Speculative Fever: Micro Evidence for Investor Contagion in the Housing Bubble PRELIMINARY

Transcription

1 Speculative Fever: Micro Evidence for Investor Contagion in the Housing Bubble PRELIMINARY Patrick Bayer Kyle Mangum James Roberts April 24, 2013 Abstract This paper examines the spread of speculative investing, or contagion, by homeowners in the recent housing bubble. Using detailed housing transaction records, we estimate the impact of speculative activity by one s neighbors and in one s neighborhood on subsequent real estate investment behavior and performance. Our research design, which isolates the impact of immediate neighbors relative to those on nearby blocks, controls for a host of potential issues that might create spurious correlation in neighbors investment activities. We find evidence of strong spillovers within neighborhoods: homeowners were much more likely to engage in speculative activity both after a neighbor had successfully flipped a home and when a home had been successfully flipped in their neighborhood. Social contagion brought amateur real estate investors into the market at an increasing rate during the boom, with their share reaching a record high just as the market reached its peak, bringing substantial equity losses, defaults and foreclosures in the subsequent crash. Keywords. Real estate bubble; investor contagion; social interactions; house flipping 1 Introduction In this paper, we examine whether new real estate investors are influenced by investing activity occurring in their own neighborhoods. While it has been hypothesized that there was a contagion effect to real estate investing during Duke University Georgia State University. Correspondence to: kmangum@gsu.edu, or Box 3992, Atlanta, GA Duke University 1

2 the housing boom, we know of no empirical evidence for its existence. In fact, we know of no micro evidence for an animal spirits effect to any recent investment bubble. Using the spatial nature of housing, this paper offers direct evidence for a role of social interactions in investment behavior. We find positive and significant effects of nearby investing activity on an individual s propensity to engage in investing behavior. Observing a flipped property is one s neighborhood or having an investor as a neighbor increases one s propensity to engage in investing by about 10 to 15 percent. Such suggestion mechanisms may have contributed to the bubble, forming a propagation mechanism that accelerated price increases. Notably, it appears amateur investors, who transacted fewer properties at lower rates of return, were more susceptible to nearby influence. 2 Data 2.1 Background Description The dataset used in the analysis is a detailed register of housing transactions in the greater Los Angeles metropolitan Area 1 from 1988 to This is a proprietary collection of public record data which was purchased from real estate research firm DataQuick Information Systems. 2 Properties contain full geographic information, including latitude/longitude, street address census designations, and a unique identification number. The transactions can be readily merged with 2011 county tax assessor data (also from DataQuick) that contains property attributes. Using the transactions information on property liens, a substantial subset of these data can be further merged with public data on reporting for the Home Mortgage Disclosure Act (HMDA) to attach information on purchaser/borrower income and race. 2.2 Designation of Investing Activity Importantly, the transactions contain the names of the buyer and seller and the sale closing date, and we use this information to construct a tenure profile for each individual name. 3 We designate a property tenure ending when we observe a transaction in which the name of the initial buyer is listed as seller. Then, for each unique name name, we construct a property purchase history with this tenure spell information. The first property which is held for more than two years is designated as the person s home. Home is an important 1 Los Angeles, Orange, Riverside, San Bernardino, and Ventura counties. 2 In 2013 DataQuick was acquired by CoreLogic. 3 Names are detailed, typically including middle initials and often names of spouse/coborrowers, assuaging concerns that we attribute a single profile to two different people of the same name. We have reviewed names in transactions to exclude purchasers by businesses, nonprofits, and various levels and agencies of the government. However, we recognize that despite our best efforts the data will contain some degree of measurement error. We have taken care to be conservative, when in doubt preferring to label a profile a non-investor. 2

3 destination, as we use the location of home as the center for a circular spatial match of investing activity in the person s neighborhood. A profile is at-risk of becoming an investor until the sooner of entry to investment activity (defined below), or the sale of the home. We then identify property investment activity in two non-mutually exclusive ways. First, if we observe the named buyer purchasing another property without selling their existing home, we designate the second home as an investment property, and the purchase date of the second home is the investor s entry to investing activity. 4 Second, we identify a set of transactions as quick sales in which a property is held for less than two years. This short tenure is indicative of the house flipping type of investment activity. Since there may be other explanations for short tenure, 5 we designate only a person who is observed to do this two or more times in our dataset as a flipper, with the purchase of the first quick sale being the flipper s entry date. A flipper can be designated without observation of a home location, since the flipper designation uses only the repeated name in quick sale activity, and not the tenure profile itself. The investor designation process is illustrated in Figure 1. The person with this tenure profile would be designated as an investor entering in year t = 2, and as a flipper entering in year t = 6. However, had the second flip in year t = 8 not been observed, the quick sale in year t = 6 would not be considered a flip. There is significant overlap in the two measures, though given our two flip restriction, the second home designation is much more common than the flipper designation. For persons in which we have identified a home location, flipping is essentially a proper subset of second home investing. Since we observe property transactions only, we do not know the owner s intentions or actions once while the property is owned. Therefore, we do not distinguish between (1), a house held vacant for purposes of price appreciation, (2) a property rented, possibly while being held for price appreciation, or (3) a property being renovated and resold. Figure 2 plots the various measures of investment activity in terms of transaction volume. Note the similarity in the time series each designation is clearly varying strongly with the housing cycle, with an upward trend during the late 1990s to late 2006, and a stark drop thereafter.. Since a home property must be observed before a person can be identified as an investor, there is a mechanical trend to the investment activity designation. However, even a detrended series exhibits the same cyclical features. Figures 3 and 4 display the time series of investor entry behavior, respectively, as counts and in terms of a hazard rate. That is, Figures 3 and 4 display the investing activity at the person level, while Figure 2 displays investing activity at the property level. Here again, there is a strong upward trend during the period of house price appreciation from the late 1990s to late 2006, with the 4 We allow for a six month overlap to account for housing search. Any person who has only one investment property, and it was held for an overlap with home for one year or less is also excluded from the investor destination. 5 job relocation, change in martial status, etc 3

