The Risk-Shifting Hypothesis: Evidence from Subprime Originations

Transcription

1 The Risk-Shifting Hypothesis: Evidence from Subprime Originations Augustin Landier David Sraer David Thesmar June 30, 2011 Abstract Using loan level data, we provide evidence of risk-shifting in the lending behavior of a large subprime mortgage originator New Century Financial Corporation starting in This change follows the monetary policy tightening implemented by the Fed in the spring of 2004, which resulted in an adverse shock to the large portfolio of loans New Century was holding for investment. New Century reacted to this shock by massively resorting to deferred amortization loan contracts ( interest-only loans). We show that these loans were not only riskier, but also that their returns were by design more sensitive to real estate prices than standard contracts. New Century was thus financing projects with a high beta on its own survival, as predicted by a standard model of portfolio selection in financial distress. Our findings shed new light on the relationship between monetary policy and risk taking by financial institutions. They also contribute to better characterizing the type of risk taken by financially distressed firms. For helpful comments, we thank Yacine Ait-Sahalia, Gadi Barlevy, Markus Brunnermeier, Robin Greenwood, Denis Gromb, Harrison Hong, Florencio Lopez-de-Silanes, Guillaume Plantin, Jose Scheinkman, Hyun Shin and Andrei Shleifer, as well as seminar particpants at Athens-Piraeus, Banque de France, Bergen, HKUST, Lausanne, LBS, LSE, Oxford, Princeton, Mc Gill, Toulouse. All remaining errors are our own. Toulouse School of Economics ( augustin.landier@gmail.com) Princeton University ( dsraer@princeton.edu) HEC and CEPR ( thesmar@hec.fr) 1

2 1. Introduction Financial institutions, when in distress, may take too much risk. Because they do not bear the losses in case of failure, shareholders of distressed banks have a natural preference for risky lending, fueling asset bubbles, banking crises and prolonged recessions (Allen and Gale, 2000). Consistently with this, the literature provides examples of distressed banks who have made value-destroying investments decisions (Esty, 1997, Gan, 2004). Most of the evidence relies on the observation of broad categories of choices by distressed firms (for instance investing in real estate ). Such a low level of granularity makes it hard to distinguish the risk shifting hypothesis from a simpler looting view, whereby shareholders simply confiscate value from debtholders, without any change in risk preferences. Yet, understanding the exact distorsion generated by financial distress is important for the design of both prudential regulation and bankruptcy law (Akerlov and Romer, 1993). Our goal in this paper is to provide direct evidence of risk shifting, using the internal records of a large subprime originator who went bankrupt in February Such data with fine granularity allows us to accurately characterize the distorsion in risk preferences induced by financial distress. Our main result is that distressed shareholders have a stronger preference for projects whose returns are comoving with their own survival. This is the natural prediction of economic theory and it is borne out by our data. Our first step is to theoretically characterize project choice in financial distress. Using a generic model, we find that the possibility of bankruptcy tilts shareholders preferences towards investments whose payoffs are correlated with firm survival. In a calibration exercise, we show that, even when the probability of bankruptcy is small, distorsions on risk preferences are large: value-destroying projects can be preferred, provided they earn a large enough return in case of firm survival. Hence, the firm need not be near bankruptcy to riskshift. We then set out to test this model on loan-level data from New Century (henceforth NC), the second largest subprime mortgage originator in the U.S. in We document that New Century was at risk financially as early as Since 2003, 2

3 NC was retaining a significant portion (about 20%) of its originations on balance sheet and was increasing its leverage. The sharp increase in interest rates initiated by the Fed in 2004 constituted a strong negative shock to the value of loans, hence to equity value. The reasons are (1) a massive balance sheet mismatch as some loans paid fixed rates, but were financed with flexible rates, and (2) flexible rate mortgages made to subprime households would become riskier, as monthly repayments would predictably increase. Regulatory shocks, competition, as well as the progressive saturation of the real estate market, further increased the pressure on New Century s shareholders. 1 An indirect indication of financial distress is the dramatic increase in payout ratio, from 6% until 2003 to 95% in 2005Q1. 2 To bring the generic risk-shifting model to our data, we assume that NC s survival is tied to the prices of the properties indirectly held through its loan portfolio. This helps us to identify time variation in NC s survival process, and derive three testable predictions. Our first prediction is that NC should issue loans whose repayment is more sensitive to local house price growth. The economic intuition is that these loans tend to default if hoouse prices collapse, a state of nature in which NC s shareholders are protected by limited liability. We find that this policy was implemented through the issuance of interest-only loans, who grew massively from 2 to 20% of total originations in Due to their delayed amortization feature, such loans typically exhibit a strong hike in due repayments 24 months after origination. As a result, they need to be refinanced into a lower interest bearing loan when the interest-only period expires. 3 This, however, is only possible if the borrower has built enough equity in its house. Hence, interest-only loans repayments depend strongly on the local real estate market. Analyzing NC s internal records on repayments, we indeed find strong evidence that interest-only loans in regions where house prices grew little over 1 The top executives of New Century had significant incentives to maximize shareholder value. Between 2001 and 2005, the dollar stake of NC s three founder-managers went up from $42M to $147M (Source: EXECUCOMP). Nor is there much evidence that the founders tried to cash out before bankruptcy. Between 2003 and 2007, they sold less stocks than they were granted (Source: SEC). 2 Source: COMPUSTAT. The company transformed into a REIT in the last quarter of In contrast to prime mortgages, prepayment was an integral part of the business model of subprime lending (Gerardi et al, 2009) 3

