Residential Real Estate Traders: Returns, Risk and Strategies

Transcription

1 Residential Real Estate Traders: Returns, Risk and Strategies Marco Giacoletti and Victor Westrupp Abstract We study the capital gains attained by investors trading residential real estate properties in the main urban areas of California over the period from January 1998 to December 212, using a unique dataset of residential real estate sales and remodeling permits. Investor performance is measured in excess of fluctuations in local housing markets. We find evidence of high gains for transactions with holding periods shorter than twelve months, mainly after 27. For these short holding periods, gains are higher for agents who have been more active in housing markets during the two years preceding the trade. Capital gains do not seem to be explained by risk compensation and are larger than rental income forgone by short term speculative traders. Finally, we find that the percentage gains earned by individual agents over multiple trades are persistent. However, persistence is weaker when we measure gains in terms of dollar values. Overall, our results provide evidence that real estate traders can extract value from local housing markets. Stanford Graduate School of Business: mgiacol@stanford.edu Stanford Graduate School of Business: victor_westrupp@stanford.edu 1

2 1 Introduction We study the capital gains attained by investors trading in residential real estate markets, using a unique dataset covering the main urban areas of California over the period from January 1998 to December 212. We find that transactions with holding period shorter than one year have delivered substantially higher percentage gains with respect to the rest of their local market. Higher gains are particularly evident for houses bought in the period following the housing bust, while they are lowest during the years of the housing boom. We also find that for short holding periods, percentage gains are higher for home buyers that have been managing multiple real estate transactions over the two years before the trade. Moreover, for home buyers that engage in multiple house trades over time, risk adjusted investment performance is persistent. Most of the housing literature studies the problem of households that buy a property in order to consume its housing services. Relatively little work has been devoted to the study of investors in housing markets. In particular, it seems reasonable to believe that investors who are experts in local real estate might be able to extract value from an highly frictional market, for example by acting as middlemen buying and re-selling properties over short periods of time. In fact, we know from previous research that the market for homes is thin (Piazzesi and Schneider (29) use data from the American Housing Survey and show that on average 6% of owner occupied houses trade every year) and segmented. Among others, Piazzesi et al. (214) build a structural model of segmented search, showing the importance of search and segmentation frictions in determining market prices and trading volumes. They document that, based on their characteristics, households in the San Francisco area focus on specific geographic and quality sub-segments and do not search the entire metropolitan market. Moreover, it has been shown that properties sold trough Real Estate Owned sales and short sales experience high price discounts, and have a negative externality on the sale price of properties sold nearby (see Campbell et al. (211) and Anenberg and Kung (214)). The relevance of intermediaries in real asset markets has been documented in the data, for example in the case of the commercial aircrafts in Gavazza (211). The presence of dealers in these decentralized has multiple general equilibrium implications. Dealers can lower trading delays by matching low and high valuation agents (see Gavazza (215)) or by removing asymmetric information (see Biglaiser (1993)). However, the impact of asset dealers on aggregate welfare is not clear. Gavazza (215) argues that the presence of dealers might decrease social welfare. In his model, dealers provide better allocations but they also extract all gains from trades. While in our study we are not able to provide counter-factuals and directly study the general equilibrium implications of middlemen activities in housing markets, we show the extent to which traders in local housing markets are able to earn rents and document that agents that hold assets for short holding periods and that potentially act as middlemen are able to earn sizable capital gains. Bayer et al. (212) have already studied the behavior of flippers (traders that buy and re-sell a housing property quickly) in the metropolitan area of Los Angeles. The authors found that these agents are on average able to buy houses at a discount. However, Bayer et al. (212) do not study the performance of 2

3 house trades. In our data, we also find that traders holding houses for shorter periods on average get a price discount with respect to comparable properties. However, we move beyond the existing literature by investigating the value extracted by house traders from local housing markets. In order to do so, we need to take into account all the cash flows involved in housing transactions. In particular, it is known that investing in real estate can require relevant remodeling expenses. We therefore use a proprietary dataset from Buildzoom 1 that collects information from remodeling permits for the state of California. We match this information with micro data on house sales from the main urban areas of California that we obtain from CoreLogic 2, creating a unique database of house transactions and remodeling. In our study, we want to sort traders in different groups based on whether they are professional investors in real estate markets. Since there is no disclosure of trading motives, the approach in previous papers has been to infer the type of a home buyer based on past investment behavior. Haughwout et al. (211) look for home owners who hold mortgages on multiple properties. Bayer et al. (212) identifies a home buyer as a flipper if she bought and resold at least two properties in the two years prior to the current purchase. In our work, we want to identify the agents that actively manage a real estate portfolio. We therefore classify investors at the moment of purchasing a new house into groups based on the number of properties that they bought in the last two years, including the current transaction. The window of two years guarantees that we are sorting investors based on their current degree of activity in the market. These agents may be trading properties either by buying and re-selling them or by using rents extracted from existing properties to acquire more real estate. When assessing trades performance, we base our analysis on the literature in financial economics that studies the abnormal returns earned by professional money managers and private equity firms. In particular, each trade in our dataset involves multiple, irregularly spaced cash flows and performance can be actually measured only at the time when the trade is closed and the housing unit is sold. In this sense, we face challenges similar to the ones present in the private equity literature. Researchers in this area have developed a variety of return measures (see for example Harris et al. (214)). Kaplan and Schoar (25) develop the Public Market Equivalent (PME) to calculate the excess performance of a private equity fund with respect to the public equity market. We introduce the Local Market Equivalent (LME) that applies the same principle in order to calculate the excess performance of real estate investment with respect to local ZIP code house price indexes (provided by Zillow). Under the assumption that housing units have an exposure of one to the corresponding local house price index, the LME is a measure of the abnormal performance of the real estate trade with respect to the index. While this assumption is to some extent reasonable, it is clear that different house types will bear different exposure to house price fluctuations. It is also reasonable to believe that such exposure will change over time. Aware of these issues, when studying LMEs we introduce controls for characteristics of the housing unit (for example number of bedrooms, number of rooms, size in square feet) along with time and ZIP code fixed effects. It is important to note that abnormal performance is measured ex-dividend for 1 We thank Issi Romen and Buildzoom for providing access to the database. 2 Licensed by Stanford University. 3