4 Figure 1: Illustration of Designation of Investors Inv"Prop"1" Flip"1" Flip"2" Home" 0" 2" 4" 6" 8" 10" 12" 14" t" falloff thereafter. 2.3 Spatial Match of Investing Activity Using the designations of at-risk tenures, second-homes, and flips, and their dates and locations, we conduct spatial matches to create measures of nearby investment activity that a person may observe in his/her neighborhood. Using latitude and longitude information, we match all at-risk tenures to (1) flips and (2) investors. A match occurs when a property A is within a distance ring of X of the center property, B, with each home forming the center of its own circle. The two separate matches are each designed to reflect a particular information diffusion mechansim. The flipped property match is spatially to an at-risk home by the property location, with the flip s sale date being the relevant timing. That is, flips which occur outside the at-risk tenure are not counted. The rationale for this match is that a person who is at-risk of becoming a real estate investor may observe properties being flipped in his/her neighborhood and may consider entering real estate investing themselves. For the property match, we choose only the more selective flips designation, since a person is more likely to notice the more frequent transactions than a property being held as an investment for an extended period of time. 6 The investors homes *i.e. not the properties they transact) are matched spatially to at-risk homes, with the investor s entry date the relevant timing. 6 A person may observe information from quick sales as well, but we view this as a conservative classification of the treatment. 4

5 Figure 2: Investment Activity in Southern California, " 12000" 10000" Transac2on"Count" 8000" 6000" 4000" 2000" 0" 1990" 1991" 1992" 1993" 1994" 1995" 1996" 1997" 1998" 1999" 2000" 2001" 2002" 2003" 2004" 2005" 2006" 2007" 2008" 2009" 2010" 2011" Quick"Sales" Flips" Investment"Proper2es" Investment"Proper2es"(detrended)" NOTES: Quick sales are properties held for less than two years. Flips are quick sales executed by a buyer who has at least two quick sales. Investment properties are any properties held without the buyer selling his/her/their previous property. The rationale for this type of match is a more typical social interactions story: a person gets information, suggestion, and encouragement from one s neighbors. Table?? displays the overlap in these measures. Because of the difference in interpretation of the underlying mechanisms, we conduct analyses of these two types of treatments. Note that the same individual property or person may appear multiple times in the adjacency matrix. For example, for three homes next to one another, an investor in home A will be counted as a potential influence on both of his neighbors, B and C. 3 Research Design We want to examine the extent to which real estate investing behavior diffuses spatially. That is, we want to compare a treated at-risk tenure (one with investing/investor behavior occurring in its neighborhood), to a control with with no nearby investing, and measure their respective propensities to become real estate investors. Of course, one may be concerned that because people choose their neighborhoods, they are sorting into areas with similar people, and any spatial correlation would be attributable to a selection effect, possibly unobserved. We propose a 5

6 Figure 3: Entry to Investment Activity (Counts), " 600" 4500" 4000" 500" Investor"Entry"Count" 3500" 3000" 2500" 2000" 1500" 400" 300" 200" Flipper"Entry"Count" 1000" 500" 100" 0" 1990" 1991" 1992" 1993" 1994" 1995" 1996" 1997" 1998" 1999" 2000" 2001" 2002" 2003" 2004" 2005" 2006" 2007" 2008" 2009" 2010" 2011" 0" Investor"Entry"(le="axis)" Flipper"Entry"(right"axis)" research design to address this concern. Using our spatial match, we split the treatment into inner and outer rings of activity. We will measure the effect of an inner ring a hyperlocal effect conditioning on a larger rings. This technique is illustrated in Figure 5. This design is an appeal to search frictions present in a characteristically thin market such as housing. While one may be able to choose a neighborhood, one is limited from choosing exact locations by the list homes for sale when one is searching. Thus, while sorting may happen at a broader neighborhood level, search frictions impede it from happening at a block-by-block level. In practice, we will measure the effect of activity within 0.10 of a mile, conditioning on 0.30, 0.50, and/or 1.0 mile rings, as well as ZIP code dummies. Of course, as is typical with social interactions, we cannot actually observe how information is dispersed, processed, or diffused. While we believe we can plausibly identify the impact of spatial interaction, we are left to infer a mechanism of influence that would be consistent with our definitions of investing and their spatial patterns. The research design will work if there is no block-by-block sorting. We examine this through robustness checks in which we measure the inner and outer rings to be most similar on a number of observables. As an initial check, Table?? below displays a summary of differences between the innermost (0.10 mile) and 0.30 mile rings in a number of attributes. One can see that inner/outer ring differences are slight when compare to metro wide variation. For all attributes, the 95th percentile the homes with the largest differences between the 0.10 and 0.30 mile rings display differences well below the metro wide stand deviation. 6

7 Table 1: Absolute Differences in Home Attributes Between <0.1 mile and mile Rings No. Rooms Living Area No. Beds. Year Built Pct Equity Value Pct Value (sqf) (1-initial LTV) ($2000) Metro- Wide SD ,316 Differences x i,0.1 x i, mean p , p , p , p p , p , p , Abs Diffs x i,0.1 x i, mean , p , p , p , p , p , p , p , NOTES: The statistics reported are for the distribution of d i = x i,0.1 x i, , i.e. the absolute difference between the inner ring (< 0.1 mi) and outer ring ( mi) averages for each house i. 7