4 were more likely to become delinquent. A natural consequence of this analysis is that NC s competitors who were in financial distress should also have massively issued deferred amortization loans. This should be particularly the case for originators who held an important quantity of loans as long-term investment, i.e. whose balance sheet was most exposed to the interest rate increase of Using information from 10K filings, we do find that originators who had deviated the most from the pure originate-to-distribute model in 2003 originated larger amounts of deferred amortization loans in 2005 (Figure 1). 4 In this small cross-section, exposure to the monetary shock clearly correlates with subsequent propensity to issue house price contingent loans. Going back to our model, the second prediction states that New Century should be issuing a larger fraction of its loans in regions whose house prices comove most with its own assets. The economic intuition follows the same logic as the first prediction. The lender survives if regional markets to which it is exposed via its loans are doing well enough (prices go up). Since shareholders only care about payoffs in case of survival, they issue more loans in regions who will pay off when the other loans in their portfolio will also perform. To test this, we calculate, for each region, the regression coefficient (β NC ) of local house price change on the price change of properties underlying NC s portfolio in This coefficient is shown to be a strong predictor of the geographic dispersion of loan issues. The third prediction is an interaction of the first two. To maximize exposure to its own survival risk, we expect that New Century would issue more price-sensitive loans in regions whose prices are the most correlated with its own assets. If it survives (i.e. its own assets do well), real estate markets in these regions are going to do well, which means the pricesensitive loans are going to do even better. In line with the third prediction, we find that NC issued relatively more interest-only loans in regions with higher β NC. Our paper contributes to three separate strands of finance literature. First, it belongs to 4 Origination of these loans was very small in 2003, both for New Century and the rest of the industry (Mayer et al, 2009; Demanynk and Van Hemert, 2009) 4

5 the large corporate finance literature that seeks to find evidence of costs of financial distress. Cost of financial distress are invoked to explain, in theory and in practice, why companies hold so little debt (Almeida and Philippon, 2007, Ju, Parrino, Poteshman and Weisbach, 2005). The bulk of these costs are indirect (Andrade and Kaplan, 1998): financial distress distorts risk preferences of shareholders so much that they do not maximize firm value any more. In some studies, firms lower their risk taking too much (Rauh, 2009, Gormley and Matsa, forthcoming). In others, shareholders often tend to favor risk taking to gamble for resurrection (Esty, 1997, Gan, 2004, Fischer et al, 2011, for banks; Eisdorfer, 2008, and Becker and Stromberg, 2010, for non-financial firms). The main contribution of this paper with respect to this literature is to provide a precise characterization of risk preferences of shareholders in financial distress. We first derive detailed predictions based on a model, which we then test using unique project-level data of a distressed institution. In studying a subprime mortgage originator, we also add to the growing literature analyzing the US mortgage crisis. Our analysis provides a supply-side explanation for the massive origination of toxic loans in (Demyanyk and Van Hemert, 2009, Mayer, Pence and Sherlund, 2010). We argue that originators became financially distressed as a result of their long term investment in mortgages. This led them to issuing more survival sensitive loans, i.e. loans that would only payoff if real estate prices continued to go up. This explanation is complementary to Barlevy and Fisher (2010), who attribute the rise of these loans to risk-shifting by borrowers. It also complements the existing papers on mortgage originators, which have so far focused on the perverse incentives of the originate-to distribute-model. 5 Our paper sheds light on an unexplored part of these originators activities: investment in their own loans. 6 It suggests that keeping skin in the game may be ex ante desirable, but may also distort incentive ex post. The narrative of the onset of the crisis that emerges from 5 Most papers in this literature focus on securitization (Keys, Mukerjee, Seru and Vig, 2009, Pikorski, Seru and Vig, 2010, Purnanandam, 2010, Demiroglu and James, 2011). Others investigate the role of mortage brokers (Jiang, Nelson and Vytlacil, 2010, Berndt, Hollifield and Sandas, 2009). 6 In the same vein, Acharya, Schnabel and Suarez (2010) show that securitization was not always riskless for the banks that securitized. In some vehicles, the issuing bank provided investors an explicit guarantee over their investment, in case underlying loans would default. 5

6 our analysis is thus different from a looting view of the crisis, whereby banks executives, salesmen and traders engaged in risk-taking, negligence or other forms of rent extraction to the detriment of shareholders (Akerlof and Romer, 1993, LaPorta et al., 2003, Biais et al., 2010). Third, our results suggest a new relationship between monetary policy and risk shifting. A popular view is that the Fed s persistent policy of low interest rates in the early 2000s fuelled risk taking by financial institutions reaching for yield (see for instance Rajan, 2005); Greenwood and Hanson (2011) provide evidence consistent with this view. As we show in Section 3, the monetary tightening did severely weaken NC s balance sheet, and may be a large factor behind the subsequent risk taking that we document. The behavior of the other originators, which we report on Figure 1, suggests that this consequence of monetary policy extended beyond NC. The remainder of the paper is organized as follows. Section 2 provides a simple project choice model of a levered financial institution that maximizes shareholder value. Section 3 shows that New Century was financially distressed in Section 4 describes the data. Section 5 tests the prediction of the risk shifting model. Section 6 concludes. 2. Investing in financial distress: Some theory In this section, we present a simple model of project choice under distress and derive testable predictions Basic insight We start with a general framework. There are two periods: 0 and 1. S, the survival dummy, equals 1 when the company does not default at date 1 and zero else. p = E S is the date 0 probability that the company survives at time 1. At date 0, the firm can invest in either one of two small projects: project 1 or project 2. Each project i = 1, 2 will yield a random 6

