The momentum effect for equity and mortgage REITs in the US, a comparative study

(1)

Master Thesis

The momentum effect for Equity and Mortgage REITs

in the US, a comparative study

Author: Spike W. Gontscharoff

Supervisor: Dr. Erasmo Giambona

MSc Business Economics: Finance & Real Estate Finance Track July 2016

(2)

i

Statement of Originality

This document is written by Student Spike W. Gontscharoff 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)

ii

Acknowledgements

First I would like to express my gratitude to my thesis supervisor Dr. Erasmo Giambona for his support and advice while writing my dissertation. I would also like to thank him for the course Real Estate Finance and Portfolio Management which introduced to me to the topic of REITs. After being introduced to the topic, further looking into this asset class sparked my interest and ultimately led to the creation of this research.

Finally my thanks go out to my friends, family and fellow students for their support and giving me the opportunity to express my complaints on setbacks during the writing process.

(4)

iii

Abstract

This study investigates the momentum effect for Real Estate Investment Trusts (REITs) to determine if momentum varies across REIT- and property type. Although there exists a significant body of literature on the momentum effect for REITs in general, until now it has not been studied at a more detailed industry level. Fama-Macbeth (1973) cross-sectional regressions are used to test if momentum is a determinant of expected US REIT returns over the 1994-2014 period. Then, by extending the basic cross–sectional model with dummy variables, the variation in momentum across REITs is tested for significance. The performance of momentum is tested by making use of a portfolio analysis based on the techniques of Chui et al. (2003a). Both the outcomes from the cross-sectional regression and the portfolio analysis show REIT- and property types provide significant momentum profits. However, additional testing shows there is no significant variation in the momentum effect across these REITs.

Keywords: Real Estate Investment Trusts (REITs), Momentum predictability, portfolio analysis, Fama Macbeth regression

(5)

iv

1. Introduction

There has always been a considerable interest in simple trading strategies that have the potential to generate superior returns. According to the Efficient Market Hypothesis it is not possible for an investor to outperform the market using such a strategy on a risk-adjusted basis. Jegadeesh and Titman contradicted these beliefs in their 1993 paper. They discovered going long in stocks that had performed well in the past and going short in stocks that performed poorly generated positive returns over a 6- to 12-months holding period. This continuation in stock returns came to be known as momentum and yielded an annual return of 12% on the US stock market for the 1965-1990 period. In the same paper they show these momentum returns are not the result of systematic risk as correcting for the traditional one-factor CAPM still generated significant profits. Since then, research has found pervasive evidence for the momentum effect across different markets and time periods.

The momentum effect did not remain purely a theoretical approach. Fund trading based on this strategy was very popular in the late 1990s. Recent examples such as the AQR Large Cap Momentum Style Fund that started in September 2009, still exist to this day. Since traditional risk factors have been unable to explain the momentum profits, research began to look for possible explanations in other fields of finance as well. Theories from the field of behavioural finance in particular have been used to explain these market irrationalities. Over time, part of the momentum literature started to focus solely on real estate investment trusts (REITs). Chui et al. (2003a) find that in the 1990-2000 period, momentum is the only significant predictor of REIT return. A momentum portfolio based on the techniques of Jegadeesh and Titman (1993) resulted in a monthly momentum profit of 1.2%. These results were later confirmed by other papers: (Chui et al., 2003b; Hung & Glascock, 2008; Goebel et al., 2013). Although the momentum effect for REITs has been proven by multiple papers there are reasons to believe the strategy has become less powerful in recent times. Goebel et al. (2013) for the 1993-2009 period and Hung and Glascock (2008) for the 1972-2000 period find lower momentum profits on the REIT market compared to earlier papers. Further, Chordia et al. (2014) mention recent improvements in trading technology have made markets more efficient, resulting in lower momentum profits. Finally, Jegadeesh and Titman (2011) show the recent financial crisis had a negative influence on momentum with a loss of 36.5% for the momentum strategy in 2008. In this study, we first examine whether the momentum effect is still significantly positive for our more recent 1994-2014 sample period.

(7)

2

There exists a significant body of literature on the momentum effect for REITs in general. However, it has not been studied at a more detailed industry level. This study ads to the existing literature on REIT momentum by determining if momentum varies across REIT- and property types. To answer this question we turn to the literature on industry level momentum from the stock market. Moskowitz and Grinblatt (1999) conclude momentum profits are no longer significant at an industry level as the industry effect itself is the driving factor behind the momentum effect. On the contrary, Grundy and Martin (2001) and Chui et al. (2003b) find significant momentum profits at an industry level when the bid-ask bounce is accounted for. However, results from both papers show the momentum effect is not significantly different across these common stock industries.

There are a number of reasons why we expect the results to be different for the REIT market. First, Chui et al. (2003b) show earnings and return volatility increased significantly in the post-1990 period for REITs, whereas this was not the case for common stocks industries. As we will discuss in more detail below, the overconfidence theory proposed by Daniel et al. (1998) shows this increased volatility causes higher momentum profits for REITs. Moreover, Sun et al. (2015) show the return volatility differs between property types, which could result in relatively weaker/stronger momentum profits for certain REITs. Second, not all REITs were affected equally by the recent financial crisis. For example, Lodging/Resort REITs experienced stronger losses compared to less cyclical REITs such as Healthcare.

Apart from these two more general reasons there are separate arguments to be made at a REIT type or property type level. Ro and Gallimore (2014) show the level of herding for Mortgage REITs is significantly higher compared to Equity REITs. According to Scharftein and Stein (1990), higher levels of herding indicate decreased market transparency. Which based on the overconfidence theory of Daniel et al. (1998) should result in stronger momentum profits for Mortgage REITs compared to equity REITs. Chui et al. (2003a) provide suggestive evidence that the same firm characteristics predicting expected stock return also influence the strength of the momentum effect. Using a double-sort procedure they find momentum strategies for REITs are non-significantly related to size, but decrease with analyst coverage. Further, momentum strategies are stronger when implemented on REITs with higher book-to-market (BM) ratios and those with higher turnover.

We argue high BM ratios cause investors to be overly optimistic about firm performance. Based on the representativeness heuristic of Barberis et al. (1998) this optimism causes stronger momentum. According to Hong and Stein (2007) higher trading volume means there is more disagreement among investors about the true value of the stock. According to the

(8)

3

information diffusion theory by Hong and Stein (1999), this disagreement results in a stronger momentum effect. The descriptive statistics for our REIT sample show that firm characteristics vary across property types. Lodging/Resort- and Industrial/Office REITs in particular have higher BM and turnover ratios, which is why we expect to find a stronger momentum effect for these property types. An advantage of studying the momentum effect within a single industry is that we control for systematic effects potentially influencing our results (Feng et al., 2014).

The main sample consist of all REITs trading on the NYSE, AMEX and Nasdaq for the January 1994 – December 2014 period. In the beginning of the 1990s, the REIT market changed fundamentally. These changes caused the momentum effect to be significantly higher in the post-1990 period (Chui et al., 2003a; Hung & Glascock, 2008). This specific timeframe, starting in 1994, prevents the structural changes from influencing our results. The significance of a momentum effect is investigated using two separate models based on the methodology by Chui et al. (2003a). The first is a cross-sectional Fama Macbeth (1973) regression. Factors known to influence stock return: past six-month cumulative return, BM, size and turnover are regressed on six-month cumulative return. REIT- and property types are investigated by splitting the entire sample into smaller samples. A drawback of this method is that we reduce the sample size for each subsample. To control for this, a second regression using dummy variables for the respective REIT- and property types is used as well.