8 Figure 4: Entry to Investment Activity (Probability), " 0.03" 1.000" 0.025" Investor"Entry"Probability"(%)" 0.800" 0.600" 0.400" 0.02" 0.015" 0.01" Flipper"Entry"Probability"(%)" 0.200" 0.005" 0.000" 1990" 1991" 1992" 1993" 1994" 1995" 1996" 1997" 1998" 1999" 2000" 2001" 2002" 2003" 2004" 2005" 2006" 2007" 2008" 2009" 2010" 2011" 0" Investor"Entry"(leA"axis)" Flipper"Entry"(right"axis)" 4 Results Given the similarity in the investment measures, and to avoid an unwieldy number of results, we focus discussion on the entry of individuals to investing behavior. Entry is defined as the first of either second home investment behavior or flipping. The relevant entry date is the purchase of the first secondhome/flipped property. We then focus on two types of treatment, one at the person level and one at the property level. At the property level is flipping activity, with the sale of the flip being the relevant date and the property location the relevant spatial designation. At the person level is investment activity of either type, with the investor s first entry to investing the relevant date and the investor s home location. That is, we investigate two potential treatments: the neighborhood (structures) and the neighbors themselves. We note that, when conducted separately, our results for different designations of investing activity have yielded quantitatively very similar conclusions. The investment activity is measured by counts of flips or counts of investor within given rings. 4.1 Summary Statistics We begin by presenting simple differences in means in Table 2. Nearby activity is measured at the 0.10 mile ring, and there is a panel for each type of treatment effect (flipped properties in Panel A, investors homes in Panel B). The atrisk tenures are split into those who become investor and this who do not, presented in three-year intervals. In all years, investors have more investing activity occurring in their neighborhoods, evidenced by differences in means 8

9 Figure 5: Illustration of Inner/Outer Ring Research Design ZIP * * * * * * * * * * * * * Neighbors* * * * Close*neighborhood* * * Neighborhood* * * * * * ** and in the frequency of zeros (that is, the differences in means are not being driven by a few with large a pockets of activity). We further divide investors into those who have already entered and those who will enter in the future; that is, the at-risk tenures who are active in the interval but do not enter in the interval or beforehand. Note that those who have yet to invest have nearby activity more similar to non-investors. This suggests that these are not permanently different neighborhoods or individuals, and that influence or contagion may be occurring. 4.2 Baseline Specifications Table 2 is illustrative, but not a formal test. We know turn to regressions exploiting our inner/outer ring research design. Our level of observation is the monthly at-risk tenure for each, which we define as an active tenure in which the owner has not yet engaged in investing activity (of any type). An at-risk tenure is ended by sale of the property or entry into investing activity. Our primary specification is a linear probability hazard regression. In each monthly observation, the at-risk tenure may be ended by 0.4* 0.6* 0.8* 1* 1.2* 1.4* 1.6* 9

10 entry to investing (an indicator variable outcome); we measure the extent to which recent nearby investing activity (measured by counts of flips/investors nearby) is correlated with such entry. We focus on the period , which bookends the period of house price appreciation. In this hazard specification, there is both spatial and temporal variation in the level of housing activity. Thus, we can identify an effect by comparing two at-risk tenures, one with investors neighbors and one without, or by comparing the propensity of an at-risk individual to enter when there has been recent investing activity to a period when there was not. We note from the outset that entry is an uncommon event. Over the entire period of , the average monthly entry rate was 0.07 percent. Before proceeding, a note on nomenclature. The explanatory variables are expressed as wixx t, where wi stands for within, the XX is the distance ring in hundredths of a mile, and t refers to the length of time. For example, wi10 1 refers to activity within 1/10 of a mile of the at-risk tenure, occurring within the last year (12 months up to and including the current month). The rings are defined inclusively, so that any activity in the inner ring is also measured in the outer ring. Thus, coefficients can be interpreted as the additional impact of the inner ring beyond its average impact of being included in the outer ring. Table 3 reports results from our main specifications. It includes results from each type of treatment flips and investors. Coefficient estimates are followed by hazard ratios, the change in propensity to enter attributable to the explanatory variable(s). In column 1, we see there is a positive and significant effect of activity within 0.10 mile on the propensity to enter as a real estate investor in a given month. Measured as a percentage increase in the baseline hazard, having a flipped property in one s immediate neighborhood increases the propensity to enter in a given month by 17 percent, while having an investor neighbor increases the propensity by nearly 12 percent. Column 2 utilizes our inner/outer ring reach design. Controlling for the broader neighborhood reduces the coefficient somewhat, but the propensity effects are of similar magnitude. Column 3 adds additional rings, while columns 4 and 5, respectively, add year-quarter and ZIP code dummies. None of these substantially changes the estimated effect. Column 6 uses an indicator variable for the inner ring instead of a count, again to little effect. Columns 7 and 8 widen the inner ring to larger radii. The effect size drops considerably, suggesting that the effect gets more diffuse as the neighborhood widens. Henceforth we use specification 5 as our baseline. Table 4 examines how the effect size changes over time by running the hazard regressions separately for three year intervals. The hazard ratios are also conducted separately, since the baseline entry rate changes over time. Despite an increase in the baseline rate, the effect of flip activity (Panel A) within 0.10 is strongest during the periods of largest house price appreciation; the effect size roughly tracks with the price cycle. The effect of a nearby investor (Panel B) is more constant over the time period. 10

