56|$\%$| of high-balance loans are still included in TBA-eligible MBS. Thus, the average rate spread between conventional conforming loans and high-balance loans should be about 12 bps (= 0.44 |$\times$| 28 bps).

3.1.5 Other considerations

Density of other loan characteristics. One of the identifying assumptions in our RKD framework is that the density of unobserved characteristics that are correlated with mortgage rates are smooth around the cutoff. Because expected prepayments are a major determinant of loan values for MBS investors in the cross-section, we test whether there is discontinuity of ex post prepayments at the cutoff.²¹ We find that ex post prepayments do not discretely change at the cutoff. The Internet Appendix reports the details of this test.

Another testable identifying assumption of the RKD is that the density of observable characteristics is smooth at the cutoff. We consider credit score, loan-to-income ratio, whether a loan is originated through a broker, and whether a loan is originated through a correspondent bank. We find all four variables to be smooth around the cutoff. The Internet Appendix reports the details.

Multidimensional mortgage pricing. In principle, mortgage pricing is multidimensional. For example, borrowers can purchase points to reduce their mortgage rates, as highlighted in Bhutta, Fuster, and Hizmo (2020) and Bhutta and Hizmo (2020). Thus, we may not capture the full effect of TBA eligibility on mortgage pricing by focusing on mortgage rates alone. For example, if TBA-ineligible loan borrowers purchase more points to reduce mortgage rates, our estimates would understate the impact of TBA eligibility. We look into whether TBA eligibility affects points purchased using another data set that provides information about points purchased. Because this data set does not provide information about whether each loan is included in a TBA-eligible MBS, we instead look at whether high-balance loan borrowers purchase more or less points.

Appendix A provides the details of this analysis. Depending on the specification, we find either that TBA eligibility does not affect the amount of points purchased or that high-balance loan borrowers tend to purchase additional points. Whenever we have statistically significant estimates, they indicate that high-balance loan borrowers tend to purchase additional points. Thus, if anything, our estimates on mortgage rates potentially understate the effect of TBA eligibility on mortgage pricing. We conclude that our main finding is robust even when considering multidimensional mortgage pricing.

GSE securitization. Our estimate is potentially subject to selection bias if lenders choose to sell certain high-balance loans to the GSEs. In the Internet Appendix, we look at the relationship between the appraisal value and the probability of selling the loan to the GSEs. Since 2013, there has been a visible jump or kink around the cutoff, which led to our decision to restrict the sample to data up to 2012.

Comparison within high-balance loans. One might consider estimating the benefit of TBA eligibility in an alternative way by comparing mortgage rates between high-balance loans in TBA-eligible MBS and high-balance loans in TBA-ineligible MBS. We find a much smaller estimate (6 bps) with this alternative research design. In the Internet Appendix, we discuss in detail why the magnitude is very different from our main estimate. In sum, one of the necessary conditions for this alternative design to provide an unbiased estimate is that lenders know perfectly at the time of interest rate lock, which happens weeks or months before loan closing, whether the loan will be included in TBA-eligible or TBA-ineligible MBS. In reality, this condition is unlikely to be satisfied. In contrast, our Fuzzy RKD estimator does not require econometricians to take a stance on whether lenders know perfectly which loans will be included in which MBS, as long as the identifying assumptions are met.

3.2 Loan-to-value of 105|$\%$|

3.2.1 Identification strategy

Figure 6, panel A, shows that the density of HARP loans has discontinuity at LTV of 105|$\%$|⁠, suggesting the possibility of sorting of borrowers around the cutoff based on unobservables. As in our analysis with the CLL cutoff, this bunching suggests higher rates for loans above the cutoff, but it also makes identification challenging.

As in the CLL test, we use the RKD for this analysis. If |$MaxLTV_{i}\leq105$|⁠, we expect that the actual LTV of a new HARP loan is not greater than 105|$\%$|⁠. If |$MaxLTV_{i}$| is slightly above 105|$\%$|⁠, the actual LTV will be above 105|$\%$| if and only if the closing cost rolled into the new loan is close to the upper limit. As |$MaxLTV_{i}$| moves away from the cutoff value, the probability that the actual LTV is higher than 105|$\%$| will increase because borrowers would have to pay a higher amount upfront in order to bring the new LTV to 105|$\%$| or lower.

RKD requires that there is neither a jump nor a kink in the density of |$MaxLTV$| at the cutoff. Some borrowers may delay refinancing to pay down remaining balances to lower their updated LTVs below 105, resulting in a kink in the density of |$MaxLTV$| at the cutoff. Therefore, we focus on a subset of borrowers that are likely financially constrained and may not want to delay refinancing: borrowers with lower credit scores, that is, below the median (733).²⁴

Figure 6, panels B and C, plots the density of |$MaxLTV$| for the full sample and the low-credit-score subsample, respectively. While the density seems smooth and without a kink for both figures, a more formal test shows that there may be a kink at the cutoff for the full sample depending on the specification, but not for the low-credit-score subsample.²⁵|$^, $|²⁶ Thus, for the rest of Section 3.2, we use the low-credit-score subsample, but the results remain similar with the full sample as we will discuss in Section 3.2.5.