For the second method, REITs are sorted based on six-month cumulative return, where the top (bottom) 30% performing REITs is assigned to the winner (loser) portfolio. Portfolios are held for six months, skipping one month between the formation and holding period. A momentum portfolio is created by shorting the stocks in the loser portfolio and using the proceeds to buy the stocks from the winner portfolio. Two different methods are used to assign REITs to the winner and loser portfolio. A value-weighted strategy following Chui et al. (2003a) and an equally-weighted strategy following Jegadeesh & Titman (1993). To analyse the various subsamples, intra-industry portfolios are formed based on the methodology of Moskowitz and Grinblatt (1999).

The findings are as follows: The cross-sectional regression shows a dominant momentum effect with only past six-month cumulative return having a significant influence on expected return. For the entire sample the momentum coefficient is smaller compared to previous literature. Portfolio analysis confirms this results with an average monthly profit of 0.28%. Further, cross-sectional regression results indicate both Equity and Mortgage REITs having a significant momentum effect. Although the momentum effect seems higher for Mortgage REITs, the dummy approach shows this difference is not significant. Finally, the

(9)

4

results for property type are similar to that of the REIT type. Both the cross-sectional regression and the portfolio analysis show a significant momentum effect on a property type level. However, we cannot conclude they are significantly different.

In order to properly asses the research question the remainder of this thesis is structured as follows. Section 2, presents the existing literature and the resulting hypotheses. Section 3 discusses the used data and provides some descriptive statistics, while Section 4 describes the used methodology. Section 5 provides the empirical results from the main analysis and Section 6 provides the results of several corresponding robustness checks. Finally, Section 7 concludes and discusses the results.

(10)

5

2. Literature Review

This section starts by discussing the relevant literature on momentum for the stock market followed by the momentum literature for the REIT market. Next, the recent performance of the momentum strategy is shown after which we form the first hypothesis. Subsequently, explanations for the persistent momentum effect are discussed. Finally, the previous findings for momentum strategies at an industry level are given together with the remaining hypotheses.

2.1 Momentum effect observed in stocks

Proving the existence of momentum is part of a bigger debate on market efficiency. According to the Efficient Market Hypothesis (EMH) by Malkiel and Fama (1970) prices of tradable assets should reflect all available market information. Fama (1998) acknowledges that two robust and persistent anomalies potentially challenge this efficiency. The first, known as earnings momentum is defined in two ways. The first, called the standardized unexpected earnings momentum (SUE) looks at the difference between quarterly earnings and expected quarterly earnings. Firms whose earnings growth exceeds expectations subsequently outperform competitors (Foster et al., 1984; Bernard & Thomas, 1989). The second measure is the revision of analyst forecasts. Stocks whose earnings expectations are revised upwards outperform peers between 5% and 7% (Chan et al., 1996; Stickel, 1991). Although there is substantial literature on earnings momentum for stocks, empirical results are still limited for REITs. To make our results more comparable with previous literature, whenever we mention the momentum effect hereafter we are referring to the price momentum effect.

The second anomaly violating the EMH, known as (price) momentum has long been investigated starting with DeBondt and Thaler (1985), who found a long-term reversal in stock returns. For the 1933-1980 period, stocks that performed badly in the past three-five years outperformed the best performing stocks in the following three-five years by 25%. Jegadeesh and Titman (1993) found opposing results regarding the relative performance of past winners and losers. Stocks performing well in past continued to perform well in the short term future, which came to be known as momentum. To investigate the presence of a momentum effect their strategy selects stocks based upon their lagged returns over the past 3- to 12 months (J-month return) and holds them for another 3- to 12 months (K-month holding period). Stocks are sorted based on their past J-month performance where the top (bottom) 10% performing stocks are assigned to the winner (loser) portfolio. A momentum portfolio is created by shorting the stocks in the loser portfolio and using the proceeds to buy the stocks from the winner portfolio. A

(11)

6-6

month/1-month/6-month strategy provided average monthly profits of 1%.

Since the publication of their paper in 1993 other researchers have documented the existence of a momentum effect across different markets and data periods. Rouwenhorst (1998), investigates the momentum effect for an internationally diversified portfolio consisting of stocks from 12 European countries between 1978 and 1995. He finds an average monthly profit of 1%, confirming the results of Jegadeesh and Titman (1993). Griffin et al. (2003) look at a sample of 39 countries which studies the momentum effect on a global basis for the 1975-2000 period. They report average monthly momentum profits of 0.77% for Europe, 0.59% for the US, 1.63% for Africa, and 0.32% for Asia. Only the results for the Asia region are not statistically significant.

2.2 Momentum effect observed in REITs

Over time, some of the momentum literature started to focus solely on REITs. This development started due to a growing REIT market in the post-1990 period but also because REITs are a relatively homogeneous industry group. The latter makes it, in contrast to most industries, fairly straightforward to determine if a firm classifies as a REIT. Using a single industry also reduces possible confounding effects as a result of differences in risk, intangible assets and growth potential (Feng et al., 2014). Moreover, according to Case and Shiller (1989) underlying real estate assets in REITs create a favourable environment for momentum, since real estate returns exhibit strong serial correlation.

Chui et al. (2003a) were among the first to investigate the determinants of expected REIT return. Using a Fama Macbeth (1973) procedure they find that in the 1990-2000 period, momentum is the only significant predictor of REIT return. In addition, they create momentum portfolios using the portfolio formation technique of Jegadeesh and Titman (1993). Monthly momentum profits of 1.2% for the 1990-2000 period are higher compared to the 0.44% profits in the pre-1990 period. They argue this difference is caused by structural changes on the REIT market, such as the evolution of the UPREIT structure. Chui et al. (2003b) show momentum profits were stronger for REITs (1.33%) compared to stocks (0.50%). They attribute this difference to the stronger serial correlation in real estate returns mentioned by Case and Shiller (1989). Further, empirical findings show momentum returns differ greatly depending on the length of the holding period, with a six-month holding period providing the strongest results. For the 1972-2000 period, Hung and Glascock (2008) find a portfolio return of 0.81% following Chui et al (2003a) and a monthly return of 0.46% using the methodology by Jegadeesh and Titman (1993). Moreover, according to their results assigning the top (bottom)

(12)

7

10% of REITs to the winners (loser) portfolio provides higher momentum profits compared to a 30% approach. These results are in line with Jegadeesh and Titman (1993) who show lower ranking deciles lead to higher momentum profits on the stock market. Finally, Hung and Glascock (2008) report REIT momentum is higher when the market is going up (bull market) and when REITs have a higher dividend/price ratio. Goebel et al. (2013) find a profit of 0.2% per month for the 1993-2009 period based on the methodology by Chui et al. (2003a). Since they look at a one-month holding period, momentum returns are lower compared to most other studies using a six-month holding period. They argue the significance of other cross-sectional determinants such as size and illiquidity on REIT return depends on the monetary policy environment these REITs face.

