F o r e c l o s u r e s, R e t u r n s, a n d B u y e r I n t e n t i o n s

F o r e c l o s u r e s, R e t u r n s, a n d B u y e r I n t e n t i o n s A u t h o r Herman Donner A b s t r a c t Using data from Stockholm, Sweden, I examine whether foreclosed properties are sold at a discount. This institutional setting differs from the United States and studies of real estate owned properties. The findings show that properties experience an eight times higher turnover subsequent to a foreclosure compared with the general market, indicating that professional buyers are taking advantage of a discount. That a substantial fraction of buyers are found to have bought more than one foreclosed property provides further support towards such a pattern. Holding period returns prior and subsequent to a foreclosure also supports a price discount. When households default on their debt, a typical outcome is that owned real estate enters foreclosure; that is, it is repossessed on behalf of the creditor. If no arrangement that results in the loan becoming current is made, such a scenario ends with a property being sold through a foreclosure sale. The probability of such a mortgage default is typically modeled through an option theoretical framework in which the option to default is in the money when housing equity becomes negative (Jones, 1993; Deng, Quigley, and Van Order, 2000; Ambrose, Capone, and Deng, 2001). In such a setting, a deficit between the outstanding loan and the sales price will cause a loss for the lender. Therefore, sales of foreclosed properties should ideally achieve prices as close as possible to market value so as to minimize losses. In this study, I examine the impact on price and turnover caused by such a sale mechanism using several quantitative methods. A foreclosure sale is associated with substantial transaction costs that influence both parties in a credit relationship. Costs of moving, impact on reputational value such as a lower credit rating, and the stigma of default will affect a borrower s decision to default (Elul et al., 2010). The responsibility of repayment subsequent to a foreclosure also affects borrower default probabilities, with recourse debt decreasing the probability of default compared with non-recourse debt (Ghent and Kudlyak, 2011). Correspondingly, costs facing the lender will influence the decision about whether to foreclose on a delinquent mortgage or reach an agreement with the borrower. Mian, Sufi, and Trebbi (2011) find that lengthier judicial foreclosures decrease the lender s probability of foreclosing on a J R E R V o l. 3 9 N o. 2 2 0 1 7

1 9 0 D o n n e r delinquent mortgage compared with settings with nonjudicial foreclosures. As substantial discounts on the sale price would constitute a major transaction cost, the question of whether or not foreclosure status causes a discount on price is highly relevant for a large number of stakeholders. Despite the relevance of this issue, earlier researchers have paid little attention to defining and explaining a foreclosure sale. If a foreclosure discount is to be attributed to foreclosure status itself, it is important to understand how such a sale mechanism works and differs from unforced sales. Researchers have provided sound reasons for why sales prices are influenced by laws and regulations of sales of foreclosed property, as exemplified by Pennington-Cross (2006) finding larger discounts in U.S. states with judicial foreclosures compared to nonjudicial states. As the regulatory setting varies considerably across the U.S. and even more amongst countries, a foreclosure discount should be viewed in the light of the sale mechanism itself. This study is based on sales of single-family houses sold in the greater Stockholm area of Sweden from 2005 to mid-2014. In Sweden, ownership of foreclosed properties is transferred directly from the initial owner in default to the buyer that won the foreclosure auction, which is conducted in accordance with the Swedish Enforcement Code (1981, p. 774), and administered by a government agency. As the transaction price at such an auction is for the foreclosed property itself, this institutional setting is considerably different from the U.S. where lenders typically acquire the property for the foreclosure amount and subsequently sell the property as so-called real estate owned (REO) property. As noted by Chinloy, Hardin, and Wu (2016), foreclosure sales and REO sales are often referred to as though they are interchangeable terms, despite being considerably different sale mechanisms. The importance of distinguishing between these forms of sales is exemplified by a large number of earlier studies of REO sales attributing large discounts on price to seller s willingness to achieve a quick sale at the expense of maximizing sale price (Shilling, Benjamin, and Sirmans, 1990; Forgey, Rutherford, and VanBuskirk, 1994; Hardin and Wolverton, 1996; Springer, 1996; Campbell, Giglio, and Pathak, 2011). In a Swedish setting, this explanation is unlikely to hold as lenders have limited opportunities to influence the sale process. A benefit of this setting is that variations in discounts across observations are less likely to be caused by unobserved variation in seller incentives (i.e., some sellers preferring a quick sale, while others attempt to maximize the sale price). Besides providing empirical research in a non-u.s. setting, several measures aimed at addressing the issue of endogeneity are applied. Most previous studies use hedonic regression to estimate the impact of foreclosure status on sale price (e.g., Forgey, Rutherford, and VanBuskirk, 1994; Campbell, Giglio, and Pathak, 2011). As property characteristics among foreclosed properties and other unforced sales are likely to diverge (Clauretie and Daneshvary, 2009), a central issue is that the impact of foreclosure status on price is overestimated. Such status is typically associated with unmeasured property characteristics that negatively influence price, causing foreclosure status to proxy for characteristics, such as neglected maintenance or a less attractive location.