Figure 7, panel A, shows that |$MaxLTV_{i}$| predicts the probability that a HARP loan would be included in a TBA-ineligible MBS with a kink in the relationship at 105|$\%$|⁠. Figure 7, panel B, shows a kink for the residual mortgage rate, suggesting a causal effect of TBA eligibility on mortgage rates.²⁷ The slope in Figure 7, panel B, is slightly positive even for TBA-eligible loans below the cutoff and those far above the cutoff (⁠|$MaxLTV$| higher than 110). A possible reason for the upward trend is because |$MaxLTV_{i}$| is highly correlated with the actual LTV and loans with higher LTVs have higher default risk, which is costly for the lenders and servicers.²⁸ However, default risk does not change discretely at the cutoff, so the kink is due to changes in TBA eligibility for loans around the cutoff.²⁹

3.2.2 Summary statistics

Table 3 presents summary statistics of the subsample. 41|$\%$| of loans in the subsample are above the cutoff based on |$MaxLTV$|⁠. 34|$\%$| of loans have actual LTV greater than 105, which are always included in TBA-ineligible MBS as shown in Figure 3. Next, the fraction of loans paid off by 36 months is very low at 1|$\%$|⁠, compared with that of loans in the sample for the CLL test (33|$\%$|⁠) in Table 1. Average |$MaxLTV$| is about 104 and slightly higher than the actual LTV. The table also shows that the loan balance increases by $3,600, on average, after HARP refinancing, which is consistent with our identification strategy of accounting for closing costs.

3.2.3 Regression specification

3.2.4 Regression results

Table 4, Panel A, shows that the estimated slope for the first stage is more positive for |$MaxLTV_{i}$| above 105, which is consistent with Figure 7, panel A. Table 4, Panel B, reports the second-stage regression results. Our preferred estimate (column 3) shows that TBA eligibility reduces the mortgage rate by 7.2 bps for HARP loans with LTVs around 105|$\%$|⁠. Note that because the relationship between |$MaxLTV_{i}$| and |$NoTBA_i$| is nonlinear, the estimate can be misleading when we include only up to first-degree polynomials. In column 9, the point estimate of the impact of TBA eligibility on mortgage rates is comparable to estimates from other columns, but it is statistically insignificant because the standard error of the estimate becomes very large with the third-degree polynomials and a relatively small number of observations. Overall, these results show that the value of TBA eligibility for loans around the LTV 105|$\%$| cutoff is 3.3–7.6 bps. As discussed earlier, the TBA option value is likely lower for such loans due to their lower expected prepayments. These estimates are much lower than those for loans near the national CLL. We discuss the difference in the estimates between the two cutoffs in more detail in Section 3.3.1.

3.2.5 Other considerations

Density of other loan characteristics. As in the national CLL test, we test and confirm whether ex post prepayments and exogenous variables are smooth around the LTV cutoff. For exogenous variables, we consider the credit score, whether a loan is originated through a broker, and whether a loan is originated through a correspondent bank, and the interest rate of the previous loan. The Internet Appendix reports the details of these tests.

Estimation with the full sample. For our main analysis, we restrict our sample to borrowers with relatively low credit scores. To address the concern that our main results are driven by this sample selection, we estimate our main regressions with the full sample, and Table C.1 in Appendix C shows that the results are very similar to our main results. This shows that our main results are not very sensitive to our sample selection.

3.3 Discussion

So far, we have found that TBA eligibility reduces the mortgage rate by 28 bps for loans near the national CLL and by 7 bps for loans with LTVs near 105. Thus, mortgage borrowers benefit from a liquid TBA market, suggesting that policy makers should take into account the effect on the TBA market liquidity when implementing changes in the housing finance system.

3.3.1 Heterogeneous benefit of TBA eligibility and prepayment speed

We argue that the difference in the magnitudes is due to the difference in prepayments. Because of adverse selection in the TBA market, cheaper MBS—typically MBS with higher prepayments—are more likely to trade in TBA than in SP (Huh and Kim, 2021). Hence, the option to trade in the TBA market would be more valuable for MBS with higher prepayment.³¹ Since loans with sizes around the national CLL have higher prepayments than loans with LTVs around 105|$\%$| as shown in Section 2.3, we would expect the TBA eligibility benefit to be higher for loans with sizes around the national CLL. Thus, the difference in magnitudes of the estimates is consistent with this prediction.