The results from these papers on common stocks and REITs suggest momentum profits depend on the techniques used to construct the momentum portfolios. Therefore, several robustness checks are performed throughout this paper.

2.3 Recent performance of momentum strategies

The previous two sections show the momentum effect is not an incidental occurrence. It has been proven to remain significant under various model assumptions and/or sample periods. However, Goebel et al. (2013) for the 1993-2009 period and Hung and Glascock (2008) for the 1972-2000 period find lower momentum profits on the REIT market compared to earlier papers. The results suggest the strength of this momentum effect has decreased over time. Further, Chordia et al. (2014) mention a gradual decrease of prominent return anomalies for the 1983-2011 period on the US stock market as well. Although most anomalies in their sample such as size, turnover, and momentum are still statistically significant, hedge portfolio returns and FM coefficients on these anomalies attenuate towards zero. An exponential decay model illustrates the decreasing strength of these anomalies was strong in the post-2000 period, which they attribute to more efficient markets. Improvements in trading technology such as algorithmic trading and increased hedge fund activity are thought to have made markets more efficient.

There is another event that took place during our sample period that could result in lower and potentially insignificant momentum profits, the financial crisis. Momentum strategies fro the stock market performed badly in the 2000-2010 period. The main reason is a loss of 36.5% for the zero-cost portfolio in 2009. The stock market decline in 2008, followed by a market recovery in 2009 induced negative return serial correlation, resulting in these momentum losses (Jegadeesh & Titman, 2011). These results were confirmed by Sun et al. (2015) for the REIT

(13)

8

market. The NAREIT1_{All Equity REITs index went from 10,256 in January 2007 to a low of}

3,337 in February 2009, with the biggest drop of 60% between September 2008 and February 2009. Overall, these findings show it is possible the momentum effect has become less dominant and potentially insignificant as a predictor of REIT return in recent times. Although the momentum effect is likely to be less powerful compared to earlier studies, we still expect it to be significant for our entire REIT sample. Goebel et al. (2013) find positive and significant momentum profits for the 1993-2009 period which incorporated the recent financial crisis. Moreover, Chui et al. (2003b) show a stronger momentum effect for REITs compared to non-REITs in the post-1990 period. Therefore we do not expect the results of Chordia et al. (2014) for the stock market to apply on our REIT sample. To conclude, we still expect to find significantly positive momentum effect for our 1994-2014 REIT sample:

Hypothesis 1: US REITs experienced a significant momentum effect between 1994 and 2014

2.4 Risk- and behavioural based explanations of momentum

The persistence of the momentum effect over the past decades has led researchers from different fields of finance to develop theories for its existence. One part of this literature has turned towards the more traditional risk based explanation. Some argue momentum can be explained by the cross-sectional dispersion in expected stock returns. Stocks with high expected returns in this period are likely to have higher average realized returns in the next period as well. It is this return continuation that causes momentum (Conrad & Kaul, 1998; Boothra, 2011). Other papers have looked into the risk based explanation by investigating if risk-adjusted returns still generate positive and significant momentum strategies. Jegadeesh and Titman (1993) compute risk-adjusted momentum returns using the traditional one-factor CAPM. Others have done the same using the three-factor CAPM model (Jegadeesh & Titman, 2001; Grundy & Martin, 2001). For all these cases the regressions intercepts of the momentum strategies remain significantly positive, suggesting risk factors are unlikely to explain the momentum profits2. Since traditional risk factors were unable to explain the momentum profits, researchers from the field of behavioural finance took an alternative approach in explaining this phenomenon. At the core of all these theories is the thought that investors act irrationally due

1_{The National Association of Real Estate Investment Trusts (NAREIT)}

2_{The computation of risk-adjusted returns is beyond the scope of this thesis. However, based on previous literature we expect}

(14)

9

to a number of behavioural biases (Moskowitz & Grinblatt, 1999). According to DeMarzo and Skiadas (1998) investors act rationally if they correctly interpret and use their private information, as well as information obtained from the market and trades of others. Since in reality investors have limited processing ability, they often rely on ‘’hunches’’ and ‘’intuition’’ which are easily influenced by these behavioural biases.

Jegadeesh and Titman (2011), mention serial correlation in individual stock return as the driver of the momentum effect. Although the exact cause of this serial correlation is still a point of debate, two types of investor irrationalities, underreaction and overreaction, are the most agreed upon. An underreaction means that in contrary to the EMH, information is only slowly incorporated into the price. Barberis et al. (1998) explain this underreaction based on a conservatism bias. When new information comes out, individuals are overly conservative in adjusting their price expectations, since they believe part of the effect is temporary. As a result, stock prices underreact to new public information. Shefrin and Statman (1985) introduce the disposition effect. Investors hold on to their losing stock for too long but sell their winning stock too soon, due to their risk averseness. Kahneman and Tversky (1979) add that this behaviour reflects the reluctance of admitting a mistake by taking the loss. As a result, stocks with high past returns have unrealized capital gains while low past return stocks have unrealized capital losses, causing an underreaction.

On the other hand, overreaction causes positive abnormal returns in the short run, followed by negative returns in the long run. The overconfidence theory of Daniel et al. (1998) argues that due to biased self-attribution investors place more emphasis on their abilities when actions turn out good compared to actions that turn out bad, making the investor overestimate his or her abilities. This overestimation of abilities causes stock prices to be pushed above their fundamental values. The information diffusion theory of Hong and Stein (1999) considers two groups of investors who trade based on different sets of information. The informed investors buy assets in period t based on their private information. Because not everyone has this information it is only partially incorporated in the price, resulting in an initial underreaction. However, the uniformed traders buy assets (in period t) that performed well in the previous period (t-1), pushing the asset above its fundamental value. Finally, Barberis et al. (1998) developed the idea on a representativeness heuristic. According to this theory, investors become too optimistic about a firms performance when it follows a sequence of good news, and vice versa. As a result, security prices rise above their equilibrium price.

(15)

10

2.5 Cross-sectional determinants of momentum

Asset pricing literature on common stocks has documented that past stock return is not the only determinant of expected stock return. A firm’s stock return is negatively related to its size, measured by its market capitalization (Banz, 1981). Lee and Swaminathan (2000) show that turnover in the past 3- to 12 months is negatively related to future stock return. On the other hand, Rosenberg et al. (1985) and Fama and French (1992) find a positive influence by the BM ratio. Due to high dividend pay-outs these cross-sectional determinants may be different for REITs. Chui et al. (2003a) find that the REIT sample deviates from the stock market as the size effect has a non-significant but positive coefficient. Further, in the post-1990 period, momentum is the only significant predictor of REIT return.

Moreover, a number of studies have documented that the same firm characteristics predicting expected stock return also influence the size of the momentum profits. First, momentum results are lower for large firms and firms with high analyst coverage (Hong et al., 2000). Daniel and Titman (1999) find that momentum profits are higher for low BM stocks. Finally, Lee and Swaminathan (2000) show that momentum profits increase monotonically with the turnover ratio and trading volume. Chui et al. (2003a) re-examine these theories for REITs using a double-sort procedure. All REITs in the sample are sorted into three equal groups according to their firm characteristic. These characteristic portfolios are then further divided into three equal groups based on past six-month cumulative return. REIT momentum is not related to size but momentum profits are significantly higher for REITs with no analyst coverage. Surprisingly, REITs with low BM ratios have lower momentum profits which goes against the results by Daniel and Titman (1999) for the stock market. Finally, the positive relation between turnover and momentum profits is confirmed for the REIT market.