11 4.3 Robustness Tables 5 through 7 contain robustness checks of our main specifications. Tables 5 and 6 impose a stricter standard on the inner/outer ring method, subsetting the data to tenures that exhibit the least amount of difference between the mile and mile rings in a number of property and purchaser attributes. This checks whether the results are subject to hyperlocal sorting, in violation of our research design assumptions. Table 5 uses tenures below the median inner/outer difference (hence using half the data with the most similarity). Table 6 uses tenures with differences smaller than 1/10 of the metrowide standard deviation in the attribute. Table 7 checks the sensitivity of our results to our ability to identify at-risk tenures in the data. We have inferred whether the purchaser of the home was at-risk. The data contain two sources additional information on whether the property was owner-occupied. First, the HMDA data match includes a flag for whether loan application was for an owner-occupied home; column (1) uses only at-risk tenures that contain the owner-occupied flag. Second, the assessor data match includes information on the owner s home mailing address; when this is the same as the property address, the home is owner-occupied. Column (2) uses only at-risk tenures with this owner occupied flag. Note that because the assessor data comes from 2011, only tenures that persist into this assessment year will be included, hence limiting the data to later purchases. Effect sizes are marginally smaller when limiting the data to at-risk tenures with the least differences between rings, ranging from 5-12 percent instead of 10 to 17. This suggests that some sorting may still be occurring, even at the block level. However, even at this most conservative estimate, the effect sizes are still significant and economically meaningful. 4.4 Amateurism While the preceding realists indicate spatial interaction is occurring, we have not yet addressed whether the information mechanism is useful or harmful. A contagion story seems to presume some type of harmfulness, with amateurs being left for fools after imitating others. It is admittedly difficult to answer such a normative question, but we attempt to unpack the question of whether nearby investing activity has heterogenous effects on different types of investors. First, we will test separately whether the treatments has heterogeneous effects on professional investors versus amateurs. 7 Then we will measure the effect of nearby investing activity of various measures of investing success. We need a working definition of amateurism. Since we cannot observe investors intentions when he/she enters, we are left to infer amateurism ex post from behavior. We work under two definitions: first, that a professional is an investor who buys at least four investor properties and a novice less, displayed in Table 8, and (2) a professional has an investing careerthe time from first 7 Professional still means an individual identified by name, since business entities were excluded from the analysis. 11

12 investment property purchase to lastof at least two years, and a novice less, displayed in Table 9. The measured effect sizes are clearly and statistically significantly smaller for professionals than for amateurs. However, this is somewhat misleading since their baseline hazard rates are much lower. Looking at the hazard ratios ( HR ), we cannot reject a null that professionals and amateurs propensities are increased at the same rate. An informal comparison of the effect of nearby flips suggests that professionals are less affected by property flipping occurring nearby. This is especially evident in the triennial comparisons, as most of the average effect is driven by late-entering professionals. Perhaps those late entrants who were designated professionals were simply exuberant amateurs. Finally, we measure the effect of nearby investing activity on investing success, conditional on entry. To do so, we run cross-sectional regressions of investors (or, in some specifications, investment properties), comparing those with investment activity near them at the time of entry to those without. We also include specifications at the property level with the investing activity occurring just prior to property purchase (which will be after entry for 2nd and 3rd properties). Results for flipped properties are displayed in Table?? while results for investor neighbors are displayed in??. First we consider the likelihood that the investor becomes a professional. The outcome variable in column 1 is the number of investment properties the investor will purchase in his/her career, and column 2 is the length of the career (defined as above). Column 3 is an indicator for whether the investor becomes a professional by either definition. These all control for entry date and ZIP code dummies. We find that nearby flips, but especially nearby investors, are associated with fewer properties and shorter careers, suggesting that nearby activity fosters amateurism. The remaining columns measure financial success. Column 4 is a regression at the investor level, measuring total nominal earnings from the sale of investment properties. The regression controls for the number of properties transacted over their career. Observing a nearby flip, and especially, having an investor nearby reduces an investor s earnings by thousands of dollars. Columns 5 through 8 are at the property level. Columns 5 and 7 measuring the rate of return on the sale relative to the market rate (measured by the county-level home price index); column 5 uses investing activity prior to the investor s entry, whereas column 7 uses investment activity prior to the purchase of the property. Having nearby activity reduces returns by an average of 1.5 to 3 percent, relative to a mean of 14 percent. Columns 6 and 8 look at the probability that a purchased property was held past the price peak in Having investment activity nearby increases the probability of holding beyond the optimal selling time by 1.5 to 3 percentage points. This suggests investors possibly influenced by their neighborhood were slightly less informed about prices than other investors. These results suggest that those investors with nearby activity more likely to be influenced fare worse than those who invest without a nearby influence. 12

13 Table 2: Summaries of Nearby Investment Activity, by Investor Behavior Panel A: Match of Flipped Properties Tenure Group N Mean SD Pct with active in: or more Non-investors 1,065, Investors, all 85, by entry year , > , Non-investors 1,137, Investors, all 90, by entry year , > , Non-investors 1,110, Investors, all 90, by entry year , > , Non-investors 1,085, Investors, all 84, by entry year , > , Panel B: Match of Other Investors primary residences Tenure Group N Mean SD Pct with active in: or more Non-investors 1,065, Investors, all 85, by entry year , > , Non-investors 1,137, Investors, all 90, by entry year , > , Non-investors 1,110, Investors, all 90, by entry year , > , Non-investors 1,085, Investors, all 84, by entry year , > , NOTES: Each panel reports the flipping activity within a 0.1 mile radius of the at-risk tenure (a primary residence). The at-risk tenures are split into non-investors, who are never identified to engage in flipping/investment activity, and those that do, who are subdivided by time of entry. 13