7 return R i at time 1. We focus on the choice between two marginal projects, i.e. projects that have no effect on survival S. 7 In this model, shareholders receive cash flows only when the firm survives. The NPV of project i per dollar of investment to shareholders is thus: E[ S R i ]. By definition of the covariance, project 1 is preferred over project 2 if and only if: E[ R 1 ] > E[ R 2 ] + [cov( R 2, S/p) cov( R 1, S/p)] (1) This decision rule does not coincide with NPV maximization. In the absence of leverage (i.e. when S is always equal to 1), shareholders compare the net present values of projects (here, E[ R 1 ] and E[ R 2 ]) to decide on the marginal investment. When the firm is levered, S is not always equal to 1, and shareholders prefer projects whose returns covary with S/p. Shareholders s preferences are tilted towards projects that tend to pay off in states where the company is afloat, even if their unconditional expected returns are lower. To understand the economic magnitude of this distorsion, let us consider the following numerical example. Assume that New Century, expects home prices to collapse with probability Unless this happens, NC will stay in business, so p = The firm can issue safe or risky mortgages. Whatever the state of the real estate market, safe loans pay an interest rate of 6% with probability With probability 0.03, they default and the return is -20% (recovery rate is 80% of loan value). Such a mortgage has expected return equal to 5.2%; from the viewpoint of the shareholder, the expected payoff is % = 5%. Risky mortgages are more likely to default in case of real estate meltdown. In normal times, risky mortgages yield 8% with probability.97 and -20% with probability In case of home price collapse, however, the risky loan defaults with probability 80%, with a recovery rate of 7 This assumption shuts down the countervailing effect according to which, if projects are large enough to affect overall firm survival, shareholders may want to choose safe projects in order to increase the likelihood of survival. The marginal project assumption thus allows us to focus on risk-shifting. 8 This is a reasonable estimate. In a 2005 report, the investment bank Lehman Brothers was placing a probability 5% on a meltdown scenario, under which nationwide home prices would fall by 5% annually over the following 3 years (Gerardi et al, 2008). 7

8 50%. Even if bankruptcy is very unlikely (5%), shareholders have a strong preference for risky mortgages. Given the above numbers, the expected return of the risky mortgage is 4.9%, lower than the safe loan. From the viewpoint of NC s shareholders, however, the expected return of the risky mortgage is 0.95 (0.97 8% %) = 6.8%, which is much higher than the 5% the safe mortgage delivers. Hence, even if the firm is far from being insolvent, investment distorsions can be large Testable Implications In this Section, we impose structure to the above general model in order to test it on our dataset. Since we do not observe many observations of NC s survival process, we assume that S = 1 RA >ρ A, where R A is asset returns. We label financial distress a reduction of the probability of survival p = P ( S = 1), i.e. an increase in the threshold ρ A. Secondly, we assume that each mortgage issued in region s has a payoff structure that is correlated with local home price growth g s. When prices go up, leverage goes down, so it is easier for borrowers to refinance the loan and lower monthly payments. Housing capital gains thus reduce default rates. To simplify exposition, we assume that return to loan i issued in region s can be approximated by the following linear relationship: R i,s = a i + b i. g s + ũ i,s (2) where ũ i,s has mean zero. Last, house price growth in region s is assumed to have a stable and measurable correlation with New Century s asset returns, so that we can write: g s = α s + β s R A + ɛ s (3) where ɛ s has mean zero. High β s regions are the regions whose home prices are more likely 8

9 to collapse if NC goes bankrupt. Combining (2) and (3), the NPV of loan i is given by: E( S. R i,s ) = pa i + pb i α s + b i β s pe( R A R A > ρ A ). (4) When the firm is financially distressed, pe( R A R A > ρ A ) increases, i.e. NC s assets have to perform very well to save the firm from bankruptcy. The first implication of equation (4) is that distressed firms prefer to issue loans with high exposure to local house prices b i. The intuition is that such loans are indirectly exposed to New Century s asset (in the case where β s 0 which is the empirically relevant case). Since, conditional on survival, RA is large, these loans will have high returns and will therefore be desirable: Prediction 1. When NC is in distress, it is more likely to issue high b i loans, i.e. loans whose payoff is sensitive to local real estate prices. The second prediction is also a direct consequence of equation (4). For given loan exposure to house price growth b i, the distressed firm has a preference for regions whose house prices are more correlated with its own assets (high β s ). As above, this is a way for New Century to achieve higher exposure to its own portfolio of loans, which performs well conditional on S = 1. Hence: Prediction 2. When NC is in distress, it is more likely to issue loans in high β s regions, i.e. regions whose home prices covary with its own assets. The third prediction takes advantage of the complementarity between b i and β s. Assume NC originates loans in both high and low β s regions, and that it seeks to issue one low and one high β i loan. Equation (4) suggests that NC creates more shareholder value if it issues the house price-sensitive loans in the own asset-sensitive region. Hence, the share of high b i loans should be higher in high β s regions: Prediction 3. When NC is in distress, it issues a bigger fraction of high b i loans in high β s regions. In words, asset-sensitive regions receive a larger fraction of price-sensitive loans. 9

10 : New Century in financial distress This Section shows that New Century was in financial distress in New Century became an leveraged investor Initally, all mortgages issued by New Century were destined to be sold to third parties. These mortgages belonged to one of the following three categories: Standard fixed rate mortgages (FRMs): interest rate and monthly payments are fixed for the lifetime of the loan. Hybrid Adjustable Rate Mortgages (hybrid ARMs). Loan repayment starts with a 2 year-period ( teaser period), during which monthly payments are fixed. Then, they follow the fluctutations of money market interest rates. Interest-Only ARMs: these are hybrid ARMs, with the added feature that during the 2-year teaser period the principal is not amortized.at the end of the teaser period, the monthly payment is therefore expected to increase a lot. Like most of its competitors, the firm pursued an agressive growth strategy in the first half of the 2000s. Annual loan production increased from $4.7bn in 2001 to $35.1bn in 2005 (numbers from the 10k filings). NC retained larger amounts of loans on its own balance sheet: loans held for investment went up from nothing in 2002 to $4.7bn in 2003 and $13.2bn in These investments were financed through bond issues. As its balance sheet grew, the firm experienced a sharp increase in its gearing ratio, from 5 in 2002 to 11 in Such debt-fueled growth made the firm vulnerable to a negative shock to the value of its investments. The incentive to risk-shift was particularly strong given that its bonds had a reasonably long duration. 9 Hence, the interest rate paid on these bonds, while variable 9 According to the June 10-Q filing, as of June 2004, over the $9.1bn of outstanding bonds issued to finance these loans, $1.8bn (20%) were due in less than 1 year, $2.5bn (27%) were due in 1 to 2 years, and $2.3bn (25%) were due in 3 to 4 years. Only $2.4bn (26%) were due in more than 5 years (NC s 10-Q filing, June 2004). 10

11 (a fixed margin over LIBOR), would not reflect the increase in risk taking by the firm, which is a necessary condition to make risk-shifting value maximizing from the equityholders viewpoint Monetary policy tightening The Federal Reserve increased its baseline interest rates from 1.5% in mid-2004 to 5.25% in mid As illustrated in Figure 2, this was anticipated by professional forecasters in the second half of This tightening is believed to be an important macro factor triggering the crisis (Mayer, Spence and Sherlund, 2009); we show here that it had a large adverse impact on NC s shareholders wealth. First, loans held for investment paid fixed rate, but were financed with variable rate bonds. In June 2004, NC held $9.2bn of loans as investments. Some of these loans were paying fixed rates until maturity (FRMs). The rest was made of ARMs issued between mid 2003 and mid 2004, who would pay fixed rate during their first two years of existence. All these loans were financed through flexible rate bonds indexed on LIBOR. 10 For $9bn of loans, a linear increase in LIBOR by 4 percentage points over the next 2 years could be expected to reduce total cash flows by some $360m. 11 This is a large effect: June 2004 book equity was only $743m. New Century s accounts confirm this back-of-the-envelope calculation. Interest income to interest expense ratio dropped from 3 in 2003 to 2.5 in 2004 and 1.8 in This suggests that NC did little to hedge its interest rate exposure, as explictly acknowledged by the company. 12 Second, beyond the teaser period, ARMs held as investments became riskier (Mayer, Spence and Sherlund, 2009). As interest rates would go up, monthly payments on ARMs were 10 Source: 10-K filing, Assume that the interest rate increases linearly with time, by 4 percentage points over two years. Assume no discounting to simplify. Then, total loss due to interest increase is given by $9bn 2 (2t/100)dt = $360m New Century reports hedging for some of this interest rate exposure, using derivative contracts such as Euro Dollar futures or interest rate caps. However, positions were very limited in size. As of December 2004, the fair value of Eurodollar contracts was a mere $26.1m and the fair value of interest caps was $7.4 million, compared to an interest income was almost $1bn as of December

