Can online search queries improve the forecasting efficiency of the short-term house pricing dynamics? : an empirical analysis of the owner-occupied house market in Amsterdam

(1)

“Can Online Search Queries improve the forecasting efficiency

of the Short-Term House Pricing Dynamics? An Empirical

Analysis of the Owner-Occupied House Market in Amsterdam”

Todor Ristov

University of Amsterdam

Supervisor: Prof. Dr. Marc K. Francke

Master Thesis Business Economics

(2)

STATEMENT OF ORIGINALITY

This document is written by Student Todor Ristov who declares to take full responsibility for the contents of this document.

I declare that the text and the work presented in this document is original and that no sources other than those mentioned in the text and its references have been used in creating it.

The Faculty of Economics and Business is responsible solely for the supervision of completion of the work, not for the contents.

(3)

ABSTRACT

The following thesis investigates the possibility of Search Query Index (SQI) obtained from online search engines such as ‘Google Trends’ in improving the short-term forecasting price dynamics in the owner-occupied housing market in Amsterdam. The results from the empirical analysis provide evidence that SQI ‘Huis Te Koop’ (House To Buy) influences price developments one and four months prior the actual transaction on the market, which is estimated thorough an ARIMA model. Furthermore, the analysis indicates that SQI is also able to forecast the price and liquidity in the housing market four months before a transaction occurs. This is proven by simultaneously modeling the house prices and the volume traded in a VAR model. In fact, the approach implemented in the thesis provides an innovative insight in explaining the price dynamic in the owner occupied market in Amsterdam as it employs an alternative online data source, Google Trends, directly related to the activity of the participants in the housing market.

(4)

ACKNOWLEDGEMENTS

I am would like to express my sincere appreciation to my supervisor Prof. Dr. Marc Francke. I am thankful to his guidance, comments and encouragement during this process without which the completion of this thesis would not have been possible. I would also like to thank my friends for providing all those creative ideas and for inspiring me to conduct research in the current topic. I am also truly grateful to all professors lecturing MSc Real Estate Finance at the University of Amsterdam for their knowledge, for increasing my understanding in Real Estate and for their support during the completion of my studies.

Also special thanks to the Dutch Association of Realtors for providing me with an access to the vast amount of data without which this research area would have clearly be impossible to write.

Finally I would like to express gratirude to my parernt for the support and their true believe in my potential. At the end I would like to thank my amazing sister for her great mind and help throughout the process.

(7)

1. INTRODUCTION

The recent financial crisis has forced policy makers and governmental institutions to take a proactive approach in tracking the price dynamics of financial assets. In fact, the asset price developments have direct impact over the economy and the welfare of the society. Hence, potential creation of asset price bubbles might lead to financial imbalances and wealth destruction. Real Estate (RE) is the largest asset class that provides both investment incentives and consumption benefits. As such it widely affects wealth creation at institutional and social level. The literature to date provides strong evidence of positive correlation in the Real Estate markets and due to its illiquid nature it increases the importance of closely tracing RE price formulation. Measuring RE price pressure and identifying trends can be used as signal to indicate future trends and the price direction of real estate assets. By closely analyzing these movements stakeholders might better create their optimal market positions and better evaluate the risks and returns associated with RE investments.

The empirical paper provides an alternative data approach to forecasting developments in the owner-occupied housing market. By employing a data that takes into consideration the initiatives of the participants in the housing market through the Search Queries Indices created on the online Search Engines might increase the understanding of the house price dynamics. The importance of this research topic arises due to the particularity of Real Estate as an asset class, such as the heterogeneous nature of each individual asset, the lumpy investment costs, and especially the highly illiquid nature. In contrast to existing empirical evidence, price developments in this thesis are traced irrespective of explicit changes in the fundamental characteristics on the market. Advances in IT and the increased application of Online Search Engines (OSE) provide access to a high scale of disaggregated data on which trillions of decisions are made (Wu and Brynjolfsson, 2015). The possibility to browse through the query history on these OSE at no cost further increases the attractiveness of the

(8)

about users’ intentions to proceed with an economic transaction. In fact the thesis research topic tests the hypothesis:

“The activity in the search markets available from Google Trends can improve the forecasting ability of the owner-occupied house price developments”

The available literature on Internet Search Behavior in relation to the effect on house prices is limited. From the literature review can be seen that this topic is growing in popularity. However, there are many aspects that can increase the understanding of the causal relationship between the Online Search Behavior, the Real Estate price dynamic and trade volume. This thesis research is motivated by behavioral finance theories signifying the fact that fundamental factors fail to capture real estate price changes in the short run. The empirical evidence to date does not incorporate an aggregate unbiased measure on the attitudinal and behavioral actions of participants in the search markets. This empirical paper analyzes the owner occupied residential market in the urban area of Amsterdam. In order to challenge the research and to provide increased understanding in the area, this empirical paper tests the following questions:

‘Can Internet Search Behavior help in identifying short term price dynamics within the Amsterdam urban area?’

‘What is the magnitude of Internet search markets on the prices in the owner occupied housing market in Amsterdam?’

In the following thesis the activity of online search engines is received in the form of Query Index obtained from Google Trends. The Query Index represents the search interest over time expressed relative to the highest interest point in time. In providing a robust estimate of the impact of the Search Query Indices, the ‘Huis Te Koop’ Search Query Index is taken as a proxy of the activity in the search housing market. In examining the relationship between the House Price and Online Search activity the data is modeled in two econometric models,

(9)

Autoregressive Integrated Moving Average (ARIMA) model and Vector Autoregressive Model (VAR) due to the autocorrelation present in the house prices as indicated in the literature review. In order to estimate the pressure on the market and control for the different market conditions a Market Tightness variable is introduced estimated as the ratio of the number of properties sold during the period against the number of properties available for sale as provided by the Dutch Associations of Realtors (NVM database). Both ARIMA and VAR incorporate the SQI variables where their significance is further tested within the previously established framework. The sample period of the empirical evidence is limited to the data availability of Google Trends starting from 2004 and is applied on monthly frequency for the period of 2004-2014. Considering the data availability constraint the research is trying to explain the short-term house dynamics in the owner-occupier housing market in Amsterdam.

The empirical research is structured in the following manner. The second chapter outlines the relevant literature overview, relating to different market pricing theories, behavior and attitude theories in relation to the pricing dynamics and liquidity. In the third chapter an overview of the hosing market in Amsterdam is given and current market trends are identified. The fourth chapter describes the different data sources and provides descriptive statistics on the variables used. The succeeding chapter five underlines the applied methodology and elaborates in detail the econometric modeling of the research question. Chapter six discusses the main results and compares them with the main hypotheses of the research question. In the following chapter seven the main implications and limitations of the research topic are addressed. The thesis is concluded at chapter eight along critical discussions on the key findings compared to the literature review outlined in chapter two.