A common criticism against research designs based on discontinuities is that the resultant estimate might be difficult to extrapolate to the rest of the population. This concern would apply to our setup if we estimated the impact on the mortgage rate using only one of the two cutoffs. In our case, the estimates of TBA eligibility benefit for mortgage rates with the two cutoffs are likely to be close to the upper and lower bounds because the two cutoffs are at opposite ends of the spectrum of prepayment, as discussed in Section 2.3.

To the best of our knowledge, our finding that the impact of TBA eligibility on mortgage rates varies with prepayment speeds has not been pointed out in the literature. Our finding implies that not all borrowers will benefit from the TBA market similarly. Borrowers with fast prepayment speeds, for example, those with large loan sizes or low LTVs, will receive much larger benefit. Because such borrowers are likely more affluent, policy makers should take into account potential distributional consequences of resources they spend for the TBA market.

One may argue that other differences in the estimation samples for the two cutoffs may explain the difference in the estimated benefit of TBA eligibility. To rule out this possibility, we reestimate both regressions using the subsamples that are as similar as possible to each other: Freddie Mac loans originated in 2012 with below-median credit scores. We find that the new estimates are very similar to our baseline results: 29 bps for the CLL test and 7 bps for the LTV test.³²

3.3.2 Economic mechanism

In this section, we do back-of-the-envelope calculations to gauge whether our 7 and 28 bps estimates seem reasonable and also to explore potential economic mechanisms through which TBA eligibility benefits MBS investors and mortgage lenders. We highlight a few potential channels. First, TBA eligibility gives an MBS holder an option to trade in the more liquid TBA market that has lower trading costs than SP markets. Second, SP trading costs are lower for TBA-eligible MBS than for TBA-ineligible MBS because market makers can hedge their TBA-eligible MBS positions more cheaply (Gao, Schultz, and Song, 2017). Lastly, TBA-eligible MBS are easier to hedge during the loan origination period. Lenders are subject to interest rate risk during the rate lock-in period in the course of loan origination, which they can hedge by selling TBAs.

The first two mechanisms related to trading cost differences should result in higher prices for TBA-eligible MBS, whereas the last mechanism of the hedging benefit during the loan origination process should not. Therefore, we first estimate the effect of TBA eligibility on MBS prices to tease out the effect of the trading cost differences from the hedging benefit. Unfortunately, it is not possible to estimate the impact of TBA eligibility on MBS prices in a rigorous manner because of the reasons we detail in Appendix B. Therefore, this analysis should be taken as suggestive evidence.

After controlling for MBS characteristics that affect prepayment speeds, TBA eligibility increases MBS price by $1.1–1.5 (for $100 par value) around the CLL cutoff, and by $0.08–0.27 around the LTV cutoff.³³ This large difference in magnitude in the estimates between the two cutoffs is consistent with the difference in the impact of TBA eligibility on the mortgage rate between the two cutoffs. The larger estimate for the CLL cutoff reflects the fact that TBA eligibility is more valuable for MBS with high prepayments.

We then translate this difference in MBS price into mortgage rates. Because a 100 bps difference in MBS coupons is associated with a difference of $4.403 per $100 in MBS prices in the data, we assume that a dollar increase in the MBS price translates into a decrease in the mortgage rate of 22.7 |$(=100/4.403)$| bps. Therefore, the estimated impacts of TBA eligibility on MBS prices translate to between 25.0 bps (⁠|$=1.1 \times 22.7$|⁠) and 34.1 bps (⁠|$=1.5 \times 22.7$|⁠) around the CLL cutoff, and between 1.8 bps (⁠|$=0.08 \times 22.7$|⁠) to 6.1 bps (⁠|$=0.27 \times 22.7$|⁠) around the LTV cutoff. These numbers are roughly in line with the 28 and 7 bps estimated for the TBA eligibility benefit on mortgage rates. Therefore, if lenders fully pass on the TBA eligibility benefit to the mortgage borrowers, these findings suggest that the variation in the TBA eligibility benefit on mortgage rates is mostly driven by the TBA liquidity benefit and that the hedging benefit during origination plays a limited role because the latter benefit would not be reflected in the MBS price. Since Fuster, Lo, and Willen (2017) find that 92|$\%$| of a dollar change in the MBS price is passed through to the primary mortgage market, it is reasonable to assume that lenders fully pass on the TBA eligibility benefit to mortgage borrowers. Moreover, the difference in the transaction cost between the TBA and SP markets is likely the main driver for the large liquidity benefit for loans near the CLL cutoff because MBS containing such loans are typically traded in the TBA market (Huh and Kim, 2021). For loans near the LTV cutoff, the difference in the SP transaction cost between TBA-eligible and TBA-ineligible MBS likely plays some role, as MBS with high average LTVs trade in the SP market more frequently.