F o r e c l o s u r e s, R e t u r n s, a n d B u y e r I n t e n t i o n s 1 9 1 To overcome the difficulties caused by the heterogeneous nature of real estate, in this study I examine the price effect of a foreclosure using the transaction history of properties. Substantial discounts on the price of foreclosed properties would allow buyers of such properties to earn excess returns by quickly reselling them in the general market. Measurement of the rate of turnover subsequent to a foreclosure therefore offers an effective way of assessing whether buyers are taking advantage of a discount on price. The question of whether or not foreclosed properties are priced efficiently is further examined using holding period returns, both prior and subsequent to a foreclosure. Because property characteristics that influence property value will impact the sales price every time the property is sold, this controls for unobserved property characteristics. Lower returns compared with the general market between an unforced sale and a foreclosure will indicate a discount on price, as will excess returns between a foreclosure and an unforced sale. Consistent results for both periods provide robust support for the idea that the difference in returns is attributable to a discount on price due to a foreclosure, as opposed to being a consequence of unmeasured changes in property characteristics between transactions. The remainder of this paper is structured as follows: A theoretical background is provided next. I then describe the data, and discuss the methodological approaches as well as the results. The paper closes with concluding remarks. T h e I m p a c t o f F o r e c l o s u r e S t a t u s o n P r i c e P r e v i o u s M e a s u r e s o f a F o r e c l o s u r e - R e l a t e d D i s c o u n t The impact of a foreclosure on the real estate sale price has been estimated with hedonic regression models in several studies, with most studies finding that properties sold by lenders, so-called REO properties, are sold at considerable discounts. Shilling, Benjamin, and Sirmans (1990), Forgey, Rutherford, and VanBuskirk (1994), and Hardin and Wolverton (1996) estimate REO discounts of 24%, 23%, and 22%, respectively. A considerably smaller discount of 4% to 6% is estimated by Springer (1996), while Carroll, Clauretie, and Neill (1997) find no statistically significant discount at all. More recently, Clauretie and Daneshvary (2009) find that properties with foreclosure status are sold at a 7.5% discount. They state that previous research studies tended to overestimate the price discount because they did not separate a proxy effect from a stigma effect on price the former being attributed to foreclosed properties having unobserved characteristics that decrease property value and the latter being described as an impact on price that is solely attributed to the foreclosure status itself. Campbell, Giglio, and Pathak (2011) estimate the price impact on a foreclosure to be considerably higher at 27%. The aforementioned studies have applied measures such as inclusion of time-on-market in the model and application of spatial J R E R V o l. 3 9 N o. 2 2 0 1 7

1 9 2 D o n n e r econometrics to handle the possibility that foreclosure status is endogenously related to characteristics that negatively influence price. Using an automated valuation model on properties sold over the 2000 2012 period across 16 U.S metropolitan areas, Zhou et al. (2015) estimate an average discount of 14.7% for REO properties. They find that properties in bad condition are sold at larger discounts and that a high concentration of REO sales increases the discount, whereas recent house price appreciation decreases it. Measurement of holding period returns in order to estimate a foreclosure discount offers a way of controlling for unobserved property characteristics. Using a sample of over 12,000 REO properties across the U.S., Pennington-Cross (2006) finds that foreclosed properties on average experience a 22% lower cumulative value appreciation prior to a foreclosure compared with metropolitan area price indices. The author states that the lower returns might be attributable to foreclosures being located in areas experiencing less value appreciation compared with a metropolitan area as a whole. The findings are also in agreement with studies showing that foreclosed properties are sold at substantial discounts. The author does not claim to offer insight into how much of this difference in returns can be attributed to foreclosure status itself. The author also finds that the price discount increases with the time a property spends in foreclosure and that the price discount is higher in states requiring a judicial foreclosure compared with states in which the lender has a statutory right to redeem the property. Other findings include a decrease in discounts with increasing loan-to-value (LTV) ratios and an association between a higher mortgage rate risk premium and a larger discount. Through matching of repeat sales data from across several U.S. metropolitan areas, Harding, Rosenblatt, and Yao (2012) find that during a holding period of roughly seven years, buyers of REO properties earn a 1.4% higher annual return compared with those buying through unforced transactions. This fairly small estimate is an upper bound of a true excess return, as the possible costs of repair that might be associated with foreclosed properties are unknown. When applying a hedonic regression model, Harding, Rosenblatt, and Yao (2012) also find evidence of diverging attribute coefficients regarding foreclosed properties compared with other properties. This finding indicates that those who buy foreclosed properties place different values on property attributes compared with those who buy other properties. Lower sales prices could therefore be partially attributed to diverging pricing functions. Across the examined metropolitan areas, investors account for 14% to 18% of buyers of foreclosed real estate, compared with 4% to 6% in the non-distressed market. Additional support that those who buy foreclosed properties diverge from market participants in the general market is provided by Brasington and Sarama (2008) in their examination of the deeds of real estate transactions so as to detect correlations between sale price and seller characteristics. In addition to a 36% discount on price, foreclosures are found to be associated with lower mortgage rates. This finding indicates that foreclosed properties are bought by more sophisticated buyers.