14 Table 3: Linear Probability Hazard Models, Panel A: Match of Flipped Properties (1) (2) (3) (4) (5) (6) (7) (8) wi *** 5.02e-05*** 7.04e-05*** 7.82e-05*** 7.50e-05*** (7.63e-06) (8.60e-06) (8.64e-06) (8.67e-06) (8.67e-06) wi e-05*** 5.57e-05*** 3.67e-05*** 2.19e-05** 2.14e-05** (7.79e-06) (8.77e-06) (8.83e-06) (8.86e-06) (8.86e-06) wi e-05*** 1.85e-05*** 1.70e-05*** 1.63e-05*** 1.96e-05*** (3.23e-06) (3.60e-06) (3.64e-06) (3.64e-06) (3.58e-06) wi e-05*** 5.82e e e e-06 (3.20e-06) (3.61e-06) (3.64e-06) (3.64e-06) (3.58e-06) wi e-05*** 3.15e-06*** -7.65e-06*** -7.81e-06*** -6.84e-06*** -9.55e-06*** (9.11e-07) (9.75e-07) (1.02e-06) (1.02e-06) (9.77e-07) (1.17e-06) wi e-05*** -3.04e-05*** -2.77e-05*** -2.78e-05*** -2.79e-05*** -2.87e-05*** (8.80e-07) (9.33e-07) (9.59e-07) (9.57e-07) (9.10e-07) (1.12e-06) d wi e-05*** (1.05e-05) d wi e-05* (1.07e-05) wi e-05*** (4.95e-06) wi e-05*** (4.97e-06) wi e-05*** (2.41e-06) wi e-06*** (2.40e-06) Constant *** *** *** *** *** *** *** -4.48e-05* (2.60e-06) (2.99e-06) (3.64e-06) (4.81e-06) (4.14e-05) (4.36e-05) (2.76e-05) (2.71e-05) ZIP dummies Y Y Y Y Y Year-qtr dummies - Y Y Y Y Joint Est ( ) ( ) Inner Ring hazard ratio ( ) ( ) ( ) ( ) ( ) ( ) ( ) ( ) Joint hazard ratio ( ) ( ) ( ) ( ) ( ) ( ) ( ) Observations 109,204, ,204, ,204, ,204, ,204, ,204, ,204, ,204,253 Panel B: Match of Investors Residences (1) (2) (3) (4) (5) (6) (7) (8) wi e-05*** 4.33e-05*** 7.04e-05*** 7.94e-05*** 8.15e-05*** (4.63e-06) (5.29e-06) (5.46e-06) (5.49e-06) (5.49e-06) wi e-06** 3.79e-05*** 7.25e e e-06 (4.66e-06) (5.36e-06) (5.62e-06) (5.65e-06) (5.65e-06) wi e-05*** 1.23e-05*** 1.38e-05*** 1.47e-05*** 2.37e-05*** (1.97e-06) (2.70e-06) (2.71e-06) (2.71e-06) (2.65e-06) wi e-05*** -2.38e e e e-06 (1.95e-06) (2.75e-06) (2.77e-06) (2.77e-06) (2.65e-06) wi e-05*** 9.67e e-05*** -1.44e-05*** -1.12e-05*** -2.15e-06* (1.51e-06) (1.55e-06) (1.58e-06) (1.58e-06) (1.33e-06) (1.17e-06) wi e-05*** -3.70e-05*** -3.81e-05*** -3.68e-05*** -3.98e-05*** -3.40e-05*** (1.51e-06) (1.54e-06) (1.56e-06) (1.54e-06) (1.29e-06) (1.13e-06) d wi e-05*** (7.29e-06) d wi e-05*** (7.47e-06) wi e-05*** (3.28e-06) wi e-05*** (3.34e-06) Constant *** *** *** *** -4.13e *** 1.50e e-05 (2.76e-06) (3.44e-06) (3.89e-06) (4.50e-06) (4.56e-05) ( ) (3.01e-05) (1.41e-05) ZIP dummies Y Y Y Y Y Year-qtr dummies Y Y Y Y Notes: All specifications Joint Est ( ) ( ) Inner Ring hazard ratio ( ) ( ) ( ) ( ) ( ) ( ) ( ) Joint hazard ratio ( ) ( ) ( ) ( ) ( ) ( ) ( ) Observations 109,869, ,869, ,869, ,869, ,869, ,869, ,869, ,869,535 Robust standard errors in parentheses *** p<0.01, ** p<0.05, * p<0.1 include 109,868,535 tenure-month observations, within which there are 2,182,550 separate tenures. Standard errors are clustered at the tenure level. Joint Est refers to the total effect of a flipped property or investor entry, i.e. the sum of the inner and outer rings. The hazard ratio is the ratio of the coefficient to the baseline propensity, i.e. the constant from a regression with the activity rings but no dummies. 14

15 Table 4: Linear Probability Hazard Regression Results, Triennially, Panel A: Flipped properties as explanatory variables. (1) (2) (3) (4) Entries 19,728 24,662 33,745 6,345 At-Risk Tenures 34,349,549 41,343,524 41,626,274 37,442,584 Entry Rate (1) (2) (3) (4) wi e-05** 3.49e-05*** *** 4.94e-05*** (1.56e-05) (1.31e-05) (1.43e-05) (1.84e-05) wi e-05** 3.51e-05** 1.47e e-05 (1.86e-05) (1.49e-05) (1.49e-05) (1.21e-05) wi e-05** 1.26e-05** 1.93e-05*** 9.42e-06 (6.50e-06) (5.77e-06) (5.89e-06) (7.38e-06) wi e e e e-06 (7.54e-06) (6.01e-06) (6.23e-06) (4.88e-06) wi e-06*** -1.37e-05*** -1.23e-05*** 1.53e-06 (1.83e-06) (1.70e-06) (1.72e-06) (2.15e-06) wi e-06*** -1.19e-05*** -4.76e-05*** -1.10e-05*** (1.94e-06) (1.54e-06) (1.76e-06) (1.33e-06) Constant *** *** *** *** (1.52e-05) (1.31e-05) (1.54e-05) (1.62e-05) Inner Ring hazard ratio ( ) ( ) ( ) ( ) Joint hazard ratio ( ) ( ) ( ) ( ) Observations 34,349,549 41,343,524 41,626,274 37,442,584 ZIP dummies Y Y Y Y Year-qtr dummies Y Y Y Y Robust standard errors in parentheses *** p<0.01, ** p<0.05, * p<0.1 Panel B: Investors primary residences as explanatory variables. (1) (2) (3) (4) wi e-05*** 6.67e-05*** *** 3.73e-05*** (1.01e-05) (8.95e-06) (8.72e-06) (9.39e-06) wi e e-05** -2.42e-05*** 1.19e-05 (1.12e-05) (9.90e-06) (9.13e-06) (8.73e-06) wi e e-06** 2.14e-05*** 8.89e-06* (5.13e-06) (4.42e-06) (4.27e-06) (4.76e-06) wi e-05*** -8.12e-06* 6.97e e-06 (5.66e-06) (4.90e-06) (4.48e-06) (4.23e-06) wi e-05*** -2.19e-05*** -1.82e-05*** -1.68e-05*** (2.98e-06) (2.58e-06) (2.51e-06) (2.76e-06) wi e-05*** -1.91e-05*** -5.93e-05*** -1.58e-05*** (3.16e-06) (2.78e-06) (2.57e-06) (2.39e-06) Constant *** *** *** *** (1.54e-05) (1.25e-05) (1.50e-05) (1.55e-05) Inner Ring hazard ratio ( ) ( ) ( ) ( ) Joint hazard ratio ( ) 15( ) ( ) ( ) Observations 34,454,494 41,526,947 41,932,953 37,723,539 ZIP dummies Y Y Y Y Year-qtr dummies Y Y Y Y Robust standard errors in parentheses *** p<0.01, ** p<0.05, * p<0.1 NOTES: The outcome is whether the at-risk homeowner enters the investment market; i.e. begins to engage in either flipping or investing activity. The notation wix0 t refers to the flipping activity within X tenths of a mile t year(s) ago. For instance, wi30 2 refers to