(10)

2. LITERATURE REVIEW

2.1. Rational Expectations Theory (RET) and Efficient Market Hypothesis (EMH)

Many concepts in Finance are based on the theory that markets in general do not misuse information when pricing assets. In fact expectations are formed as a consequence on the actions occurring on the market. Muth (1961) indicated that the nature of the information and the framework in which they are used in forecasting future market condition is an important remark in understanding the dynamics of the upcoming events on the market. According to Muth (1961): “expectations are informed predictions of future events” (p.316). Thus, speculations on the market triggered outside the range of available information, unless it is an inside information, impacts the market in a rational way. As a result, firms on the market are guiding their strategic decisions based on the “rational” expectations on the market.

Assuming rationality in terms of an economic decision, price expectations are corresponding to the equilibrium point of the market price and the available information. As Muth (1961) illustrates that if only supply shocks occur in the price then the quantity equilibrium will be reached along the demand curve. Since the housing market is considered to be supply inelastic in the short run, it is reasonable to assume that shocks in demand triggered by either changes in the fundamentals on the market, expectations and/or demographics forces, the equilibrium is going to be found through movements along the long run supply curve.

Additionally, Fama (1970) indicates the imperative importance of information in providing an accurate indication of capital allocation on the market. In accordance, a market is considered to be ‘efficient’ only when the market price of securities ‘fully reflects’ all available information concerning the security. Fama (1970) tests if the information is incorporated in asset prices in three information subsets such as: Strong-form tests examining whether monopolistic information available to insiders is relevant in formatting asset prices; Semi-strong-form tests examining whether information that is publicly available is incorporated in

(11)

asset prices; and Weak-from test examines the relevance of historical price data in price formation.

In defining ‘efficient’ price adjustment on the market, Fama (1970) recognizes that transaction costs occurring as a result of trading as well as the acquisition of relevant transaction information should be costless. He also supports the idea of frictionless markets with a uniform opinion on the impact of the information on the price of the securities. Under such an environment, which is in line with the ‘rational expectations theory’, all the information available on the market will certainly be reflected in the price of the securities.

2.2. Market View of RET and EMH

The academia considers rational expectations, efficient markets and rational behavior as the main building block in finance, whose assumptions support the fact that asset prices are the reflection of market fundamentals. Any deviation from such factors is considered as an ultimate instance of irrationality. Many empirical studies have concluded that fundamental factors partially determine asset pricing. Blanchard and Watson (1982) indicate ‘crowd psychology’ as an additional potential price determinant. In turn, they indicate that investors voluntarily execute their strategies, neglecting the possibility of investors holding private information that will help them increase their utility through portfolio relocation. However, Blanchard and Watson (1982), under the assumption that investors are risk averse, imply that volatile markets and possible bubbles on the market will increase the expected returns due to the anticipated increase in risk on the market. Hence, they suggest that with an increased risk expectation on the market, the expected prices (returns) should grow at an even faster rate.

Blanchard and Watson (1982) propose a more realistic market assumption incorporating information asymmetry between investors. As a result, investors assess differently the

(12)

‘imperfect’ due to heterogeneous asset nature and a significant information asymmetry between buyers and sellers. As a result, deviations from the aforementioned factors in favor of the theories provide support of the boom-burst nature in the housing market supporting the strong positive correlating in house prices.

2.3. Market Fundamental Factors

The efficient market hypothesis states that the current fundamental value represents a capitalized value of all relevant market information that provide insights about future price developments. This statement is in line with the theory of rational expectations and supports the fact that markets follow semi-strong efficiency. In testing these hypotheses, Clayton (1996) developed a forward-looking model in the Vancouver residential market over the period of 1979-1991. His empirical model tested the predictive power of fundamentals in forecasting the short-term fluctuations in the housing market. His approach incorporates the observed market fundamental value which is proxied through a function of owner-occupied rents in the housing market. In combination with the arbitrage theory relationship Clayton (1996) originates a model of the present value of single-detached houses. Clayton (1996) presented empirical analysis indicating that the market dynamics can be forecasted well in periods of less volatile markets, while the model fails to provide accurate forecasts in periods of real estate booms and bursts. This establishment advocates that prices in the housing market diverge from fundamentals at the peak and the through price cycle. As novelty, Clayton (1996) illustrates the importance of the psychology and the local market expectations in conducting an improved forecast in following house price cycles.

In line with this methodology, Meese and Wallae (1994) assess market efficiency in the residential market in Alameda and San Francisco. Their model includes several fundamental variables such as the prices in the housing markets, the rents and the implied cost of capital. In their model, they estimate the short and long term relationship of the assumed fundamental value and the actual transaction house prices. The empirical findings of Meese and Wallae

(13)

(1994) indicate negative present value relation in the short run which is inconsistent with the efficient market theory. As a main influence of the result, they indicate that the particularities of the housing market do not confirm with the efficient market theory, such as the information asymmetry, the lumpy transaction costs and the heterogeneous nature of the asset class. In particular, the optimal decision rule is largely influenced by the large transaction costs due to fact that expected gains in the housing market have to exceed transaction costs in order to achieve increased utility.

However, the empirical evidence of Meese and Wallae (1994) is significant and consistent with the literature in the long run, accounting for both taxation and the cost of borrowing. The inconsistency in the results in the short run may be impacted by the boom-burst nature of housing markets that makes prices diverge from their fundamental value. However, Messe and Wallae (1994) stress out that their evidence does not necessarily provide evidence of irrational price expectations, considering the fact that the tested joint hypothesis also incorporates the hypothesis of ‘no’ risk premium compensation in the housing market. In line with the results, they also raise an issue about possible model misspecification, non-linearity and the possible behavioral influence in the house price market.

2.4. Expectations – Behavior and Attitude

Developments in the house prices is the result of complex process which involves interaction between behavioral and attitudinal actions of the participants on the market. Therefore, expectations of different parties play an imperative role in house price formulation, especially expectations from parties that directly impact the housing market, i.e. the expectations of homebuyers.

(14)

having in consideration the fact that in this environment the homebuyer will also be compensated by an additional house prices increase in the future, this investment becomes acceptable. In fact Case and Shiller (2003) notify that the appreciating house prices are influencing the timing of first time homebuyers due to the fact that a continuous house price increase will limit the affordability of purchasing a house in the future. This concept of house price increase underestimates the risk associated with home investment. Thus, Case and Shiller (2003) state that if expectations are buyers’ motivating factor then house prices are inherently volatile. As a result Case and Shiller (2003) find that income as fundamental factor alone has explanatory power in less volatile markets, while in markets with large swings and increased volatility house prices experience inertia and the income patterns are not significant in explaining changes in house prices. Adding additional fundamental variables such as mortgage rate and unemployment rate adds little explanatory power in a state with high market volatility.