3.3.3 Decomposing the jumbo-conforming spread

We compare the TBA eligibility benefit around the CLL cutoff with the jumbo-conforming spread, which refers to the difference in mortgage rates between jumbo loans and conventional conforming loans.³⁴ This spread reflects not only the TBA eligibility benefit but also other factors related to GSE securitization, such as credit guarantee benefit, g-fees, and banks’ demand for jumbo loans.

Using the McDash data, we estimate the jumbo-conforming spread using a similar RKD approach with low-cost county loans in 2012.³⁵ Since all loans larger than the national CLL cannot be sold to GSEs in this sample, the estimate would reflect both the TBA eligibility benefit and the GSE benefit. The estimated jumbo-conforming spread is around 30 bps, and, since the estimated TBA eligibility benefit is 29 bps around the CLL for 2012, almost all of the spread is due to the TBA-eligibility benefit in this sample. This is not to say that GSE eligibility gives no benefit, rather that g-fees likely offset such benefit. In fact, g-fees were increased in 2012 to around 29–35 bps.³⁶

While such decomposition is useful, we should be cautious in comparing the estimated jumbo-conforming spread with the estimated TBA eligibility benefit. Loans for houses in low-cost counties may be materially different from those in high-cost counties in unobservable ways. These results should be taken as suggestive evidence that would be interesting to explore in future work.

4. Effects on Refinancing Behavior

So far, we have established that TBA eligibility lowers mortgage rates. One would then expect TBA eligibility to also influence mortgage demand, which may affect the real economy in various ways. First, the TBA market may affect home purchases by lowering mortgage rates. For example, Bhutta and Ringo (2021) find a lower mortgage rate increases home purchases by making the DTI requirement less binding. Because loans securitized into agency MBS in general benefit from the TBA market, we can expect that the TBA liquidity will lead to more home purchases. Second, the TBA market can also affect refinancing and consumption. Lower mortgage rates resulting from TBA liquidity benefit may increase refinancing demand, and consumer spending typically increases subsequent to mortgage refinancing (Agarwal et al., 2017, Abel and Fuster, 2021).

In this section, we investigate the second channel: whether TBA eligibility affects borrowers’ refinancing behavior. Since refinancing is important for monetary policy transmission, studying the refinancing behavior is valuable from a policy perspective. We focus on the refinancing behavior near the national CLL by comparing refinancing probabilities between borrowers who can easily refinance into loans with sizes at or below the national CLL and those who cannot. We do not consider loans with LTVs near 105|$\%$| because of data limitations. Because our analysis hinges on differences in borrowers’ refinancing behaviors above and below a cutoff, it is important to observe exactly how far the loan is from the cutoff. Our data only allow us to observe updated LTVs at the time of refinancing but provide data on the evolution of a borrower’s loan balance over time.

4.1 Refinancing and national CLL

We expect that TBA eligibility influences refinancing decisions for borrowers with outstanding mortgage balances near the national CLL in three ways. First, the TBA market may affect the extensive margin of refinancing decisions. Some borrowers that would not refinance into a mortgage rate offered for high-balance loans may refinance into a lower mortgage rate offered for conventional conforming loans. This may be because the higher rate offered for high-balance loans may either not make refinancing worthwhile or make the DTI ratio binding. Second, the TBA market may also have intensive-margin effects. Some borrowers that would have refinanced into high-balance loans if those were the only options would instead wait to refinance into conventional conforming loans. Third, some borrowers may pay off part of the outstanding balances to refinance into conventional conforming loans today.³⁷ However, as we will show later, empirically, only a small number of borrowers do so.

In the empirical analysis, we examine whether the refinancing probability is higher for borrowers right below the national CLL relative to those right above the cutoff. We interpret the difference in the refinancing probabilities as a demand response to the spread in mortgage rates.³⁸ The first two channels would result in a jump up in refinancing probability at the national CLL, while the third one would rather see a gradual increase in refinancing as the remaining loan balance reaches the national CLL from above.

One could also try to examine mortgage demand using the bunching estimator as in DeFusco and Paciorek (2017) because all three channels explained above would result in bunching at the national CLL. However, applying the bunching estimator to this setting may be problematic because the counterfactual refinance loan size distribution that would exist in the absence of the national CLL is not likely smooth. This problem arises because of the fact that previously originated loans already have bunching at the national CLL. If loans with the original balance right below the national CLL are refinanced soon after origination, their new balance will still be right below the national CLL. Thus, refinance loans may be bunched at the cutoff even in the absence of a behavioral response to the TBA market.³⁹ Thus, we instead directly examine refinancing decisions of potential refinancing borrowers.