12 expected to increase, making some loans unaffordable and eventually become delinquent. In Figure 3, we report the average, over all ARMs issued in 2003, of cumulative growth in monthly payment since origination. This is computed using observed due monthly payments from NC s internal data (see Section 4). Monthly payments do not react immediately because loans issued in 2003 are still in their teaser periods, but they start to rise in mid 2005; by the beginning of 2007, repayments have increased by 25% since origination. This made ARMs unaffordable to some borrowers. We report in Figure 4 monthly delinquency rates (fraction of loans whose repayment is more than 60 days late) separately for FRMs and ARMs issued in While it remained in the vicinity of 5% for FRMs, the delinquency rate of ARMs increased from 8% in early 2005 to nearly 30% in the beginning of Third, monetary tightening hurt growth opportunities. Prospective buyers would find it more difficult to borrow, as monthly payments would increase. In addition, higher interest rates on new loans would make refinancing (about three quarter of NC s loan applications in 2003) less attractive to borrowers already holding a fixed rate mortgage Saturation of the housing market and competition Besides interest rate increases, the progressive saturation of the real estate market and competition shocks further hurt NC s growth opportunities in Some signs suggest that the US real estate market was beginning to mature in The 10 city Case-Shiller index slows down in the first semester of This was an early warning: Existing home sales reached their maximum in mid 2005, and home prices only started to decrease in Gerardi et al. (2008) show that a nationwide collapse in HPA was anticipated to happen with a low probability. Our calibration in Section 2.1 suggests, however, that even if small, a non zero probability of distress can significantly distort project choice. Competitive and regulatory pressure intensified in early 2004 for pure originate-to-distribute issuers. Since the late 1990s, some states had adopted Antipredatory Lending Laws (APLs), 12

13 whereby the purchasers of subprime loans were made liable for wrongdoing by the originators (Bostic et al, 2008). All these factors made newly issued loans harder to sell to end-investors. In some states (e.g. Georgia and Massachussetts), APLs were so stringent that pure originators such as NC altogether stopped issuing them. At the same time, some of NC s competitors were exempted from these regulations. In August 2003, the Office of the Comptroller of the Currency decided to preempt state legislation in these matters, effectively shielding national banks and their subsidiaries from state APLs Data We describe here the three datasets we use for our tests. Detailed description of the sources and construction is deferred to Appendix A Loan characteristics dataset This dataset is used to test our first prediction. We start with all 192,973 loans issued in 2004 by New Century. For each loan, we collect the following borrower variables: a full documentation dummy, loan to value ratio, household income, FICO score. We also retrieve characteristics of the loan itself: property zip code, principal amount, an ARM dummy, and an interest-only dummy. For 45,546 of these loans (see section 4.2), we are also able to construct a dummy equal to 1 if the loan ever becomes delinquent (i.e. the borrower will be at least once late by more than 60 days on his payments) before February Last, for 25,732 of these 45,546 loans, we were able to retrieve monthly home prices at the MSA (Metropolitan State Area) level. We use this information to construct a Low growth dummy, equal to 1 if home prices grow by less than 10% between the date of issue and 24 months after the date of issue (end of the teaser period). Low growth MSAs are the areas 13 NC explictly acknowledges this new constraint in its K: Federally chartered banks and thrifts have a competitive advantage over us because the federal laws applicable to their operations can preempt some of the state and local lending laws applicable to our operations. 13

14 where we expect repayment problems to be the most stringent. We also define four quartile dummies, each equal to 1 when the first 24 months home price growth is in the relevant quartile. Cutpoints are defined the sample of MSAs to avoid composition effects. We report descriptive statistics on loan characteristics in Table 1, Panel A. 56% of all loans do not have full documentation. This raises the concern that characteristics of these loans may not be reliable, so we will check that our results also hold for the subset of full documentation loans. The loan-to-value ratio is high, on average (85%). Borrowers tend to be reasonnably well-off: Average household income is about $6,500 per month, compared to nationwide average $3, Their average FICO is 623, slightly above the 620 limit below which borrowers are usually considered subprime. Finally, for the subset of loans for which performance is available, 15% will be delinquent before February Overall, New Century issued, in 2004, 37% of FRMs, 50% of hybrid ARMs and 13% of interest-only ARMs. These figures confirm that NC s issuance policy was similar to that of the rest of the industry: Mayer et al. (2009) document that in 2004, on a large sample of securitized loans, 11% of the loans were interest-only loan, while 21% were FRMs. The fact that we do not measure the performance of all loans issued by NC may raise a representativeness concern. In Table 1, Panel B, we check the observable differences between loans that are, and loans that are not in our performance database. These loans do not differ in terms of documentation, borrower creditworthiness, or borrower income. There is a statistically significant difference in leverage, but the economic magnitude is small (85 versus 86% loan to value ratio). Loan composition differs more: Loans with performance data are 9 ppt more likely to be ARMs, and 5 ppt to have an interest-only period Loan repayment panel We will also look graphically at the time series changes in loan repayment and default. To do this, we will use a panel of monthly repayments for 45,546 loans issued in 2004, for 14 Source: Census bureau. 14