Case, Shiller and Thompson (2012) in a recent survey study conducted on 5,000 homebuyers in Los Angeles, San Francisco, Boston and Milwaukee analyzed the behavior and the market outlook of market participants after they have purchased a house. They find that in all of the different geographical locations participants were well informed about house price trends and had high positive correlation coefficients with the upcoming price trend in the short run. The short-term expectations were more volatile and sometimes negative in comparison with the long-run expectations i.e. in the next 10 years. Expectations in the long run exceed the corresponding expectations in the short run illustrating that house buyers’ expectations are more optimistic in the long run.

Case, Shiller ad Thompson (2012) find statistically significant relationship between the house price changes on the market, and the expectations on housing prices given by the participants in their survey. These coefficients have the correct sign and are larger than 1. Under the traditional rational expectations theory, these coefficients can be interpreted as containing

(15)

relevant forecast information about future house price developments. By including the expectations of the respondents from the previous period both from the same metro area and the expectations from the USA overall, these variables are not significant, which is in line with the rational expectation theory, considering the fact that this information is already reflected in the forecast in the current period. In fact, the expectations in the short run enclose significant information about future house price changes.

As an important feature of the survey is the possible positive homebuyers bias considering the fact that 70% to 80% of the participants responded with a positive answer to the statement that “It is a good time to buy a home because prices are likely to rise in the future” (Case, Shiller and Thompson, 2012; p.274). Such statistics is in line with the fact that homebuyers are excessively optimistic which decreases the robustness of the analysis. This proposition suggests that house prices and house prices bubbles are frequently motivated by irrational expectations. Modeling these irrational expectations might provide an insight on the short-term house price dynamics.

2.5. Price-Volume Correlation and Market Tightness

Conventional finance bases many theories on the ‘Efficient Market Hypothesis’ (EMH) concept developed by Fama (1970; 1965). The EMH states that asset prices follow random walk where the information available at any time is reflected in the prices. Such an ideological market environment has been extensively tested and it has been rejected with an empirical evidence of serial correlation in asset prices and non-normal price distribution (Kendall and Hill, 1953; Lux and Sornette, 2002). Recent evidence is found by De Wit et al. (2013) in the Dutch residential market. They illustrate that the price–volume correlation can be explained by the interaction of the down payment, the house price, the mobility and behavioral rationalizations.

(16)

Real Estate markets are not centralized, characterizing them as a search markets between buyers and sellers (Diaz and Jerez, 2013). Berkovec and Goodman (1996) develop search and matching model which indicates a significant predictive power in forecasting appreciations in housing prices. By studying the proportion of buyers to seller, i.e. the effect of market tightness, can illustrate both the bargaining power of sellers on the market and the probability of sale occurrence. Such a model can be used as a proxy of market tightness, and it replicates the mechanism of information asymmetry between buyers and sellers. These interactions can contain significant information, which can help in explaining expected changes in house prices over time. In fact, ‘tight’ or ‘hot’ search-and-match markets are more liquid due to the higher turnover and volume rates. Thus the increased liquidity in these markets results in higher prices. This phenomenon exists on the market due to the larger number of potential buyers in respect to the number of potential sellers. Berkovec and Goodman (1996) find a strong positive correlation between changes in price and sales volume in low-frequency markets; and they also find the sale volume to be more responsive to changes in prices than the responsiveness of house demand on price developments. In addition to this model, Carrillo et al (2015) improve the relationship of house price models by including the market tightness in a standard forecasting technique. In fact, they indicate that market tightness increases forecasting performance by 30% and 10% 1 quarter and 1 year in the future respectively.

2.6. Alternative Data from Online Search Engines

Wu and Brynjolfsson (2015) conduct analysis based on the Internet Search Behavior and their forecasting significance in explaining the dynamics of the housing markets, which is in line with the research analysis conducted in the thesis. Wu and Brynjolfsson utilize data obtained from Google Trends, which is significant in forecasting house price changes and the house sales volume in the United States. They state that Online Search Queries (OSQ) provide economic value and represent an improved forecasting tool of price trends in real estate in

(17)

comparison to the predictions made by the fundamental data available from the National Association of Relators. In fact, their seasonal autoregressive model captures 98% of the variation between Online Search Query indices and market indicators.

According to the Dutch Association of Realtors (NVM, 2016), online housing platforms, such as ‘Funda’, have achieved a record number of 31 million monthly visitors during the fourth quarter of 2015. This represents a 34% increase from the same quarter previous year. Dorinth van Dijk (2014) uses Internet data obtained from ‘Funda’ in predicting price developments in the Dutch housing markets. Van Dijk (2014) utilize the data on the number of clicks per properties listed on this brokerage website as supply/demand proxy of the housing market in the Netherland. Van Dijk (2014) employs a panel VAR model depicting transaction volume as responsive to changes in supply and demand, while he finds a gradual adjustment in prices. Van Dijk finds ‘times watched per object’ variable to explain regional house price differences; and significantly indicate a causal relationship of a 1% increase in ‘the variable’ initiating 0.25% increase in house prices.

Furthermore Van Dijk and Francke (2016) provide supportive evidence in explaining the liquidity dynamic and price developments in the Dutch housing market. They also employ Internet search data, where they also apply a panel VAR model. They find demand shocks to have a direct impact on liquidity resulting in permanent price change in urban areas. In fact, the analysis conducted by Van Dijk and Francke (2016) indicates a statistically significant relationship between the previous value of the ‘times watched per house’ variable and the house prices and sales volume.

(18)

3. MARKET CHARACTERISTICS

The population in the Netherlands is following the global trend of increasingly populating urban areas. As of 2015, 85% of the Dutch population is occupying dwellings in urban areas resulting in a rapid change in the Dutch residential market (Centraal Bureau Statistiek, 2016). The population age group between 20-34 increasingly moves to the biggest cities in the Randstad (Amsterdam, Utrecht, Rotterdam and The Hague) due to the better economic prospects and many amenities they provide such as education, culture and better quality of living. On the other side, the Dutch housing market has the particularity to be dominated mainly by two groups, the large owner-occupiers and the social housing sector. The government has undertaken series of actions and schemes in stimulating owner-occupier dwellings, such as tax deductibility of the mortgage interest. On the other side, as the result of the recent financial crisis, the regulators are tightening the mortgage terms by imposing stricter Loan-to-Value lending in order to lower the risk of default. Such actions directly impact housing affordability in the low to middle income households. As a result of these policy imperfections, subsidies are provided to low income households by offering hosing in the social (regulated) rental sector with a monthly rental limit to €711. These different actions have divided the Dutch housing market into an unsustainable market structure invoking changes to reposition the commercial and unregulated housing segment. The growing demand of these sectors and the limited housing supply translates in growing shortage, imposing quite some volatility in house prices (CBRE, 2014).