F o r e c l o s u r e s, R e t u r n s, a n d B u y e r I n t e n t i o n s 1 9 3 Greater insight into the linkage between a foreclosure discount and seller incentives is provided by Chinloy, Hardin, and Wu (2016). They analyze differences in sales prices between foreclosure auctions, REO sales, and subsequent unforced transactions. In the U.S., most foreclosed properties offered at auction are bought by the lender and subsequently sold as REO properties. However, it is possible that the foreclosure auction results in a third-party buyer. Such third-party buyers have received little previous attention by researchers, with most examining REO properties. Using foreclosure data from Miami spanning 2010 to 2013, Chinloy, Hardin, and Wu find that foreclosure auctions at which the property is sold to such a third-party buyer results in a larger discount compared with REO properties, which are in turn sold at lower prices compared with unforced sales. These differences are rational and due to costs taken by the lender that are associated with holding the property and by varying levels of buyer uncertainty regarding property condition. In addition, third-party buyers are found to achieve higher cumulative capital gains as compared to REO properties. The setting of the present paper can be likened to third-party buyers at a foreclosure auction. There is, however, a substantial difference between auctions at which very few properties are sold to bidders other than the lender and the Swedish setting in which the lender does not participate as a bidder. Donner, Song, and Wilhelmsson (2016) examine foreclosures in the Swedish setting, and estimated discounts of 20% 29% of foreclosed apartments and single-family houses after having applied a hedonic spatial Durbin model. M e a s u r i n g H o l d i n g P e r i o d R e t u r n s Although the repeat sales approach requires less knowledge about the physical attributes of the properties, some issues do emerge. The primary concern is that property characteristics might have changed during the holding period. Between an unforced sale and a foreclosure, it is likely that the owner has underinvested in maintenance, therefore lowering the value of the property. The risk of biased results is exemplified by the findings of Wilhelmsson (2008), who estimated an annual depreciation rate of 0.77% for well-maintained properties and 1.10% for properties without indoor or outdoor maintenance over a 20-year period. Because foreclosed properties are likely to have been owned by highly indebted households, Harding, Miceli, and Sirmans (2000) finding that homeowners with high LTV ratios invest less in maintenance provides sound reason to believe that foreclosed properties are less maintained compared with other properties. However, the impact of neglected maintenance should be of significance only when the holding period is fairly long. Because the mean holding period prior to a foreclosure is 4.3 years in the current study, the level of maintenance is unlikely to have had a significant impact on property value. Substantially lower returns prior to a foreclosure compared with those achieved by similar properties sold in the general market should therefore provide strong support that foreclosed properties are sold at a discount. J R E R V o l. 3 9 N o. 2 2 0 1 7

1 9 4 D o n n e r A more difficult problem arises when the returns between a foreclosure and a subsequent unforced transaction are measured. These properties are likely to have been improved in-between the transactions, causing an overestimation of the earned excess returns. Measurement of both periods does, however, add to the robustness of the findings. If foreclosed properties are sold at a discount, lower returns prior to a foreclosure should be matched with higher returns subsequent to a foreclosure. F o r c e d S a l e s i n S w e d e n The Enforcement Agency is the government body in Sweden that handles all final debt enforcement and sales of foreclosed property. This institutional setting differs from that of the U.S., in which the creditor holding a mortgage with a lien to the property typically will sell it as a REO property after assuming ownership through foreclosure proceedings. Sales of foreclosed real estate are regulated by the Swedish Enforcement Code (1981, p. 774), which stipulates that the auctioneer cannot accept a price that does not cover the costs of the sale and all debt that is more senior than the debt triggering the foreclosure. Furthermore, no price should be accepted if the auctioneer deems it likely that a considerably higher price can be achieved at a later time. Although creditors with liens to the property are asked whether they accept the highest achieved bid, it is the Enforcement Agency that decides whether the highest bid meets the requirements of the law and, therefore, whether a foreclosed property is to be sold. The costs associated with the sale are charged by the Enforcement Agency and are taken from the achieved price before any other creditors receive payment. Because all Swedish debt is recourse debt, a deficit between the achieved sale price and claims by creditors is still owned by the homeowner after a foreclosure. Therefore, no incentive for strategic default is provided. The sales mechanism of foreclosures deviates considerably from sales in the general market. The property is sold as is, and there is typically only one viewing of the property. The sale is done at a physical auction at the Enforcement Agency. A bidder with the highest bid is required to provide an immediate down payment of 10% or an amount equal to all claims with a more senior right to the house than the claim that initiated the foreclosure, whichever is higher. This will therefore exclude buyers with little equity. Full payment is to be provided within a month, after which full ownership of the property is transferred. Forgey, Rutherford, and VanBuskirk (1994) and Clauretie and Daneshvary (2009) find that cash sales have a negative impact on price. In the Swedish setting, buyer financing is typically resolved before bidding, through either loan commitments or cash. Bidding without such commitment is very uncommon. Sellers, therefore, have no incentive to choose between buyers in Sweden. As a consequence, no such data are collected for sales in the general market.

F o r e c l o s u r e s, R e t u r n s, a n d B u y e r I n t e n t i o n s 1 9 5 D a t a Data regarding foreclosed properties were manually obtained through the Enforcement Agency s physical archive in Stockholm, Sweden. The sample covers 249 foreclosed single-family houses sold through physical auction over the 2006 2013 period in the county of Stockholm. Foreclosures are a fairly rare event in Sweden and these data encompass all, or almost all, foreclosures of single-family houses in Stockholm over this period. The data were matched with information in Sweden s Real Property Register covering the date and price of transactions registered by May 2014. Through the proportion of foreclosed properties having been resold subsequent to a foreclosure, these data provide a measure of the rate of turnover subsequent to a foreclosure. The transaction history of foreclosed properties also enables the creation of a dataset with repeat sales pairs. This allows for calculation of holding period returns between a prior unforced sale and a foreclosure, as well as between a foreclosure and a subsequent unforced sale. In order to compare the rate of turnover of foreclosed properties with the turnover in the general market, transaction data on single-family houses sold through realtors is needed. These data are provided by Valueguard, a company that constructs property indices and has considerable market coverage. The unique combination of the judicial identity and municipality enables identification of each time a property is sold. When holding period returns are compared, repeat sales pairs are constructed from the aforementioned transaction data. Because the data of unforced transactions cover the 2005 through May 2014 period, this timeframe constitutes the boundary of included transactions when holding period returns are measured. 1 The property characteristics of foreclosed and other unforced sales deviate considerably, as can be seen in Exhibit 1. Foreclosed properties can be viewed as having considerably lower sale prices, being smaller, and having a lower quality (measured in standard points assigned by the Swedish Tax Authorities), in addition to having larger plot sizes, which is consistent with foreclosures being located in more peripheral areas. The occurrence of foreclosures increased over the study period, as illustrated by the year of transaction, showing that the foreclosed sample is weighted more toward the later years of the 2006 2013 period. M e a s u r e s o f a F o r e c l o s u r e D i s c o u n t P r o p e n s i t y S c o r e M a t c h i n g When one estimates the effect of some treatment with observational data in this case, the effect of a foreclosure on holding period return and turnover one should be aware of the concern that the assignment to the treatment group is not random. J R E R V o l. 3 9 N o. 2 2 0 1 7