2.6 Industry momentum

There are currently no papers investigating the difference of the momentum effect for REITs on a REIT- or property type level. However there is some evidence on industry level momentum from the stock market. Moskowitz and Grinblatt (1999) were the first to investigate this for the 1963-1995 period using common stocks. First, stocks are attributed to a certain industry based on two digit SIC3 codes, after which momentum portfolios are created. Industry portfolios buying (selling) the top (bottom) three performing industries provided an average momentum profit of 0.43%. Due to the significance and the size of the industry portfolios, they concluded

(16)

11

industry effects were responsible for a large portion of the observed momentum effect. In order to test this further they created industry-neutral portfolios buying the winners and shorting the losers within each industry. These portfolios produced an insignificant average profit of 0.1% which confirmed their beliefs of industry effects driving the momentum profits.

Grundy and Martin (2001) show the strong performance of the industry momentum found by Moskowitz and Grinblatt (1999) disappears when formation- and holding period are non-contiguous. Meaning, if one month is skipped between the holding and formation period, industry-neutral momentum profits are significant. Chui et al. (2003b) later confirmed the results of Grundy and Martin (2001) for the post-1990. However, an F-test whether momentum profits differ across these portfolios is not rejected in both studies. This suggest there is no significant variation for the momentum effect in each of these common stock industries. Overall, the papers discussed in the previous section have documented the momentum strategy to work well when implemented over a large and diverse number of stocks/REITs from different industries. On the contrary, results for common stocks from this section suggest the same strategy does not work as well when implemented within a single industry. To explain why we do expect to find a difference for the REIT- and property types, literature on the different sources of momentum is used.

Chui et al. (2003b) show that monthly momentum profits for REITs remain positive until one year after portfolio formation, but start to become negative on average in the months thereafter. This return reversal in the long term provides suggestive evidence that momentum profits are caused by a delayed overreaction. Based on these results the theories on investor overreaction from (Daniel et al., 1998; Hong & Stein, 1999 and Barberis et al., 1998) are used to form our hypotheses. Further, Chui et al. (2003b) argue earnings- and return volatility increased significantly for REITs in the post-1990 period, whereas for common stock industries it did not. Return- and earnings volatility can be seen as proxies for valuation uncertainty (Dichev & Tang, 2009). According to Daniel et al. (1998) this causes increased investor overconfidence, leading to stronger momentum effects. Since REITs showed increased volatility we expect the momentum effect to be stronger compared to the stock market. Figure A2 based on NAREIT data reinvestigates the results from Sun et al. (2015). The total return changes on a property type level for REITs further illustrate our point. Overall, return volatility is high for REITs, but more importantly the volatility differs between the property types. This suggests momentum profits could differ between the property types.

Finally, we discuss for which REIT- and property type the momentum effect is strongest. Ro and Gallimore (2014) investigate the herding behaviour for REITs as a whole and

(17)

12

various REIT subsectors. Herding occurs when individuals imitate each other’s actions, for instance because they believe others have superior information (Scharfstein & Stein, 1990). When there is less information available (decreased market transparency) decision makers are more likely to follow the actions of others, causing an inversed relationship between herding and transparency (Bikhchandani et al., 1992). Ro and Gallimore (2014) show the level of herding for Mortgage REITs is significantly higher compared to the Equity REITs. Using the behavioural theory of Daniel et al. (1998) this lower transparency should generate increased investor overconfidence and therefore stronger momentum profits for the Mortgage REITs. Moskowitz and Grinblatt (1999) describe this in the form of ‘’hot’’ and ‘’cold’’ sectors in the economy. Investors may simply herd toward hot industries/sectors and vice versa, causing return persistence and possibly momentum:

Hypothesis 2: The momentum effect is stronger for Mortgage REITs compared to Equity REITs

Chui et al. (2003a) show that REITs with higher BM ratios, which are often seen as value stocks, have higher momentum profits on average. Barberis et al. (1998) give a possible explanation for this effect with their representativeness heuristic. REITs that do well consistently over a number of periods can cause investors to be overly optimistic about the REITs performance, pushing the price above its equilibrium value. However, it could simply be the case that this good performance is the result of chance instead of new value enhancing activities. REITs with high BM ratios are seen as value stocks which often have a proven track record. Therefore, investors are more likely to interpret a sequence of good performance for high BM REITs as good performance instead of luck. Investors are too optimistic about the performance of REITs with high BM ratios, resulting in stronger momentum profits.

Next, REITs with higher turnover ratios have higher momentum profits as well (Chui et al., 2003b). This positive relation is a bit surprising since the increased trading should provide more public information, reducing underreaction. On the other hand, higher trading volume means there is more disagreement among investors about the true value of the stock (Hong & Stein, 2007). As a result the knowledge difference between informed traders and uniformed traders as described by Hong and Stein (1999) becomes bigger, resulting in higher momentum profits. Based on these arguments we expect the momentum profits to be higher for REITs with higher BM- and turnover ratios. The descriptive statistics in Section 3 show Lodging/Resort and Industrial/Office have the highest levels of turnover and BM:

(18)

13

One final note, Jegadeesh and Titman (2011) show the recent financial crisis had a negative effect on the momentum profits. As we described earlier, Figure A2 shows not all property types were affected in a similar manner. Less cyclical property types such as Health Care and Residential experienced less of a downfall, which could result in a stronger momentum profit compared to the other REIT types in our sample. Since the sample covers a 20 year period we expect turnover and BM to have a stronger influence than the crisis effect.

(19)

14

3. Data

Before we start analysing and describing the used data it is important to understand that there have been some structural changes on the REIT market. Table A1 shows that in the 1992-1994 period both the number of Equity REITs and their market capitalization increased substantially due to an initial public offering (IPO) boom. These IPO’s were driven by real estate entities going public in order to restructure their balance sheet and the introduction of the Umbrella Partnership REIT (UPREIT) structure (Parker, 2012). In order to prevent these structural changes from influencing our results, the sample consist of REITs that are traded on the NYSE, AMEX and Nasdaq for the January 1994 - December 2014 period. This still allows for 20 years of data to be investigated, providing a sufficient sample size.

First, monthly price data on the following variables is collected from the CRSP/Ziman

REIT database: ticker, REIT type, property type, share code, market capitalization, shares

outstanding (both recorded in thousands) and total return. The share code variable is used to drop all REITs in our sample issuing shares other than ordinary common shares, indicated by share code 18. This is done for two reasons, first it makes our results comparable with earlier findings of momentum on common stocks (Chui et al., 2003a). Second, as can be seen in Table A2 the majority of observations have share code 18 or 48. REITs with share code 48 issue shares of beneficial interest. Since this type of share is often infrequently traded, it could give distorted results and is therefore not included for further analysis.