3.1. Amsterdam Housing Market

Amsterdam, being the capital of the Netherlands is an attractive destination for international corporations attracting a lot of young and highly skilled work force. The economic activity of Amsterdam is of a high importance to the Dutch economy due to the fact that the region contributes circa 17% of the Dutch GDP. The Amsterdam Metropolitan Area is among the top five leading European metropolitan regions inhabiting nearly 2.4 million people (MRA,

(19)

2016). The overall demographics continue to be strong resulting in tighter housing demand against the supply on the market in Amsterdam and experience an economic growth above the 2013 national average (Centraal Bureau Statistiek, 2016). Having into consideration the limited geographical expansion in Amsterdam the housing market is experiencing growing housing prices due to the continuous housing shortage.

3.2. House Prices Dynamics

The Dutch hosing market has experiences a 21% decrease in prices from peak to trough i.e. form August 2008 – June 2016 (Gemeente Amsterdam, 2016). In perspective, the current national house price average, as of the end of 2015, is 16% bellow the peak level of 2008, while the rising price level of the housing in Amsterdam almost fully recovered, outpacing the national average (Gemeente Amsterdam, 2016). Going out of the financial crisis indicated that Amsterdam has a ‘sponge capacity’ absorbing an ever-growing number of inhabitants. In fact, this feature signifies the house price dynamic and the importance to closely monitor any house price developments.

3.3. Demographics and Households – Pressure on the inner ring city

Currently there are more than 800.000 people living in Amsterdam (i.e. within the A10 ring road). Forecasts conducted by Central Statistics Bureau, (Centraal Bureau Statistiek, 2016) indicate that the population in Amsterdam is expected to increase by more than 12% by 2020 resulting in more than 900.000 inhabitants. The growth is mainly facilitated by the strong expected increase in population from the 20-34 and 65+ age group. As the population continues to grow so will the scarcity of single and multi family dwellings, as these are the most attractive settlement-housing sector of these age groups. In fact the number of dwellings in single- and multi family dwellings is only expected to increase for nearly 10% (Gemeente

(20)

3.4. Market Structure - scarcity and Rent Regulated segment

The housing stock in Amsterdam is consisted of circa 400,000 dwellings from which 46% is owned by the Social Hosing Corporations, 32% is owner-occupied and 22% by private investors (CBRE, 2014). Another feature of the housing market in Amsterdam is that mostly the rent is government regulated with a small share of 10% in the non-regulated segment. The regulated rent segment is based on the so called ‘Points System’ or WWS based on the certain physical characteristics of the dwelling and its location. The maximum rent that can be charged is determined by the government and it is set at €711 per month as of 2016. As a consequence the natural occupancy flow is obstructed as many high income families occupy these homes and have no incentives to move within the regulated market segment. This particularity of the rental market in Amsterdam makes it impossible to apply the rent level as a fundamental factor influencing the house prices as it is extensively applied in other housing markets.

3.5. Importance for Econometric Modeling

As it was previously mentioned, Amsterdam attracts many highly skilled workforce and expats occupying the mid-liberalized housing segment which is relatively limited. In recent years can be seen both an increased housing development and increased demand for single person households. As it was explained previously, a large share of the housing stock is social hosing whose rent is fully governmentally controlled. Having into consideration the relatively small development in the affordable Owner-Occupied Hosing stock, raises the importance for understanding and monitoring the house prices trends in Amsterdam. Therefore, developing an accurate forecast model can help in better welfare allocation and more prudent estimations of the directions in the house prices.

(21)

4. DATA

The following chapter describes the data and the data collection process, providing statistical and spatial description of the variables used in the research. This chapter also describes the main variable transformations implemented in the research.

4.1. Data Sources

The empirical research employs two main datasets. The first database is obtained from the Dutch Brokerage Association (NVM) providing information of house prices and transaction volume used as a representative of the owner-occupied housing market in Amsterdam. The NVM data set also provides information on individual property characteristics listed in the database, as well as characteristics of the neighborhood where the property is located. All these information provide an insight on the factors influencing the market value of the property. This information is of a prime importance in modeling house prices due to the heterogeneous nature of this asset class. In addition, information is available for the original asking price and the transaction price providing indication of the general market condition, allowing a general house price index to be estimated. The data is available for the period from January 2004 until December 2015. During this period, the NVM data set includes 122,113 transactions from which 90,059 properties were sold, 31,691 properties were withdrawn from the market and 363 properties were rented out. In order to ensure data robustness, the analysis excludes 12 periods (months), i.e. properties transacted in 2015, considering the fact that the data available only contains data on properties that were either sold, withdrawn or rented out. In fact, the real amount of properties listed on the market will not be a truthful market representation and the market indicators will not be reliable due to the fact that properties that were listed on the market but did not transacted until the end of 2015 are not included in the database. The database provides the zip codes of the properties, based on which they can be

(22)

‘completed’ when a contract is signed, which provides a clear indication about the market timing.

The realtors in the Netherlands represent an important element in the housing market due to the fact that they are the intermediaries between buyers and sellers, and their participation is registered in nearly 90% of all owner-occupied transactions (Op't Veld, Bijlsma and Van de Hoef, 2009). The major bulk of transactions completed are done by the NVM realtors, i.e. 75% of all transactions, providing a reliable representation of the overall owner-occupied housing market in the Netherlands (Kerste et al., 2012). This fact is supported by the relatively low market share of other brokerage associations, such as VBO which is the second largest brokerage association, having less than 10% market share (Kerste et al, 2012).

As aforementioned, the empirical research examines the relationship between house prices and the online search behavior expressed through ‘Search Query Indices’. According to Funda (2014), online search markets are becoming a more important vehicle in completing transaction in residential real estate. The number of online visitors of ‘Funda’, the most widely used brokerage website in the Netherlands, resulted in 72.4 million visits per month in 2014 which represents an increase of 37% from 2013. The search engine technology provides a precise information tool consisting the aggregate consumers’ intentions represented through their digital searches. Wu and Brynjolfsson (2015) illustrate that ‘Internet Search Queries’ obtained form Google Trends can provide a reliable forecast variable in predicting prices and volume in the housing market in the United States several month prior to the actual market transaction. In fact, they support the idea that changes in the house price index are associated with the general supply-demand dynamics through which a ‘honest signal’ through the online search engines can provide inferences about the housing market mechanism. Quantitatively, they conclude that an increased search frequency of one percentage point by real estate agents leads to additional 67,700 housing sales in the upcoming quarter. In addition Wu and Brynjolfsson (2015) find the correlation between a Google Search Engine related to the

(23)

housing market to be statistically significant in explaining the house price index, in particular with the Case-Shiller index.