1 9 6 D o n n e r Exhibit 1 Descriptive Statistics of Foreclosed and Unforced Properties Sold, 2006 2013 Panel A: Descriptive Statistics of Foreclosed Single-Family House Sales Panel B: Descriptive Statistics of Unforced Single-Family House Sales Variable Mean Std. Dev. Mean Std. Dev. Sales Price, SEK 2,419,248 1,778,807 3,373,582 1,817,088 No. of Rooms 4.590 1.903 5.078 1.508 Living Area, m 2 109.627 55.015 123.233 43.466 Other Area, m 2 26.747 39.201 28.663 36.960 Year of Construction 1968.671 22.187 1967.516 23.893 Semidetached 0.052 0.223 0.214 0.410 Terraced House 0.181 0.386 0.143 0.350 Leasehold 0.064 0.246 0.052 0.221 Plot Size, m 2 1,840.526 2,523.615 1,119.740 1,331.312 Standard Points 26.490 5.981 28.173 4.742 Northwest 0.116 0.321 0.148 0.355 Northeast 0.285 0.452 0.235 0.424 North 0.044 0.206 0.042 0.200 Central 0.096 0.296 0.175 0.380 East 0.185 0.389 0.155 0.362 Southeast 0.169 0.375 0.142 0.349 Southwest 0.104 0.306 0.103 0.304 Year of Sale 2010.526 2.085 2009.462 2.334 Notes: In Panel A, the number of observations is 249; in Panel B, the number of observations is 56,559. As foreclosed properties diverge from other properties, a comparison of the rate of turnover and the value appreciation between foreclosed and other unforced transactions might result in biased estimates. Rosenbaum and Rubin (1983) propose propensity score matching as a method to correct for confounding factors that can impact an estimated treatment effect. The propensity score is defined as each observation s conditional probability of receiving treatment, given the following pretreatment characteristics: p(x) Pr(D 1X) E(DX), (1)

F o r e c l o s u r e s, R e t u r n s, a n d B u y e r I n t e n t i o n s 1 9 7 with D (0,1) referring to the observation being treated and X being the multidimensional vector of pretreatment characteristics. Probit regression is used to determine the probability of treatment. To estimate the effect of treatment (in this case, a foreclosure), I assume that the assignment to treatment is independent of the potential outcomes, given the covariates X. This central assumption of exogenous assignment of treatment is that of unconfoundedness (Rosenbaum and Rubin, 1983). It is also known as the conditional independence assumption (Caliendo and Kopeinig, 2008). Formally, this is stated as follows: Y(0), Y(1) DX, (2) with Y(1) denoting the outcome for treated observations and Y(0) for control observations and with signifying independence. When this assumption holds, all variables that influence treatment assignment and outcomes simultaneously are observed so that differences in outcomes between treatment and control observations with the same covariate values are attributable to treatment. Propensity score matching is most effective when substantial overlap between the subgroups of treatment and control observations exists so that observations with the same variable values on X have a probability of being both treatment and control observations. This requirement is referred to as the common support or overlap condition and ensures that characteristics X do not perfectly predict treatment assignment D (Caliendo and Kopeinig, 2008). As a consequence, the probability of treatment is bounded in the interval between zero and one: 0 Pr(D 1X) 1. (3) When both the aforementioned assumptions hold, Rosenbaum and Rubin (1983) state that treatment assignment is strongly ignorable, therefore allowing for estimation of the treatment effect by comparison of the mean outcomes of treatment and control observations. This effect is referred to as the average treatment effect on the treated (ATT) and is an estimated counterfactual outcome: ATT E(D 1) E[Y(1)D 1] E[Y(0)D 1], (4) with being the effect of treatment (i.e., a foreclosure). When estimating the impact that a foreclosure has on holding period returns, three matching schemes are used to estimate the ATT. First, each foreclosed property is matched with its nearest control observation in terms of the propensity score. Second, matching on J R E R V o l. 3 9 N o. 2 2 0 1 7