Table A2 further shows the different REIT and property types present in the dataset. For our main analysis Equity and Mortgage REITs are studied in more detail by creating subsamples and dummy variables. Since Hybrid REITs do not consist of one particular asset class it would be difficult to explain any potential differences found in the momentum effect. Therefore they are not investigated as a separate subsample. For the property type we look further into: Healthcare, Office/Industrial, Lodging/Resort, Residential and Retail. REITs with property type Unknown, Unclassified and Diversified are not included for the same reason as the Hybrid REITs. Mortgage, MBS and Self-Storage REITs are not analysed due to limited data availability in Compustat. Fortunately, our five property types of interest cover the majority of property type data with 74% of total observations, matching the results by Feng et al. (2014).

To determine the book value of equity, data on total assets and total liabilities, (both recorded in millions), is collected from the COMPUSTAT North America Fundamentals Annual

(20)

15

File. Finally, the monthly risk free rate in the form of one month treasury bills is collected from

the Kenneth R. French Data Library.

3.1 Data processing

Raw data files are merged and further adjusted in a number of ways to make them suitable for the planned analysis and comparable with the results found in earlier literature. Each of these steps means the sample size changes, an overview of these adjustments can be found in Table A3. The Ziman, CRSP, COMPUSTAT and Kenneth R. French datasets are merged together. In the process, REITs with missing book value data are dropped. We only include REITs that have at least two years of data in COMPUSTAT, since we need two years of data to calculate the book-to-market (BM) ratio following Fama and French (1992). Next REITs with negative BM ratios are excluded4_{. Following Atkins and Dyl (1997) the reported trading volume of REITs}

trading on the NASDAQ is reduced by 50% to make them comparable with REITs trading on the NYSE and AMEX, the so called dealer effect. Finally, manual inspection of the dataset showed a number of merging errors in the data. As a result eight more REITs were deleted and for two REITs the ticker was adjusted in order to properly merge all datasets5.

3.2 Variable construction

This section describes how the basic variables used in the cross sectional regression are created. Firm size (Size) of a REIT in month t is set equal to its market capitalization at the end that month. The BM ratio for an individual REIT is calculated by dividing the book value in December of year τ-1 by the market capitalization in December of year τ-1. Next, the turnover ratio (Turn) of a REIT is calculated by dividing the shares traded in a month by the total number of shares outstanding in that same month. Lastly, excess return (risk-adjusted return) is calculated by subtracting the one month T-bill rate, used as a proxy for the risk free rate from the individual REIT returns. To avoid the influence of extreme values we check for outliers using distribution plots for the main variables: BM, Size, Turn and excess return. After inspecting the data it becomes clear that BM has a number of unusually high values that do not coincide with annual reports. Therefore, the BM ratio is winsorized and ratios higher than the 0.9975 fractile or lower than the 0.0025 fractile are replaced with those boundary values respectively. Next, the Turn variable shows outliers but is only adjusted for values higher than

4_{We assume investors cannot face a negative value due to the company’s limited liability structure}

5_{The following tickers were removed: "APO","CFR","CLR","COP","GSL", "IAC", "MPG", "MT", "PAG", "RA", "SSI",}

(21)

16

the 0.9975 fractile6_{. Finally, excess return and Size show some extreme values. However, these}

do match financial reports data and are therefore left in the dataset.

3.3 Descriptive Statistics

Table 1 reports the descriptive statistics for the entire sample of 183 REITs for the 1994-2014 period as well as the subsamples based on REIT- and property type. In Panel A the mean Size is equal to roughly 2.1 billion while the median value is around 0.9 billion, indicating that the variable is skewed. The mean BM and turnover ratio are 89.5% and 1.43% respectively. The mean excess return for the total REIT sample seems to be normally distributed based on the mean of 0.95% and the median of 1.1%. These findings are fairly in line with results found in earlier literature such as Chui et al. (2003a). One notable difference is that average market capitalization is higher in our results, which can be explained by the fact that we study the sample period 1994-2014 compared to the sample period 1984-2000 by Chui et al. (2003a). Table A1 shows that market capitalization increased a lot after the year 2000, partially explaining the differences.

Comparing the mean and median values together with visual inspection of the data shows that the variables Size, BM and Turnover are skewed. We control for this skewness by taking the natural logarithm which makes the variables normally distributed as shown in Figure A1 (Veenman, 2011). We briefly analyse the results for our subsamples based on property type and REIT type. Panel B and C show there are roughly 3.5 times more Equity REITs than Mortgage REITs. The Equity REITs have an average Size and excess return of 2.5 billion and 0.97% respectively which is larger than that of the Mortgage REITs. On the other hand, the average BM value of 1.18 and turnover ratio of 2.38% for the mortgage REITs are larger than those of the Equity REITs. Panels D-H show the sample is broadly distributed across the five property types. The turnover ratios are the highest for Industrial/Office (1.57%) and for Lodging/Resort(1.64%). These two property types also have the highest BM ratios. Residential REITs have a low excess return of 0.86% while their turnover ratio of 1.36% is above average. With an average market capitalization of 2.5 billion for Healthcare REITs and 2.6 billion for Retail REITs these two property types are relatively large.

(22)