16 Table 5: Regression Results for Properties Below the Median Difference in Each Attribute (1) (2) (3) (4) (5) (6) (7) Attribute value initial equity income race(pct nonwhite) size(sqft) no. beds year built Entries 36,385 35,119 35,195 35,230 35,966 35,146 33,530 Panel A: Flipped properties At-Risk Tenures 57,724,139 53,581,994 52,220,836 51,952,659 53,274,515 54,187,590 54,426,920 Entry rate as explanatory variables. (1) (2) (3) (4) (5) (6) (7) Attribute value initial equity income race(pct nonwhite) size(sqft) no. beds year built wi e-05*** 5.61e-05*** 5.63e-05*** 6.77e-05*** 5.03e-05*** 4.30e-05*** 7.27e-05*** (1.20e-05) (1.19e-05) (1.24e-05) (1.18e-05) (1.26e-05) (1.27e-05) (1.21e-05) wi e e-05* 1.67e e e-05* 6.07e e-05 (1.20e-05) (1.21e-05) (1.26e-05) (1.19e-05) (1.29e-05) (1.30e-05) (1.25e-05) wi e-05*** 2.21e-05*** 2.52e-05*** 1.52e-05*** 1.57e-05*** 2.19e-05*** 1.66e-05*** (4.73e-06) (5.09e-06) (5.04e-06) (5.05e-06) (5.06e-06) (5.15e-06) (5.06e-06) wi e e e e e e e-06* (4.73e-06) (5.05e-06) (4.99e-06) (5.04e-06) (5.09e-06) (5.14e-06) (5.08e-06) wi e-06*** -9.08e-06*** -1.05e-05*** -7.46e-06*** -8.01e-06*** -9.59e-06*** -7.95e-06*** (1.32e-06) (1.42e-06) (1.42e-06) (1.43e-06) (1.42e-06) (1.43e-06) (1.44e-06) wi e-05*** -2.80e-05*** -2.83e-05*** -2.79e-05*** -2.79e-05*** -2.78e-05*** -3.04e-05*** (1.25e-06) (1.33e-06) (1.33e-06) (1.34e-06) (1.33e-06) (1.33e-06) (1.37e-06) Constant *** *** *** -3.42e *** *** *** (3.43e-05) (6.13e-05) (4.15e-05) (4.80e-05) (4.25e-05) (4.44e-05) (4.88e-05) Inner Ring hazard ratio ( ) ( ) ( ) ( ) ( ) ( ) ( ) Observations 57,724,139 53,617,113 52,220,836 51,987,889 53,274,515 54,187,590 54,426,920 Panel B: Investors primary residences as explanatory variables. (1) (2) (3) (4) (5) (6) (7) Attribute value initial equity income race(pct nonwhite) size(sqft) no. beds year built wi e-05*** 8.43e-05*** 6.89e-05*** 7.74e-05*** 5.47e-05*** 6.31e-05*** 8.07e-05*** (7.72e-06) (7.71e-06) (7.98e-06) (7.57e-06) (8.30e-06) (8.15e-06) (7.63e-06) wi e-05*** -1.07e e e e-05** -1.41e-05* -5.16e-06 (7.95e-06) (7.95e-06) (8.16e-06) (7.72e-06) (8.44e-06) (8.39e-06) (7.89e-06) wi e-05*** 1.48e-05*** 1.57e-05*** 1.44e-05*** 1.94e-05*** 1.33e-05*** 1.38e-05*** (3.64e-06) (3.76e-06) (3.81e-06) (3.82e-06) (3.89e-06) (3.82e-06) (3.75e-06) wi e e e e e e e-06 (3.71e-06) (3.85e-06) (3.90e-06) (3.87e-06) (3.95e-06) (3.92e-06) (3.80e-06) wi e-05*** -1.63e-05*** -1.79e-05*** -1.36e-05*** -1.71e-05*** -1.48e-05*** -1.12e-05*** (2.12e-06) (2.22e-06) (2.23e-06) (2.26e-06) (2.28e-06) (2.24e-06) (2.20e-06) wi e-05*** -3.53e-05*** -3.86e-05*** -3.65e-05*** -3.98e-05*** -3.81e-05*** -4.01e-05*** (2.13e-06) (2.19e-06) (2.24e-06) (2.22e-06) (2.25e-06) (2.21e-06) (2.16e-06) Constant * *** -5.39e e e ** -1.08e-05 (8.20e-05) (7.42e-05) (5.13e-05) (7.30e-05) (4.58e-05) (5.85e-05) (3.61e-05) Inner Ring hazard ratio ( ) ( ) ( ) ( ) ( ) ( ) ( ) Observations 58,062,732 53,937,856 52,553,742 52,301,538 53,602,003 54,508,915 54,741,633 Robust standard errors in parentheses *** p<0.01, ** p<0.05, * p<0.1 NOTES: The specifications exclude at risk tenures in which di > med(di), where di is defined as in Table (3). 16