As such, a proxy of the behavior in the real estate search market is obtained from a second online database, Google Trends, providing monthly data reports on ‘Search Queries’ statistics. Index for the ‘North Holland’ region of the Netherlands is obtained for the pre-defined ‘Search Query’ “Huis Te Koop” (House To Buy). This index is incorporated in the analysis as it represents a proxy for the demand/supply activity in the residential real estate search market in the Netherlands. In fact, Google Trends queries represent an aggregate search volume index disaggregated by different geographic location over time. The dataset is available on monthly basis which also allows for an empirical research with monthly frequency. One major limitation of the ‘Search Query Index’ is the fact that the index is only available for major provinces at an inter-country level. Therefore, ‘North Holland’, which is used in the further analysis, is inclusive but not fully matching representation of the housing search-market in Amsterdam. On the other side, since the information search for owner-occupied housing is increasingly conducted on the Internet, and the usage of the Internet increases through time, the robustness of the model should increase over time. Nevertheless, the systematic efficiency of the aggregate level of the index is not improving over time, as the index contains data originated for the whole North Holland Region.

Table 1. Variables Description

Variable Description Source

(Jan 2004 – 2014 Dec)

HPA the natural logarithm of the house price index in Amsterdam NVM

PropertiesMarket # of listed properties on market at beginning of the period NVM

TV Transaction Volume - # of properties sold during the period NVM

PropertiesWithdrawn # of properties withdrawn from the market in the period NVM

PropertiesAdded # of properties added during the period NVM

(24)

4.2. Housing Market Indicators

In analyzing the relationship between house prices and the online search behavior, an aggregate data on the owner occupied housing market transactions in Amsterdam is obtained from the Dutch Association of Realtors. As a by product from the database can be estimated an indication about the housing market condition in different time periods by estimating the number of houses listed for sale, the number of houses unlisted from sales (withdrawals), and number of sales in the corresponding period. These variables are graphically represented in Figure 1.

Figure 1. Market Indicators during the period 2004 – 2014

a) Number of properties sold b) Number of properties listed on the market

c) Number of Properties added in the period d) Number of properties withdrawn from the market

The ‘number of properties sold’, represented in the Figure 1. a) is estimated as the number of contracts signed in the period i.e. when a transaction is initiated. As it can be seen from Figure 1. a) the variable clearly depicts the pre- and post-effect of the financial crisis in 2007-2008 with a significant drop in transaction volume, bottoming up in 2009 and a clear recovery

200 400 600 800 1000 N u mb e r o f Pro p e rt ie s So ld 2004 2005 2006 2007 2008 2009 2010 2011 2012 2013 2014 Year 2000 4000 6000 8000 10000 N u mb e r o f Pro p e rt ie s o n t h e Ma rke t 2004 2005 2006 2007 2008 2009 2010 2011 2012 2013 2014 Year 400 600 800 1000 1200 1400 N u mb e r o f Pro p e rt ie s Ad d e d 2004 2005 2006 2007 2008 2009 2010 2011 2012 2013 2014 Year 100 200 300 400 500 N u mb e r o f Pro p e rt ie s W it h d ra w n 2004 2005 2006 2007 2008 2009 2010 2011 2012 2013 2014 Year

(25)

afterwards. In line with this, the graph ‘number of properties listed on the market’ as presented in Figure 1. b) also illustrates the same effect. As it can be seen, the number of listed properties on the market follows the same trend of bottoming and then increasing post-2008. This is a clear representation of the cyclicality in the property.

As introduced by Berkovec and Goodman (1996), market tightness indicator provides an insight about participants on the market based on which inferences can be made about market condition and the probability of occurrence of a potential market transaction. The variable ‘Market Tightness’ represents the ratio of the number of properties sold during the period against the number of properties available for sale listed at NVM. In fact, an increase in the ratio of properties sold against the number of properties listed on the market provides a clear indication about market attractiveness, i.e. the hotness of the market, where the number of interested buyers is increasing in respect to the supply on the market. From Figure 2. can be seen that the Market Tightness indicator inversely follows the same pattern as the number of properties sold but with a lower noise. In fact its reliability can also be visually confirmed by the fact that it almost perfectly captures the market crash in 2008. However, a limitation of the market tightness indicator is the fact that it is created only by properties that were listed in the NVM database and were either sold, withdrawn or rented out during the period. The limited representative of the indication might also provide bias in the analysis limited to NVM representation.

Figure 2. Market Tightness Indicator - Demand Supply Ratio

.1 .2 .3 .4 N u mb e r o f Pro p e rt ie s So ld / o n t h e Ma rke t

(26)

4.3. Google Trends Search Query

The Search Query Index ‘Huis Te Koop’ obtained from Google Trends, as presented in Figure 3, follows quite volatile pattern. This variable is majorly affected by the increased usage of the Internet, with an increased robustness through time. This fact is supported graphically by the fact of increased volatility in the first half of the data set and a reduced volatility in the second time set. The clear house price trend cannot be depicted from the graph. However there can be seen an increased search behavior in the recent years.

Figure 3. Google Trends - Search Query Index – ‘Huis Te Koop’ 2004 -2014

Variables descriptive statistics and their correlation matrix are presented in Table 2 and Table 3 respectively.

Table 2. Variables Descriptive Statistics

Variable Mean Median Deviation Standard 25

th Percentile 75 th Percentile Observations # of (Jan 2004 – 2014 Dec) PropertiesMarket 5049.05 4889.18 1946.73 3491.27 6563.18 132 TV 613.29 603.40 124.22 516.43 694.31 132 PropertiesWithdrawn 230.33 218.84 87.00 159.634 293.81 132 PropertiesAdded 835.94 821.42 187.41 696.54 961.24 132 MarketTightness 0.1481 0.1221 0.0787 0.08297 0.2021 132 HPA 0.18711 0.2021 0.0870 0.1444 0.2506 132 SQI-HTK 53.55 53.47 8.23 47.42 59.05 132 30 40 50 60 70 Se a rch Q u e ry In d e x - 'H u is T e Ko o p ' 2004 2005 2006 2007 2008 2009 2010 2011 2012 2013 2014 Year

(27)

From the correlation matrix presented in Table 3 can be seen that the all of the independent variables are positively correlated with the dependent variable, HPA, but the PropertiesMarket. This negative correlation is in line with the basic supply/demand economic concept indicating that an increase in supply will lead to decease in the house prices. Therefore the negative sign is as expected. Among the independent variables, the MarketTightness variable is highly positively correlated with the Transaction Volume (0.7345) and highly negatively correlated with the properties on the market (-0.949). This is the case as the MarketTightness is a ratio variable constructed from these two variables.