1 9 8 D o n n e r the four nearest matches is applied. Third, kernel (Gaussian) matching is applied, meaning that foreclosed properties are matched with a weighted sample of all unforced transactions, based on the inverse distance of their propensity scores. These different matching schemes deviate in relation to the number of controls and, therefore, also the closeness in terms of propensity scores. Matching on the nearest neighbor yields very close matches using few controls, whereas the kernel uses all potential controls (although assigning smaller weights to distant matches). Matching on the four nearest neighbors is somewhere in between the other two matching schemes in terms of the closeness of the controls and the number of controls. When estimating the impact that a foreclosure has on a property s rate of turnover, only kernel matching is applied. This as a non-smooth estimation based on a small number of controls would be unreliable when estimating a binary occurrence (a property being resold subsequent to a transaction) that has occurred for a very small proportion of all potential controls. Recent work demonstrates that bootstrapping standard errors is invalid for nonsmooth estimators (Abadie and Imbens, 2008). Therefore, standard errors of the ATT are estimated following Abadie and Imbens (2006), for matching on the nearest and four nearest control observations. Bootstrapping with 500 replications is applied for the kernel estimates. Tu r n o v e r o f F o r e c l o s e d P r o p e r t i e s C o m p a r e d w i t h t h e G e n e r a l M a r k e t Based on the judicial identity of each property, a comparable measure of the rate of turnover of properties in the general market is constructed from the transaction data. As each time a property is sold it is identified, it is possible to compare holding periods between foreclosed properties and those bought in the general market. The proportions of properties resold within 6, 12, and 18 months of the initial transaction are compared between foreclosed and unforced properties. The period with foreclosed transactions is from 2006 to 2013, with subsequent transactions known until May 2014. To ensure that all transactions carried out within 18 months subsequent to the initial sale are known, this estimation is based on transactions carried out over the period of 2006 to November 2012. 2 As seen in Exhibit 2, a large proportion of foreclosed properties are resold shortly after the foreclosure sale, with 29.8% of foreclosed properties being resold within 18 months, whereas the equivalent rate of turnover is 3.0% for properties bought through unforced transactions in the general market. This finding provides evidence that foreclosed properties are being sold at a discount and that professional buyers are taking advantage of this mispricing. It is possible that foreclosed properties have characteristics that in turn are associated with a higher (or lower) rate of turnover. A potential bias could be

F o r e c l o s u r e s, R e t u r n s, a n d B u y e r I n t e n t i o n s 1 9 9 Exhibit 2 Percentage of Properties Sold from 2006 to November 2012 That Are Resold Shortly after the Initial Transaction Foreclosed Unforced Resold Within % Resold % Resold 6 months 10.11 0.86 12 months 25.53 1.82 18 months 29.79 3.02 induced by the foreclosure status being associated with smaller and less expensive houses, which tend to be bought by younger buyers who move more frequently. To correct for any such influence on the results, the propensity scores of each property s probability of being foreclosed are estimated so as to compare the turnover of foreclosed properties with that of unforced properties that are as similar as possible in terms of property characteristics, location, 3 and date of transaction. Foreclosed properties are then matched based on propensity scores through kernel matching. One foreclosed property is excluded due to being outside the range of common support, meaning that its propensity score perfectly predicts it to be a foreclosure. As shown in Exhibit 3, the differences between the samples of foreclosed properties and matched unforced properties are not statistically significant for the matched variables. The average date of sale deviates by only 26.7 days for the period of 2006 to November 2012. Time-varying liquidity is therefore controlled for. Estimating the turnover in six-month increments (at 6, 12, and 18 months after the foreclosure sale) controls for the probability of being resold increasing over time since its last transaction. The estimated differences shown in Exhibit 3 are slightly smaller than the unmatched comparison provided in Exhibit 2. This finding indicates that foreclosed properties have attributes that are associated with a higher rate of turnover. Controlling for property characteristics, location, and date of sale through kernel matching reveals that 3.7% of similar properties bought in the general unforced market were resold within 18 months of the initial transaction. The corresponding proportion among foreclosed properties is 29.9%, which is equal to an eight times higher rate of turnover. E s t i m a t i o n o f t h e I m p a c t o n H o l d i n g P e r i o d R e t u r n s A measure of a foreclosure discount is provided by holding period returns between two transactions of the same property, of which one of the transactions is a J R E R V o l. 3 9 N o. 2 2 0 1 7

2 0 0 D o n n e r Exhibit 3 Estimated ATT on Turnover by a Foreclosure Kernel (Gaussian) Foreclosed Unforced Difference (ATT) Std. Err. a Resold within 6 months 0.1016 0.0107 0.0909 0.0234 Resold within 12 months 0.2567 0.0214 0.2353 0.0344 Resold within 18 months 0.2995 0.0374 0.2620 0.0364 No. of Rooms 4.7005 4.5668 0.1337 Living Area, m 2 110.9198 107.7487 3.1711 Other Area, m 2 25.8289 25.6791 0.1497 Year of Construction 1,968.7754 1,968.9037 0.1283 Semidetached 0.0695 0.0695 0 Terraced House 0.2086 0.1551 0.0535 Leasehold 0.0749 0.0749 0 Plot Size, m 2 1,559.4118 1,554.6150 4.7968 Standard Points 26.6524 26.6096 0.0428 Northwest 0.1230 0.0963 0.0267 Northeast 0.2620 0.2834 0.0214 North 0.0535 0.0695 0.0160 Central 0.1176 0.0802 0.0374 East 0.1979 0.1818 0.0160 Southeast 0.1551 0.2032 0.0481 Southwest 0.0909 0.0856 0.0053 Date of Sale 18,325.4064 18,298.6791 26.7273 Notes: The number of observations is 48,955. a Bootstrapped. *Two-sample t-test of difference in means is significant at the 10% level. **Two-sample t-test of difference in means is significant at the 5% level. ***Two-sample t-test of difference in means is significant at the 1% level. foreclosure. The impact of a foreclosure on holding period returns is estimated by matching such observations with as similar as possible repeat sales pairs (between two unforced transactions), in terms of property characteristics, location, date of transaction, 4 and holding period length. Before estimating the propensity scores, observations that are likely to have been influenced by new construction during the holding period are excluded. This is done by the omission of repeat sales pairs with a year of construction that is subsequent to the year of the first transaction. Observations with holding periods shorter than 60 days are also excluded.