17 Table 6: Regression Results for Properties With Difference Less Than One-Tenth Standard Deviation in Each Attribute value initial equity income race(pct nonwhite) size(sqft) no. beds year built Entries 32,654 32,972 44,724 30,054 20,216 17,353 28,487 At-Risk Tenures 51,719,342 50,333,478 66,917,651 44,256,497 29,764,065 26,997,770 46,567,080 Entry rate Panel A: Flipped properties as explanatory variables. (1) (2) (3) (4) (5) (6) (7) Attribute value initial equity income race(pct nonwhite) size(sqft) no. beds year built wi e-05*** 5.03e-05*** 6.43e-05*** 6.87e-05*** 5.17e-05*** 3.65e-05** 6.87e-05*** (1.27e-05) (1.23e-05) (1.09e-05) (1.26e-05) (1.70e-05) (1.83e-05) (1.31e-05) wi e e-05* 1.81e e e e e-05* (1.26e-05) (1.25e-05) (1.12e-05) (1.29e-05) (1.74e-05) (1.88e-05) (1.34e-05) wi e-05** 2.42e-05*** 2.26e-05*** 1.33e-05** 1.41e-05** 1.80e-05** 1.64e-05*** (4.96e-06) (5.24e-06) (4.49e-06) (5.44e-06) (6.75e-06) (7.22e-06) (5.45e-06) wi e e e e e e e-06 (4.99e-06) (5.22e-06) (4.46e-06) (5.45e-06) (6.73e-06) (7.19e-06) (5.45e-06) wi e-06*** -9.71e-06*** -9.52e-06*** -7.59e-06*** -7.83e-06*** -9.45e-06*** -7.85e-06*** (1.38e-06) (1.46e-06) (1.27e-06) (1.55e-06) (1.89e-06) (2.02e-06) (1.55e-06) wi e-05*** -2.79e-05*** -2.91e-05*** -2.76e-05*** -2.74e-05*** -2.70e-05*** -3.05e-05*** (1.32e-06) (1.36e-06) (1.18e-06) (1.44e-06) (1.76e-06) (1.87e-06) (1.48e-06) Constant *** *** *** *** ** 3.89e ** (3.58e-05) (5.63e-05) (4.19e-05) (7.90e-05) (5.58e-05) (5.73e-05) (4.70e-05) Inner Ring hazard ratio ( ) ( ) ( ) ( ) ( ) ( ) ( ) Observations 51,719,342 50,366,450 66,917,651 44,256,497 29,764,065 26,997,770 46,567,080 R-squared Panel B: Investors primary residences as explanatory variables. (1) (2) (3) (4) (5) (6) (7) Attribute value initial equity income race(pct nonwhite) size(sqft) no. beds year built wi e-05*** 8.28e-05*** 7.21e-05*** 7.80e-05*** 5.96e-05*** 6.23e-05*** 8.28e-05*** (8.17e-06) (7.96e-06) (7.01e-06) (8.18e-06) (1.12e-05) (1.17e-05) (8.20e-06) wi e-05*** -7.57e e e e e e-06 (8.41e-06) (8.14e-06) (7.17e-06) (8.35e-06) (1.14e-05) (1.19e-05) (8.47e-06) wi e-05*** 1.51e-05*** 1.60e-05*** 1.49e-05*** 1.36e-05*** 1.49e-05*** 1.35e-05*** (3.85e-06) (3.88e-06) (3.39e-06) (4.14e-06) (5.21e-06) (5.41e-06) (4.05e-06) wi e e e e e e e-06 (3.91e-06) (3.97e-06) (3.45e-06) (4.20e-06) (5.30e-06) (5.47e-06) (4.09e-06) wi e-05*** -1.61e-05*** -1.65e-05*** -1.47e-05*** -1.68e-05*** -1.50e-05*** -1.09e-05*** (2.23e-06) (2.29e-06) (1.99e-06) (2.43e-06) (3.06e-06) (3.16e-06) (2.38e-06) wi e-05*** -3.54e-05*** -4.07e-05*** -3.66e-05*** -4.11e-05*** -3.70e-05*** -4.11e-05*** (2.25e-06) (2.26e-06) (1.97e-06) (2.40e-06) (3.02e-06) (3.11e-06) (2.33e-06) Constant -8.38e *** 6.36e e * 8.67e e-05 (8.16e-05) (8.14e-05) (4.97e-05) (5.12e-05) (6.45e-05) (5.58e-05) (3.81e-05) Inner Ring hazard ratio ( ) ( ) ( ) ( ) ( ) ( ) ( ) Observations 52,020,597 50,666,975 67,335,235 44,526,983 29,944,873 27,153,919 46,834,735 R-squared Robust standard errors in parentheses *** p<0.01, ** p<0.05, * p<0.1 NOTES: The specifications exclude at risk tenures in which di > 0.1 σx, where di is defined as in Table (3), and σx is the metro-wide standard deviation in the attribute. 17