15 which New Century was in charge of monitoring the repayments beyond their 12th month of existence. These data therefore start in January 2005, and stop with NC s bankruptcy in February These loans are mostly loans that New Century chose to keep on its balance sheet. For each loan i, we collect, in month t after the 12th month, the following variables: the cumulative growth in due monthly payment between origination plus 13 months, and t, a dummy variable equal to 1 if the borrower is, in month t, more than 60 days late on his or her payments (a standard measure of delinquency), and a dummy equal to one if the loan is refinanced in month t. The data does not directly reports refinancing, so we label refinanced at date t a loan that (1) exits the sample in month t + 1 and (2) is not delinquent in month t. We will also use the price growth quartile dummies defined above. They will be loan-specific and therefore will not vary with t. We report summary statistics by loan type (FRM, ARM and Interest-only) in Table 2. We focus on 4 cross sections: June 2005, December 2005, June 2006, and December The number of observations falls over time as loans are refinanced or defaulted upon and therefore exit the sample. In Panel A, we show the average monthly payment growth since the 13th month after origination. As expected, average growth is zero for FRMs. For ARMs it is zero in 2005, as the 24 month teaser period is not finished. In december 2006, monthly payments have grown, since origination, by about 6%. This comes from the fact that the benchmark interest rate increases during the period. For interest-only loans, the movement is stronger, which reflects the beginning of principal amortization. Such a payment shock leads to a clear increase in refinancing, as borrowers who have earned enough home equity can shift to lower interest rate loans. This can be seen in Panel B. For ARMs and FRMs, the refinancing rate tends to decrease over time, while it doubles, from 6% ot 12% for interest-only loans. So the loan category with the stronger payment shock is the one that experiences the stronger increase in refinancing. But some interest-only borrowers find themselves unable to refinance, and their loans become delinquent. As shown in Panel C, delinquency rates increase for all types of loans, but much more in relative terms for 15

16 interest-only loans. Interest-only loans are therefore more toxic Geographic data Finally, in order to test the last two predictions, we use geographic variation in loan issuance. We thus construct loan issuance data, aggregated at the MSA level. This dataset is made of 352 MSAs. For each of these MSAs, we calculate the log of 1 plus the dollar value of all NC loans issued both in 2003 and 2004, as well as the log of 1 plus the total amount of interest-only loans per MSA. We also use the 2000 census to compute, per MSA: the fraction of inhabitants that did not reach the 9th grade of high school, the fraction of households below the poverty line (POVERTY variable), as well as the average home price to income ratio in We report summary statistics for these variables in Table 3. The average log total origination in 2004 is 16.13, which coresponds to about $10m per MSA. The share of households below the poverty line is 15%, slightly above national average (13%). The 2000 price to income ratio is on average 3.6. Central to our hypothesis testing is βs NC, which measures the sensitivity of MSA level home prices to NC s assets. Let R HOME s,t be the home price quarterly growth rate in MSA s. Then, we define R NC t, the return on NC s assets, as: R NC t = s w s.r HOME s,t (5) where w s is the share of MSA s in total dollar loan issuance by NC in R NC t is thus not exactly the return on NC s loan portfolio, but the average price growth of the underlying homes. β NC s is the linear regression coefficient of R HOME s,t on R NC t over the period. It captures the extent to which home prices in a particular MSA s behave like home prices of loans already in NC s portfolio. As Table 3 shows, the cross MSA average β NC s is

17 5. Evidence of risk-shifting This Section tests the three implications of the analysis of Section The shift toward home price-sensitive loans In the first quarter of 2004, New Century started to massively issue interest-only ARMs. We show in this Section that the return of these loans is, in effect, more sensitive to local home prices. This is in line with our Prediction 1. In Figure 5, we report the evolution of the fraction of interest-only loans issued. Until February 2004, 99% of NC s loan production was made of FRMs and ARMs. Then, the share of interest-only loans starts to take off and reaches nearly 40% of originations in mid Then, these loans are progressively substituted for by balloon loans. This new category, like interest-only loans, also defers amortization of the principal: their maturity is 30 years, but amortization is apread over 40. Hence, a sizeable fraction of the principal has to be repaid at maturity. Together, interest-only and balloon ARMs acount for about 60% of total loan production in January Because we need to observe loans at least 24 months after their issue (end of the teaser period), we need to focus on the 2004 vintage; hence, we will leave balloon loans out of the subsequent analysis. Before turning to the econometric analysis, we explain why interest-only loans are by design more sensitive to home prices than other loans. Interest-only loans were not supposed to be held to maturity, but only until the end of the teaser period. At this stage, refinancing was necessary because monthly payments would increase dramatically, as the borrower began to repay the principal. But refinancing required a strong home market: if in the first 24 months the borrower has built enough home equity, his wealth could be used to borrow at a lower LTV and therefore a lower rate. 15 In a weak real estate market, refinancing with better terms becomes impossible, and some borrowers have no other choice but to default. 15 An alternative was to use the capital gains to increase the amount of principal in exchange for a lower interest rate. This practice is referred to in the industry as the purchase of lender s points. 17

18 In short, issuing these loans was a bet on future home price increases. We can check, using NC s monthly payment data, that early refinancing is the normal fate of an interest-only loan. First, the payment shock occuring at the end of the teaser period was very large compared to standard ARMs. Figure 6 reports, for both ARMs and interest-only loans, the average cumulative increase in monthly payment since origination, around the end of the teaser period. For interest-only borrowers in 2004, monthly payments increase by more than 50% after the end of the teaser period. 16 Second, we can document that many borrowers refinance their loan to avoid the payment shock. In Figure 7, we calculate, for each loan type (ARM, FRM, Interest-only), the mean refinancing rate each month around the end of the teaser period. In the beginning, the refinancing rate is 4% per month, but after the end of teaser period, it jumps to 20%. By comparison, the refinancing rate of FRMs remains around 4%. ARMs experience a modest increase in refinancing rate (to about 10%): monthly payments increase too, due to monetary policy tightening. So for ARMs too, borrowers have an incentive to refinance more after 24 months, but the effect is smaller. Last, we directly check that borrowers experiencing smaller capital gains find it harder to refinance, and are more likely to become delinquent. We provide evidence of this in Figure 8. Three months around the end of the teaser period, we calculate the difference in delinquency rates between loans issued in slow growing (top quartile of first 24 month home price growth) and fast growing (bottom quartile) MSAs. We then plot such excess delinquency rates separately for ARMs and interest-only loans. As shown earlier, ARMs experienced a modest increase in monthly payments, thus did not force borrowers to refinance. Home price appreciation was less crucial to avoid delinquency, therefore, the differential delinquency rate is flat for ARMs. It increases, however, by 12 percentage points for interest-only loans. For these loans, strong home price appreciation was necessary to allow for refinancing, and avoid the increase in delinquency. 16 Borrowers with ARMs experience a more moderate increase (about 20%). This comes from the interest rate rise discussed in Section 3. 18