17

Table 1 ■ Summary Statistics

#REITs # obs Mean SD Min p25 p50 p75 Max

Panel A: Total Sample

Excess Return 183 18,743 0.0095 0.081 -0.727 -0.029 0.011 0.050 2.903

Size 183 18,866 2066 3986 2.4 345 884 2099 56600

BM 183 17,474 0.895 0.769 0.005 0.532 0.762 1.048 12.030

Turn 183 18,866 1.435 1.488 0.000 0.538 1.016 1.831 10.520

Panel B: Equity REITs

Excess Return 141 15,640 0.0097 0.080 -0.727 -0.028 0.011 0.050 2.363

Size 141 15,725 2150 4123 2.4 385 973 2227 56600

BM 141 14,616 0.850 0.765 0.005 0.494 0.711 0.996 12.030

Turn 141 15,725 1.313 1.409 0.000 0.505 0.934 1.661 10.520

Panel C: Mortgage REITs

Excess Return 39 2,337 0.0083 0.087 -0.674 -0.031 0.010 0.050 2.903 Size 39 2,374 1437 2510 6.5 210 542 1547 17600 BM 39 2,138 1.183 0.730 0.083 0.895 1.045 1.230 12.030 Turn 39 2,374 2.382 1.779 0.040 1.180 1.961 3.126 10.520 Panel D: Healthcare Excess Return 12 2,171 0.0107 0.081 -0.495 -0.029 0.011 0.053 0.550 Size 12 2,174 2469 4275 16.6 504 940 2109 24800 BM 12 2,040 0.8102 0.718 0.029 0.505 0.637 0.829 6.183 Turn 12 2,174 1.187 1.045 0.092 0.502 0.870 1.508 10.520 Panel E: Industrial/Office Excess Return 33 3,304 0.0102 0.081 -0.657 -0.029 0.011 0.052 0.802 Size 33 3,330 2390 3222 7.2 567 1439 2730 21500 BM 33 3,120 0.920 0.815 0.032 0.616 0.820 1.050 12.030 Turn 33 3,330 1.573 1.486 0.012 0.681 1.159 2.007 10.520 Panel F: Lodging/Resorts Excess Return 12 746 0.0106 0.107 -0.601 -0.039 0.012 0.063 1.011 Size 12 753 910 841 42.5 307 694 1184 5264 BM 12 681 1.371 1.595 0.005 0.762 1.080 1.502 12.030 Turn 12 753 1.642 1.227 0.150 0.777 1.334 2.089 9.332 Panel G: Residential Excess Return 21 2,494 0.0086 0.067 -0.532 -0.026 0.009 0.046 0.404 Size 21 2,507 1273 1403 29.5 361 729 1700 13200 BM 21 2,327 0.740 0.381 0.006 0.480 0.687 0.920 3.137 Turn 21 2,507 1.362 1.382 0.139 0.493 0.932 1.712 10.520 Panel H: Retail Excess Return 37 5,007 0.0092 0.077 -0.418 -0.026 0.012 0.048 2.364 Size 37 5,029 2605 5522 4.9 360 958 2588 56600 BM 37 4,692 0.758 0.599 0.008 0.454 0.665 0.926 6.540 Turn 37 5,029 1.138 1.346 0.012 0.425 0.763 1.330 10.520 Notes: The sample includes all REITs listed on the NYSE/AMEX/Nasdaq between January 1994 and December 2014. This table shows summary statistics for the entire sample and subsamples based on REIT type and property type. Subsamples for REIT type are created based on the rtype variable and subsamples for Property type using the ptype variable, both found in the CRSP/Ziman REIT database. The table provides the average number of REITs, average number of monthly observations, mean, standard deviation, minimum, 25th percentile, median, 75th percentile and maximum. Excess return is the average monthly raw return minus the risk-free rate, in decimal format. Size is measured as a firm's market average capitalization, in millions of dollars. Average book to market ratio (BM) is the book value in December of year τ-1 divided by the market capitalization in December of year τ-1, in percent. The average turnover ratio (Turn) is the number of shares traded divided by the number of shares outstanding, in percent.

(23)

18

4. Method

In this section, the methodology by which the hypotheses are tested is described. First, Section 4.1 looks if past return is a predictor of future REIT return using a Fama MacBeth (1973) cross-sectional regression. Section 4.2 investigates whether we can use momentum in the form of a profitable trading strategy by creating momentum portfolios. Moskowitz and Grinblatt (1999) mention that using a combination of both a cross-sectional regression and portfolio analysis acts as a robustness check on itself. The use of individual REITs in Section 4.1 avoids the possibility that test results are sensitive to the portfolio grouping procedure in Section 4.2.

4.1 FM cross-sectional regressions

This thesis mainly follows Chui et al. (2003a) regarding the selection and construction of the dependent and independent variables although some adjustments are made. Most of the variables used in the cross-sectional regression are defined in Section 3.2. This section describes how the variables are further adjusted to make them suitable for usage in equation (1). The momentum effect is investigated using a cross-sectional regression. We want to determine if factors known to influence common stock return such as book-to-market, size, past return and liquidity are able to predict REIT returns in our sample. Chui et al. (2003a.) investigate this relationship using a single sample consisting of all REITs (Equity, Mortgage and Hybrid) and find only the past return variable to be significant in the post-1990 period. To test the hypotheses from Section 2 we perform the following cross-sectional regression:

𝑅_{𝑖,𝑡:𝑡+5} = 𝛼_0,𝑡+ 𝛽_1,𝑡𝑅_{𝑖,𝑡−2:𝑡−7}+ 𝛽_2,𝑡Ln 𝑆𝑖𝑧𝑒_{𝑖,𝑡−2}+ 𝛽_3,𝑡Ln 𝐵𝑀_{𝑖,𝑡−2} +

𝛽_4,𝑡Ln 𝑇𝑢𝑟𝑛_{𝑖,𝑡−2:𝑡−7}+ 𝜀_𝑖,𝑡, t = January 1994 to December 2014 (1)

𝑅_{𝑖,𝑡:𝑡+5} = ∑𝑡+5 (𝑟_𝑖,𝑡− 𝑟𝑓_𝑡)

𝑡 , where 𝑅𝑖,𝑡:𝑡+5 is the cumulative excess return on the REIT i from

month t till t+5, 𝑟_𝑖,𝑡 is the monthly return and 𝑟_𝑓𝑡 is the monthly risk free rate.

𝑅_{𝑖,𝑡−2:𝑡−7} = ∑𝑡−2_𝑡−7 (𝑟_𝑖,𝑡− 𝑟𝑓_𝑡), where 𝑅_{𝑖,𝑡−2:𝑡−7} is the cumulative past six-month excess return

from month t-7 till month t-2. Ln 𝑆𝑖𝑧𝑒_{𝑖,𝑡−2} is the natural logarithm of market capitalization at the end of month t-2. Ln 𝐵𝑀_{𝑖,𝑡−2} is the natural logarithm of the book-to-market ratio for each REIT at the end of month t-2.

Ln 𝑇𝑢𝑟𝑛_{𝑖,𝑡−2:𝑡−7} = [∑𝑡−2_𝑡−7 (𝑟_𝑖,𝑡− 𝑟𝑓_𝑡)]/6, where Ln 𝑇𝑢𝑟𝑛_{𝑖,𝑡−2:𝑡−7} is the log of the arithmetic mean of the turnover ratio from month t-2 till t-7.

(24)

19

As can be seen from equation (1) all variables skip month t-1, to control for the so called ask bounce effect. When momentum return is measured using contiguous months the bid-ask bounce could have a diminishing effect on momentum returns (Chan et al. 1996). We do not include analyst coverage as one of the independent variables in our model as done by Chui et al. (2003a). A large number of REITs had no data on analyst coverage, which resulted in a reduction of our sample size. Finally, previous studies have included additional control variables when performing similar regressions. Jegadeesh (1990) mentions that one-month past return, 𝑅𝑖,𝑡−2:𝑡−2 should be added to control for liquidity and microstructure effects. Moreover,

DeBondt and Thaler (1985) include long-run past return, 𝑅𝑖,𝑡−13:𝑡−36 to account for the three-

to five-year reversal effect. However, Moskowitz and Grinblatt (1999) show that cross sectional regressions using a six-month holding/formation period are unaffected by these effects. Therefore, liquidity and long term reversal effects are not included in equation (1).

4.1.2 Endogeneity issues

As mentioned, equation (1) is estimated using a Fama Macbeth (FM)7 procedure. The idea behind this procedure is to run a predictive regression were a dependent variable (monthly REIT

return) is regressed on a number of lagged independent variables (past return ,BM, Size, Turn).

A significant beta coefficient tells us whether the independent variables are able to predict REIT return. We use this particular procedure since error terms of asset pricing data are likely to be cross-sectional correlated at a given time Cov(εit, εkt)  0 (Petersen, 2009). If return for one

REIT is unusually high in a certain month, this is also likely to be the case for another REIT, since they are exposed to similar outside factors. If this correlation is not corrected for, our estimates are still consistent, however the reported standard errors are too small.

By using the FM procedure we calculate the standard error while correcting for cross-sectional correlation. One advantage is the ability to estimate marginal effects in a multivariate regression. One potential problem is that estimates can be dominated by extreme observations. This has been partially solved by handling outliers accordingly as described in Section 4.2. While accounting for cross sectional correlation the procedure assumes there is no time series correlation present Cov(εit, εit-k) = 0 (Petersen, 2009). According to Cochrane (2009) the

assumption that error terms are not correlated over time is justified on many occasions, since return data is almost always independently distributed.