4 both the house trade and the price index, since housing rents are excluded. We cannot properly account for differences in dividend rates, since micro data on rents are not available. Two issues have to be discussed, with respect to performance measurement. First, the local market equivalent is a gross return measure, in the sense that it does not provide any information on what fraction of the value extracted is earned by the trader and what fraction is earned by investors financing the trader. In fact, while in our data we have information on the agents engaging in real estate trades, we have very scarce information on their different sources of financing. Nonetheless, we do not believe this is a limitation of our study. For there to be any value to split in between traders and financiers, there needs to be a positive gross return. Second, the local market equivalent is a percentage return. As pointed out in the mutual funds literature by Berk and van Binsbergen (214), the size (in terms of dollar amounts) of an investment might be endogenously chosen by the trader, who might then be targeting a specific net present value. Thus, we extend our results by measuring performance also in terms of the dollar value added (to use Berk and van Binsbergen (214) definition) of the trade. When we use this measure, we still find evidence in favor of abnormal performance for shorter holding periods. However, differences across trader types are mostly indistinguishable. Moreover, individual trader s performance is way less persistent when it is measured in terms of value added. We interpret these results as evidence that traders in the housing market have a hard time targeting dollar profits, due to high fluctuations in house prices. One potential explanation for the measured excess returns would be that they are compensation for risks not captured by local house price indexes and the other controls that we include in our regressions. For example, liquidity effects not spanned by ZIP code indexes might make LMEs in a specific area extremely uncertain. We (admittedly very roughly) proxy for these effects by building a measure of the local volatility of LMEs for each property in our sample. While local volatility is positively correlated with higher abnormal returns, our results are robust to accounting for this risk measure. An alternative explanation would be that the abnormal returns are erased when accounting for rents that traders operating with very short holding periods are foregoing. In fact, a trader buying, remodeling and re-selling a property over a short period of time would not be able to also rent the property in the meantime. However, when we adjust the returns of trades with short holding periods by subtracting the potentially foregone rents, we still find abnormal performance. The rest of the paper is organized as follows. Section 2 discusses the structure of our unique dataset. Section 3 gives a first overview of the behavior of the real estate investors in our dataset. Section 4 discusses in detail how we define and measure performance. Section 5 presents the main empirical results and section 6 studies some potential explanations of our results. Section 7 studies performance persistence at the level of individual traders. Finally, 8 concludes and presents the final remarks. 2 Data We build a unique database consisting of micro data on both real estate transactions and remodeling activity. Information on house transactions is provided by Corelogic Solutions 3. The main dataset consists 3 4

5 of real estate deeds, which are both house regular sales, REO sales and short sales, as well as mortgage refinances. The deeds are collected from county recorders and provides details on transaction prices, owner and seller names, mortgage amounts. Housing units are identified using Assessor Parcel Numbers (APN, assigned to each plot of land by tax assessor of a specific jurisdiction for purposes of record-keeping) as well as addresses (including apartment and housing unit numbers). Corelogic also provides a snapshot of house characteristics based on tax assessment data for the fiscal year ending in 214. These data allow to match each APN to its ZIP code, as well as to geo-spatial information (latitude and longitude). Moreover, they provide details on the size and characteristics (number of bedrooms and bathrooms, age) of the property. The deed and tax files are merged using APNs as identifiers. In our analysis, we define house sales as arm s length transaction and exclude nominal sales. We discuss the several issues that emerged in the treatment of the data in appendix A. While Corelogic data extend back to the the eighties and seventies, the quality of the records is higher for more recent entries and seems to be best starting from the late nineties. We therefore decide to use for our analysis the sample between January 1998 and December 213. We then decide to focus on the main urban areas of California, both because these are some of the most important and highly studied housing markets in the United States and because these are some of the areas with the best data quality from Corelogic. In particular, we study four main areas. The first one is composed of Los Angeles county and Orange county. The second and third one are respectively the counties of Sacramento and San Diego. Finally, the fourth one consists of the Metropolitan Statistical Area (MSA) of San Francisco (which includes Alameda county, Contra Costa county, Marin County, San Francisco County, San Mateo County) and Santa Clara County. Data on remodeling contracts are provided by Buildzoom 4. The company is an intermediary matching house owners to contractors for commercial and residential remodeling. To perform its business, Buildzoom has collected from local census authorities information on house remodeling and building permits. The data report the kind of job that was performed (for example, kitchen renovation, plumbing, roof repairs and so on), the job cost and the fees paid to the local authority by the contractor. Our dataset in then built by merging information on each single housing unit from Corelogic and Buildzoom. In the merge we identify each property based on its APN and address. Details on the methodology and the issues faced in the merge can be found in appendix A. 3 Trading Residential Real Estate 3.1 Identifying Traders Groups There is evidence of speculative activity in real estate markers. Bayer et al. (212) study the behavior of flippers, investors who buy residential real estate with the aim of quickly reselling it for a profit. The behavior of flippers has also been tracked by companies providing information and analysis on real estate markets 5 and we have even witnessed the raise of companies that provide training to individuals 4 Information on the company is available at 5 See for example or 5