19 Given above evidence, we expect the delinquency rate of interest-only loans to be more sensitive to home price appreciation, consistently with prediction 1. We test this econometrically, on our cross section of loans, by running the following probit regression: EV ERDEL i = a + blg i + b ARM ARM i LG i + b IO IO i LG i + controls i + ɛ i (6) where EV ERDEL i = 1 if loan i is at least once delinquent. LG i = 1 if the loan is issued in an MSA where home price grow by less than 10% over the first 24 months if the loan s life. ARM i = 1 if the loan is an ARM, and IO i = 1 if it is interest-only. The marginal effects, reported in Table 4, show that interest-only loans exhibit a higher sensitivity of unconditional default to real estate prices than other loans. Column 1 shows that origination in a slow price growth area increases the unconditional probability of default of interest-only loans by 7.1ppt relative to FRMs. The difference in sensitivity between IO and FRM loans (and between IO and ARMs) is significant at the 1% confidence level. This effect is economically very large, since the average everdelinquency probability of interest-only loans is 12,8%. Column 2 shows that controlling for LTV and borrower FICO does not change our estimate. In column 3, we show that our results are unchanged when we focus on full doc loans. Columns 4-6 repeat the estimation, but take the four quartile dummies of price growth. The results appear monotonous in home price appreciation, giving confidence in the robustness of our estimates The shift toward asset sensitive regions We now test our prediction 2, namely that New Century invested more in regions who covary more strongly with its own returns. To test this prediction, we use MSA level data and run the following regression: 19

20 log(total $ issues in 2004) s = a + b β NC s + controls s + ɛ s (7) where log(total $ issues in 2004) s is the log of 1 plus total loan amount issued in MSA s by NC in β NC s measures the comovement of home prices in MSA s with NC s assets (see Section 4.3). We also control for potential determinants of the geographic dispersion of loans issued by NC. In particular the log of 2003 issues in MSA s is designed to capture existing biases of NC s issuance policy toward particular MSAs. The share of low income and low education households in the MSA should capture the propensity to issue subprime loans in low financial literacy areas. We report the results in Table 5. We find that in 2004, New Century tends to issue loans in regions that have higher βs NC. The effect is statistically significant at 1% even though we only have 352 observations. In our multivariate specification (columns 2-3), an increase in β NC s by 0.28 (one sample s.d.) leads to an increase in log issues by 0.5, which corresponds to about 30% of the sample s.d. of this variable. Last, we exclude California, Florida and Texas, which account for a disproportionate share of issues and could be argued to be outliers (column 4). The estimates are unchanged Geographical dispersion in loan composition We finally test our prediction 3, which states that, in distress, New Century should issue relatively more interest-only loans (high local price sensitivity b i ) in MSAs that comove the most with its own assets (high βs NC ). To do this, we estimate the following model, for each loan type k (k=frm,arm,interest-only): log total $ issues k s = a k + b k β NC s + controls s + ɛ s (8) where log total $ issues k s is log total issues of loans of type k in MSA s. Controls are identical to Table 5. We control for the total amount of loans issued in 2003 in MSA s, to capture the 20

21 fact that NC may issue more loans of type k in regions where it was previously lending more. We also control for education and poverty, in order to capture the fact that some loans types (interest-only, to some extent ARMs) may have targeted financially illiterate households. Table 6 reports the results. Columns 1-4 confirm that New Century issued a larger fraction of interest-only loans in high β NC s regions. The effect is statistically significant, and quite sizeable. In column 1, the estimate suggests that a 0.28 increase in β NC s (one s.d.) is typically accompanied by a 1.1 increase in the log issues of interest-only loans (about a third of the sample s.d.). This effect is barely affected by the financial literacy controls (columns 2). In column 3, we remove abnormal states (California, Florida, Texas), the estimate weakens somewhat (we lose more than 60 observations), but remains statistically significant and has the same order of magnitude. Last, in column 4, we split β NC s by quartiles and find a monotonous relation, which confirms the robustness of our results. Looking at columns 5-8, we also observe a positive and statistically significant effect of β NC s on the share of ARMs, but it is smaller than for interest-only lans (the difference between coefficients in column 2 and 5 is statistically significant at 1%). For given 2003 issues, however, β NC s has no predictive power on 2004 issues of FRMs. All in all, high β NC s MSAs received much more interest-only loans, but the same amount of FRMs, consistent with prediction 3. The effect on ARMs is intermediate, consistent with evidence of Section Robustness: An other measure of comovement between loan returns and firm survival The tests of predictions 2 and 3 rest on our ability to measure the parameter β NC s through the past comovement between local home prices and the prices of NC s underlying assets. One possible concern with this approach is that past price movements over may not be helpful to predict future loan return comovements as anticipated in Another concern is that loan weighted home price changes may not be a good measure of NC s asset returns. 21