However, since in our analysis the dependent variable uses overlapping returns by

(25)

20

looking at returns from month t till t+5, serial correlation is introduced. While some papers like Chui et al. (2003a) correct for this autocorrelation, others do not. To compare the outcomes with both strands of literature, results are presented with and without a Newey-West correction. Since Chui et al. (2003a) do not explicitly mention the number of lags they correct for and no data-dependent method to pick the number of lags is available in Stata, we use the Stock and Watson (2007) guideline. Based on this formula8 and the suggestions by Lewellen (2014) we correct up to four lags to account for possible autocorrelation in the slopes.

4.1.3 Implementing the FM procedure

Instead of estimating a single cross-sectional regression with sample averages, the FM procedure runs a cross-sectional regression for each month between January 1994 and December 2014:

𝑅_{𝑖,𝑡:𝑡+5} = 𝛼_0,𝑡+ 𝛽_1,𝑡𝑅_{𝑖,𝑡−2:𝑡−7}+ 𝛽_2,𝑡Ln 𝑆𝑖𝑧𝑒_{𝑖,𝑡−2}+ 𝛽_3,𝑡Ln 𝐵𝑀_{𝑖,𝑡−2} +

𝛽_4,𝑡Ln 𝑇𝑢𝑟𝑛_{𝑖,𝑡−2:𝑡−7}+ 𝜀_𝑖,𝑡, t = 1,2,…,233 (2) The FM estimators are calculated as time-series averages of the monthly cross-sectional parameter estimates, which is done as follows:

𝛽̂ = _𝐹𝑀 1

𝑇∑ 𝛽̂𝑡

𝑇

𝑡=1 (3)

Next, we use the time-series standard deviation of 𝛽̂ to estimate the standard error of 𝛽_𝑡 ̂ : _𝐹𝑀 𝜎̂(𝛽̂ ) = √𝐹𝑀 1 𝑇∗(𝑇−1)∑ (𝛽̂ − 𝛽𝑡 ̂ )𝐹𝑀 2 𝑇 𝑡=1 (4)

Fama and MacBeth (1973) use rolling five year regressions to estimate the beta coefficient. We implement a simpler technique used by many other papers, in the form of full-sample betas.

4.1.4 Subsample analysis

The creation of subsamples based on property- and REIT type is done in two different ways. The first option is to construct subsamples by dividing the entire sample. Equation (1) is regressed on this newly created sample to analyse the effect for this particular group. A drawback of this method is that it reduces the sample size, which means a small group of outliers can influence individual results. The descriptive statistics in Section 3.3 show the sample size

(26)

21

for Lodging/Resorts could give potential problems when performing these regressions.

For the second method we create indicator variables set equal to one when a REIT is classified as a certain property type: Healthcare, Industrial/Office, Lodging/Resort, Residential,

and Retail or REIT type: Equity, Mortgage. Using these dummy variables serves as a control

for the subsample approach since the sample remains intact. Observations of REIT- and property types that are not analysed in detail are removed to avoid over specification of the model. Apart from the dummy variables themselves we also create interactions between each dummy and the momentum variable. These interaction variables allow us to test if the differences in momentum for the REIT- and property types are statistically significant. To avoid perfect multicollinearity, n-1 dummy variables are included. The full regression equations including the interaction terms and other firm characteristics are given in the Appendix.

4.2 Portfolio Analysis

In this section we present the methodology needed to determine whether a trading strategy based on momentum actually generates significant momentum profits. Similar to Section 4.1 we investigate momentum strategy for the entire sample as well as for the REIT type and property type subsamples.

4.2.1 Momentum portfolio construction

As done by Hung and Glascock (2008), two different methods are used to assign REITs to the winner and loser portfolio: (1) A value-weighted strategy following Chui et al. (2003a) and (2) an equally-weighted strategy following the technique by Jegadeesh & Titman (1993).

(1) Value-weighted momentum portfolios

In order to create the momentum portfolios, all REITs are ranked at the end of month t based on their cumulative return, of the preceding J-months. For our main analysis we look at the past six-month cumulative return, so J takes on the value of six. In general, past J-month cumulative return is calculated as follows:

𝑅_𝑖,𝑗 = ∑ 𝑟_𝑖,𝑡

𝑡−1

𝑡−𝑗

(5)

Here 𝑟_𝑖,𝑡 is the return for REIT i in month t. We use a 30% breakpoint which means based on six-month cumulative return, the top (bottom) 30% performing REITs is assigned to the winner (loser) portfolio. The majority of literature on momentum strategies for the stock market,

(27)

22

including Jegadeesh and Titman (1993) uses decile portfolios. We have chosen to create portfolios based on a 30% breakpoint due to the smaller size of the REIT sample compared to the stock market (Hung & Glascock, 2008).

Value-weighted9 portfolios are created using the market capitalization of the individual REITs at the end of month t as the weight. Similar to the use of lagged regressors in the cross- sectional regression, cumulative returns on the winner/loser portfolios are measured one month after the ranking takes place. After skipping one month each of these portfolios is held for K-months. Our main analysis looks at a six-month holding period, so K takes on the value of six. General post formation returns are calculated as follows:

𝑅_{𝑝,𝑡,𝑘} = ∑ 𝑟_𝑖,𝑡

𝑡+𝑘

𝑡 (6)

Here 𝑅𝑝,𝑡,𝑘 is the cumulative monthly return for portfolio p in month t with a holding period of

k months. Chui et al. (2003a) replace missing return values for the individual REITs with

value-weighted market returns. Although the dataset shows few missing return values we do not follow this approach. The problem with mean substitution is that it creates average data points which can change the value of the underlying variance, potentially causing biased standard errors (Graham, 2009). Finally, monthly zero-cost momentum portfolios are created by taking a long position in the winner portfolio and shorting the loser portfolio, creating a so called winner minus loser (W-L) portfolio. These momentum portfolios are said to be zero-cost since the winners are purchased using the proceeds from short selling the losers.

Overlapping momentum portfolios are created by taking the average six-month return of the six individual momentum portfolios. For example, the momentum return on December 1994 is the average monthly return of the six momentum portfolios formed on May, June, July, August, September and October 1994 (Hung and Glascock, 2008). This procedure increases the number of observations, which is further illustrated in Figure A3. To calculate the monthly momentum profits we use equation (6) only now on the momentum portfolio. Equation (7) presents K-month cumulative return, so monthly momentum profits are calculated as the simple mean, dividing cumulative return by K:

𝑅_𝑡,𝑘 = [∑ 𝑟_𝑖,𝑡 𝑡+𝑘 𝑡 ] /𝐾 (7)

(28)

23

Here 𝑅_𝑡,𝑘 is the monthly momentum return for a holding period of K months. (2) Equally-weighted momentum portfolios

Instead of value-weighting the portfolios based on market capitalization, Jegasdeesh and Titman (1993) use equally-weighted portfolios. Momentum returns based on equally-weighted portfolios are included to make the results comparable with other previous literature as well. With the exception of the procedure used to assign REITs to the winner/loser portfolios all steps are similar to the value-weighted technique.

