As we accumulate more data at a fine local level, the opportunities to evaluate policies derived from insights into lender, borrower and supervisor behavior grow massively. Our recent post looking at the potential impact of FHFA’s new rate mod policy is our most recent example of the application of digital tools in the policy space, but as noted previously, the new policy framework is designed to focus on wealth creation and housing sustainability at the local level.
To look at this issue, we need to have data at hand that tells us which local areas have a preponderance of low-income borrowers on which we can overlay the HMDA data set, which reports census tract level indicators related to the policy issue at hand.
It turns out that the income data can be found in the American Community Survey (ACS). A key facet of this survey is information regarding the share of every census tract where low and moderate income (LMI) people comprise less than or equal to 51% of the total population. This data is available through the HUD Exchange.
With that, we now have a robust tool for analysis. An immediate challenge in this regard is to come up with specific queries out of the myriad of possibilities that demonstrate their power. Below finds a chart that provides a big-picture view of lender behavior in LMI neighborhoods broken down between banks and nonbanks. Specifically, we look at the trend in purchases from third-party originators of both FHA and conventional loans:
Recently, researchers at the Federal Reserve published a blog about trends in first time home buyers (FTHB). They utilize a dataset that is a 5% sample of credit files from a credit bureau. It appears from their analysis that there was a bit fewer than 5 million loans originated in 2020, so it would seem they are using a sample of around 250,000 loans. It is natural to ask how their measure of first-time home ownership compares with our calculation using Agency loan level delivery data consisting of over 4 million loans for that year.
To start, we can look at the trend of purchase mortgage delivery for the past seven years:
And second, we can compare estimates of first-time homeownership:
There are several distinctions to be made between the two data sets that are worth pointing out. First, the loan level delivery data only goes back through 2014 on an annual basis. Second, the Agency delivery data does not include loans originated but held on investor balance sheets, including any non-QM loans. A third difference is that the Fed researchers consider FTHB to be the first-ever mortgage file for a household while the Agencies consider a household to be “first-time” if they have not owned a house in more than three years.
The number of loans delivered to the Agencies in 2020 increased by considerably more (15.1%) in Chart 1 than that from the credit bureau (10.5%) in Chart 2, which does not reflect a difference in the growth in the total number of loans originated, but rather an increase in the share of originated loans delivered to the Agencies. During Covid, investors preferred to hold mortgages with insured credit risk. This can be seen in the sharp decline seen in the “not sold” category from the recent 2020 release of HMDA data.
The share of FTHB in our data is slightly higher than in the Fed report, which may be due to a low FTHB share in segments such as Jumbos and non-QM loans not included in our datasets.
But a key difference is that our FTHB share declined slightly in 2020 from the prior year (52.5% from 53.1%) while it rose slightly in the Credit Bureau data (48.8% from 48.2%). A rigorous attribution of the difference is beyond the scope of this note, but a few comments based on the composition of our delivery data can be made. First, in our data, the decline in the first-time homebuyer share can be attributed to a rise in the market share of conforming loans with lower FTHB shares at the expense of GNM loans with higher FTHB shares. Second, our data comes in monthly so we can look a bit at the dynamics within the year:
Interestingly, we show the FTHB share picked up in a normal seasonal pattern in 2020 Q2 and Q3, when the pandemic hit with full force. However, in Q4 the share declined to pre-pandemic levels, and rebounded only slightly in the first half of this year. It will be very interesting to monitor these trends going forward.
Last August we reported that we had downloaded 2019 HMDA and detailed queries were accessible to our clients via HMDA Analyzer. Recently, the CFPB provided a preliminary release of 2020 data, with information from smaller reporters coming a bit later in the year. Nevertheless, the new data is available on HMDA Analyzer and several insights can already be gained.