6 interested in becoming real estate middlemen 6. Moreover, it is known that large private equity groups like Blackstone have directly undertaken large speculative investments in distressed US housing markets after the financial crisis. Figure 1 plots the fraction of houses that were bought and resold with a holding period respectively less than six months and in between six and twelve months in our data. The fractions are computed out of all properties that were transacted in a specific year. Since our dataset ends in December 213, we end the sample with the transactions that were undertaken in December 212. It is interesting to note that the relative importance of transactions with short holding period is highly cyclical. It expands in the years of the housing boom and in the post crisis recovery, while it is at its minimum in 28 and < 6 Months 6 to 12 Months 8 (%) Frac of Transactions Figure 1: Fraction of houses bought and resold with holding period smaller or equal than 12 months. Fractions are computed out of all arms-length house transactions in a specific year over the counties covered in our sample. However, we cannot just argue that an individual is following a speculative trading strategy just because she has held a property for a short period of time. In fact, some households may end up buying and reselling properties for completely idiosyncratic reasons related to their job or family situation. However, it is not clear whether these job or family related motives could generate the cyclical pattern that we observe in figure 1. Ideally, we would like to know the agent s motivations for trading at the moment that they bought a property. The approach in the literature so far has been to infer the type of a buyer based on past investment behavior. Haughwout et al. (211) identify real estate investors as owners who hold mortgages 6 An example is FortuneBuilders, as described at 6

7 on multiple properties. Bayer et al. (212) identify a home buyer as a flipper if she has bought and resold at least two properties in the two years prior to the current purchase. Our aim is to identify agents that actively manage a real estate portfolio. We therefore sort house buyers at the moment of purchasing a new house into groups based on the number of properties that they bought in the last two years, including the current transactions. The window of two years guarantees that we are sorting investors who have been recently active in real estate markets, either by buying and reselling housing units or by using rents extracted from existing properties to expand their real estate portfolio. When sorting investors, we consider the total number of transactions across all the main urban areas in our study, since it is reasonable that professional real estate investors might be focusing on different counties and cities. However, there might be highly diversified investors who trade across different states in the US. For these investors, we are only tracking the part of their portfolios which is invested in the main urban areas of California. We define three groups for our analysis: agents that over the last two years have only bought the current housing unit, agents that have bought between two and ten properties including the current one and agents who have bought more than ten. Figure 2(a) shows the fraction of all houses bought in arm s length transactions by the different investor groups in each vintage year from 1998 to 212. The fractions are based on the entire dataset, including the urban areas of Los Angeles, Sacramento, San Diego and San Francisco. We can see that the most passive traders conduct the vast majority of transactions. However, the presence of more active traders is detectable. Traders that have transacted at least two properties in the last two years take part to in between 2 and 3% of yearly transactions in our data. However, the most active traders (with more than 1 trades) represent only between 1.7 % and 4.7 % of transactions. The fractions of active traders seem to move with cycles in the housing markets. They are highest in the housing boom of 24, 25 and 26 and in the post crisis period, in particular in the years 211 and 212. If we break down trader groups by holding period, we find that active traders are highly involved in transactions resulting in short holding period. In subfigure 2(b) we split transactions with holding period smaller or equal than 6 months across the different trader groups. The least active traders undertake here only between 4 and 25 % of transactions. On the other hand, the most active group of traders (with more than 1 trades over the last two years) accounts for between 15 and 4 % of transactions. The least active traders are clearly predominant in subfigure 2(d), for holding periods longer than twelve months, while subfigure 2(c) shows that results for the holding period between six and twelve months are very similar to the ones for the holding period up to 6 months. Thus, active real estate traders dominate the part of the market that involves short holding periods (less than one year). A related interesting finding that emerges from Corelogic deed data is that most active investors do not get a mortgage at the time of buying a house. Anecdotal evidence 7 suggests that real estate traders do not get financing from banks, but rather from non-bank Private Money Lenders (PMLs). One of these institutions in the area of San Jose (in Santa Clara county) is Grand Coast Capital, which offers 7 We attended information sessions held by FortuneBuilders to gain more insight in the industrial organization of this industry. 7