4.2 Sample selection and data

We restrict the sample to loan-month pairs of 30-year fixed-rate mortgages that were originated in 2006 or later with remaining balances greater than the national CLL at any point during our sample period (2009–2012).⁴⁰ Many of these loans in the sample were not securitized by the GSEs but were jumbo loans or portfolio loans.⁴¹ We only consider borrowers in high-cost counties with their high-cost CLLs larger than the national CLL by at least $50,000.

We use the CRISM data, which provide loan-level mortgage performance information matched to borrower-level credit records. CRISM allows us to discern different reasons for a voluntary payoff of a mortgage, such as rate refinancing or cash out refinancing, among other reasons.

4.3 Regression specification

The unit of analysis is a loan-month pair. |$z_{it}$| refers to the difference between the remaining balance of loan |$i$| in month |$t$| and the national CLL. |$\alpha$| measures the jump in the refinancing probability at the national CLL and is the coefficient of interest. |$g^{+}$| and |$g^{-}$| are polynomials of |$z_{it}$| depending on whether or not |$z_{it}$| is greater than zero. Vector |$K_{it}$| refers to a set of loan characteristics.⁴² Lastly, |$\xi_{zip\times t}$| refers to the interacted fixed effects at the level of loan |$i$|’s ZIP code and time |$t$|⁠.⁴³

The dependent variable, |$y_{it}$|⁠, is an indicator variable that equals one if loan |$i$| is refinanced at time |$t$|⁠. We consider rate refinancing and cash out refinancing. We view refinancing as a cash out if a borrower increases the loan balance by more than 5|$\%$| of the ending balance of the previous loan. All other refinancing is classified as rate refinancing. We expect that the probability of rate refinancing increases discontinuously at the cutoff. In contrast, we do not expect to see a similar pattern for cash out refinancing because cash out refinancing of a loan with the remaining balance right below the cutoff will make the new loan balance greater than the cutoff.

4.4 Results

Figure 8 displays the relationship between the remaining loan balance and the residual probability of rate and cash out refinancing in panels A and B, respectively. Panel A clearly shows a jump in the probability of rate refinancing at the cutoff, the size of which is about 1 pp. Given the average monthly rate-refinancing probability of 1.14|$\%$|⁠, the size of the jump is economically significant. As the remaining balance increases away from the national CLL, the probability decreases only slightly; this pattern suggests that only a few borrowers make extra mortgage payments to refinance into conventional conforming loans. As expected, moreover, panel B shows that the probability of cash out refinancing does not exhibit a similar pattern around the cutoff.

We report the estimates of the regression for rate refinancing in Table C.3 and for cash out refinancing in Table C.4 in Appendix C. The estimates are consistent with Figure 8 with the estimate of the jump at around 1 pp for rate refinancing and the statistically insignificant estimate for cash out refinancing.

Mortgage refinancing is an important channel of monetary policy transmission, in which the accommodative monetary policy increases consumer spending through mortgage refinancing. In the Internet Appendix, we confirm that consumer spending indeed increases subsequent to refinancing in our sample. Implications of our result for monetary policy transmission depend on whether the demand response at the national CLL is driven by the extensive- or intensive-margin effects. On the one hand, if the demand response is driven by the extensive-margin effect, the TBA market will increase total refinancing and facilitate monetary policy transmission. On the other hand, if the demand response is driven by the intensive-margin effect, the TBA market will still result in a larger savings for borrowers who refinance but will delay refinancing and monetary policy transmission.

Although disentangling these two effects empirically is difficult, we expect that the extensive-margin effect plays a considerable role, because delaying refinancing is likely costly for many borrowers. Even paying down the balance of $5,000 under the amortization schedule would take about a year in our sample, so the borrowers would have to wait a significant period of time for their balance to reach the national CLL. Because many borrowers experienced economic hardship during this time, many would have preferred to refinance into high-balance loans sooner despite a smaller savings. Thus, we conclude that the TBA market increases total refinancing activities and facilitates monetary policy transmission at least to some extent.

5. Conclusion

Our paper overall shows that the liquidity and trading structures of the secondary market can significantly affect the cost of funding in the primary market and on real economic outcomes. The option to trade in the liquid TBA market reduces mortgage rates by 7–28 bps, and the effect is larger for loans with higher expected prepayment. We also show that TBA eligibility affects refinancing and thus the transmission of monetary policy through refinancing.

Since the TBA trading structure lowers mortgage rates for borrowers, preserving a liquid TBA market is important in the context of housing finance reform. Our finding also implies that advantages of the GSEs over private mortgage financing include not only their ability to distort an allocation of credit by borrowing at a below-market rate because of the government backstop of the GSEs but also the liquidity benefits of the TBA market. Thus, the TBA benefit will also make it difficult for private mortgage financing channels to compete with the GSEs despite ongoing efforts to reduce the government’s footprint in the mortgage market, for example, gradual increases in guarantee fees.