1. Total Origination loan count grew to its highest level since the runup to the Global Financial Crisis, driven by a surge in refinancings:
2. Nonbanks Rule – Covid 19 accelerated the long-term trend increase in nonbank lending share:
3. The held on book share collapsed, as banks preferred to hold mortgage risk in the form of MBS to avoid the potential for credit losses:
Much more can be found through with just a few clicks of the button.
The Covid-19 pandemic has resulted in a great economic shock that has been met with a tremendous policy reaction in the form of interest rate cuts and MBS purchases by the central bank. Prepayments have picked up substantially during the year. The question arises as to whether the magnitude of the response is unusual compared to previous episodes of rate declines, and whether they can rise further should borrowing costs fall further.
The magnitude of the relationship between rates and refis is complex and depends crucially on a number of factors. First, the relationship is path-dependent. That is, it doesn’t just matter if rates fall 1.0%, but whether this decline takes rates to new lows so that the biggest possible set of borrowers can profitably refinance.
Great gobs of sophisticated statistics and modeling go into forecasting prepays on the part of lenders and investors. But a look at prepays over time shows three main waves of refis over the past 20 years. The first (A) is 2000-2003, (B) 2008-2013, and (C) 2019-2020. In all three instances, rates reached new lows. The dates correspond roughly to the times when the mortgage rate broke to a new low until a new trough was formed (or the present in the case of “C”)
Here are the corresponding periods for prepays:
Here are the corresponding periods for originations:
Here is a summary of the three periods:
Notably, the biggest jump in prepays occurred in the early 2000s, reaching a record high of 60 CPR. Rates had fallen substantially based on aggressive Fed ease in the wake of the bursting of the tech stock bubble. This passed right through to refis. There was also a substantial decline in rates with the Global Financial Crisis(GFC), but the refi response was more muted due to declining house price, but perhaps also more stretched out in time (off and on through 2013). The current episode with Covid-19 has resulted in record-low mortgage rates, and a substantial spike in refis, but still well below the experience of 2002-04.
What might explain the differences? A clear place to look is at credit conditions. If rates drop the same amounts in time periods X and Y, but credit conditions are tighter in Y than X, we can reasonably expect a bigger refi impact in X than Y. Below shows the Urban Institute’s Housing GSE Credit Availability Index, which is used to evaluate lender’s risk tolerance:
Note: Urban Institute’s Housing Credit Availability Index for GSE Chanel. Adapted from Urban Institute Housing Finance at a Glance (August 2020). (https://www.urban.org/research/publication/housing-finance-glance-monthly-chartbook-august-2020)
As can be seen, credit was extremely loose before the GFC, very tight thereafter, followed by a period of modest loosening until 2019. This correlates well with the magnitude of the response of prepays to interest charges in the three regimes. Too hot, too cold and just right? Maybe.
All of this is important not just for investors but for the central bank as the Fed attempts to steer the economy through this uniquely uncertain period. Recently, credit conditions appear to have started tightening. It’s unclear whether this is a “normal” market reaction as volumes rise and capacity is constrained (good credits are easier to process) or whether lenders are becoming more cautious based on a more pessimistic view of the economic outlook. There is a big difference between the two in terms of choosing a successful investment strategy or an optimal monetary policy. The answer is unlikely to come from attempts to model borrower and lender behavior in a nuanced way and more likely to be discerned by careful observation of emerging trends in big data sets.
In our third look at 2019 HMDA characteristics we look at mortgage originations by income bracket. Lending to low- and moderate-income households is an important regulatory requirement of banks. The definition of “low” and “moderate” depends on the local area in which the bank operates. HMDA data is well-suited to regulators looking to track the performance of the institutions they oversee and allows banks to benchmark their performance against their competition. If banks need to add low- to moderate-income loans to their portfolio to meet requirements, HMDA can provide direction regarding which institutions might be a source of product that meets needed characteristics.