8 12 1 Trade 2-1 Trades >1 Trades 12 1 Trade 2-1 Trades >1 Trades 1 1 (%) Frac of Transactions (%) Frac of Transactions < 6 Months (a) Entire sample of trades (b) Holding period less than 6 months 12 1 Trade 2-1 Trades >1 Trades 12 1 Trade 2-1 Trades >1 Trades 1 1 (%) Frac of Transactions 6 to 12 Months (%) Frac of Transactions 12 Months (c) Holding period between 6 and 12 months (d) Holding period higher than 12 months Figure 2: Fraction of houses bought by different investor groups. Fractions are computed out of houses transacted in a specific year debt financing with a 75 % loan to value ratio for loan sizes between $ 1, and $ 3,, 8. Grand Coast Capital seems to be also willing to provide equity-like financing under specific circumstances. So far, we have talked of the real estate traders as individuals, without disentangling whether they are persons or legal entities, such as LLCs or investment trusts. These institutions could be both real estate investor corporations as well as individuals who have created an entity in order to shield their personal wealth from losses generated by their trading activity. Figure 3(a) shows the fraction of deals undertaken by legal entities 9 for different holding periods. The fractions are computed out of all the transactions in a specific year. We can see that very few residential real estate buyers are corporate entities for holding periods longer or equal than twelve months. However, corporate entities make up a fair fraction of trades when looking at short holding periods. Interestingly, the relative importance of corporations seems to have also increased over time. In the group of properties held for at most six months, corporations moved from approximately 25 % in the 24 period to between 8 Information on loan programs can be found at 9 Corelogic deed data include a flag that splits buyers in between corporations and individuals. To address potential weaknesses in the data, we also use text-search functions to select all buyers whose name contains corporate denominations, such as for example LLC, & Co., Trust, Investment, Management and other relevant key words. 8

9 4 % and 5 % in the housing boom and bust and up to almost 7 % in 212. In figure 3(b) we look at the same split in between individuals and corporate entities across trader groups. Not surprisingly, we find that the most active traders are way more likely to be legal entities. Again, the fraction of legal entities is increasing over time. 3.2 Real Estate Traders as Middlemen We have evidence that active traders play a dominant role for transactions that result in holding periods of less than a year. Bayer et al. (212) argue that professional traders with short holding periods act as middlemen. These traders will look for sellers who have high holding cost for their houses (due, for example, to financial distress), or who own houses that are in bad conditions and will buy properties at a discount with respect to comparable housing units in the same neighborhood. The traders then aim at re-selling the property quickly at full price (sometimes after remodeling the house), in order to cash in the price gain. Do we see in our data professional traders who trade quickly buy discount properties? We investigate this question through regression equation 1: ( log P buy i ) = a z + a yq + B house X house i + β 12m I(h i < 12m)+ (1) + B type short I(h i < 12m, type) + B type long I(h i 12m, type) + v i Where P buy i is the price at which a property is bought, a z and a yq are respectively a ZIP code fixed effect and fixed effect for the year/quarter date in which the housing unit is bought. The column vector Xi house contains house characteristics: number of bedrooms, bathrooms, age and size of the house. The variable I(h i < 12m) is a dummy variable for houses held up to twelve months. The coefficient β 12m captures the price discount (or premium) for housing units held for less than one year and bought by a trader that has not bought any other housing unit in the two previous years. The column vector I(h i < 12m, type) contain dummies obtained by interaction between the dummy for holding period smaller than one year and dummies for the different types of active traders. Consistently with our analysis in the previous section, we identify traders based on the number of residential real estate investments that they have undertaken recently. Thus, we introduce dummies for traders that have bought between 2 and 1 and more than 1 housing units over the two years before buying property i (including property i itself). The coefficients in B type short will then capture the incremental price discount for short holding period trades undertaken by the more active traders. Along the same lines, vector I(h i 12m, type) contains interactions between a dummy for holding period equal or larger than one year and trader types. The coefficients in B type long capture the price discount for properties bought by active traders and held for a year or longer. Table 1 collects parameter estimates over the entire sample from January 1998 to December 212. We estimate the regression separately in each one of the main urban areas that we are considering in our study: Los Angeles, Sacramento, San Diego and San Francisco. Discounts for short holding periods are significant and economically large. We find that when the holding period is less than one year, the least active traders already pay housing units in between 12.5 % and 14 % less across different urban areas. The discount is even higher for traders that have bought in between 2 and 1 houses in the last two years. 9