Table 3. Correlation Variables

Variable HPA Properties Market TV Properties Withdrawn Properties

Added MarketTightness SQI-HTK

HPA 1.000 Properties Market -0.1858 1.000 TV 0.253 -0.483 1.000 Properties Withdrawn 0.0981 0.8019 -0.5293 1.000 Properties Added 0.2522 0.3074 0.1571 0.0678 1.000 Market Tightness 0.2351 -0.949 0.7345 -0.8121 -0.1816 1.000 SQI-HTK 0.1617 0.0618 0.1284 0.1479 0.0151 -0.0017 1.000

(28)

5. METHODOLOGY

5.1. Autocorrelation in House prices

Case and Shiller (2003; 1989) indicate that fundamental factors adjustments that bring house prices into equilibrium on the market are not as responsive as it is the case with the most asset classes. They indicate the bid-ask spread as being more responsive to changes in market conditions than the shocks in housing supply and demand. In fact the bid-ask spread increases when the demand decreases, triggering lower number of transactions. This process Case and Shiller (2003; 1989) illustrate as the outcome of sellers resistance to adjust prices downwards. Therefore, their model portraits house price adjustments following a sticky pattern, which is in favor of employing a forecasting model contain an autoregressive component.

Historically, house prices have been proven to follow a boom-bust cycle. In modeling such a pattern, a moving average component intuitively seem as better suited in forecasting house prices, considering the fact that the error term in previous period contain sufficient price adjustment information in both rising and falling markets. Muellbauer and Murphy (1997) point out that investors are facing credit limits which is the main driver of the trade-off in the timing of consumption. In turn, when yields on housing investments are high, reduced consumption in the present enables increased investment in housing, which will result in greater consumption in the future. Nevertheless, housing investment entails higher transaction costs than other asset classes. These lumpy costs also contribute towards the market timing in the housing market due to the threshold requirement in housing demand. Higher house price appreciation increases the transaction volume facilitates the transactions cost hurdle to be met quicker.

An empirical analysis by Crawford and Fratantoni (2003) compares the forecasting ability of the owner-occupied housing market by employing three different univariate models: i) ‘Autoregressive Integrated Moving Average’ (ARIMA) model, ii) the ‘Generalized Autoregressive Condition Hetroskedastic’ (GARCH) model and iii) the ‘Regime-Switch’

(29)

(RS) model. Having into consideration the limited data availability, the regime switch model cannot be applied in this empirical analysis since real estate business cycles last between 6 to 10 years. Having a monthly data for 10 years, a change in the underlying trends in the market cannot be identifies. As a representative of the univariate forecasting technique in the empirical analysis only the ARIMA model will be applied.

In addition, De Wit et al. (2013) employ a Vector Error Correction (VEC) model in modeling the change of the growth in house sales and the price level in the Netherlands. As the main research topic of their analysis is the factors influencing changes in the housing price/liquidity relationship in the Dutch housing market. As a result they indicate a gradual adjustment of participants to changes in the fundamental factors in the Dutch house markets. In their modeling they find house prices and volume traded to follow a significant first order positive correlation. As De Wit et al. (2013) indicate, house prices tend to exhibit strong positive correlation to the volume of houses transacted on the market. As such, the following empirical paper incorporates a Vector Autoregressive Model (VAR) which simultaneously models House Prices and Trading Volume.

5.2. House Price Index Estimation

Having into consideration the intransparency of the housing market, the heterogeneity of the individual assets due to different house qualities and location, it arises the necessity to estimate a house price index that takes into account all these different house-specific characteristics. In doing so the hedonic price model is applied in which the house price index is estimated based on individual transactions on the housing market. The model allows house price index estimation by isolating individual housing characteristic from the transacted prices and estimating monthly house price changes in Amsterdam. As it is explained in the data section, this technique is possible due to the information provided by the Dutch

(30)

The house price index is estimated by introducing Time-Dummy Variables that captures the house price change controlled for property size characteristics, additional house specific features, the quality of maintenance, house type, construction year and location. The regression output, data and variables description is available in Appendix A.

Equation 1. Hedonic Price Model

𝐿𝑛𝑃𝑟𝑖𝑐𝑒 = 𝛽!+ 𝛽!∗ 𝐿𝑛𝐻𝑜𝑢𝑠𝑒𝑆𝑖𝑧𝑒(𝑚!) + 𝛽!∗ 𝐿𝑛𝑃𝑙𝑜𝑡𝑆𝑖𝑧𝑒(𝑚!) + 𝛽!∗ 𝐿𝑛𝐺𝑎𝑟𝑑𝑒𝑛𝑆𝑖𝑧𝑒(𝑚!) (Size Features)

+ 𝛽!∗ 𝐿𝑖𝑓𝑡 + 𝛽!∗ 𝐴𝑡𝑡𝑖𝑐 + 𝛽!𝑀𝑜𝑛𝑢𝑚𝑒𝑛𝑡 + 𝛽!∗ 𝑃𝑜𝑜𝑙 + 𝛽!∗ 𝑃𝑎𝑟𝑘𝑖𝑛𝑔 + 𝛽!𝑃𝑜𝑟𝑐ℎ (Additional Features)

+ 𝛽!"∗ 𝐵𝑒𝑙𝑜𝑤𝐴𝑣𝑒𝑟𝑎𝑔𝑒 + 𝛽!!∗ 𝐴𝑏𝑜𝑣𝑒 𝐴𝑣𝑒𝑟𝑎𝑔𝑒

(Quality and Maintenance)

+ 𝛽!"∗ 𝑅𝑜𝑤𝐻𝑜𝑢𝑠𝑒 + 𝛽!"∗ 𝑆𝑒𝑚𝑖𝐷𝑒𝑡𝑎𝑐ℎ𝑒𝑑 + 𝛽!"𝐶𝑜𝑟𝑛𝑒𝑟 + 𝛽!"∗ 𝐷𝑒𝑡𝑎𝑐ℎ𝑒𝑑

(House Type)

+ 𝛽!"∗ 𝐶𝑃!"#$!!"#$+ 𝛽!"∗ 𝐶𝑃!"#!!!"##+ 𝛽!"∗ 𝐶𝑃!"#$!!"#"+ 𝛽!"∗ 𝐶𝑃!"#$!!"#$ + 𝛽!"∗ 𝐶𝑃!"#!!!!"#+ 𝛽!"∗ 𝐶𝑃!"#!!!""#+ 𝛽!!∗ 𝐶𝑃!""!!!"""+ 𝛽!"∗ 𝐶𝑃!!""#

(Construction Period)

+ 𝛽!"∗ 𝑁!"#$+ 𝛽!"∗ 𝑁!"#$+ 𝛽!"∗ 𝑁!"#$%#&' (Neighborhood) + 𝛽!"∗ 𝑁!"#$+ 𝛽!"∗ 𝑁!"#$+ 𝛽!"∗ 𝑁!"#$%&'( + 𝛽!"𝐷!""#!!+ ⋯ + 𝛽!"!∗ 𝐷!"#$!!" (Time-Dummy Variable) +𝜀 (Error Term) 5.3. Econometric Modeling 5.3.1 ARIMA model

The generalized ARIMA model specification in provided in Equation (2). An ARIMA model incorporates three different components: i) an autoregressive component of the dependent variable applied by including the lagged values of the dependent variable whose magnitude effect is estimated by the estimator coefficient ‘𝛽!’ for ‘k’ lags; ii) a moving average

component i.e. the lagged values of the error term, whose magnitude is estimated by the coefficient estimate ‘𝜑_!’ for ‘n’ lags; iii) and the co-integration component necessary to ensure that the dependent variable is stationary. The model presented in Equation (2) also incorporates the intendant variable SQI-HTK with ‘j’ lags.