Below we present a quick high-level example. HMDA data operates down to the census tract level, but for our purposes here let’s look at two distinct states: California and Oklahoma. In 2018, median income in the two states was $70,500 and $54,400, respectively. According to Zillow data, the median house prices in California and Oklahoma that year were $550,000 and $122,000 respectively. Clearly housing is relatively unaffordable for households at or below median income in California compared to Oklahoma. So it is not surprising that the homeownership rate in Q2 2018 for California, at 54.3%, is substantially below that of Oklahoma, at 69.1%.
Confirming this, the following table from 2018 and 2019 HMDA show that there is a substantially greater share of lower- and moderate- income loans available in Oklahoma than in California. Interestingly this share declined in 2019 relative to 2018, particularly for Oklahoma. It is not clear whether this is due to fundamental factors or technical issues related to an increase in the share of “N/A” responses between the two years.
Finally, to be consistent with prior posts we look at the share of conforming loans originated by banks that are sold to the GSEs, broken down by income brackets:
A few interesting observations pop up. First, in California the loans that banks keep on their book are almost entirely made to the highest-income households. For Oklahoma, it’s a mixture of highest income and lowest income. This suggests that policy requirements regarding serving poorer communities plays a relatively greater role in Oklahoma than California.
 The first two 2019 HMDA blogs are available at
 See for example, https://www.fdic.gov/regulations/resources/director/virtual/cra.pdf
 Data from 1984 – 2018 can be found https://www2.census.gov/programs-surveys/cps/tables/time-series/historical-income-households/h08.xls
 Taken from June 2018 data at https://www.zillow.com/ok/home-values/ and https://www.zillow.com/ca/home-values/
In a recent post we looked at the differences in bank underwriting characteristics between those conforming loans held on book compared to those delivered to the GSEs using data pulled from Recursion HMDA Analyzer. We now extend this into another dimension via the addition of LTV.
Below find the difference in share of such deliveries between sold loans and those held on book:
With the release of 2019 HMDA data, we now have two years of loan-level information that contains both demographic and credit characteristics. Demographic information in HMDA includes income, race, and geography down to the census tract level, while credit characteristics include DTI. Our agency loan level databases contain a richer set of information regarding lending characteristics, but limited data on geography and demographics. For institutions looking to benchmark their performance in affordable and minority lending for regulatory purposes, 2019 HMDA, with data on thousands of lenders, is an invaluable tool. If you are interested in finding out more, please reach out.
There are of course policy uses for this data as well. A significant difference between HMDA and the agency pool loan-level data is that HMDA contains data for loans held on book, the so-called “Unsold” category. This allows a comparison of loans that banks originate and keep and those they deliver. We can break this down in any number of ways, but let’s look at it for conforming loans broken down by DTI.
In the table above, we can readily observe that banks tend to keep higher-quality loans (as measured by DTI<=43) compared to those they deliver to the Enterprises. Of course, this is not a complete picture of this issue; there are many other ways to slice the data (credit score, LTV, loan size, geography). Moreover, as there is a correlation between low LTV and desirable loan characteristics for regulatory purposes (minority status, low income), we cannot simply conclude that it’s a matter of keeping the best for themselves.
A second interesting question is: did behavior in this regard change between 2018 and 2019? Below you can find a chart of the change in the distribution between unsold and delivered loans between these two years.
It appears that banks kept more of the loans associated with very low levels of indebtedness (DTI<35) in 2019 compared to 2018, while they distributed a small share of higher-risk loans across the spectrum of DTIs above that level.
Explanations for such behavior are the subject of future research.
2019 HMDA data has been released and is loaded into Recursion’s HMDA Analyzer so clients can perform consistent queries back to 1990. As always, a vast wealth of information is available. Below are several high-level observations.
First, total originations rose by over $700 billion compared to 2018, a 13-year high. The bank share fell for the eighth consecutive year, reaching a record low of 37%. This was down 1 percentage point from 2018, the smallest decline posted for 8 years. Nonetheless, banks have suffered a remarkable 30-point drop in market share since 2008.