Changing the Scope of GSE Loan Guarantees: Estimating Effects on Mortgage Pricing and Availability

Abstract

Using a novel combination of mortgage datasets, we analyze the effects of two policy levers influencing the scope of Fannie Mae and Freddie Mac’s (Government-Sponsored Enterprises, GSEs) involvement in the U.S. residential mortgage market. First, we find that small changes in mortgage guarantee fees charged by the GSEs were essentially fully passed through to consumers, with limited effects on mortgage demand. This implies that small fee changes are primarily transfers between mortgage consumers and taxpayers while the GSEs remain in federal conservatorship. Second, the data suggest that marginally lowering maximum conforming loan size limits would cause most affected consumers to reduce their loan amounts to the new maximum. Our findings provide new detailed evidence on how GSE policy shapes mortgage availability and contrast the differing effects on consumers of two potential policy levers to reduce (or increase) the scope of GSE lending. Additional survey data indicate that borrowers’ shopping behavior and incomplete information may also influence the effects of GSE policy changes.

Appendix

Background and Detailed Literature Discussion

GSEs Before the Early 2000s

The federal government took substantial steps during the 1930’s to promote homeownership and liquidity in housing finance markets, creating the Federal Home Loan Banking System, the Federal Housing Administration (FHA), and Fannie Mae (Federal National Mortgage Association).²⁷ Fannie Mae, originally a federal agency created to purchase and sell FHA-insured loans, was eventually fully privatized in 1968 and Freddie Mac was created with a similar private structure in 1970.²⁸ Changes at this time also allowed Fannie Mae and Freddie Mac to purchase mortgages not secured by the FHA or by any other government agency.²⁹ These non-FHA loans purchased and securitized by the GSEs are known as “conforming” loans and continue to be used as collateral for mortgage backed securities. In exchange for assuming the credit risk on these loans, the GSEs charge two types of guarantee fees; ongoing monthly fees that are largely fixed at the lender-level, and publicly-posted upfront fees which depend on the risk characteristics of a loan (e.g., credit score and LTV).³⁰

Although they were not government agencies, the GSEs’ debt was often treated as close to par with the debt of federal agencies, due to a widely held perception that the GSEs were effectively backed by the U.S. government.

As the GSEs’ market share grew, there was considerable research into the interest rate spread between conforming and jumbo loans—which were non-conforming due to their size—starting from at least the late 1980s (Hendershott & Shilling, 1989) and continuing through the early 2000s (Congressional Budget Office, 1996; Lehnert et al., 2008; McKenzie, 2002; Passmore, 2005; Passmore et al., 2005). The more recent papers estimated that jumbo loans were priced approximately 15–30 basis points per year above conforming loans during this period.³¹

GSEs From the Early 2000s, and the 2015 Guarantee Fee Changes

The GSEs’ role in the housing boom and financial crisis has been analyzed extensively, with large losses leading them to be placed in conservatorship in 2008.³² The Federal Housing Finance Agency (FHFA) acts as conservator, with the U.S. Treasury Department effectively receiving all GSE profits during conservatorship.

Conservatorship was intended as a temporary arrangement, and early reform proposals aimed to “pave the way for a robust private mortgage market by reducing government support for housing finance and winding down Fannie Mae and Freddie Mac on a responsible timeline” (Treasury Department and HUD, 2011).³³ Two of the four suggested actions were increasing guarantee fees and reducing conforming loan limits.³⁴

More Recent Literature on GSEs, Mortgage Pricing, and Loan Limits

More recent research has used applied econometric techniques to measure borrowers’ sensitivity to fee changes, rate differences, or loan limit adjustments for mortgages guaranteed by the GSEs or the FHA. DeFusco and Paciorek (2017) use bunching at the GSEs’ loan limit to derive the elasticity of mortgage demand to changes in interest rates. The authors estimate a pricing difference between jumbo and non-jumbo loans using CoreLogic data, finding a difference of 18 basis points per year—a result consistent with the previous literature, but higher than our estimates on the order of 5 basis points.³⁵ The difference is likely explained by the time period studied (we analyze 2015, while DeFusco and Paciorek (2017) analyze the early 2000s) and by the differences in data (CoreLogic has interest rates, but neither points nor fees that the consumer paid). Important changes over time are demonstrated by Fisher et al. (2021), who identify increased bank supply of jumbo loans and higher guarantee fees as direct causes of the reduced jumbo-conforming spread in recent years. Bhutta and Hizmo (2021) highlight the importance of controlling for discount points in pricing comparisons, finding that interest rate differences across racial groups can be explained in part by borrower sorting along pricing menus with trade-offs between interest rates and upfront points paid. By analyzing offered retail interest rates, we provide updated findings on rate spreads and pass-through not possible in prior studies.

Grundl and Kim (2021) study the effects of GSE and FHA loan limit increases in the wake of the financial crisis, finding small effects on overall homeownership levels despite substantial increases in loan volume backed by the federal government. These findings are consistent with related literature studying FHA loan limit changes, indicating that borrowers may respond to such constraints by adjusting home size (Park, 2021) and location (MacDonald, 2019), depending on the availability of alternative—in this case conventional—products (Park, 2017). We find many borrowers are constrained by the conforming loan limits despite small pricing differences, suggesting changes in these limits may substantially shift borrowers’ sorting into homes and locations.

Other recent studies have also measured credit availability through a focus on underwriting constraints, rather than pricing. Laufer and Paciorek (2022) and Anenberg et al. (2019) both study the characteristics of observed loan originations to infer the boundary or frontier of credit availability, the former focusing on the role of credit scores and the latter taking a more holistic, multi-variate approach. Both find important roles for lender underwriting restrictions on non-pricing dimensions. This paper complements these approaches, measuring credit availability not just through the boundary of originated loans, but through the boundary of loans offered according to lenders’ rate sheets.

Demand Model Details and Results

Our demand analysis uses a dataset of Informa rates merged to the Y-14 M supervisory loan-level data collected as part of the Federal Reserve’s Comprehensive Capital Analysis and Review.³⁶The Y-14 M data have the benefit of containing both credit scores (unlike HMDA) and also lender identifiers (unlike CoreLogic). We thus use a matched sample of 15 lenders that appear in both the Informa and Y-14 M datasets to estimate demand effects from lender-specific changes in relative prices. The Federal Reserve Y-14 M data contain information for loans held on portfolio or serviced by large banks and bank holding companies. While not representative of the market as a whole, reporting is mandatory for banks meeting the asset threshold, so the data provide complete portfolio coverage for the matched lenders. Our sample includes all 2015 one-to-four unit, conventional purchase loans originated through the retail channel by lenders in our matched Y-14 M-Informa sample.

We interpret any changes as likely a demand—rather than supply—response given the evidence of full pass-through and the widespread availability of conforming loans at these credit score and LTV levels (as shown in Table 1).

Figure 5 shows the share of loans in the treated category—those with a guarantee fee decrease— by week of origination. Because LTV is the key choice variable for consumers, treated share is shown separately by credit score bin. The share of loans in the treated category is quite consistent across all weeks, ranging from about 80% of sample originations for credit scores between 700 and 719 down to 55% for credit scores between 800 and 819.³⁷ The stability of these shares confirms there were no large demand shifts due to the guarantee fee change.

Fig. 5

Share of Loans Originated with a Credit Score and LTV Combination that had a Guarantee Fee Decrease in 2015. Note: 2015 Y-14 M first-lien, conventional purchase mortgage originations by lenders in the Informa data (identified by lender name). The vertical line at April 17th indicates the announcement date of the upfront guarantee fee change, while the vertical line at September 1st indicates the effective date. Within credit score bin, treated share determined by variation in LTVs. For credit scores below 700, treated share equals 1. See Table 2 for grid of treated credit score and LTV combinations

To more precisely estimate possible demand effects, we estimate regression models mirroring those for our pass-through estimates (specifications 4 and 5) but where the unit of observation is an originated loan and the outcome is an indicator equal to one for loans within a treated combination of credit score bin and LTV bin.

Identification comes from variation in the implementation date of the fee change across lenders, as shown in Fig. 1. We estimate the following model specifications for the 13 lenders in the Y14M-Informa matched sample where we can identify the exact date the fee change was implemented:

$${\displaystyle \begin{array}{rl}{y}_{i(jtsfl)}& ={\beta}_1{\mathrm{Post}}_{jt}+{\theta}_j+{\theta}_s+{\theta}_f+{\theta}_{w(t)}+{\epsilon}_i\\ {}{y}_{i(jtsfl)}& ={\beta}_1{\mathrm{Post}}_{jt}+{\theta}_j+{\theta}_s+{\theta}_{fw}+{\epsilon}_i\end{array}}$$

where y_i is an indicator equal to one for loans within a treated combination of credit score (f) and LTV (l) bins, Post_jt indicates dates after the fee change implementation by a given lender (j) and fixed effect controls for lender, state (s), credit score bin, and origination week (w(t)).³⁸ The second specification includes credit score bin-week interacted fixed effects θ_fw.

The treatment coefficient of interest, β₁, captures changes in the probability that an originated loan falls within a treated bin after the originating lender implements the guarantee fee decrease. Seasonal effects and persistent differences in loan choices are controlled for with week, lender, state, and credit score bin fixed effects. The specifications with interacted credit score bin-week fixed effects (θ_fw) control for possible differential time trends among borrowers with high and low credit scores.

This model assumes all loans originated after the guarantee fee change are subject to the new lower rates. To account for the possibility that some borrowers use rate locks—thereby potentially choosing their loan prior to the fee decrease coming from the guarantee fee change—we also estimate specifications in which we drop all loans originated within the 45 days following a lender’s fee implementation date.³⁹ We refer to this as the Alternative Post (Alt. Post_jt) definition in these specifications:

$${\displaystyle \begin{array}{rl}{y}_{i(jtsfl)}& ={\beta}_1\mathrm{Alt}.{\mathrm{Post}}_{jt}+{\theta}_j+{\theta}_s+{\theta}_f+{\theta}_{w(t)}+{\epsilon}_i\\ {}{y}_{i(jtsfl)}& ={\beta}_1\mathrm{Alt}.{\mathrm{Post}}_{jt}+{\theta}_j+{\theta}_s+{\theta}_{fw}+{\epsilon}_i.\end{array}}$$

For all specifications, we assume that borrowers do not manipulate their credit score bin in anticipation of or in response to the fee change, and thus the choice variable for consumers is largely restricted to LTV. With the included fixed effects, the identifying variation then comes from changes over time in the distribution of LTVs within lender, state, and credit score bin.

Identification also requires that the implementation date for a given lender is not strategically chosen based on anticipated demand patterns across loan types. In practice, all lenders appear to implement fee changes across all loan types simultaneously, suggesting the timing may be driven by lenders’ expected turnaround time to securitize the loans with the GSEs (when the guarantee fee is actually paid, as discussed in Hogan (2016)) or operational constraints.

Table 7 Effect of Guarantee Fee Decrease on Share of Originations in Treated Bins, Retail Channel Only

Appendix Table 7 presents the estimated demand effects of the guarantee fee change on retail purchase originations.⁴⁰ For the sample including all originations (columns (1)–(2)), the 95% confidence intervals range from a 1-percentage point decrease to a 4-percentage point increase in the probability that an originated loan is made in a treated category. The estimates are larger for the alternative sample which excludes loans originated during the 45 days following the fee change to account for the possible use of rate locks (columns (3)–(4)), with 95% confidence intervals encompassing a null effect and a 5-percentage point increase.⁴¹ Stronger effects for this alternative sample would be consistent with rate locks preventing some borrowers from responding to the price decrease.⁴²

We find no statistically significant demand effects, but the point estimates have the expected positive sign for a demand response to a fee decrease and imply price sensitivities comparable to those found in prior studies.⁴³ For our preferred specification which accounts for borrowers’ use of rate locks, the point estimates suggest a 2.5 percentage point increase in the share of originations in treated bins, from a base of 77%.

Table 8 Refinance Loans, Effect of Guarantee Fee Decrease on Share of Originations in Treated Bins, Retail Channel Only

Table 9 Effect of Guarantee Fee Decrease on Share of Originations in Treated Bins, Retail Channel Only, Logit Model

Table 10 Effect of Guarantee Fee Decrease on Share of Originations in Treated Bins, Retail, Wholesale, and Correspondent Channels

Conforming Loan Limit Simulation Details

Counterfactual Loan Limits

Our model is estimated to minimize the difference between the missing mass of borrowers just above the limit and the excess mass bunching at the limit.

For our calculations, all dollar loan amounts are measured in logs and normalized to zero at the conforming loan limit in the county of origination. All loans are assigned to bins j = {−J,...,L,...,0,...,U,...,J}, reflecting their distance from the loan limit, in intervals of a fixed width. We use a flexible polynomial of degree p to fit the distribution of loan counts n_j within each bin, with additional indicator variables for those bins immediately surrounding the loan limit, {m_L,...,m_U}, to estimate the amount of bunching relative to the smooth polynomial. The full regression is

$${n}_j=\sum_{i=0}^p{\beta}_i{\left({m}_j\right)}^i+\sum_{k=L}^U{\gamma}_k1\left({m}_k={m}_j\right)+{\epsilon}_j.$$

The estimated polynomial, excluding the indicators, yields the predicted smooth distribution of loan amounts without bunching

$${\hat{n}}_j=\sum_{i=0}^p{\hat{\beta}}_i{\left({m}_j\right)}^i.$$

Both the missing mass of borrowers above the limit, $\hat{M}$, and the excess mass of borrowers bunching at the limit, $\hat{B}$, are calculated as differences between the borrower counts predicted by the smooth distribution and the data:

$$\hat{M}=\sum_{j=1}^U\left({\hat{n}}_j-{n}_j\right)=-\sum_{j=1}^U{\hat{\gamma}}_j$$

and $\hat{B}=\sum_{j=L}^0\left({n}_j-{\hat{n}}_j\right)=\sum_{j=L}^0{\hat{\gamma}}_j$.