Additionally, this TBA benefit likely helps to keep the system that heavily depends on the GSEs from unraveling despite their imperfect risk-based pricing.⁴⁴ The GSEs’ return on capital for high-quality (low LTV, high credit score) borrowers is higher than for low-quality borrowers, which implies that high-quality borrowers are essentially cross-subsidizing low-quality borrowers. If high-quality borrowers switch to non-GSE loans, cross-subsidies cannot be sustained, which would increase mortgage rates for low-quality borrowers or unravel the GSE loan market. On the other hand, these high-quality borrowers tend to prepay faster; thus, the value of TBA eligibility is high for these borrowers. Hence, the benefit of having a liquid TBA market partially or fully offsets the higher guarantee fees. Therefore, this TBA benefit likely helps to keep high-quality borrowers from switching to non-GSE loans and to sustain the GSE-dependent housing finance system.⁴⁵

Appendix

A. Multidimensional Mortgage Pricing

In principle, mortgage pricing is multidimensional. Borrowers can purchase points to reduce their mortgage rates, as highlighted in Bhutta, Fuster, and Hizmo (2020) and Bhutta and Hizmo (2020). Thus, the difference in mortgage rates could be due to a difference in points purchased. In this section, we look into whether TBA eligibility also affected the amount of points purchased. Our analysis uses data from Optimal Blue (OB), which is an industry platform that connects mortgage lenders with whole loan investors. Thus, this data set captures information about loans originated through the correspondent channel, in which mortgage originators (typically smaller financial institutions) sell loans to larger financial institutions, who then typically securitize them. The data provide information about loan-level characteristics such as mortgage rate, credit score, home sales price, loan amount, occupancy status, and location. Importantly for our purpose, they also provide information about how many points a borrower purchased or received as credit. Mortgages processed through OB accounted for about 25|$\%$| of national loan originations in 2017 (Bhutta, Fuster, and Hizmo, 2020).

Unfortunately, this data set has a few limitations. First, it does not link a refinance loan to its previous loan. Thus, we cannot apply the same identification strategy used for the test with the LTV cutoff, so we only investigate the issue of multidimensional pricing for loans near the national CLL. Second, OB does not provide information about whether a loan was eventually included in a TBA-eligible MBS. Thus, we instead look at whether the rate difference between conventional conforming loans and high-balance loans can be explained by differences in points purchased between the two loan types. Third, the data provide information about loans originated in 2013 or later, although our sample for the analysis with the CLL cutoff ends in 2012. Thus, our analysis with the OB data only use loans originated in 2013 to keep the sample as close as possible to the sample period in our main analysis.

Other than the aforementioned differences, we create the OB sample in the same way that we create the main sample used for the analysis for loans around the national CLL in Section 3.1. We keep 30-year fixed-rate purchase mortgages that were purchased by Fannie Mae or Freddie Mac. We further restrict the OB sample to loans for owner-occupied single-family houses in high-cost counties, whose CLLs are at least $50,000 above the national CLL ($417,000).

Table A.1 presents summary statistics for the estimation sample whose home appraisal value is within $150,000 of the cutoff. Although the OB data draw from loans originated through the correspondent channel, loan characteristics in the estimation sample are still comparable to those in the estimation sample for the main CLL test (Table 1 in Section 3.1).

Figure A.1 displays relationships between the home appraisal value and the probability of being a high-balance loan (panel A), mortgage rates (panel B), and points purchased (panel C). Panel A shows that the relationship between the appraisal value and the probability of being a high-balance loan changes discretely at the cutoff. Note that the slope for appraisal values above the cutoff is much steeper in this figure than that in Figure 5, panel A, because Figure A.1, panel A, plots the probability of being a high-balance loan, not the probability of being included in a TBA-ineligible MBS. A large fraction of high-balance loans can still be included in TBA-eligible MBS. Panel B also exhibits a visible kink at the cutoff for mortgage rates, which is consistent with our findings in Section 3.1. Panel C shows that there may be a very small change in the relationship between appraisal values and points at the cutoff. If there is indeed a kink, the direction of the change suggests that borrowers of high-balance loans tend to purchase more points to reduce their higher mortgage rates, which would lead to underestimation of the effect of the TBA eligibility benefit on mortgage rates.

This system of equations is very similar to Equations (1) and (2) in Section 3.1. Instead of a dummy variable for TBA eligibility, the dependent variables in Equations (A1) and (A2) are a dummy variable that equals to one if loan |$i$| is a high-balance loan and its predicted value (⁠|$HiBal_i$|⁠), respectively. The dependent variable in Equation (A2) is either the mortgage rate or the points purchased. Instead of the month fixed effects, we include fixed effects for the date when each mortgage is locked to more precisely control for potential confounding factors. We also add how long a borrower locked the mortgage rate for to |$K_i$|⁠.