(31)

Equation 2. ARIMA Model ∆𝐻𝑃𝐴!= 𝛽!+ 𝛽!𝐻𝑃𝐴!!! ! + 𝛼! ! 𝑆𝑄𝐼 − 𝐻𝑇𝐾!!!+ 𝜑!𝜀!!! ! + 𝜀!

Equation (2) illustrates the monthly house price growth rate in period ‘t’, ∆HPAt. The error term of the model ‘𝜀!’ is normally distributed i.i.d. with variance ‘ℴ!’. The Search Query

Index, SQI-HTK, as obtained from Google Trends is tested for ‘j’ lags of the query index variable.

Crawford and Fratantoni (2003) apply the ARIMA model analysis only with limited degrees of freedom, i.e. up to two lags. However, in their empirical analysis they find statistically significant evidence of different house price dynamic in different markets. In fact, they illustrate that housing markets that have more volatile and extreme boom-burst cycles are better explained with an ARIMA model than the housing markets that experience more stable price developments.

As it might be assumed, a long-run relationship between online search markets and house prices may exist. However, the limited data availability restricts the horizon of the analysis. Hence, the long-run relationship cannot be examined. However, it can be expected that from the data available can help explain and forecast the short-term house price dynamics. In doing so, the most applicable ARIMA framework is structured to models short run monthly house prices changes.

The data incorporates 11 years of data for the period of 2004-2014, or 132 monthly observations. Having into consideration the monthly frequency of the data, the data is found to follow seasonal patterns. In ensuring data robustness, the variables are deseasonalized. More information on data deseasonalizing is available in appendix B.

(32)

time series analysis, the variables should be stationary and as such no trend should be possibly identified (Stock and Watson, 2011). The Ducky-Fuller Test is employed to test for stationary. Based on the Ducky-Fuller outcome available in Appendix C, it is assumed that house prices are non-stationary but their first difference is stationary. As a result an ARIMA model is applied accounting for stationarity instead of ARMA model.

In addition, an ARIMA model is fitted by determining the number of auto-regressive terms and the moving average component of the dependent variable ΔHPA. This can be estimated in a more systematic way by illustrating the autocorrelation function (ACF) and the partial autocorrelation function (PACF) through which can be identified the number of terms necessary to fit the model.

Figure 4. Identifying AR & MA terms

a) Autocorrelation Function b) Partial Autocorrelation Function

Figure 4. a) illustrates the correlation coefficients of the lags of the dependent variable ΔHPA. From the figure can be seen that the first autocorrelation is statistically significant which is supported by the fact that the first lag is placed outside the 95% confidence interval. The other lags are within the confidence interval except for the 8th_{lag. Although it is outside the}

interval, only the first lag is taken as significant. In addition, from Figure 4. b) can be seen the partial autocorrelation of the dependent variable ΔHPA and its lags. This graph illustrates the correlation magnitude that is not explained by all lower–order lags of the variable. From the figure 4. b) can be noticed that that there is a spike after the first lag which is outside the 95% confidence interval indicating that it can partially explain the variability of the lower order

-0 .4 0 -0 .2 0 0.00 0.20 0.40 Pa rt ia l a u to co rre la ti o n s o f H o u se Pri ce Amst e rd a m 0 10 20 30 40 Lag

(33)

lagged values. Also from the graph can be seen that the second lag is within the confidence interval but relatively close to the border. After several modeling and fitting techniques, it was concluded that the second moving average term has also explanatory power over the dependent variable, hence the second order was taken in the model.

5.3.2 Vector Autoregressive (VAR) model

As it might be expected and was indicated by De Wit et al. (2013) there might be a mutual long-term causality relationship between the prices and the transaction volume in the housing market. As it is abovementioned, the limited availability of the data does not allow to precisely examine the long-term relationship between these two variables. Therefore, the current empirical research focuses on explaining the short-term supply-demand dynamics in the housing market. From the autocorrelation function presented above in Table 3 can be seen that the returns in house prices and the transaction volume variables are not heavily correlated between periods due to the correlation coefficient of 0.253. However, as indicated in the Literature Review section, existing literature suggest that these two variables mutually influence each other. In testing their mutual causality i.e. in the econometric modeling in the price-volume relationship, a VAR model is employed. In fitting the VAR model the first lag of the dependent variables ΔHPA and ΔTV are included. The VAR model to be tested is graphically presented in Equation (3).

Equation 3. VAR Model

Δ𝐻𝑜𝑢𝑠𝑒𝑃𝑟𝑖𝑐𝑒! Δ𝑃𝑟𝑜𝑝𝑒𝑟𝑡𝑖𝑒𝑠 𝑆𝑜𝑙𝑑! = 𝑦! 𝛿! 𝑦! 𝛿! ! !!! Δ𝐻𝑜𝑢𝑠𝑒𝑃𝑟𝑖𝑐𝑒!!! Δ𝑃𝑟𝑜𝑝𝑒𝑟𝑡𝑖𝑒𝑠 𝑆𝑜𝑙𝑑!!! + 𝛽! 𝛽! 𝑓𝑆𝑄𝐼 − 𝐻𝑇𝐾!!! ! !!! + 𝜀!

From the model presented can be seen that it is expected the house price changes and the transaction volume to be dependent on ‘q’ lags of both variables i.e. the value of the dependent variables in period ‘t-q’. In addition, the SQI-HTK variable is included with ‘p’

(34)

5.4. Hypotheses

There is limited empirical research to date that explicitly models the behavior of market participants and estimates their impact on the marker over the asset pricing. Owing to the increasing Internet usage, the magnitude of the behavioral finance can be quantified and modeled with an increased precision. In the following empirical research several hypothesis in regard to Online Search Query Indices are tested:

I. “Participants using Online Search Engines have explanatory power in the house price dynamic”;

II. “Market tightness positively correlates house prices”;

III. “There is a positive correlation between House prices and Transaction volumes”;

Hypotheses (I) investigates whether the Online Search Query Index is significant in explaining the house price dynamics on the market. The hypothesis is tested by introducing the lagged values of SQI-HTK up to order six and the estimated coefficients are tested for their statistical significance.