F o r e c l o s u r e s, R e t u r n s, a n d B u y e r I n t e n t i o n s 2 0 1 Foreclosed transactions occurred over the 2006 2013 period. Repeat sales pairs that measure returns prior to a foreclosure are matched with control observations, with a second transaction taking place during the 2005 2013 period. Repeat sales pairs that measure returns subsequent to a foreclosure are matched with control observations, with transactions during the period of 2006 through May 2014. The two samples overlap, with 25 foreclosed properties experiencing a prior unforced sale, a foreclosure, and a subsequent unforced sale over the period. One property measuring prior returns is excluded because it is outside the range of common support. After the aforementioned exclusions, there are 71 observations of prior returns and 81 observations of subsequent returns. The results of the probit model for estimation of the propensity scores are in the Appendix. Estimation of the impact of a foreclosure on holding period returns is carried out by matching foreclosed properties with their closest match based on the propensity score as well as the four closest matches, in addition to kernel matching. Matching is done with replacement. Some unforced control observations can then be matched with several foreclosed properties if they are the closest match on the propensity score. As seen in Exhibit 4, the annualized return prior to a foreclosure is negatively influenced by a foreclosure, with results in the range of 10.7% to 7.6%. For a mean holding period of 4.3 years, the greatest difference in returns is found when kernel matching is applied, while the smallest difference in returns is found when matching on the four nearest neighbors. Subsequent to a foreclosure, the results show a 37.8% 48.6% positive impact on the annualized returns, with a mean holding period of 1.2 years, with matching on the nearest neighbor producing the smallest estimate and with kernel matching estimating the largest difference in returns. The consistency in the estimates when matching schemes are compared is reassuring. There is also consistency in the sense that holding period returns prior to a foreclosure are substantially lower than those of comparable unforced repeat sales pairs and that an equivalent positive difference is found subsequent to a foreclosure. Property characteristics, location, date of transaction, and holding period of foreclosed and unforced control observations are shown in Exhibit 4. There are no statistically significant differences in means between foreclosed and control observations when matching is applied on the nearest and four nearest unforced observations. Most important, the differences between foreclosed and matched unforced transactions are small with regard to the date of transaction and holding period length when nearest-neighbor matching is applied. When estimating the impact on returns prior to a foreclosure, the average date of sale deviates by 77.9 days when matching on the nearest match is applied and by 46.5 days when matching on the four nearest unforced transactions. These are very small differences considering that the time period extends from 2005 to 2013. When estimating returns subsequent to a foreclosure, the equivalent values are 47.3 and 106.0 days, respectively, for the period of 2006 to May 2014. Holding J R E R V o l. 3 9 N o. 2 2 0 1 7

Panel A: Prior return Foreclosed Exhibit 4 Estimated ATT on Holding Period Returns Nearest Neighbor 1:1 Nearest Neighbor 1:4 Kernel (Gaussian) Unforced Difference (ATT) Std. Err. Unforced Difference (ATT) Std. Err. Unforced Difference (ATT) Std. Err. 2 0 2 D o n n e r Annualized Return 0.0144 0.0715 0.0858 0.0209 0.0619 0.0762 0.0093 0.0929 0.1072 0.0123 No. of Rooms 4.3944 4.2113 0.1831 4.3239 0.0704 5.0663 0.6719** Living Area, m 2 105.9154 101.8732 4.0423 105.8204 0.0951 121.7571 15.8416* Other area, m 2 24.56338 21.2676 3.2958 24.0669 0.4965 25.7679 1.2045 Year of Construction 1966.38028 1964.4085 1.9718 1967.2289 0.8486 1968.5513 2.17102 Semidetached 0.0423 0.0563 0.0141 0.0599 0.0176 0.2583 0.2161*** Terraced House 0.2254 0.2535 0.0282 0.2289 0.0035 0.1712 0.0542 Leasehold 0.0704 0.0704 0.0000 0.0916 0.0211 0.0599 0.0106 Plot Size, m 2 1,819.8169 2,000.5916 180.7747 1,816.0880 3.7289 976.7738 843.0431*** Standard Points 26.9718 26.5493 0.4225 26.6901 0.2817 28.3160 1.3441 Holding Period, days 1,579.9296 1,565.8451 14.0845 1,583.4155 3.4859 1,163.8318 416.0977*** Date of Sale 18902.6620 18824.7746 77.8873 18856.1479 46.5141 18562.4431 340.2189*** Northwest 0.1268 0.1127 0.0141 0.1232 0.0035 0.1686 0.0418 Northeast 0.3380 0.3239 0.0141 0.3486 0.0106 0.2290 0.1090 North 0.0141 0.0141 0.0000 0.0070 0.0070 0.0365 0.0224 Central 0.0845 0.1409 0.0563 0.1021 0.0176 0.1613 0.0768 East 0.1549 0.1831 0.0282 0.1479 0.0070 0.1633 0.0084 Southeast 0.1831 0.1268 0.0563 0.1549 0.0282 0.1402 0.0429 Southwest 0.0986 0.0986 0.0000 0.1162 0.0176 0.1011 0.0025