Table A.2 presents the estimates of the first-stage (panel A) and the second-stage estimates for mortgage rates (panel B) and points (panel C). The first-stage estimates show that the kink in Figure A.1, panel A, is statistically significant. The second-stage estimates for mortgage rates are consistent with our main results in Section 3.1. Mortgage rates for high-balance loans are higher than conventional conforming loans by 12.2 bps to 14.6 bps. Because 56|$\%$| of high-balance loans in our main sample are included in TBA-eligible MBS, these estimates are comparable to our main estimate in Section 3.1.

Depending on the specification, we find either that TBA eligibility does not affect the amount of points purchased or that high-balance loan borrowers tend to purchase additional points in the second stage, which is consistent with the less visible kink shown in Figure A.1, panel C. The effect of high-balance loans on points is positive but statistically significant only for relatively large bandwidths. We conclude from these estimates that there is weak evidence that high-balance loan borrowers purchase more points. Therefore, our results on mortgage rates in Section 3.1 is potentially an underestimate: If all borrowers had purchased the same amount of points, high-balance loans would have even higher mortgage rates.

By how much do we potentially underestimate the effect of TBA eligibility on mortgage rates? Bhutta, Fuster, and Hizmo (2020) find that an additional point is associated with a reduction of the mortgage rate by 21 bps. Thus, the largest estimate (in column 4 of Table A.2, panel B) implies that the spread between high-balance loans and conventional conforming loans would increase by about 5 bps. Because 56|$\%$| of high-balance loans are still included in TBA-eligible MBS, this increase would be translated to an increase for mortgage rates by about 11 bps (= 5/.44). However, note that this result is not very robust to different model specifications and bandwidth sizes. Thus, the 11 bps should be thought of as the largest possible extent to which our main estimate in Section 3.1 is underestimated.

The findings on points suggest that, if anything, high-balance loan borrowers purchase slightly more points to reduce their mortgage rates. These borrowers probably did not have enough money to reduce their loan size below the national CLL, but some likely had just enough money to buy additional points to reduce mortgage rates. Note that, even if borrowers of high-balance loans indeed purchase more points, it is still due to higher mortgage rates resulting from the possibility of being securitized into TBA-ineligible MBS. If anything, our 28 bps estimate of the effect on mortgage rates for loans around the national CLL is an underestimate.

B. Estimating the Magnitude of Various Economic Channels

In this section, we provide the details of the back-of-the-envelope calculation done in Section 3.3.2.

We first estimate the effect of TBA eligibility on MBS prices. Unfortunately, unlike our main analyses using loan-level data, it is difficult to estimate the effect accurately using a discontinuity-based method here because we do not have a reasonable number of MBS with the average LTV or loan size around the cutoffs. For example, few MBS have an average LTV around 105|$\%$|⁠, because loans with LTVs slightly above (below) 105|$\%$| are always pooled with other loans with LTVs above (below) 105|$\%$|⁠. This is the case even though there is a reasonable number of loans with LTVs around 105|$\%$|⁠. This issue also exists for estimating the effect of TBA eligibility around the national CLL.

However, given that the value of TBA eligibility with respect to mortgage rates is smaller at the LTV cutoff, the impact of TBA eligibility on MBS price is likely also smaller at the LTV cutoff, making it more difficult to detect a jump at the LTV cutoff. Furthermore, the fact that MBS price generally increases with LTV may also offset the negative impact of TBA ineligibility on prices of MBS with average LTV above 105|$\%$|⁠, which may result in a less visible jump at the cutoff. Thus, controlling for LTV correctly is important in estimating the impact of TBA eligibility on MBS prices. Moreover, if there are nonlinearities, estimates may depend on model specifications. Combining these issues with the lack of data around the cutoff, accurately estimating the effect around the LTV cutoff is likely much more difficult compared with estimating the effect around the CLL cutoff.

With the above caveats in mind, we estimate the effect of TBA eligibility on MBS prices. We try to keep the samples similar to what is used in the loan-level tests. For estimation around the CLL (“CLL sample”), we use 30-year Fannie Mae and Freddie Mac MBS issued between 2009 and 2012, with 3.5|$\%$|⁠, 4|$\%$|⁠, 4.5|$\%$|⁠, and 5|$\%$| coupons, and average original loan balances between $300,000 and |$|600,000; for estimation around the LTV cutoff (“LTV sample”), we use Freddie Mac MBS issued in 2012 and 2013, with 3|$\%$|⁠, 3.5|$\%$|⁠, 4|$\%$|⁠, and 4.5|$\%$| coupons, and average original LTV between 85 and 120. We drop high-LTV MBS from the CLL test and drop high-balance MBS from the LTV test.