18 Table 7: Regression Results for Properties With Owner-Occupied Designations (1) (2) (3) Owner-occ flag: HMDA or Assessor HMDA Assessor Entries 47,953 39,346 8,607 At-Risk Tenures 76,569,241 56,182,266 20,386,975 Entry rate Panel A: Flipped properties as explanatory variables. (1) (2) (3) Owner-occ flag: HMDA or Assessor HMDA Assessor wi e-05*** 6.93e-05*** 3.57e-05** (1.04e-05) (1.27e-05) (1.66e-05) wi e-05** 3.44e-05** 1.30e-05 (1.08e-05) (1.36e-05) (1.68e-05) wi e-05*** 2.14e-05*** 9.54e-06 (4.34e-06) (5.37e-06) (6.84e-06) wi e e e-06 (4.34e-06) (5.51e-06) (6.46e-06) wi e-06*** -6.92e-06*** -4.41e-06** (1.20e-06) (1.47e-06) (1.94e-06) wi e-05*** -2.25e-05*** -1.42e-05*** (1.12e-06) (1.37e-06) (1.81e-06) Constant -2.16e *** *** (6.46e-05) (7.04e-05) (4.90e-05) Inner Ring hazard ratio ( ) ( ) ( ) Observations 76,569,241 56,182,266 20,386,975 R-squared Robust standard errors in parentheses *** p<0.01, ** p<0.05, * p<0.1 Panel B: Investors primary residences as explanatory variables. (1) (2) (3) Owner-occ flag: HMDA or Assessor HMDA Assessor wi e-05*** 7.87e-05*** 3.20e-05*** (6.60e-06) (8.15e-06) (1.02e-05) wi e e e-06 (6.76e-06) (8.44e-06) (1.07e-05) wi e-05*** 1.62e-05*** 1.23e-05** (3.17e-06) (3.92e-06) (5.03e-06) wi e e e-06 (3.26e-06) (4.10e-06) (5.03e-06) wi e-05*** -1.11e-05*** -9.62e-06*** (1.84e-06) (2.28e-06) (2.89e-06) wi e-05*** -2.94e-05*** -1.83e-05*** (1.83e-06) (2.28e-06) (2.88e-06) Constant *** *** -3.04e-05 (2.70e-05) (2.74e-05) (7.56e-05) 18 Inner Ring hazard ratio ( ) ( ) ( ) Observations 77,093,098 56,547,658 20,545,440 R-squared Robust standard errors in parentheses *** p<0.01, ** p<0.05, * p<0.1 NOTES:

19 Table 8: Effects on Professional and Amateur Investors, as Defined by Number of Investment Properties Transacted (1) (2) (3) (4) Years Pro Entries 4,589 1,566 2, Amateur Entries 67,067 12,346 25,951 28,770 At-Risk Tenures 109,204,253 26,405,140 41,923,877 40,875,236 Pro Entry Rate Amateur Entry Rate Panel A: Flipped properties as explanatory variables. (1) (2) (3) (4) Years Professionals wi e e e e-05*** (2.15e-06) (4.79e-06) (3.75e-06) (3.05e-06) Amateurs wi e-05*** 3.10e-05** 7.65e-05*** 8.95e-05*** (8.38e-06) (1.50e-05) (1.36e-05) (1.50e-05) Observations 109,106,745 26,358,655 41,880,353 40,867,737 Pros Inner Ring hazard ratio ( ) ( ) ( ) ( ) Amateurs Inner Ring hazard ratio ( ) ( ) ( ) ( ) p-value (wi10 pro = wi10 am) (HR pro = HR am) Panel B: Investors primary residences as explanatory variables. (1) (2) (3) (4) Years Professionals wi e-06** 1.05e-05*** 7.25e-06*** 4.17e-06** (1.29e-06) (3.55e-06) (2.35e-06) (1.66e-06) Amateurs wi e-05*** 5.45e-05*** 7.14e-05*** 8.68e-05*** (5.16e-06) (1.08e-05) (8.38e-06) (8.84e-06) Inner Ring hazard ratio ( ) ( ) ( ) ( ) Amateurs Inner Ring hazard ratio ( ) ( ) ( ) ( ) p-values (wi10 pro = wi10 am) (HR pro = HR am) Observations 109,759,538 26,449,450 42,095,528 41,214,560 Robust standard errors in parentheses *** p<0.01, ** 19 p<0.05, * p<0.1 NOTES:

20 Table 9: Effects on Professional and Amateur Investors, as Defined by Length of Time in Investing Activity (1) (2) (3) (4) Years Pro Entries 12,705 4,018 5,642 3,045 Amateur Entries 58,951 9,894 22,570 26,487 At-Risk Tenures 109,204,253 26,405,140 41,923,877 40,875,236 Pro Entry Rate Amateur Entry Rate Panel A: Flipped properties as explanatory variables. (1) (2) (3) (4) Professionals wi e-05*** 3.27e e-05*** 1.93e-05*** (3.72e-06) (8.58e-06) (6.28e-06) (5.31e-06) Amateurs wi e-05*** 2.67e-05** 6.30e-05*** 8.12e-05*** (7.85e-06) (1.34e-05) (1.27e-05) (1.44e-05) Pros Inner Ring hazard ratio ( ) ( ) ( ) ( ) Amateurs Inner Ring hazard ratio ( ) ( ) ( ) ( ) p-values (wi10 pro = wi10 am) (HR pro = HR am) Panel B: Investors primary residences as explanatory variables. (1) (2) (3) (4) Professionals wi e-05*** 1.99e-05*** 2.01e-05*** 1.18e-05*** (2.24e-06) (6.18e-06) (3.99e-06) (2.99e-06) Ameteurs wi e-05*** 4.58e-05*** 5.94e-05*** 7.96e-05*** (4.85e-06) (9.61e-06) (7.80e-06) (8.51e-06) Pros Inner Ring hazard ratio ( ) ( ) ( ) ( ) Amateurs Inner Ring hazard ratio ( ) ( ) ( ) ( ) p-values (wi10 pro = wi10 am) (HR pro = HR am) Robust standard errors in parentheses *** p<0.01, ** 20 p<0.05, * p<0.1 NOTES: