Price discounts at residential real estate auctions : Tte effect of liquidity and quality

Price discounts at residential real estate auctions:

-The effect of liquidity and quality-

Hugo Roosjen

10662707 July 2014

Master thesis:

MSc Business Economics: Real estate Finance University of Amsterdam

First supervisor: dr. Martijn Dröes Second supervisor: prof. dr. Marc Francke

Acknowledgements

Completing this thesis was a challenging period in which I could use the knowledge and skills, which I acquired during my bachelor’s and master’s degree courses. I would like to thank Martijn Dröes and Alex van de Minne for their help and support during the writing of this thesis. They both assisted me with questions concerning the repeat sales model and acquiring the data that were necessary to conduct this research. I would also like to thank the NVM, for providing the dataset. Special thanks go out to my parents, who supported me throughout my academic career.

Abstract

This thesis investigates the extent to which price discounts at Dutch auctions exist compared to regular (negotiated) sales and it elaborates on the liquidity and quality of dwellings as underlying factors that could cause an auction discount in the Dutch housing market. By examining 2,250 transactions, of which 751 are auction transactions, it can be concluded that an average discount of 16.3% at Dutch auctions exist. Evidence is found that illiquid dwellings at auction sell for 8% additional discount compared to liquid dwellings. Weak evidence is found for lower quality dwellings to sell for an additional discount of about 6% compared to higher quality dwellings. The auction discount cannot be fully attributed to quality and liquidity of dwellings; the remaining ‘auction effect’ is about 10%.

Table of contents Page: Chapter 1 Introduction………4 Chapter 2 Literature review………..7 Chapter 3 Methodology………..11 3.1

Repeat sales regression: The average auction effect………...……….13

3.2

The effect of illiquidity…….………...………...…….14

3.3

The effect of quality………...…16

3.4

A model with both illiquidity and quality………....17

Chapter 4

Data and descriptive statistics………18

4.1 Transaction prices………..20 4.2 Liquidity………21 4.3 Quality………. .23 4.4

Representativeness of the sample………...25

Chapter 6

Results………...26

Chapter 7

Conclusion………...31

Chapter 1 Introduction

Real estate auctions are used to sell residential real estate, either for foreclosure or voluntary sales. Not only in the United States, but also in The Netherlands, residential real estate auctions are common practice. Since the outburst of the financial crisis, auction houses experienced a large increase in auctions, mainly due to the extra dwellings that were sold due to foreclosure1_{. In 2013, 1,863 houses were sold at foreclosure}

auctions in the Netherlands2_{. Moreover, 34% of the owners of dwellings have a mortgage}

higher than the value of the asset, which makes them more vulnerable for foreclosure and the probability of an involuntary sale at an auction more likely. The Nationale Hypotheek Garantie (NHG), a Dutch guarantee fund for underwater mortgages opted for the ban of using auctions as a method of sale, since according to NHG, dwellings at auctions sell on average some 15,000 euros less when the properties are sold on the regular market.3

Conventional wisdom suggests that residential prices fetched at auctions are lower than actual property values simply due to the fact that the dwelling is auctioned. The question is to what extent an auction effect exists, or whether the house price discount at auctions is the result of particular characteristics of the dwelling that is auctioned. Economic theory would suggest that, ceteris paribus, auction prices and regular sales prices are equal. Otherwise, an arbitrage opportunity would exist. Two factors that might determine auction prices to be lower than regular negotiated sales are the liquidity and quality of a dwelling.

Private bidders at auctions could be hesitant to bid on low quality dwellings, since uncertainty about renovation costs could exist, whereas dealers see the opportunity to renovate and resell the dwelling to make a profit. Therefore, quality could be an explanatory factor for auction discounts. Auctions are known for their limited information provision and uncertainty about the quality could give rise to a negative price pressure on auctions. Also liquidity could influence auction discounts, since for an illiquid dwelling more time on average is needed to find the optimal buyer compared to a

1_{BNR, Executieveilingen explosief toegenomen. Available at:}

http://www.bnr.nl/nieuws/beurs/135549-1201/executieveilingen-explosief-toegenomen (accessed: 01-07-2014)

2 _{Kadaster Dashboard, Vastgoedbericht Executieveilingen. Available at:}

http://www.kadaster.nl/web/pagina/Vastgoed-Dashboard.html (accessed: 15-04-2014)

more liquid dwelling. Since auctions occur in a short amount of time, illiquidity could be an explanatory factor for auction discounts to exist.

The aim of this thesis is to investigate the extent to which price discounts at Dutch auctions exist compared to regular (negotiated) sales and it elaborates on the underlying factors that could cause an auction discount in the Dutch housing market. Specifically, this thesis focuses on two determinants of auction discounts 1) the liquidity of the dwelling and 2) the quality of the dwelling. The following three hypotheses are examined:

Hypothesis 1:

- Dwellings offered at Dutch auctions sell on average at lower prices in comparison to regular

negotiated sales

Hypothesis 2:

- The auction discount is partly attributable to the fact that illiquid dwellings transact at higher discounts than liquid dwellings at auction.

Hypothesis 3:

- The discount at auction is influenced by the quality of a dwelling, where low quality dwellings

transact at an additional discount compared to higher quality dwellings.

To examine these hypotheses, data from the Dutch Association of Realtors (NVM) is used. The NVM data contains 760 auctioned dwellings located in the Netherlands for the period 1973-2014. The total number of transactions before and after auctions equals 2,250, of which 751 are auction transactions. The dataset contains data on transaction dates, condition of sale and location, but also the perceived quality of a dwelling, assigned by real estate brokers. With multiple negotiated sales prices and auction transaction prices available, a repeat sales regression is used to analyze the discounts and to investigate the liquidity and quality effect at auctions.

If the conventional wisdom is correct, owners ending up in foreclosure, who are already in financial distress, will even be hit harder, because of the discount for which their house will probably transact at auction. Therefore, insights in factors influencing the auction discount are desirable to investigate whether arbitrage opportunities at auctions exist, or if a stigma effect is present.

This is not the first study on auction discounts. Studies like Mayer (1989), investigate whether dwellings at auctions sell for discounts or premiums. Most studies find evidence that an auction discount exists. Brounen et al. (2009) makes a distinction for the Dutch market between voluntary and foreclosure auctions for which he finds 2.5% and 34% discount respectively in the period 2006-2007. Interesting is the difference between the magnitude in discounts for foreclosure and voluntary auctions, as Harding et al. (2012) conclude that a foreclosure discount does not appear to exist (i.e. they argue it is a myth). Brounen et al. (2009) assign the discount at auctions to information asymmetry about the condition of the dwelling and transaction costs. The authors state that the higher underpricing is present for older dwellings, due to the fact that the bidder, who does not have the opportunity to inspect the dwelling, will probably count on a lower quality, and therefore, higher renovation costs. Where most studies focus on finding proof for discounts, insights in which factors influence this discount have not been thoroughly investigated yet. Although assumptions that for instance quality and the limited time to arrange finance when buying at auction could be reasons that the auction discount exists, evidence that these assumptions are significant factors have not been proven yet.

This thesis contributes to the existing literature by examining the effect of liquidity, which has not been previously examined in the existing literature on auction discounts. In line with the paper of Brounen et al. (2009), stated quality is also taken into account. However, instead of the age of the dwelling to measure the quality effect, a direct measure of quality will be used.

The results in this thesis show that a substantial housing auction price discount in the Netherlands exists. In particular, the results show an average auction discount of 16.3%. Illiquid dwellings at auction have an 8% higher discount compared to liquid dwellings. Low quality dwellings transact at 6% extra discount compared to a very high quality dwelling. Although illiquidity and low quality explain a substantial part of the discount we find that dwellings with good liquidity and very high quality still sell at a price discount at auction. This suggests that there may either be other important factors that determine auction discounts or the fact that the dwelling is auctioned leads to a discount, but that two of the hypothesised main determinants of auction discounts cannot explain the full auction discount. These results also imply that people, who are forced to sell their house at an auction, will probably be left with an additional loss on their house.

The remainder of this thesis is structured as follows. Chapter 2 discusses the previous literature on auctions discounts/premiums. Chapter 3 describes the repeat sales methodology and Chapter 4 gives an overview of the data. Chapter 5 discusses the results. Chapter 6 provides a summary and conclusion.

Chapter 2 Literature review

Real estate auctions are a wide spread phenomenon, and so is the rationale that prices at auctions turn out to be lower than as if they were privately negotiated (i.e. regular sale). In the existing literature, researchers find discounts and premiums at auctions, but they vary widely in their explanations why these discounts or premiums exist. Many researchers address information asymmetry as main reason why discounts exist, although the magnitude of these discounts and premiums are different due to the methodology researchers use and due to differences in the examined periods and location. The methods to investigate discounts vary from using assessed values, hedonic pricing models or repeated sales regressions. Table 1 gives an overview of the results from the previous literature.

Table 1. Literature overview

Author Market Period Discount/Premium Method

Discounts

Mayer (1998) Los Angeles/Dallas 1983-1990 0-9% - 9-21% Repeat sales Allen & Swisher (2000) Florida 1998 13,45% - 34,21% Assessment ratios Brounen & Rijks (2009) The Netherlands 2005-2008 2,5% - 37% Hedonic

Wright (1989) U.S. 1985 25% Hedonic

Premiums

Ashenfelter & Genesove (1992) New Jersey 1990 13% Direct comparison Dotzour et al. (1998) New Zealand 1991-1992 5,9%-9,5% Hedonic

Lusht (1996) Melbourne 1988-1989 8% Hedonic

Dotzour, Moorhead & Winkler (1998) investigated prices at auctions versus negotiated sales in Christchurch, New Zealand, between 1991 and 1992. Using a hedonic approach, the authors find that in the studied sample of New Zealand, there are no significant discounts at auctions present. In fact, the authors find that auctioned dwellings are sold at a premium instead of a discount, especially for above average priced categories. The fact that auction prices equal the negotiated sales prices in the moderate price ranges is

assigned to more market activity and more information of comparable sales. Dotzour et al.(1998), emphasize the fact that the New Zealand auction market differs from auction markets in other countries. Whereas in other countries like the U.S. the auctions mostly contain foreclosure dwellings, New Zealand auctions are seen as a good alternative for negotiated sales. Due to the different conditions of sale with respect to foreclosure and voluntary auction sales, differences in discounts or premiums exist in markets across locations. Harding et al. (2012) conclude that with respect to negotiated sales, no discount is found between foreclosure sales and normal sales. In accordance with Dotzour et all. (1998), Lusht (1996) finds premiums on the New Zealand auction market. The author finds a premium at auctions of 5.6% in the 1990 paper and a premium of 8% in the paper from 1996. In both studies, Lusht uses middle to high priced dwellings. This data selection could introduce bias, since the question arises whether the sample is representative compared to the auction market and therefore could introduce a sample selection bias in the hedonic approach that has been used. Ashenfelter and Genesove (1992) find an auction premium of 13 percent for 83 condominium units in New Jersey in 1990. The method used to assess the discounts is rather questionable. The condominiums were put up for auction and when the condominium did not fetch a price at auction, it was sold face-to-face. The face-to-face sale results were compared to the auction prices to calculate the premium at auction. The question arises whether the method to investigate the auction premium is reliable. The dwellings that are sold face-to-face have proven not to be of interest at auction and therefore it seems obvious that the price will be lower at the face-to-face sale, since it is the second time that the dwellings are put up for sale. Buyers will have a stronger negotiation position, since otherwise they would have bought the dwelling at the auction.

While the results of the above-mentioned literature show auction premiums, there is also literature that mainly finds an auction discount. Allen and Swisher (2000) investigate prices on Department of Housing and Urban Development auctions. The authors use a method of comparing assessment ratios with auction outcomes to assess discounts. The auction discounts vary with the geographical subsamples from 34.21% to 13.45%. These findings could be influenced by the fact that most auctioned dwellings involve foreclosures when owners are in need to dispose the (housing) asset in a short period of time, even for a low price. In contrast to Lusht (1994) and Mayer (1989), Allen and Swisher (2000) find an increase in prices fetched at auctions as the auction proceeds, also

known as the declining/increasing price anomaly. An increase in bidder aggression due to the decreasing opportunity of finding a bargain when the auction proceeds, is seen as a reasonable explanation for this price increase. It is important to note that in this study, it assumed that assessed values are equal to market values. When investigating price discounts at auctions, it is important to know whether the assessment values that are used for comparing auction outcomes are reliable since otherwise there could be over- or underestimation. For the Dutch market, De Vries et al. (2009) conclude that assessment values based on Dutch appraisal data are reliable. It should be noted that other methods are also likely to have some estimation error. Wright (1989) finds an average of 37% discount on auctions in the US. Wright mentions that a shortcoming in his research is that he was unable to correct for the problem that auctioned properties are mostly low-value properties, which introduce bias to the results. In the methodology part of this thesis, different methods and problems like the sample selection bias as noticed in Wrights paper will be discussed.

Research on the Dutch auction market is scarce. Brounen and Rijk (2009) find discounts on the Dutch market, by using 700 transactions in the period 2005 – 2008. By comparing 290 foreclosure auction transactions with a hedonic model, they find an average discount of 37%. With respect to 413 observations on normal auctions, they find a discount of 2.5%, a large difference to foreclosure auctions. The authors assign the difference in discounts to extra transaction costs and asymmetric information about the condition of the dwelling. Older dwellings have larger discounts due to the probability that the condition of the dwelling is poorer, but unknown to the bidder. In the case of a normal non- foreclosure auction, the bidder has the opportunity to inspect the dwelling, which leaves no uncertainty about the dwellings’ condition and therefore the auction price approaches the market value. Although the author states that the bidder takes the probability of low quality into account when forming a bid, it is important when investigating discounts at auction, the method of sale is under investigation and not the fact that lower quality dwellings are sold under average prices. Brounen et al. (2009) assign this finding of higher discounts to the probability of a lower quality, by assuming that older dwellings are lower quality dwellings. However, it is interesting to see that older dwellings in voluntary auctions actually sell for premiums. The question arises whether the assumption that older dwellings are lower quality dwellings holds. Since in

the dataset of this thesis quality is known, it is investigated whether the statement that the quality of a dwelling could explain a part of the auction discount holds.

Mayer (1994) describes the cost of liquidity as a reason for discounts at auctions, due to the fact that selling through other channels provides more time to wait for a better deal. By waiting in a negotiated sale, the seller might find a buyer who is a better match for the dwelling, and is willing to pay a higher price. In the case of auctions, there is a quick sale, and it is rather unlikely to find a perfect match within the short time period. The author uses a partial equilibrium model, in which the value of a dwelling is calculated by investigating the buyer’s valuation of that dwelling, which is determined by the number of available units. Higher discounts in down markets are examined by using simulations of short run variations in vacancy rates, since in down markets the vacancy rate will be higher and consequently time on markets will be longer. In a negotiated sale, it is assumed that when a valuation by a potential buyer is higher than the seller’s reservation price, there will be a transaction and the surplus will be divided between the seller and buyer. In the case of an auction, the price will only be equal to the valuation of the second highest bidder, the so-called underbidder. Adams, Kluger and Wyatt (1992) also conclude that an auction will always transact at a lower price than a negotiated sale, due to the available time on market.

In contrast to the hypothesis that a longer time on market leads to a higher price, since more time is available to find the buyer with the optimal valuation, one could also form the hypothesis that a longer time on market leads to a lower price, due to increased bargaining power. Next to increased bargaining power, a long time on market could imply that a dwelling is just not attractive to buy. Krainer (2008) finds that prices and transactions are negatively correlated to the time on market. This research uses the above-mentioned findings on the time on market effect, to further examine if a long time on market for previous sales, which is used as a proxy for illiquidity, has an effect on the magnitude of the discount or premiums received at an auction.

It is important to note, that most of the above-mentioned literature is dated and based on auctions in several different countries. The outcomes should therefore be interpreted with caution when comparing with the results in this thesis. For instance, the New Zealand and Australian auction markets have very different characteristics than for

instance the U.S. market or the Dutch market. The Dutch housing market is characterized by on average high loan to value and recourse mortgages. According to CBS, 34% of the owners of residential dwellings in The Netherlands had a mortgage higher than the value of the asset for the year 2013 and young people have the largest share in these underwater mortgages.4

As shown in the literature, auction markets, which are known by a wider public, show more positive results than markets where auctions are mostly used to obtain a sale within a short period of time. Apart from geographic differences, markets develop over time. In comparison to auctions markets in the 1990s, anonymous bidding via Internet has become more widely available. The availability of Internet bidding could open up the bidders’ market and make buying at auctions more attractive, both for buyers and for sellers. Due to anonymity, unfair competition that results in price agreements is less easily accomplished by dealers who are buying at auctions. These developments could affect auction results and their discounts or premiums.

Chapter 3 Methodology

As shown in previous literature, different kinds of methodologies are used to investigate price discounts at auctions. Some researchers use assessed values to compare to prices at auctions; others use hedonic- or repeated sales models. This research uses the repeat sales approach, which uses data on the same dwelling in two or multiple periods.

Bailey, Richard, Muth and Nourse (1963) argue that multiplicative chain indexes as proposed by Laurenti (1960), which compares average sales prices for a given time period and location, face two difficulties: variation in quality among observed properties and a change in quality over time. According to the authors, the solution to overcome these difficulties is to use regression analysis. When the index is based on transaction prices of the same property at multiple times, the difficulty of quality differences could be reduced. Bailey et al. (1963) were among the first in describing the repeat sales method.

The advantage of the repeat sales model in comparison to the hedonic model is the relative few variables that are needed in order to do the regression analysis. In the basic

4 _{NVM, CBS: 1,4 miljoen huiseigenaren ‘onder water’ Available at:}

http://www.nvm.nl/actual/maart_2014/cbs_1_4_miljoen_huiseigenaren_onder_water.aspx (accessed: 08-04-2014)

repeat sales regression, only prices, addresses and sale dates are used. Data on other variables are not necessary, since for each price pair, the same dwelling is taken into account over time and constant characteristics are captured in the fixed effects. On the contrary, for the hedonic model, one needs an extensive dataset on the dwellings and its characteristics. For some variables, the data might be easily retrievable. For instance, location, square metres and number of rooms are examples that are registered by the land registry. The variable quality is more difficult to measure, since quality is a measure that depends on the subjectivity of the appraiser. Therefore, omitted variable bias in hedonic models is likely to occur. Mayer (1989) uses a repeat sales model and a hedonic model to assess auction discounts. Although the results show a similar sign of the coefficient, they are not similar in magnitude. Whereas discounts between 9-21% are found for the Dallas market with a repeat sales index, discounts between 21-31% are found by using a hedonic model. Mayer attributes this difference to omitted variables (house characteristics) that account for the more negative result in the hedonic model. Since characteristics are correlated with the method of sale, he concludes that the repeat sales model is a more appropriate model to investigate auction discounts (i.e. to avoid omitted variable bias).

In the repeat sales model, it is assumed that only dates of sale and sales prices differ between the sales. For instance, square feet and number of rooms are assumed to stay constant over time. The problem arises whether the dwelling indeed remains the same over time, since depreciation and renovations could influence the condition or quality of the dwelling. The longer the time period between sales, the more uncertainty exists whether the constant quality assumption holds. Case and Shiller (1987) therefore introduced an addition to the existing repeat sales model, the weighted repeated sale index, which gives weight to the time periods between the two sales. Case and Schiller state that the variance of the noise is not constant for house-specific factors, but increases over the time period between the transactions and is therefore not homoscedastic. Mayer (1989) uses the weighted repeated sales method proposed by Case and Shiller (1987), to correct for possible heteroskedasticity, but notes that the estimated coefficients do not differ much and there is almost no effect on the auction dummies. Jansen et al. (2008) conclude that the standard model of repeat sales as proposed by Bailey et al. (1963), is adequate for constructing a house price index for the Netherlands and no evidence is found for heteroskedasticity. Therefore, the weighted repeat sales

method is not chosen to use in this thesis. A disadvantage of the repeat sales model is the inefficient use of available transaction data, since the model requires at least two transactions and ignores the information available on single sales (i.e. sample selection problem).

In this thesis, the application of the repeat sales model as proposed by Bailey et al. (1963) is used. Apart from the time fixed effects, an auction dummy is included in the regression to estimate the auction discount. Interaction effects with illiquidity and quality are used to investigate additional discounts. Each following subsection describes the regression models estimated in this thesis. In all of the regressions, robust standard errors are used to adjust for heteroskedasticity.

3.1 Repeat sales regression: The average auction effect

The basic repeat sales regression with the additional differenced auction dummy variable takes the following form:

𝑝𝑖,𝑠− 𝑝𝑖,𝑏 = 𝜇𝑠− 𝜇𝑏+ 𝛼(𝐴𝑖,𝑠− 𝐴𝑖,𝑏) + 𝜀𝑖𝑡 (1)

where 𝑠 and 𝑏 are subscripts for time within a transaction pair for dwelling 𝑖, 𝑠 indicates a second occurrence and 𝑏 indicates a first occurrence (i.e. buy versus sell), with 𝑠>𝑏, 𝑝_𝑖,𝑠− 𝑝_𝑖,𝑏 are differences in log prices between the second and first transaction for a transaction pair of dwelling 𝑖, 𝜇𝑠− 𝜇𝑏 is the difference between time fixed effects (i.e.

differenced time dummies are used), this difference takes a value of -1 when the dwelling is sold at time 𝑏; the difference equals 1 when the dwelling is sold at time 𝑠 and equals 0 otherwise5_{, 𝐴}

𝑖,𝑠− 𝐴𝑖,𝑏 is the differenced auction dummy6 takes a value of -1 when the

dwelling was auctioned at time 𝑏 ; the difference equals 1 when the dwelling was auctioned at time 𝑠 and equals 0 when there has been no auction, or both at time 𝑏 and time 𝑠 an auction of dwelling 𝑖 occurred, and 𝜀𝑖𝑡 is the error term (i.e. captures the

differenced error term between 𝑠 and 𝑏).

5 _{The fixed effects term equals 0, when no transaction (𝑏 or 𝑠) in that specific month and year occurred, or}

when both 𝑏 and 𝑠 occurred in the same month and year.

6 _{The auction dummy takes the value of 1 when the dwelling was auctioned in that specific month and year}

Equation (1) is used to measure whether there is an auction discount on average, where 𝛼 is the parameter of interest. It is expected that 𝛼 will show a negative sign, which indicates that a difference between transaction prices at auction and negotiated sales exists (i.e. in line with the first hypothesis).

The duration of a month is chosen for the periods between the time fixed effects. This implies that for every month between 1973 and 2014 a dummy variable is created. The reason for choosing a month instead of for instance a year is the following: When a time span of a year is chosen, the problem arises when two transactions take place in the same year, (which occurs quite frequently) the price pair is not taken into account, due to the fact that the differenced dummies for that specific price pair will be equal to zero. Dwellings that are resold within a short time period are kept in the dataset, since the tendency exists that dealers resell a dwelling quickly after an auction.

3.2 The effect of illiquidity

For the investigation whether illiquid dwellings have a higher impact on the auction discount than liquid dwellings (the second hypothesis), an interaction term for illiquidity and auctions needs to be added to equation (1). By adding illiquidity, the equation takes the following functional form:

𝑝_𝑖,𝑠− 𝑝_𝑖,𝑏 = 𝜇𝑠− 𝜇𝑏+ 𝛼(𝐴𝑖,𝑠− 𝐴𝑖,𝑏) + 𝛽[(𝐴𝑖,𝑠− 𝐴𝑖,𝑏) ∗ (𝐿𝑖,𝑠− 𝐿𝑖,𝑏)] + 𝜀𝑖𝑡 (2)

where (𝐿𝑖,𝑠− 𝐿𝑖,𝑏) is the difference between illiquidity dummies. The difference takes a

value of -1 when the dwelling is assigned as illiquid at time 𝑏; the difference equals 1 when the dwelling is assigned as illiquid at time 𝑠 and equals 0 when either at time 𝑏 and time 𝑠 the dwelling was assigned as illiquid, or no illiquidity was assigned at all. The term (𝐴_𝑖,𝑠− 𝐴_𝑖,𝑏) ∗ (𝐿_𝑖,𝑠− 𝐿_𝑖,𝑏) is the interaction term between the auction and illiquidity term. Note that illiquidity is a variable that does not vary over time and as such is differenced out (although the interaction with the auction dummy remains).

The parameter of interest is 𝛽, which we estimate to examine whether illiquid houses sell for an additional discount at auction. The parameter 𝛼 captures the remaining effect for those houses that are liquid. With respect to the second hypothesis, it is expected that

that 𝛽 will show a negative sign, which implies that illiquid dwellings sell at additional discounts in comparison to liquid dwellings.

As time on market is used as a proxy for illiquidity, it is assumed that when a dwelling has a high time on market the dwelling is illiquid. The time on market is defined as the time between the date when the dwelling was put up for sale and the transaction date. Using only the calculated time on market for each transaction would not be an appropriate estimator for liquidity of a certain dwelling, since different factors can influence time on market over time. Therefore, several steps are necessary to investigate whether a dwelling could be defined as liquid or illiquid. A measure of liquidity is created using a three-step approach.

A first step is the creation of an index for the time on market. For each transaction the time on market is calculated and divided by the average of the time on market for each year and province. The reason for the use of averages per year and province is the variation of time on market over time and location7_{. For instance, the time on market}

could differ over time due to an economic downturn as seen in previous years and therefore the time on market in an up market cannot be compared to the time on market in a down market. To correct for market conditions in time on market, indexation is used based on the year of sale. The same logic applies for the indexation by location. When an average time on market per year is calculated, location should also be taken into account, as time on market differs geographically. The formation of the index takes the following functional form:

𝐼

=

𝜂𝑦,𝑝

̅̅̅̅̅̅ (3)

where 𝐼𝑖,𝑛 is the index value for the time on market per transaction 𝑛, 𝜂𝑛 is the time on

market in days per transaction and 𝜂

̅̅̅̅̅ is the average time on market across dwellings,

The second step is the creation of an average (across time/transactions) index value of time on market for each dwelling. As the time on market will differ between transactions

7_{One could argue that adjusting the average not only by year and province, but also by house type would}

be desirable, to make an optimal distinction in time on market to assign liquidity to a dwelling. Adjusting the average time on market for house type did not improve the results.

for the same dwelling, the average index value for time on market will give insights about the liquidity of the dwelling. When the index takes the value of 1, the dwelling has an average time on market. When the index value is above 1 selling the dwelling taken a relatively long period to sell.

𝑄𝑖 = 𝐼̅̅̅̅ 𝑖,𝑛 (4)

where 𝑄_𝑖 is the average index value for time on market per dwelling.

The third step is to generate a dummy variable 𝐿, which indicates whether the dwelling can be seen as illiquid or liquid. It is assumed that a dwelling is illiquid when the time on market is 20% higher than average. Therefore, the dummy for illiquidity 𝐿 takes a value of 1 when the average of the index for time on market for a certain dwelling takes a value of 1.2 or higher.

3.3 The effect of quality

Quality dummy variables are added to the model to investigate the effect of quality on the auction discount. Three variables, low quality, good quality and perfect quality are indicators of the quality and condition of a dwelling. The equation including quality dummies takes the following functional form:

𝑝_𝑖,𝑠− 𝑝_𝑖,𝑏 = 𝜇_𝑠− 𝜇_𝑏+ 𝛼(𝐴_𝑖,𝑠− 𝐴_𝑖,𝑏) + 𝛾[(𝐴_𝑖,𝑠− 𝐴_𝑖,𝑏) ∗ (𝑀_𝑖,𝑠− 𝑀_𝑖,𝑏)] + 𝛿[(𝐴𝑖,𝑠− 𝐴𝑖,𝑏) ∗ (𝐺𝑖,𝑠− 𝐺𝑖,𝑏)] + 𝜀𝑖𝑡 (5)

where (𝑀𝑖,𝑠− 𝑀𝑖,𝑏) is the difference between low quality dummies. The difference takes

a value of -1 when the dwelling is assigned as low quality at time 𝑏; the difference equals 1 when the dwelling is assigned as low quality at time 𝑠 and equals 0 when either at time 𝑏 and time 𝑠 the dwelling was assigned as low quality or no low quality was assigned at all. The term (𝐴𝑖,𝑠− 𝐴𝑖,𝑏) ∗ (𝑀𝑖,𝑠− 𝑀𝑖,𝑏) is the interaction term between the auction

and low quality term, (𝐴𝑖,𝑠− 𝐴𝑖,𝑏) ∗ (𝐺𝑖,𝑠− 𝐺𝑖,𝑏) is the interaction term between the

auction and good quality term (perfect quality is the reference category). Note that the quality indicators themselves are variables that do not vary over time and are differenced out.

The parameters of interest are 𝛾 and 𝛿, which are estimated to examine whether low quality and good quality dwellings sell for an additional discount at auction compared to perfect quality dwellings. The parameter 𝛼 captures the remaining effect for those houses that have very good quality. It is expected that both parameters 𝛾 and 𝛿 will show a negative sign, since lower quality dwellings will sell for an extra discount at auction compared to perfect quality dwellings, as proposed in the third hypothesis.

In the repeat sales model, it is assumed that quality remains constant over time. Therefore, the average quality is calculated per dwelling after which the dwelling can be assigned as low, good or perfect quality. The data and descriptive statistics Chapter (Chapter 4) elaborates further on this issue.

3.4 A model with both illiquidity and quality

In the next equation, interaction terms for illiquidity and quality are added into the model. It can be investigated whether there remains an auction discount, even after there is controlled for both illiquidity and quality.

𝑝_𝑖,𝑠− 𝑝_𝑖,𝑏 = 𝜇_𝑠− 𝜇_𝑏+ 𝛼(𝐴_𝑖,𝑠− 𝐴_𝑖,𝑏) + 𝛽[(𝐴_𝑖,𝑠− 𝐴_𝑖,𝑏) ∗ (𝐿_𝑖,𝑠− 𝐿_𝑖,𝑏)] +

𝛾[(𝐴_𝑖,𝑠− 𝐴_𝑖,𝑏) ∗ (𝑀_𝑖,𝑠− 𝑀_𝑖,𝑏)] + 𝛿[(𝐴_𝑖,𝑠− 𝐴_𝑖,𝑏) ∗ (𝐺_𝑖,𝑠− 𝐺_𝑖,𝑏)] + 𝜀_𝑖𝑡 (6) The interaction terms used in equation (6) are the same as previously described. The parameters of interest are 𝛽 , 𝛾 and 𝛿 , which are estimated to examine whether low quality, good quality and illiquid dwellings sell for an additional discount at auction. Parameter 𝛼 captures the remaining effect for those dwellings that are of perfect quality and have proven to be liquid.

Chapter 4 Data and descriptive statistics

To conduct this research, a dataset on repeat sales has been provided by NVM, the Dutch association of real estate brokers. The NVM database is one of the largest databases in the Netherlands containing information about residential real estate.

The dataset contains the repeat sales of 760 dwellings that have been sold at voluntary or foreclosure auctions8_{. The total number of transactions is 2,250 with an average number}

of 3 transactions per dwelling throughout the period 1973-2014. The dataset contains 751 auction transactions, which leaves 1,499 regular sales. Besides transaction prices, the dates of the transactions are known, as well as the date when the dwelling was put up for sale. With respect to the location, the variables postal code, municipality and province are known. Data on house types are available, as are the size and condition of the dwelling. Table 2 provides the descriptive statistics of the dataset.

This thesis uses transaction prices, the method of sale (auction/no auction), transaction dates and quality to form the repeat sales regression. By using transaction dates, the time on market of a dwelling is calculated to assess the liquidity of a dwelling. The dataset does not distinguish between voluntary or foreclosure auctions and we can therefore not separately identify their effects (by contrast to Brounen et al. (2009)). The stated quality is divided into three categories, low quality, good quality and perfect quality. The following subsections provide more insight in the statistics of each variable, and the last subsection provides a comparison of the sample dataset and the full NVM database.

4.1 Transaction prices

The dataset requires adjustments with respect to outliers. Transaction prices lower than 8000 euros have been removed. This value seems quite low, although one has to take into account that the dataset contains transactions from the 1970s, after which periods of steep price increases and decreases occurred. Also transaction prices higher than 2.500.000 euros have been removed. Before 2002, transaction prices were observed in Dutch Guilders. The dataset contains transactions already corrected for the exchange rate of Dutch Guilders to Euros. By removing the transaction prices below the threshold, the dataset is automatically corrected for other issues like condition of sale. The dataset contained dwellings, which were put up for sale, but appeared to be unsold. The transaction prices for the retrieved dwellings took a transaction value equal to -1, which were removed by the above-mentioned threshold. Also dwellings that were rented out are removed since the rent prices are below the threshold.