Table B.1 presents the regression results. TBA eligibility increases the MBS price by $1.1–$1.5 (for $100 face value) for the CLL sample, and the estimates are statistically significant and relatively similar across various specifications in columns 1 through 4. In the LTV sample, on the other hand, the estimates vary between $0.08 and $0.27 depending on the specification. Also, the estimates are not statistically significant in columns 5 and 8. It is more difficult to estimate the effect accurately in the LTV sample for the reasons mentioned earlier.

An important observation is that the impact of TBA eligibility on MBS prices is much smaller for the LTV sample than for the CLL sample. These differential impacts reflect that TBA eligibility is more valuable for MBS with high prepayments, which is consistent with the estimated differential impacts on mortgage rates for the CLL and LTV tests in Section 3.

Next, we estimate how the higher MBS price translates into mortgage rates. To do so, we look at differences in prices across the cheapest-to-deliver MBS with various coupons.⁴⁷ We estimate that the MBS price is higher by $4.403 for the MBS with the coupon rate that is higher by 100 bps. This result indicates that an increase in MBS price by $1 decreases mortgage rates by roughly 22.7 bps (⁠|$=100/4.403$|⁠).⁴⁸

Putting this estimate together with earlier results, the estimates of TBA eligibility on MBS prices translate into a mortgage rate benefit ranging from 25.0 bps (⁠|$=1.1 \times 22.7$|⁠) to 34.1 bps (⁠|$=1.5 \times 22.7$|⁠) in the CLL sample. This estimate is in line with the 28 bps estimate of the TBA eligibility benefit on mortgage rates. Therefore, if we assume that TBA eligibility benefits are fully passed down to mortgage borrowers in the form of lower mortgage rates,⁴⁹ the benefit from better hedging during origination is relatively small around the CLL. In the LTV sample, the estimated range is from 1.8 bps (⁠|$=0.08 \times 22.7$|⁠) to 6.1 bps (⁠|$=0.27 \times 22.7$|⁠). The relatively large standard errors associated with the estimates make it difficult to determine whether the TBA eligibility benefit on mortgage rates around the LTV cutoff of 7 bps is driven by the liquidity and hedging during origination. However, both types of benefits are unlikely very large in magnitude for loans near the LTV cutoff given that the TBA eligibility benefits for such loans are relatively small.

Our findings so far indicate that our main estimates of the TBA benefit on mortgage rates in Section 3 are reasonable, compared with the estimated TBA eligibility impacts on MBS prices. We also find that the difference in the TBA eligibility benefit on mortgage rates between the CLL sample and the LTV sample is mostly driven by the liquidity benefit and that the hedging benefit during origination plays a limited role. As mentioned earlier, the liquidity benefit of TBA eligibility results from two sources: (1) the difference in the transaction cost between the TBA and SP markets and (2) the difference in the SP transaction cost between TBA-eligible and TBA-ineligible MBS (Gao, Schultz, and Song, 2017). Which mechanism drives the TBA liquidity benefit for loans near the CLL cutoff? Note that the second mechanism will be important only to the extent that MBS including such loans are traded in the SP market. However, Huh and Kim (2021) find that such MBS are mostly traded in the TBA market, which suggests that the former channel is likely the main driver of the TBA liquidity benefit for loans near the CLL cutoff.⁵⁰ For loans near the LTV cutoff, the difference in the SP transaction cost between TBA-eligible and TBA-ineligible MBS likely plays some role as MBS with high average LTVs trade in the SP market more frequently.

C. Additional Tables

Acknowledgement

We thank Itay Goldstein (the editor), two anonymous referees, Neil Bhutta, Nina Boyarchenko, Özlem Dursun-de Neef, Marco Giacoletti, Mike Fratantoni, Aurel Hizmo, Edith Hotchkiss, Sanket Korgaonkar, Tess Scharlemann, Paul Schultz, Zhaogang Song, and Clara Vega, as well as the seminar and conference participants at SFS Cavalcade, Stern Microstructure Meeting, Fixed Income and Financial Institutions Conference, AREUEA Annual Meeting, ITAM Finance Conference, Women in Market Microstructure Meeting, European Finance Association Annual Meeting, Western Finance Association Annual Meeting, Federal Reserve Board, Federal Reserve MBS Analytical Forum, Korea University, Yonsei University, Baruch College, and University of Washington for helpful comments. We thank Ben Gardner for excellent research assistance. The analysis and conclusions set forth are those of the authors and do not indicate concurrence by other members of the staff, the Board of Governors, or the Federal Reserve System. The data manager at the Federal Reserve Board has reviewed the paper to ensure that this paper follows the user agreements set by data providers. Equifax also reviewed the paper to ensure that this paper follows the user agreement with regard to the Equifax Credit Risk Insight Servicing and Black Knight McDash Data (CRISM).

Footnotes

References