Exhibit 4 (continued) Estimated ATT on Holding Period Returns J R E R V o l. 3 9 N o. 2 2 0 1 7 Panel B: Subsequent return Foreclosed Nearest Neighbor 1:1 Nearest Neighbor 1:4 Kernel (Gaussian) Unforced Difference (ATT) Std. Err. Unforced Difference (ATT) Std. Err. Unforced Difference (ATT) Annualized Return 0.5794 0.2019 0.3775 0.0838 0.1509 0.4286 0.0539 0.0931 0.4863 0.0576 No. of Rooms 4.7531 4.8642 0.1111 4.8889 0.1358 5.0395 0.2864 Living Area, m 2 118.0741 117.2099 0.8642 119.5370 1.4630 121.7152 3.6412 Other Area, m 2 29.7346 36.6296 6.8951 36.4753 6.7407 26.7929 2.9417 Year of Construction 1973.4568 1973.9506 0.4938 1972.8426 0.6142 1969.5653 3.8915 Semidetached 0.0988 0.1111 0.0124 0.0988 0.0000 0.2364 0.1377** Terraced House 0.2469 0.1852 0.0617 0.2068 0.0401 0.1768 0.0701 Leasehold 0.1234 0.0988 0.0247 0.0926 0.0309 0.0672 0.0563 Plot Size, m 2 1,204.0370 1,514.2222 310.1852 1,269.4753 65.4383 1,032.8782 171.1588 Standard Points 28.6543 29.2469 0.5926 28.8056 0.1512 28.3277 0.3267 Holding Period, days 442.9383 478.4444 35.5062 472.9352 29.9969 1,033.7094 590.7712*** Date of Sale 18780.5432 18827.8889 47.3457 18886.5031 105.9599 18840.5600 60.0168 Northwest 0.1852 0.1975 0.0124 0.1883 0.0031 0.1734 0.0118 Northeast 0.2099 0.1852 0.0247 0.1512 0.0586 0.2292 0.0193 North 0.0494 0.0617 0.0124 0.0617 0.0124 0.0402 0.0092 Central 0.0864 0.0864 0.0000 0.0833 0.0031 0.1498 0.0634 Std. Err. F o r e c l o s u r e s, R e t u r n s, a n d B u y e r I n t e n t i o n s 2 0 3

Exhibit 4 (continued) Estimated ATT on Holding Period Returns 2 0 4 D o n n e r Nearest Neighbor 1:1 Nearest Neighbor 1:4 Kernel (Gaussian) Foreclosed Unforced Difference (ATT) Std. Err. Unforced Difference (ATT) Std. Err. Unforced Difference (ATT) Std. Err. Panel B: Subsequent return East 0.2346 0.2840 0.0494 0.2809 0.0463 0.1704 0.0641 Southeast 0.0988 0.0494 0.0494 0.1142 0.0154 0.1291 0.0303 Southwest 0.1358 0.1358 0 0.1204 0.0154 0.1080 0.0278 Notes: In Panel A, the number of observations is 7,203; in Panel B, the number of observations is 6,465. * Two-sample t-test of difference in means is significant at the 10% level. ** Two-sample t-test of difference in means is significant at the 5% level. *** Two-sample t-test of difference in means is significant at the 1% level.

F o r e c l o s u r e s, R e t u r n s, a n d B u y e r I n t e n t i o n s 2 0 5 periods are also similar in duration when matching on the nearest and four nearest unforced transactions. When estimating the impact on returns prior to a foreclosure, holding periods deviate by only 14.1 and 3.5 days, respectively. The equivalent values are 35.5 and 30.0 days when returns subsequent to a foreclosure are estimated. These differences are still fairly small in relation to the average holding period of 442 days for resold foreclosures. As expected, the differences in means between foreclosed and control observations are greater with kernel matching (as the number of controls increases), with several variables exhibiting statistically significant differences. It is important to keep in mind that a pure foreclosure discount is going to be independent of the holding period length, both prior and subsequent to the foreclosure. A property that has been owned for a short period prior to a foreclosure will experience a significantly greater negative impact on annualized returns compared to a property owned for a longer period, even if the price discount is the same for both properties. The measured differences in annualized returns should therefore be viewed in light of holding period lengths and be interpreted as the discount on value appreciation caused by a foreclosure or, alternatively, the excess returns earned due to foreclosed properties being bought at a discount. It is possible that the very large annualized excess return that is achieved subsequent to a foreclosure is overestimated due to property improvements. Professional buyers who buy foreclosures and resell them within a very short time period are likely to have taken value-increasing measures. If one excludes observations with a holding period shorter than one year, 5 the annualized excess return that is achieved subsequent to a foreclosure ranges from 15.9% (kernel) to 19.0% (1:1) for a mean holding period of 2.5 years (1:4 matching results in a 16.3% excess return). Given the longer holding period, this is a small difference in returns compared with the full sample. E s t i m a t i n g t h e P r o p o r t i o n o f P r o f e s s i o n a l B u y e r s Because the transaction data include information about the buyer(s) of each property, the number of buyers who had purchased several properties in the sample could be determined. Among the properties that had been sold as foreclosures, 38 of 247 properties 6 (15.4%) were bought (or partially bought) by someone who had bought more than one property in the sample. Of these, two buyers had bought four foreclosed properties each, and 15 buyers had bought two foreclosed properties. These properties also experienced a high rate of turnover, with 29 of the 38 properties (76.3%) being resold after the foreclosure. When comparing holding period returns between the subgroups of properties that had likely been bought by professional buyers with the residual sample, there is not much deviation in the returns subsequent to a foreclosure. The annualized return for the total sample of resold foreclosed properties is 57.9%, compared to J R E R V o l. 3 9 N o. 2 2 0 1 7