10 These agents get an additional 7 % to 4 % discount across the different areas. Traders that have bought more than 1 properties get an average discount larger than the one for the least active agents in the areas of Los Angeles, Sacramento and San Diego, even if for the last one the difference is not significant. In San Francisco, the most active traders pay on average higher prices with respect to the least active traders. There are price discounts for more active traders against the least active traders for holding periods longer than twelve months, but the overall price discount is way smaller than the ones found for shorter holding periods. We investigate the issue a bit further by estimating equation 1 in three different subperiods: the preboom/early boom years (from January 1998 to December 22), the housing boom (from January 23 to December 27) and the housing bust, including the most recent years (from January 28 to December 212). We can see that the large discounts for the short holding period are mostly driven by observations belonging to the most recent period. Also, while it is a true across all periods that traders who bought between 2 and 1 houses in the last two years were able to get higher discounts, this is not true for the traders that bought more than 1 properties. During the boom years and across all urban areas, the most active traders were on average engaging in trading properties that given ZIP code, time and house characteristics controls, were more expensive than the ones bought by the least active traders. Dependent Variable is log Buy Price: sample 1998/1 to 212/12 Los Angeles Sacramento San Diego San Francisco I(h < 12m).143 [ ] I(h < 12m, Ntr 2, Ntr 1).74 [ ] I(h < 12m, Ntr > 1).46 [ ] I(h 12m, Ntr 2, Ntr 1).251 [ 6.127] I(h 12m, Ntr > 1).439 [ 4.346].1426 [ ].72 [ ].767 [ ].592 [ 5.579].828 [ ].1251 [ ].723 [ ].32 [ ].155 [ ].45 [.3758].1281 [ ].4 [ 3.261].3 [1.3322].522 [ ].468 [ ] House Char Y es Y es Y es Y es Zip FE Y es Y es Y es Y es Year/Quart Start FE Y es Y es Y es Y es Adj R % 82.6 % % % N. Obs. 545, ,47 51,45 321,29 Table 1: Regressions on log prices. Standard errors clustered by Zip code and start quarter. 1

11 7 < 6 Months 7 6 to 12 Months 1 12 Months (%) (a) Separating by holding period Trade Trades 9 >1 Trades (%) (b) Separating by trader group Figure 3: Fraction of corporations and trusts out of all transactions by year. 11

12 Dependent Variable is log Buy Price: sample 1998/1 to 22/12 Los Angeles Sacramento San Diego San Francisco I(h < 12m).1329 [ 8.336] I(h < 12m, Ntr 2, Ntr 1).987 [ ] I(h < 12m, Ntr > 1).157 [ ] I(h 12m, Ntr 2, Ntr 1).173 [ 6.195] I(h 12m, Ntr > 1).534 [ ].118 [ ].98 [ ].111 [ ].221 [ ].438 [ ].473 [ ].884 [ 2.978].484 [.8299].14 [.1725].196 [.7711].191 [ 6.191].225 [.8651].866 [.9716].266 [ ].427 [ 2.957] House Char Y es Y es Y es Y es Zip FE Y es Y es Y es Y es Year/Quart Start FE Y es Y es Y es Y es Adj R % % % % N. Obs. 16,651 26,641 1,98 71,145 Dependent Variable is log Buy Price: sample 23/1 to 27/12 Los Angeles Sacramento San Diego San Francisco I(h < 12m).81 [ ] I(h < 12m, Ntr 2, Ntr 1).243 [ 3.147] I(h < 12m, Ntr > 1).514 [3.6974] I(h 12m, Ntr 2, Ntr 1).17 [.8822] I(h 12m, Ntr > 1).124 [ ].876 [ ].258 [ ].147 [.4836].98 [ 1.813].19 [.9112].628 [ ].379 [ ].75 [4.3496].8 [.2253].7 [.549].814 [ ].11 [.569].956 [6.6288].77 [ ].9 [.131] House Char Y es Y es Y es Y es Zip FE Y es Y es Y es Y es Year/Quart Start FE Y es Y es Y es Y es Adj R % % % % N. Obs. 183,412 41,317 22, ,513 Dependent Variable is log Buy Price: sample 28/1 to 212/12 Los Angeles Sacramento San Diego San Francisco I(h < 12m).2269 [ ] I(h < 12m, Ntr 2, Ntr 1).41 [ ] I(h < 12m, Ntr > 1).129 [.7236] I(h 12m, Ntr 2, Ntr 1).423 [ ] I(h 12m, Ntr > 1).814 [ ].2123 [ 1.751].56 [ ].546 [ 2.375].854 [ ].1269 [ ].1987 [ ].427 [ 4.833].11 [.7983].567 [ ].833 [ ].273 [ ].429 [ ].593 [2.1714].799 [ 1.152].1458 [ 7.181] House Char Y es Y es Y es Y es Zip FE Y es Y es Y es Y es Year/Quart Start FE Y es Y es Y es Y es Adj R % % 86.6 % % N. Obs. 21,473 66,512 17, ,371 Table 2: Regressions on log prices. Standard errors clustered by Zip code and start quarter. 12

13 As pointed out above, price discounts can be explained by the presence of households with high holding costs for their house, for example due to distress. We show in figure 18 in Appendix B that starting from 28 a large fraction of the transactions in our database involve the acquisition of a property that is sold as Real Estate Owned (REO) or through a short sale. This is particularly the case if we look at the two categories of more active traders. We want to explore whether houses bought in distress entirely drive the price discount effect highlighted in tables 1 and 2. In order to do that, we extend regression equation 1 and include a dummy variable for houses bought as distress sales. That way, the discounts will be estimated only for properties that were not bought in distress. We estimate the regression equation in each urban area and each subperiod as in table 2. Table 13 in Appendix B reports the results. Even if the magnitude of the price discounts is now smaller, the nature of our results does not change. This interestingly highlights that the price discount is not entirely explained by the decision to buy distressed properties. Thus, traders working on the short holding periods seem to be able to exploit both search and information frictions that go beyond the ability of identifying distressed sales. Of course, the existence of these average price discounts does not have any direct implications for the returns earned by house traders. First, middlemen trades may need further investment in remodeling, which could be substantial. We find that for 14.8 % of house transactions in our database, a remodeling job is carried out before the housing unit is sold again. For slightly more than 5% of these remodeling contracts, Buildzoom provides the cost of the renovation job. Panel (a) of table 3 reports median purchase prices and remodeling costs for the houses over the entire sample between January 1998 and December 212. All dollar figures are in terms of December 213 dollars. Panels (b) and (c) report respectively median prices for houses that were remodeled, but for which the cost of the remodeling job is not available, and houses that were not remodeled. Since information for the transactions in panel (b) is incomplete, we will drop them from our further analysis. Second, in performing their activity middlemen are exposed to a variety of risks, which include global and local fluctuations of housing markets, as well as risks idiosyncratic to each specific property that they are trading. For example, the time needed for renovation might end up being longer than expected, or the condition of the house could turn out to be worse than what assessed at the time of buying. The assessment of trades performance are the focus of the next two sessions. 4 Measuring Performance Assessing the performance of real estate traders requires disentangling the part of returns that is determined by compensation for the risks involved in trading this specific asset from the value that skilled traders might be able to extract from housing markets. The financial economics literature has faced the same problem when studying the performance of professional investors in the mutual funds, hedge funds and private equity industries. The approach from the mutual funds requires to determine the key risk factors that drive risk compensation in the economy (or in the context of a specific asset class) and the loadings of these factors on the returns 13

14 Panel (a): transactions with remodeling permits (job value available) Type of Trader Value at Purchase Remodeling Costs N 1 trade $ 479,581 $ 1,188 67, trades $ 43,186 $ 9,282 16,256 >1 trades $ 283,444 $ 9,6 1,622 Panel (b): transactions with remodeling permits (job value not available) Type of Trader Value at Purchase Remodeling Costs N 1 trade $ 518, , trades $ 46, ,61 >1 trades $ 283,691. 2,581 Panel (c): transactions without remodeling permits Type of Trader Value at Purchase Remodeling Costs N 1 trade $ 446, , trades $ 391, ,381 >1 trades $ 347,868. 3,727 Table 3: Median transaction values and remodeling costs, all transactions taking place between January 1998 and December 212. Values are expressed in terms of December 213 dollars. generated by a specific investment fund (see for example Fama and French (21)). The part of returns that is not explained by risk factors represents the abnormal or risk adjusted performance of the manager, which is not driven by compensation for risk. This methodology has also been interpreted as a horse race of the fund against an alternative strategy, which is invested in portfolios that replicate exposures to the risk factors. However, as pointed out by Berk and van Binsbergen (214), this interpretation is flawed when factors cannot be replicated using existing investment opportunities. Figure 4 presents a simplified example of how the basic conceptual framework can be applied to real estate trades in our database. A trader acquires property i at time t with initial investment X t. In the example, the investment delivers over horizon T a total return R i,t+t, while R b,i,t+t is the return determined by the exposure of the trade to the relevant risk factors. The difference R i,t+t R b,i,t+t represents the gross abnormal return of the trade. With gross we mean that it captures the sum of the abnormal return earned by the real estate trader and the compensation gained by any investors providing the trader with capital. In other words, if the capital X t is provided by the trader herself, the gross abnormal return is just the risk adjusted compensation earned by the trader. If X t is provided by a third party, we can assume that there is a scalar τ i > so that (1 τ i ) (R i,t+t R b,i,t+t ) is the risk adjusted compensation earned by the trader, and τ i (R i,t+t R b,i,t+t ) is the share earned by the investor. The latter one is called in the literature net abnormal return. The split τ i is determined by the nature of the contractual relationship between investor and trader and will therefore depend on the market power of the two sides. Thus, as pointed out in Berk and Green (24) and Berk and van Binsbergen (214), the net abnormal return is not a measure of the value created by professional investors. We have limited and incomplete information on sources of financing for real estate traders in the data. Along the same lines, we do not know whether professional traders are earning fees when performing their 14

15 X t R b,i,t+t + X t (R i,t+t R b,i,t+t ) }{{} } Gross Abn. Ret. {{ } Value Added X t Figure 4: Simplified diagram of trade returns. activity. Thus, we are not able to compute a measure of net abnormal returns. Nonetheless, we don t believe this is a shortcoming of our work. The fact that gross abnormal returns are greater than zero is a necessary condition for there to be any value to share. Berk and van Binsbergen (214) highlight that also gross returns can be a misleading measure of the rents extracted by traders from the market. Their argument, which is applied to the specific case of mutual funds, is that the money manager chooses the dollar size of the rent that she plans to extract by selecting fund size. Thus, the authors argue that gross returns are a reliable measure of the rents extracted only when mutual funds are of very similar size, which is not the case in the data. An alternative measure introduced by Berk and van Binsbergen (214) is value added, which is just equal to the gross abnormal return times investment size. In our simplified example, we can calculate value added, as shown in figure 4 as X t (R i,t+t R b,i,t+t ). While our analysis is related to performance measurement in the mutual funds literature, the specific market that we are studying carries many additional challenges. First, the construction of returns for real estate investments is a non trivial problem, since these trades involve multiple unevenly spaced cashflows. Moreover, it is problematic to define R b,i,t+t. Real estate markets are characterized by frictions, incomplete information and are in general illiquid (see Piazzesi and Schneider (29) and Piazzesi et al. (215)). Thus, the difference R i,t+t R b,i,t+t could be driven to a large extent by risks and frictions that are not explained by the chosen R b,i,t+t, rather than by the ability of the trader to extract rents. Figure 5 shows a diagram of the cashflows from an investment in housing unit i, bought at time t and sold at time t + T. The variable X i,t is the initial investment equal to the amount paid to buy the house. There can be H additional intermediate investments X i,t+τh which capture remodeling costs, with < τ h < T. Moreover, the landlord can collect dividends from the asset, in terms of the monthly rents D i,t+1, D i,t+2,... Finally, X i,t+t is the final payoff from selling the house. D i,t+1 D i,t+2 D i,t+τh D i,t+t + X i,t+t X i,t X i,t+τh Figure 5: Cashflows diagram for real estate investments. 15

16 As already discussed in the previous sections, the investment horizon T is highly heterogeneous across different transactions and tends to be relatively long (at least a few months and up to many years). To measure the risk adjusted performance of private equity funds, Kaplan and Schoar (25) develop the Public Market Equivalent (PME), which is a measure of the excess return of a fund with respect to an investment in the S&P 5 with the same inflows. Under the assumption that private equity funds have a beta of 1 with respect to the market index, the PME is a measure of risk adjusted performance, or in other words a measure of the component of returns which is not explained by systematic or market risk. Of course, if private equity funds have a beta larger than 1 (as it is reasonable to assume), the PME is a biased measure of abnormal performance. For the purpose of our analysis, we believe the performance of local housing markets to be a reasonable benchmark for the returns of real estate trades and we define the Local Market Equivalent (LME) in the same fashion as the PME, but with respect to ZIP code house price indexes. We download the series of Zillow Home Value Index (ZHVI R ) for ZIP codes in the California counties covered in our study 1. Zillow uses a statistical model to track the valuation of 1 million homes in the United States 11. The ZIP code indexes are computed as median value of homes in the area and are published monthly. The LME for a transaction involving property i between time t and t + T is then computed as: LME i,t = RZIP,i t+t X i,t+t H h=1 R ZIP,i t+τ h X i,t+τh 1 (2) X i,t Where R ZIP,i is the return on the price index for the ZIP code where property i is located. Notice that we omit rents from the performance calculation and house price indexes do not include rent payments. Thus, we are consistently comparing ex-dividend performance for the trade 12 and the benchmark index. Moreover, if we make the (admittedly strong) assumption that rent rates are uniform within ZIP code, we can interpret LME as a measure of total excess return with respect to the ZIP code index. In line with what discussed above for PMEs, under the assumption that the loading of real estate returns on the local price indexes is equal to one, the LME is the component of returns which is not explained by fluctuations in local housing markets. However, it is reasonable to believe that houses with different characteristics will have different exposures to local risks. Moreover, loadings on the local house price index might change over time. Following Kaplan and Schoar (25), in our empirical analysis we try to deal with these issues by controlling for housing unit characteristics and for time fixed effects. Still, there can be other risks, related to liquidity, or the likelihood of performing a specific trade within a certain time window, that are not captured by general fluctuations in the level of ZIP code indexes. We will address these issues in a later section of the paper. Finally, the LME is a measure of gross abnormal returns, but can be easily transformed into a value added 1 The data are available at 11 More information on the methodology is available at 4ed2b.pdf. 12 While our data provide extensive micro level information on transaction prices and remodeling costs, they do not include rent data at the level of individual properties or even individual ZIP codes covering the entire period of our study. 16

17 measure: V LME i,t = LME i,t X $t end i,t Where X $t end i,t is the value of the initial investment needed to buy the property in terms of December 213 dollars. 5 Performance of Real Estate Traders Figure 6 shows the distribution of LMEs over the entire sample (from January 1998 to December 212) as wells as for sub-periods based on when the trade was started. The sub-periods are respectively from January 1998 to December 22, January 23 to December 27 and January 28 to December 212. The mean LME over the entire sample is %, while the median is %. The distribution of LMEs is quite similar for the pre-boom and the boom-bust years. It is slightly positively skewed, but overall centered around zero. Skewness increases in the last sub-period. Figure 7 shows the same distributions for value added. The mean value added over the entire sample is approximately -3, $, and the median is close to -4, $. Unlike the distributions of LMEs, the ones of value added are negatively skewed. In general, it appears that real estate trades in our database are slightly underperforming the Zillow ZIP code indexes : Mean :22 Mean (%) :27 Mean :212 Mean Figure 6: Distribution of LMEs. Full sample and sub-periods. In table 4 we report quantiles of the distribution of LMEs by holding period and by trader type in 17