4.2.2 Subsample analysis

Using the methodology described above, industry-neutral momentum portfolios are formed for each of the different REIT- and property types following Moskowitz and Grinblatt (1999). An industry-neutral portfolio buys the winners and shorts the losers within each industry. Since these winners and losers are from the same industry, returns from these industry-neutral portfolios reflect intra-industry momentum. Moskowitz and Grinblatt (1999) measure returns on the industry-neutral portfolio directly after portfolio formation, producing insignificant momentum profits. As discussed in Section 2, momentum profits for industry- neutral portfolios are only significant if the formation and holding period are not contiguous. Therefore we follow Grundy and Martin (2001) and skip one month between the holding and formation period.

(29)

24

5. Results

This section presents the results of the main empirical analysis performed in this study. First, the outcomes for the cross-sectional regressions of firm characteristics on REIT return are discussed. Subsequently, the profitability of the various momentum strategies is given.

5.1 Results FM cross-sectional regression

Table 2 presents the results of the FM cross-sectional regressions. By performing these regressions we determine whether known predictors of expected return are significant for our REIT sample. To extend on current literature and test our hypotheses, past six-month

cumulative return, BM, Size and Turnover are regressed on six-month cumulative return.

Coefficients, t-statistics, number of observations and adjusted R2’s are reported for all regressions. Slope coefficients are presented as the time-series averages of monthly cross-sectional regressions between 1994 and 2014. Further, all independent variables skip month t-1 to control for the bid-ask bounce. T-statistics are given with and without a lag correction.

The methodology for the cross-sectional regression presented here is based on the work of Chui et al. (2003a). Using a REIT sample consisting of Equity, Mortgage and Hybrid REITs, they find momentum to be the only significant predictor of REIT return in the post-1990 period. As discussed in Section 2, lower momentum profits reported by recent papers and the influence of the financial crisis could have resulted in a less dominant and potentially insignificant momentum effect for our sample period. The results in this section enable us to test the first hypothesis which states the momentum effect for REITs is significant in the 1994-2014 period.

Column (1) of Table 2 reports the cross-sectional regressions for the entire sample consisting of 16,428 observations. Results show a dominant momentum effect with only past six-month cumulative return having a significant influence on future return. This suggests that only momentum is priced in as one of the cross-sectional determinants. These findings confirm earlier results by Chui et al. (2003a) for the post-1990 period, who also find size, BM and

turnover to be non-significant. The coefficient of 0.076 on momentum in column (1) implies

that if past six-month return goes up by 1%, future six-month return increases with 0.076%. The size of the momentum coefficient in our sample is somewhat smaller compared to Chui et al. (2003a), who find a value of 0.126 for past six-month return in the post-1990 period. This result is in line with Hung and Glascock (2008) and Goebel el al. (2013) who also find lower momentum effects compared to earlier literature.

(30)

25

Table 2 ■ Cross-sectional regressions of firm characteristics on expected REIT returns

(1) (2) (3) (4) (5) (6) (7) (8)

Returnt,t+5 Entire Sample Equity Mortgage Healthcare

Indust/ Office

Lodging/

Resort Residential Retail

Panel A: no lag correction

Return t-2:t,-7 0.076*** 0.074*** 0.119* -0.086 0.083** -0.019* 0.179*** 0.027 (4.25) (3.67) (1.95) (-1.29) (2.38) (-1.94) (3.75) (0.82) Ln BM t-2 0.005 0.005* -0.074 0.017 0.011 0.077*** 0.013 -0.001 (1.64) (1.82) (-1.08) (1.16) (1.56) (2.83) (1.02) (-0.34) Ln SZ t-2 -0.004* -0.002 -0.006 0.013 -0.005* -0.006 -0.035*** 0.004 (-1.91) (-0.83) (-0.33) (1.24) (-1.68) (-0.22) (-3.10) (1.61) Ln TN t-2,t-7 -0.001 -0.002 0.017 -0.089*** 0.008 0.031 0.032 -0.003 (0.11) (-0.47) (0.44) (-2.69) (0.90) (0.88) (0.94) (-0.69) Intercept 0.059*** 0.047** 0.350** -0.143 0.113*** 0.054 0.460*** -0.001 (3.10) (2.41) (2.39) (-1.02) (2.98) (0.16) (3.54) (-0.04)

Panel B: 4 lag Newey-west Correction

Return t-2:t,-7 0.076** 0.074** 0.119 -0.086 0.083 -0.019 0.179** 0.027 (2.41) (2.06) (1.38) (-1.29) (1.52) (-1.54) (2.57) (0.45) Ln BM t-2 0.005 0.005 -0.074 0.017 0.011 0.077* 0.013 -0.001 (0.93) (1.03) (-0.73) (0.81) (0.93) (1.70) (0.74) (-0.20) Ln SZ t-2 -0.004 -0.002 -0.006 0.013 -0.005 -0.006 -0.035* 0.004 (-1.05) (-0.48) (-0.27) (0.86) (-1.10) (-0.16) (-1.80) (0.90) Ln TN t-2,t-7 -0.001 -0.002 0.017 -0.089* 0.008 0.031 0.032 -0.003 (0.06) (-0.26) (0.34) (-1.80) (0.54) (0.65) (0.80) (-0.44) Intercept 0.059* 0.047 0.350* -0.143 0.113* 0.054 0.460** -0.001 (1.79) (1.42) (1.65) (-0.70) (1.89) (0.12) (2.06) (-0.02) Observations 16,428 13,840 1,892 1,942 3,095 614 2,181 4,426 R-squared 0.182 0.194 0.725 0.645 0.433 0.989 0.584 0.366 # time periods 233 233 195 233 233 197 233 233 t-statistics in parentheses *** p<0.01, ** p<0.05, * p<0.1

Notes: This table reports the average coefficients from the FM cross-sectional regressions when monthly excess return is regressed on lagged firm characteristics. The sample includes all listed US REITs with share code 18 from January 1994 to December 2014. 𝑅𝑡:𝑡+5 and 𝑅𝑡−2:𝑡−7 are the cumulative excess return from months t to t+5 and months t-7 to t-2

respectively. Ln 𝐵𝑀𝑡−2 is the natural logarithm of the book-to-market ratio at the end of month t-2, computed as the book

value in December of year τ-1 divided by the market capitalization of year τ-1. Ln 𝑆𝑖𝑧𝑒𝑡−2 is the natural logarithm of the

market capitalization at the end of month t-2. Ln 𝑇𝑢𝑟𝑛𝑡−2:𝑡−7 is the natural logarithm of the average turnover ratio from

month t-7 to t-2, computed as the number of shares traded divided by the number of shares outstanding. Column (1) shows the results for the entire sample, while columns (2) and (3) give the outcomes for the REIT type subsamples and columns (4)-(8) give the results for the property type subsamples. Panel B reports the same results as panel A only now t-statistics incorporate a Newey-West correction up to four lags.

The most intuitive explanation for the lower momentum coefficients is the crisis period included in the samples here (Jegadeesh & Titman, 2011). Another explanation is the decreasing strength of anomalies like momentum due to increased market efficiency (Chordia et al., 2014). Based on the significance of the momentum coefficient in column (1) of Table 2 ,we find confirming evidence for our first hypothesis that US REITs experienced a significant momentum effect between 1994 and 2014. As discussed, the momentum effect for REITs has

The momentum effect for equity and mortgage REITs in the US, a comparative study

Master Thesis