After estimating the above model separately for borrowers with low, medium, and high credit scores, we use the smooth polynomial distribution of loan amounts and the indicator coefficients γ_k to calculate the fraction of borrowers missing from each of the U bins above the limit:

$${f}_k=-\frac{{\hat{\gamma}}_k}{{\hat{n}}_k}.$$

These fractions trace out the dip in originations above the existing loan limit and form the basis for our counterfactual predictions.

Our counterfactual shifts the loan limit downward by S bins, using f_k to model the number of borrowers bunching at given distances from the new loan limit. We assume that borrowers now bunch from the full range of loan sizes in {−S + 1,...,U}. Given the larger interval of loan sizes covered, we define a wider set of bins j_c = {1,...,U} such that it covers the same interval as j = {−S + 1,...,U}, and then apply the fractions f_k to these wider bins. This method traces out our counterfactual dip, and then allocates the total counterfactual mass of missing loans,

$${M}_c=\sum_{j_c=1}^U{f}_{j_c}{\hat{n}}_{j_c},$$

below the new loan limit on the interval from j = {L − S, …, −S}. Like our estimation algorithm, this counterfactual does not allow for the extensive margin choice to not take out a mortgage.

Simulation Details - Counterfactual LTV Ratios

As with the distribution of loan amounts, we assume that the LTV distribution of jumbo loans immediately above the new limit would equal the distribution observed immediately above the existing limit, f(l,above), using observed LTVs for loans from $417,001 to approximately $480,000.⁴⁴ We similarly assume that the LTV distribution of conforming loans below the new limit would equal the observed distribution below the existing limit, f(l,below), using observed LTVs for loans from approximately $404,000 to $417,000. While this method does not model loan choice at the borrower level, it provides an aggregate sense of where borrowers are switching from and switching to relative to the observed data.

Consider the high credit score borrowers observed choosing loan amounts between the new and existing loan limits. In our counterfactual, 19% of these borrowers remain between the limits after the simulated policy change. We allocate these borrowers according to f(l,above), as shown in Fig. 6 alongside the observed distribution. In this simulation, 75% of borrowers with an LTV of 80 or less would adjust their loan amounts as a result of the shift, compared to 95% of those with higher LTVs. Among borrowers with good credit, the GSEs have a larger effect on the availability of high LTV loans. However, many borrowers at lower LTVs adjust as well, suggesting that unobserved factors like stricter underwriting or incomplete information are economically important even for those with good credit.

Fig. 6

LTV Distribution for Borrowers Between Counterfactual and Existing GSE Conforming Loan Limits, High Credit Score Borrowers. Note: Data and simulations based on single-family, conventional, 30-year, fixed rate loan originations in 2015 CoreLogic data. Sample restricted to high credit score borrowers below existing conforming loan limit, and above new, counterfactual loan limit. For counterfactual, total number of borrowers is the model prediction from loan amount counterfactual, distributed according to the observed LTV distribution of jumbo loans above the existing loan limit

To see what loans switching borrowers obtain, we apply the same method to assign LTVs to all borrowers bunching below the new limit in our counterfactual, based on the distribution observed just below the existing limit, f(l,below). Note that if the purchase price of homes is held constant, our counterfactual limit reduction of four percentage points would force any bunching borrowers to reduce their loan amount (and hence LTV) by at most four percentage points. Figure 7 shows that the counterfactual distribution has less pronounced peaks at the focal values of 80 and 95, and increased mass elsewhere. The regions where the number of counterfactual loans exceeds that in the data are suggestive of the LTVs that bunching borrowers would choose. First, the simulation shows an increase in loans taken out with LTV above 80, but below 95. This would be expected from the population of borrowers who preferred high LTV conforming loans near the old limit, and now must marginally reduce their loan amount to maintain conforming status. The second region where we observe substantial numbers of bunching borrowers is at LTVs below 80. This would reflect many of the borrowers already bunching at the old limit, who would reduce their loan size even further. Based on credit score and LTV, these borrowers are eligible for jumbo loans priced at or below conforming rates.

Fig. 7

LTV Distribution for Borrowers Below New Counterfactual GSE Conforming Loan Limit, High Credit Score Borrowers. Note: Data and simulations based on single-family, conventional, 30-year, fixed rate loan originations in 2015 CoreLogic data. Sample restricted to high credit score borrowers immediately below new, counterfactual conforming loan limit. For counterfactual, total number of borrowers is the model prediction from loan amount counterfactual, distributed according to the observed LTV distribution of conforming loans immediately below the existing loan limit

Table 11 Percent of Lender-State-Date Combinations with Offered Retail Jumbo Rates Priced Below Conforming, 2015

Table 12 Average Interest Rates by Loan Type, 2015 CoreLogic Data

Table 13 Effect of Guarantee Fee Decrease on Offered Jumbo Retail Interest Rates (Placebo Test)

Table 14 NSMO Survey Responses of Conventional Purchase Borrowers at or Near GSE Conforming Loan Limit, 2013–2016, Retail Channel Only