Table 2. Descriptive statistics: Transaction price, Quality and Liquidity

N Mean Std Dev Min Max

Transaction price

Total sample 2,250 € 130,185 € 128,742 € 8,168 € 2,025,000 Auctions 751 € 123,895 € 138,141 € 8,168 € 1,800,000 Negotiated 1,499 € 133,336 € 123,692 € 10,891 € 2,025,000

Main independent variable

Auctions 2,250 .334 .472 0 1

Liquidity and Quality

Quality (per transaction)

Total sample 2,250 6.66 1.33 1 9 Auctions 751 6.29 1.55 1 9 Negotiated sales 1,499 6.84 1.16 1 9 Quality dummies Low quality 2,250 .507 .500 0 1 Good quality 2,250 .220 .415 0 1 Perfect quality 2,250 .273 .446 0 1 Liquidity dummies Liquid 2,250 .705 .456 0 1 Illiquid 2,250 .295 .456 0 1

Time on market (days)

Total sample 2,232 144 190.86 0 1,458 Auctions 737 205 238.13 0 1,438 Negotiated sales 1,495 115 154.28 0 1,458 Other variables Type of dwelling

Total sample: House 2,250 .595 .490 0 1

Apartment 2,250 .405 .490 0 1

Auction: House 751 .640 .480 0 1

Apartment 751 .360 .480 0 1

Regular sales: House 1499 .572 .494 0 1

Apartment 1499 .428 .494 0 1

Size

Square meters 2,250 109.64 47.2 35 400

Number of rooms 2,250 4.11 1.53 1 16

Table 3. Comparison of average transaction prices between auctions and regular (negotiated) sales

Auctions Negotiated Difference T-test N Mean Std Dev N Mean Std Dev % P-value

1970-1980 - - - 133 € 53,000 34,999 1981-1990 321 € 45,481 38,038 247 € 46,792 34,158 -3% 0.671 1991-2000 84 € 107,730 88,229 477 € 114,336 91,028 -6% 0.538 2000-2010 174 € 227,193 154,066 498 € 198,741 151,827 14% 0.034* 2010-2014 172 € 173,633 167,870 144 € 192,725 104,139 -10% 0.236 House 481 € 142,837 159,981 858 € 158,576 145,57 -10% 0.067 Apartment 270 € 90,151 75,834 641 € 99,551 73,744 -9% 0.082 Full sample 751 € 123,895 138,141 1,499 € 133,336 123,692 -7% 0.101 * Difference significant at 5% significance level

The average transaction price in the total sample equals 130,185 euros. This average is taken over an extensive period of time, where large price increases and decreases occurred. The standard deviation of 128,742 euros is therefore high. When a difference in transaction method in taken into account, auctions appear to have a 7% lower transaction average than negotiated sales, as shown in Table 3. When applying a t-test to compare the averages, it can be concluded that no difference is found between average prices at auctions and negotiated sales (T-value of -1.58). As this indicates that an auction discount, based on averages does not exist, a closer look into the differences between averages in multiple time periods and house type is desirable.

Since prices are likely to increase and decrease over time, a distinction is made between multiple time periods separated for auction and negotiated sales. Table 3 shows that auction transactions appear to be lower than negotiated sales if differences in periods are taken into account, with the exception of the period 2010. Only the period 2000-2010 shows a significant difference in transaction averages. When a distinction is made between apartments and for other houses, a difference in average prices is found of 9% and 10% respectively, as can be seen in Table 3. When a t-test is applied to investigate whether the difference between auctions and regular sales is significant, p-values above the 5% significance level are found. Therefore, it can be concluded there is no evidence that a difference between auction transactions prices and regular negotiated transactions prices exists. These finding are interesting as they contradict the finding of Brounen et al. (2009). In this example, only averages are compared over an extensive period and the

findings only give an indication of the auction discount. Whether an auction discount (keeping other factors fixed) actually exists, will be investigated by the repeat sales method.

Francke (2009) describes data adjustments for repeated sales indexes for thinly traded markets. One of the adjustments is deleting transactions for dwellings that sell within a short period for large profits, since this would bias the results, as the return is abnormal and not comparable to an average increase in prices. This principal is called flipping houses. In this research ‘flipped’ houses are kept in the dataset, since a dealer who buys at an auction will try to sell the dwelling within a short period of time for a highest possible profit. By removing multiple sales within a short period existing discounts at auctions would be neglected which would bias the results.

4.2 Liquidity

As time on market is used to create a measure of liquidity of a dwelling, as described in the methodology part, it is important to first describe the time on market. The time on market of a dwelling is calculated by subtracting the sale date from the date when the dwelling was offered for sale. When a dwelling is already on the market for a long time, the tendency exists to withdraw the dwelling for a short period and put the dwelling up for sale again after this period. Especially on brokers’ websites, the actual time on market is available and could have a negative effect from a seller’s, but also from a buyer’s perspective. Sellers think a weaker negotiation position could arise when buyers know that little interest has been shown in the dwelling. The seller could also decide to change broker. In both cases, the dwelling appears again in the dataset, while the previous period is shown as withdrawn. The actual time on market is higher than the difference in days between the two dates for that specific transaction. Therefore, a correction has been made for the time on market when a dwelling was withdrawn and put up for sale again within 90 days. The time on market from the previous retrieved transaction is added to the next sale for which a transaction did occur. This adjustment is only made for dwellings that are withdrawn from the market and not for dwellings that actually did transact and put up for sale again shortly after the sale. Another adjustment is the removal of outliers. Seven observations showed a negative time on market, which is obviously not possible. These observations are coded as missing. Eleven dwellings with a time on market longer than 4 years are also seen as outliers and are also coded as missing. The time on market for the total sample equals 145 days on average.

Figure 1. Time on market by year of sale

In Figure 1, it can be seen that the time on market follows a pattern. The mid 1980’s and years around 2010 are high compared to other years. Therefore, when investigating the effect of liquidity using time on market, one needs to adjust for differences of time on market over the years. The same holds for averaging by province. As proposed in the methodology part equation (4), an index for the time on market by year and province is used. Since time on market is a proxy for liquidity, the average of the time on market per dwelling needs to be calculated by equation (5) to state whether the dwelling can be seen as illiquid or liquid. When the time on market is calculated and the index per dwelling is created, 528 dwelling are coded as liquid dwellings and 231 are illiquid dwellings as shown in Figure 2.

Figure 2: Liquidity and Illiquidity of the dwellings

4.3 Quality

The dataset contains information about the quality on the interior and exterior of the dwelling. An overall quality variable is formed, by combining the two variables. Due to broker subjectivity and the constant quality assumption, the average quality of a dwelling is calculated. By averaging the assigned qualities for each transaction the quality per dwelling is approximated.

The quality of a dwelling is either described as: low quality, good quality or perfect quality and all three categories are formed into dummy variables. Whereas 9 subcategories for quality were available, the reason to reduce the subcategories to 3 has to do with the subjective interpretation of quality. The subcategory good quality is set as a threshold. All observations below good quality are described as low quality and observations above good quality are assumed to be perfect quality dwellings. From the 2,250 transactions, 707 are low quality, 1,183 are good quality and 360 are perfect quality. When the average quality per dwelling is calculated, 389 dwellings turn out to be low quality, 179 are good quality and 191 dwellings are perfect quality. It should be noted that the interpretation of quality is subjective to measurement error. For a broker, the distinction between each subcategory is not clear, since there are no strict assumptions about measuring quality.

Table 4. Average transaction price by quality

N Mean Std Dev Low quality 389 € 108,617 120,101 Good quality 179 € 149,817 100,859 Perfect quality 191 € 154,370 155,017 Illiquid 30% (231) Liquid 70% (528)

Interesting to see is whether the average price of low quality dwellings deviates from good- or perfect quality dwellings. Table 4 shows dwellings that are of low quality have a lower average price than good and perfect quality dwellings, as could be expected.

Table 5. Index value of Time on market by quality

N (Time on market) Index value Std Dev T-test (p-value)

Low quality 389 .933 .541

Good quality 179 1.089 .723 0.0044

Perfect quality 191 1.065 .698 0.0129

*T-test compares low quality to good and perfect quality

Apart from the differences in average prices, a closer look into time on market between the three quality categories is interesting. The question arises whether the time on market for lower quality dwellings appears to be higher than for higher quality dwellings. One should note that the general statement that unattractive dwellings suppose to have a longer time on market is not only captured by the condition of the dwelling. Location and other factors are of influence as well. Table 5 shows average index values for each quality category. The good quality category shows the highest time on market index value and perfect quality dwellings show a higher index value than low quality dwellings. This finding implies that on average, good and perfect quality dwellings have a longer average time on market than low quality dwellings. With a t-test, the time on market index values of good and perfect quality are compared to the low quality index, and the difference appears to be statistically significant at the 5% significance level. It is surprising that the time on market for low quality dwellings is shorter than for higher quality dwellings as one would expect the contrary. These findings could be due to the fact that lower quality dwellings are sold within a shorter time span, due to the fact that low quality dwellings are more attractive in price. When a low quality dwelling can be renovated, these dwellings could be attractive for dealers who search for opportunities to make a profit. In auction terms: people go on bargain hunt.

4.4 Representativeness of the sample

It is important to examine whether the sample is representative for the total market, before one can interpret results. For some variables, the averages are tested with a t-test to investigate if the used sample is significant and comparable to the total NVM database (which in itself captures the majority of market transactions). To avoid differences in

averages due to a different sample period, only data from the NVM database are taken into account for the same period as the sample period.

Table 6. Comparison Sample to NVM database

Type of dwelling Sample NVM database

Apartment _40.5% 26.1%

House _59.5% 74.0%

Table 7. Comparison: Square metres, Transaction prices, Rooms, Time on market

Mean N Std Dev Median P-value T-test

Square Metres NVM database 121.3 2,584,329 44.6 117 Sample 109.6 1,701 47.2 100 0.0001* Transaction prices NVM database 174,271 3,140,840 € 146,671 145,000 Sample 130,185 2,250 € 128,42 104,000 0.0001* Number of Rooms NVM database 4.4 3,140,840 1.1 4 Sample 4.11 2,239 1.53 4 0.0001*

Time on Market (days)

NVM database 120 3,140,840 161.0 62

Sample 115 1,494 154.3 63 0.2301

In Table 6 it can be seen that apartments are overrepresented compared to the NVM database. Single-family houses are therefore underrepresented. Table 7 provides an overview of variables that are compared to the NVM database. The average size of a dwelling in the NVM database is somewhat larger than the used sample. This difference could exist due to the fact that the sample contains relatively more apartments than houses compared to the NVM database. As apartments are on average smaller, the sample could therefore show a lower size average. When it is tested whether the NVM database and the sample averages significantly differ, the t-test statistic shows statistical significance at a 1% significance level.

When the number of rooms in sample and the NVM database are compared, an average of 4.11 and 4.4 number of rooms are calculated respectively. A t-test shows that this difference is significant. Again, this difference could be attributed to the relatively large amount of apartments within the sample. The same rationale is applicable for comparing transaction prices. Where the NVM database shows an average of 174,271 euros per transaction, the average transaction price in the sample equals 130,185 euros. With respect to the average time on market, no significant difference is found between the sample and the NVM database.

To summarize, there are some differences between the sample and the NVM data. This implies that the results are not necessarily representative for the whole population of transactions, but that the results are mostly relevant for houses that are (eventually) auctioned. What determines whether a house is auctioned (selection effect) is left for future research. To the extent that differences exist regarding quality and liquidity, this is explicitly controlled for in the regression analysis.

Chapter 5 Results

The first step in analysing whether discounts at auctions in the Netherlands exist is to use the basic repeat sales model including the auction term as described in equation (1)9_{. As}

shown in Table 8 the parameter of the first regression for the total sample period equals -.163. With the parameter significant at the 1% significance level, it can be concluded that in this sample auction sales on average result in 16.3% lower prices compared to regular negotiated sales (the reference category), which confirms the first hypothesis 10_.

The adjusted R-squared takes a value of 76%, which is satisfactory (76% of the variation in the dependent variable, the price difference, is explained by the independent variable, the auction dummy and the time fixed effects). When the sample is divided into two periods, to investigate whether a more recent periods lead to different results, the results remain the same11_.

9 _{Please note that the total number of observations used in this regression is lower than mentioned in the}

descriptive statistics section due to the use of the repeat sales model (i.e. taken differences over time).

10_{To control the model for possible misspecification, a Least Square Dummy Variable Approach is used.}

The model shows a significant 18% discount when a dwelling is sold at an auction, which is comparable to the finding with the repeat sales approach. It can therefore be concluded that the applied repeat sales model is of the correct functional form.

11_{When the sample is divided in the periods 1974-2000 and 2000-2014, the results in comparison with the}

The existence of an auction discount for the Dutch market is in accordance with the findings of Brounen et al. (2009). Although the studied period of Brounen et al. (2009) is included in the used sample, the number of observations for that period is too low to find significant results. The Nationale Hypotheek Garantie (NHG) concluded that on average an auction discount exists of 15,000 euros12_{. When the average of the total}

sample is taken into account, this research finds an average discount of 20,000 euros discount at auction. Apparently, the NHG addressed the relevant topic, whether the auction method should still be used in case of foreclosure. The finding is also in line with other research like the studies of Mayer (1989) and Allen and Swisher (2000) who also find evidence for discounts at auctions. It should be taken into account that markets differ geographically and that the nature of domestic market can have a large influence on results.

Apart from finding proof whether discounts at auctions exist, it is investigated if illiquidity could drive auctions results to be lower compared to liquid dwellings. As described in the methodology chapter, liquidity of a dwelling is measured by the time on market, indexed for year of sale and location. The second regression in Table 8 includes the interaction term for the illiquidity and the auction term as shown in equation (2). The parameter of the interaction term for illiquidity and auctions shows a negative sign. As expected dwellings that are illiquid, sell for additional discounts. The result shows that a dwelling that is illiquid, sells at an auction on average for almost 8% less, compared to a more liquid dwelling. The parameter for the interaction term is significant at the 5% significance level. The auction coefficient decreases from 16,3% to 14% and remains significant at the 1% significance level. This implies that illiquidity only partly accounts for auction discounts to exist and that there are other remaining factors that influence auction outcomes. The fact that an illiquid dwelling sells at an additional discount confirms the second hypothesis.

12_{RTL nieuws, Huizen met hypotheek garantie niet meer geveild. Available at:}

Table 8. Regression output

(1) (2) (3) (4)

(Equation 1) (Equation 2) (Equation 5) (Equation 6)

Parameter T-stat Parameter T-stat Parameter T-stat Parameter T-stat

Auction -.163*** -8.56 -.140*** -6.89 -.128*** -4.57 -.103*** -3.49 (.019) (.020) (.028) (.029) Illiquidity*Auction -.079* -2.06 -.077** -2.1 (.038) (.038) Low-quality*Auction -.057 -1.57 -.063* -1.72 (.036) (.036) Good-quality*Auction .030 -0.73 -.030 -0.74 (.041) (.041)

Time fixed effects Yes Yes Yes Yes

Adjusted R-squared 76,1% 76,3% 76,1% 76,3%

Number of observations 1,301 1,301 1,301 1,301

The existence of additional discounts for illiquid dwellings could also be related to other factors. A reason can be found in the proxy for illiquidity, the average time on market of a dwelling. Since an illiquid dwelling will need a longer time on market to find an optimal buyer, the probability that this optimal buyer, who is willing to pay the maximum price, will bid at auction is less likely, since the time span for which auctioned dwellings are on the market is typically low. The time on market of a dwelling can be influenced by factors, like asking prices and overall attractiveness of a dwelling. It could well be that those factors are not fully captured by the illiquidity term and therefore are omitted factors that could affect the results (omitted variable bias). Heterogeneity could also affect liquidity, since homogenous dwellings are more easily sold compared to heterogeneous dwellings and thereby affecting the auction results (estimates). Finally, there could be reverse causality. The question is whether prices influence the time on market, or if the time on market influences prices. As mentioned, an adjusted (index) illiquidity measure is used such that this issue is less of a problem.

If quality of a dwelling also contributes to auction discounts, it can be investigated by including interaction terms for quality to the regression as proposed in equation (6). The third regression in Table 8 provides the regression output when the term for low quality and good quality are interacted with the auction dummy and included in the regression. As expected, the parameter for low quality shows a negative sign and, but appears to be statistically insignificant. Also the parameter for the good quality interaction term is not significant. Apparently, the quality of a dwelling does not prove to have explanatory power in the existence of auction discount, at least when no other factors (e.g. liquidity) are taken into account. The parameter for the auction term remains significant but decreases in magnitude. The adjusted R-squared remains on a level of 76%.

It should be noted that the interpretation of quality could differ among appraisers or buyers. Quality remains a subjective measurement, which could also be seen in the dataset where the attribution of good quality to a dwelling is assigned in most cases. To reduce this subjectivity, the average quality of a dwelling was calculated and used in the regression13_.

13_{Although the quality of a dwelling could actually change over time, we did not take this into account in}

In the fourth regression, both interaction terms for illiquidity and quality are included in the regression as shown in equation (7), see Table 8, column 4. The parameter for illiquidity also shows in this regression a negative sign and a magnitude of 8%. The conclusion that illiquid dwellings at auction sell for additional discounts compared to liquid dwellings, still holds. Again, this finding is significant at the 5% significance level. Interesting is the parameter for the interaction term between the auction dummy and low quality. Although only weakly significant at a 10% significance level, an indication that lower quality dwellings sell at additional discounts compared to higher quality dwellings exists, but the magnitude of the parameter should be interpreted with caution. The auction parameter remains significant but decreases to 10.3% and captures the effect of dwellings that are liquid and of perfect quality. Apparently, the auction discount is not fully attributable to quality and liquidity, but other factors can influence auction discounts as well or just the fact that the dwelling appears on an auction results in a lower price14_.

A reason for the extra discount for lower quality dwellings could be the fact that private parties do not sign up for low quality dwellings, as renovation costs could be difficult to estimate. Dealers will have more market knowledge and could estimate the renovation costs with more certainty, although the risk of additional costs remains due to the information inefficient nature of the auction market (you have to buy quickly). In the regular market, you have more time to find out whether there are deficiencies in the state (maintenance) of the dwelling. Therefore, low quality dwellings are expected (and we find weak evidence for this) to sell at higher discounts. The fact that, according to the AFM, dealers heavily influence the Dutch auction market, could explain the additional low quality discount effect, since presumably low quality dwellings are mostly bought by dealers and therefore could be bound to unfair competition15_.

An important issue that is not mentioned by previous literature is a trade-off effect that exists with respect to foreclosure. In the case of an auction, the dwelling will be sold within a short time period, whereas in the case of a regular negotiated sale, it could take a

14 _{Separating the dataset for apartments and other houses and using equation (7) did not improve the}

results.

15_{Autoriteit Consument & Markt, ACM publiceert boetebesluiten in de kartelzaak huizenhandelaren bij}

https://www.acm.nl/nl/publicaties/publicatie/11963/ACM-publiceert-long time before the dwelling changes owner. For the mortgagee, it could be more beneficial to sell at a discount than taking a loss on interest payments. For the mortgagor, the discount will still be a problem, since in most cases the homeowner will be left with a loss. As mentioned in the introduction, especially nowadays when housing prices are under pressure, a discount at auctions could worsen the financial position for people who are already in financial distress.

To summarize, the findings in this thesis show that on average an auction discount exists in the Dutch market, and that the liquidity of a dwelling influences the magnitude of the total auction discount. Also, weak evidence is found for extra discounts for lower quality dwellings.

Chapter 6 Conclusion and discussion

This thesis has investigated transaction prices at auctions in comparison to regular (negotiated) sales prices in the Dutch market, in particular whether auction discounts exist. The liquidity and quality of a dwelling are also examined, as determinants of the auction discounts. Conventional wisdom suggests that discounts at auctions exist simply because a house is auctioned, which suggests that arbitrage opportunities might be present. A dataset provided by the Dutch Association of Realtors (NVM) containing data on 760 dwellings and 2,250 transactions is examined by using a repeat sales approach. The dataset contains the quality of a dwelling. The liquidity of a dwelling is measured by an indexed time on market based measure.

The findings show that auction discounts in the Dutch market exist. With an average discount of 16.3% at auction, it can be concluded that dwellings offered at auction will transact at a lower price compared to regular negotiated sales. The results also show evidence for liquidity to be a factor that has an influence on the magnitude of the auction discount. Illiquid dwellings transact at an 8% extra discount compared to more liquid dwellings, which can be related to the short amount of time available at auctions, to find an optimal buyer, where illiquid dwellings will need more time to be matched to a buyer compared to more liquid dwellings. When only quality is taken into account, no evidence for additional discounts is found, although when also illiquidity is added to the regression, low quality dwellings appear to transact at an additional discount of 6%, although this effect is only statistically significant at the 10% significance level. After