In addition, the significance of hypothesis (II) or the market tightness variable is examined by introducing to the pre-defined ratio variable of the number of properties sold during the period against the number of properties available for sale at the beginning of the period. This variable is introduced along the variables tested in hypothesis (I).

At last, from the presented literature can be concluded that there exist a positive relationship between the house price dynamics and the volume of transactions completed on the market. In order to test hypothesis (III) a VAR model is formulated including lags from both HPA and TV variables.

(35)

6. RESULTS

6.1. Time Series - ARIMA models

The coefficient outcomes from the different ARIMA models are presented in Table 4. Model (1) represents the so-called model ARIMA fitting model and does not include any additional explanatory variables but only the first lag of the dependent variable as autoregressive component and the two lags of the error term representing the moving average component. Model (2) incorporates the first six lagged monthly variables of the online Search Query Index ‘Huis Te Koop’ (SQI-HTK). In addition, models (3) and (4) incorporate only the significant SQI-HTK lags and also the market tightness indicator in included. From figures available in Appendix D can be seen that residuals from all four models are approximately normally and randomly distributed.

Table 4. ARIMA models

The table presents four alternative ARIMA (1,1,2) models having the House Prices in Amsterdam (HPA) as dependent variable. Column (1) illustrates the ARIMA model transforming the dependent variable in the Change of House Prices in Amsterdam (ΔHPA), incorporating the first lag of ΔHPA (AR1) and the first and the second lags of the error term as a moving average component (MA1 and MA2). In the models presented in columns (2), (3) and (4) include different lags of the online Search Query Index ‘Huis Te Koop’ variable (SQI-HTK) and the Market Tightness variable.

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

ΔHPA ΔHPA ΔHPA ΔHPA

Coefficient P-value Coefficient P-value Coefficient P-value Coefficient P-Value LnSQI-HTKt-1 .0225453 0.020** .0205985 0.024** .0192999 0.019** LnSQI-HTKt-2 .0092323 0.348 LnSQI-HTKt-3 -.011922 0.215 LnSQI-HTKt-4 .0188895 0.036** .0171032 0.032** .0184466 0.013** LnSQI-HTKt-5 .0084136 0.398 LnSQI-HTKt-6 -.006397 0.492 MarketTightnesst .0175117 0.003*** Constant .0021587 0.339 .0027636 0.346 .0023086 0.363 .0022501 0.369 ARMA AR (L1) .9017327 0.000*** .9213049 0.000*** .9158637 0.000*** .9281244 0.000*** MA (L1) -1.418179 0.000*** -1.55193 0.000*** -1.52889 0.000*** -1.60796 0.000*** MA (L2) .5913474 0.000*** .7214291 0.000*** .6969476 0.000*** .7530674 0.000*** sigma .0142161 0.000*** .0136682 0.000*** .0139154 0.000*** .0136269 0.000*** Log Likelihood 370.8513 358.4029 361.9654 364.4579 N 131 125 127 127 Wald Chi2 322.32 492.74 444.47 732.64 Adj. R2 _0.4934 _0.5458 _0.5316 _0.5544 RMSE .0142764 .0137551 .0140357 .0137442

(36)

From the first model presented can be seen that the House Prices in Amsterdam (HPA) follow an (1,1,2) ARIMA model. From the outcome can be seen that the House Prices are highly positively related with its first lag. The coefficient of +0.9017 indicates that the current HPA value is heavily dependent on the value from the previous period. Thus, the current HPA will retain 90.1% of the value from the last period. In fact, this coefficient indicates that values in the current period value are highly correlated with its distributed lags values. This is in line with the existing literature on the strong positive correlation of prices in the housing market (Blanchard and Watson,1982; Case and Shiller, 1989;2003; Kendall and Hill, 1953).

The moving average component is solely a finite impulse response filtering out the white noise of the dependent variable (Said and Dickey,1984). The moving average component affects the dependent variable in the current period from 2 periods in the past. Effectively the moving average term is expressed by the first and second lags of the error term. As it can be seen from Table 4, the first lag of the error term which represents the first moving average component has the largest magnitude of -1.418 and is inversely related with the current HPA coefficient. However, the second lag of the moving average component i.e. the second lag of the regression error term model is positively related with the HPA with a positive coefficient of +0.59. The significance of the moving average components indicates that there is a relationship between the change in HPA and the residuals from the previous two periods. In relative terms, this indicates that the error term of the HPA in the current period will be equal to the multiplication of -1.42 of the error term in the previous period, in addition of the multiplication of +0.59 of the error term from two periods before and an additional error due to the current period model estimate. As it is expected, the magnitude of the error term decreases as the number of lagged periods increases. All three coefficients are highly significant at 1% confidence interval.

In the second model, the SQI-HTK is added as an independent variable. In order to assess the significance of the variable and the magnitude of its forecasting ability over HPA, six

(37)

consecutive monthly lags are added in the model. The variable is expected to have a positive contribution on HPA i.e. positive sign as an increase in the current Search Query Index is associated with an increased activity on the search market which potentially will be translated to an increased housing demand in the future. From the output of model (2) available in Table 4 can be seen that the lags have the expected positive sign with the exception for the third and the sixth lag. However, these two lags are not statically different from zero therefore their interpretation is meaningfulness.

The model (3) only incorporates the significant SQI-HTK lags i.e. it includes only the first and the fourth SQI-HTK lags. These two variables have the expected positive sign and are significant at 5% confidence interval. In fact, they replicate the housing market search dynamics from the literature described by Berkovec and Goodman (1996). The SQI variables indicate that a 1% increase in the SQI-HTK in previous period, will translate in +0.0206% increase in change in House Prices in Amsterdam in the current period. In addition, a 1% increase in the SQI-HTK in the previous four months results in +0.0171% increase in the House Price in Amsterdam in the current period. The magnitude and the sign of these two variables are roughly similar with the coefficient estimated in model (2), but their magnitude has slightly decreased.

In the fourth model, the variable ‘Market Tightness’ (MT) is included as a control variable which increase the stability of the regression model. MT variable indicates the market attractiveness in the housing market at period ‘t’. As it was previously mentioned, MT is a ratio variable estimated as the ratio of the number of properties sold during the period over the number of properties available for sale at the beginning of the period. From the output available in table 4 model (4) can be seen that the variable has the expected positive sign, indicating that an increase in the ratio of sold properties in respect to properties available for sale represent an increase transaction activity in the housing market. Such an increase thus

Can online search queries improve the forecasting efficiency of the short-term house pricing dynamics? : an empirical analysis of the owner-occupied house market in Amsterdam