22 To adress these issues, we propose here an alternative measure that does not depend on how we model NC s survival process. Going back to Section 2.2, we do not write equation (3). Without this assumption, the value of loan i for the shareholder becomes: E( S. R i,s ) = pa i + b i pe( g s S = 1). (9) The crucial variable is pe( g s S = 1) which takes larger values in states where home price growth conditional on NC survival is high. To measure pe( g s S = 1) empirically, we take inspiration from rational bubble models (e.g. Blanchard and Watson, 1982). Assume there is a rational property bubble in state s, and that this bubble is expected to burst if S = 0 (we implictly assume that NC goes bust when the bubble bursts). For this bubble to be rational, expected price growth in state s conditional on S = 1 (bubble does not burst) has to be larger than in non-bubbly states. If this fails to hold, no one would want to hold property in the bubbly state s. Hence, pe( g s S = 1) can be interpreted as a measure of state bubbliness of property prices. We use MSA-level home price to income ratio as our measure of pe( g s S = 1). We report the results of our tests in Table 7. In column 1, we test prediction 2, whereby NC s total issuance should be bigger in MSAs with the biggest price to income ratio. Controlling for financial literacy and geographical loan allocation in 2003, price to income comes out statistically significant at 1%. Its effect is also quite large: a 1 s.d. increase in price to income (1.9) leads to a 0.25 increase in total lending to the MSA (about 15% of a sample s.d.). In columns 2-4, we retest prediction 3, by regressing the local loan issue of each type on the MSA s bubbliness. We find that bubbly MSAs tend to receive significantly more interest-only loans than FRMs or ARMs. 17 Hence, using this alternative measure of comovement between loal returns and NC s survival, we find results consistent with previous tests of predictions 2 and This particular piece of evidence is consistent with Barlevy and Fischer (2010) findings They interpret it as the evidence that HH use interest-only loans to speculate. Both interpretations are not mutually exclusive. 22

23 6. Concluding remarks This paper has provided forensic evidence on the risk-shifting behavior of a large mortgage originator. The sharp rise in interest rate in 2004 destroyed a large fraction of New Centurys net present value. In reaction, New Century drastically modified its business model. It started originating loans with a larger exposure to real estate price risk interest-only loans. It changed the geography of its operations selling more and more of these new loans in cities with real estate prices correlated with NC legacy assets. This new business strategy is consistent with a model of a risk-shifting by a financially distressed company who starts taking long bets on its own survival. Our paper may have implications for monetary policy. In response to a heating real estate market, policy makers thought in 2004 that increasing interest rates was the appropriate response. Our paper suggests this decision had unintended consequences: by pushing mortgage originators closer to financial distress, the monetary tightening led mortgage originators to increase risk-taking. In the case of New Century and probably of other originators who held large amounts of loans on their balance sheet risk-taking meant enhancing exposure to real estate price risk. This may well have fuelled the real estate bubble and eventually accentuate the burst of this bubble. 23

24 7. References Acharya, Schnabl and Suarez, 2010, Securitization without risk transfer, working paper NYU Allen and Gale, 2000, Bubbles and Crises, The Economic Journal, vol 110, pp Almeida and Philippon, 2007, The Risk-Adjusted Cost of Financial Distress, Journal of Finance, vol 6, pp Akerlof and Romer, 1993, Looting: the underworld of bankruptcy and profit, Brookings Papers on Economic Activity, vol 2, pp 1-73 Andrade G. and Steven N Kaplan, 1998, How costly is financial (not economic) distress? Evidence from highly leveraged transactions that became distressed, The Journal of Finance, 53(5), pp Barlevy and Fisher, 2010, Mortgage choices and housing speculation, working paper Becker, B. and Stromberg, P., 2010, Equity-Debtholders conflicts in capital structure, mimeo HBS and SIFR Berndt, Hollifield and Sandas, 2009, The role of mortgage brokers in the subprime crisis, NBER WP Biais Bruno, Jean-Charles Rochet & Paul Woolley, 2010, Innovations, rents and risk, The Paul Woolley Centre, Working Paper Blanchard, O. and Watson, M., 1982, Bubbles, Rational Expectations and Financial Markets, In Paul Wachtel (ed.), Crises in the Economic and Financial Structure. Lexington Books, 1982, pp Bostic, R., Engel, K., McCoy, P., Pennington-Cross, A., and Wachter, S., 2008, State and local Anti-Predatory Lending Laws: The effect of Legal Enforcement Mechanisms, Journal of Economics and Business, vol 60, pp Cheng, Hong and Scheinkman, 2010, Yesterday s heroes: Compensation and Creative Risk-Taking, NBER WP Demyanyk and Van Hemert, 2009, Understanding the subprime mortgage crisis, Review 24