18 :212 3 Mean , $ : Mean :27 Mean :212 Mean Figure 7: Distribution of VAs. Full sample and sub-periods. Values are in tems of thousands of December 213 dollars. different sub-periods. The sub-periods are again for trades that were started from January 1998 to December 22, from January 23 to December 28 and from January 29 to December 212. As usual, we split traders based on how active they have been over the last two years. We distinguish between traders that have bought one, between two and ten and more than ten houses. LMEs are larger for holding periods shorter than one year across all sub-periods (see panels (a), (b) and (c)). When we look at the breakdown across trader types, we find that LMEs are clearly larger for more active traders in the period from January 28 to December 212 (panel (f)). There is some evidence in the same direction for the first period, from January 1998 to December 22, especially for the right tail of the distribution. However, this is not the case during the years of the boom, when the distribution of LMEs is very similar across different trader types. Table 5 the median value added for trades with holding period smaller than one year is approximately 3, $ in the first two sub-periods and up to 85, $ in the most recent period, while the median value added is always negative for trades with longer holding period. More active traders have obtained larger value added in the most recent period, but not during the boom. 18

19 Panel (a): 1998/1 to 22/12 q9th q75th q5th q25th q1th h < 6m h 6m, h < 12m h 12m Panel (b): 23/1 to 27/12 q9th q75th q5th q25th q1th h < 6m h 6m, h < 12m h 12m Panel (c): 28/1 to 212/12 q9th q75th q5th q25th q1th h < 6m h 6m, h < 12m h 12m Panel (d): 1998/1 to 22/12 q9th q75th q5th q25th q1th nt r = nt r 2, nt r nt r > Panel (e): 23/1 to 27/12 q9th q75th q5th q25th q1th nt r = nt r 2, nt r nt r > Panel (f): 28/1 to 212/12 q9th q75th q5th q25th q1th nt r = nt r 2, nt r nt r > Table 4: Distribution of LMEs over the entire sample and by sub-periods. 5.1 Differences in Average Performance: a First Look We want to analyze differences in average performance for different trader types and different holding periods. We therefore estimate regression equation 3: y i = a z + a yq,s + a yq,e + B house X house i + β 6m I(h i < 6m) + β 6m12m I(h i 6m, h i < 12m)+ + B type 6m I(h i < 6m, type) + B type 6m12m I(h i 6m, h i < 12m, type)+ + B type 12m I(h i 12m, type) + v i (3) Where y i is either the LME for trade i or its value added. As already pointed out above, LMEs and value added are measures of abnormal performance only under the assumption that returns to the individual 19

20 Panel (a): 1998/1 to 22/12 q9th q75th q5th q25th q1th h < 6m 88,289 59,116 33,464-7,984 h 6m, h < 12m 96,845 56,937 26, ,318 h 12m 47,86 1,9-23,725-68, ,893 Panel (b): 23/1 to 27/12 q9th q75th q5th q25th q1th h < 6m 152,541 88,299 25,483-17,248 h 6m, h < 12m 174,216 91,25 32,171-3,87-39,582 h 12m 95,321 23,83-34,414-98,23-183,888 Panel (c): 28/1 to 212/12 q9th q75th q5th q25th q1th h < 6m 155, ,94 84,2 46,32 h 6m, h < 12m 197, ,855 87,953 49,764 7,59 h 12m 5,366 1,416-4,972-92, ,142 Panel (d): 1998/1 to 22/12 q9th q75th q5th q25th q1th nt r = 1 51,258 11,879-22,17-68,42-15,776 nt r 2, nt r 1 56,188 18,58-17,33-52, ,838 nt r > 1 59,149 33, -4,227-37,158-75,614 Panel (e): 23/1 to 27/12 q9th q75th q5th q25th q1th nt r = 1 16,5 25,965-34,47-1,26-186,355 nt r 2, nt r 1 87,584 15,111-41,587-12,829-18,52 nt r > 1 69,995 6,2-27,911-87, ,647 Panel (f): 28/1 to 212/12 q9th q75th q5th q25th q1th nt r = 1 67,98 11,26-35,848-88,8-173,468 nt r 2, nt r 1 98,79 38,924-15,291-61, ,8 nt r > 1 117,55 73,439 8,613-34,85-81,971 Table 5: Distribution of value added over the entire sample and by sub-periods. Values are in terms of December 213 dollars. housing units have an exposure equal to one to the local house market index. This assumption is likely to be violated in the data, and house trades risks might be driven by house characteristics, as well as the Zillow index might be a biased measure of fluctuations in housing markets for specific ZIP codes or points in time. We therefore introduce controls in the regression to account for these problems. In particular, the coefficients a z, a yq,s and a yq,e are respectively ZIP code fixed effects and year/quarter fixed effects for the starting and ending quarter of the trade. The vector X house i contains house characteristics. Then, we include dummies for holding period and trader type. The variables I(h i < 6m) and I(h i 6m, h i < 12m) are equal to one respectively for holding periods smaller than 6 months and in between 6 and 12 months. The elements of the vector I(h i 6m, h i < 12m, type) are interactions between dummies for the active trader types (between 2 and 1 and more than 1 trades in the last two years) 2