2 0 6 D o n n e r 63.5% among the 28 properties 7 likely purchased by professional buyers and 55.0% among the 53 other properties. Larger return differences may have been attained if data for a longer period of time had been available. Properties with longer holding periods would be more likely to have been purchased by unprofessional buyers, whose initial intention was to occupy the property, not to resell it. S e n s i t i v i t y A n a l y s i s As propensity score matching relies on the assumption of unconfoundedness, failure to include all relevant variables in the regression might cause biased estimates. The sensitivity to such a hidden bias is tested, in accordance with the approach proposed by Rosenbaum (2002), by estimating how inference about the treatment effect is affected by an unmeasured confounding variable, that is, a variable that impacts both treatment assignment (foreclosure) and outcome (turnover and annualized holding period return). If i is the probability of being a foreclosed property for observation i, the odds ratio is equal to the following: i. (5) 1 i As such, an odds ratio can be defined for observation j. The odds ratio between i and j is defined as : i/(1 i) j/(1 j). (6) Two observations with equal covariates have odds of receiving treatment that diverge by, which can be seen as a multiplier of the degree of departure from random assignment. Therefore, 1 corresponds to no hidden bias; 2 indicates that between two observations with the same observed covariates, one might be twice as likely to receive treatment. The upper and lower confidence bounds of the p-values of significance regarding the estimated effect at different values of indicate how much bias is required before the interval becomes uninformative. As seen in Exhibits 5 and 6, the estimates are quite insensitive to hidden bias, with the bounds of p-values exceeding.05 at values of as high as 2.9 and 3.0 when returns prior and subsequent to a foreclosure are compared. The equivalent estimates of regarding differences in the probability of a property being resold within 6, 12, and 18 months are 3, 7.5, and 5.8, respectively. As the true is unknown, the high degree of robustness to hidden bias is reassuring.

F o r e c l o s u r e s, R e t u r n s, a n d B u y e r I n t e n t i o n s 2 0 7 Exhibit 5 Rosenbaum Bounds of the Estimated ATT on the Probability of Property Being Resold Min. Max. Panel A: Resold within 6 months 1.0.0001.0001 2.9.0001.0454 3.0.0001.0507 Panel B: Resold within 12 months 1.0.0001.0001 7.4.0001.0489 7.5.0001.0511 Panel C: Resold within 18 months 1.0.0001.0001 5.7.0001.0469 5.8.0001.0506 Exhibit 6 Rosenbaum Bounds of the Estimated ATT on Annualized Holding Period Return Min. Max. Panel A: Prior return 1.0.0001.0001 2.8.0475.0001 2.9.0581.0001 Panel B: Subsequent return 1.0.0001.0001 2.9.0001.0470 3.0.0001.0579 J R E R V o l. 3 9 N o. 2 2 0 1 7

2 0 8 D o n n e r C o n c l u s i o n The findings of this study provide evidence that the market in Stockholm, Sweden for foreclosed properties diverges from the general real estate market, both in terms of the market participants and the properties sold. The finding that foreclosed properties to a significant extent are resold shortly after the foreclosure sale supports the view that professional buyers are buying the properties and intend to resell them. Such intent is likely a consequence of an incentive to capture gains provided by a price discount. Estimating the turnover rate of foreclosed properties therefore provides a robust method of assessing whether foreclosures are sold at a discount that is less reliant on data regarding property characteristics. The use of several methods increases the robustness of the findings, with the measurement of holding period returns also being concurrent with a foreclosurerelated discount on price. Over a mean holding period of 4.3 years, foreclosed properties are estimated to achieve a 7.6% 10.7% lower annualized return prior to a foreclosure, compared to similar properties sold through unforced transactions. Annualized holding period returns subsequent to a foreclosure were 37.8% 48.6% higher, compared to the returns on similar properties for a mean holding period of 1.2 years. The results support the view that foreclosure status causes a substantial discount on price and that this discount is not a proxy effect driven by property characteristics. The results are in line with previous research as they indicate substantial price discounts. It is noteworthy that the results are in agreement with previous research, considering that a different sale mechanism is used in Sweden than in the U.S., and that seller incentives, which typically influence a foreclosure discount, such as the costs of holding the property, do not apply. The finding that a fairly substantial proportion of foreclosed properties (15.4%) were bought by buyers having bought more than one such property supports the view that a lack of competition leads to lower sales prices. Therefore, marketing efforts might increase the achieved sales prices. As it is possible that foreclosure sales are associated with conditions that discourage potential buyers, the reasons for a price discount that are unrelated to property condition therefore deserves further attention.

F o r e c l o s u r e s, R e t u r n s, a n d B u y e r I n t e n t i o n s 2 0 9 A p p e n d i x Exhibit A1 Probit Regression Results for Estimation of Propensity Scores Used When Comparing the Proportion of Properties Resold Subsequent to an Initial Transaction Variable Coeff. Std. Err. Z-Statistic p-value No. of Rooms 0.0070 0.0283 0.25.805 Living Area, m 2 0.0012 0.0010 1.18.238 Other Area, m 2 0.0008 0.0009 0.97.333 Year of Construction 0.0027 0.0013 2.13.033 Semidetached 0.3811 0.0968 3.94.000 Terraced House 0.1417 0.0730 1.94.052 Leasehold 0.3041 0.1103 2.76.006 Plot Size, m 2 0.00004 0.00002 2.58.010 Standard Points 0.0139 0.0059 2.36.018 Date of Sale 0.0002 0.00004 5.04.000 Northwest 0.0157 0.1097 0.14.886 Northeast 0.0623 0.0985 0.63.527 North 0.2119 0.1399 1.51.130 Central 0.0288 0.1123 0.26.798 East 0.1524 0.1022 1.49.136 Southeast 0.0874 0.1065 0.82.412 Southwest Constant 10.8799 2.575 4.22.000 p-value.0000 Pseudo R 2.0418 Notes: The dependent variable is Property Being a Foreclosure. The likelihood ratio of the chisquare test is 102.71. There are 48,955 observations. J R E R V o l. 3 9 N o. 2 2 0 1 7