Seasonal anomalies in REIT returns : the Halloween effect and the January effect

(1)

Seasonal anomalies in REIT returns:

The Halloween effect and the January effect

Name: Sven Pijlman

Student number: 10650563

MSc: Real Estate Finance & Corporate Finance Supervisor: Martijn Dröes

(2)

2 Abstract:

In this paper, calendar anomalies in the US property share market are examined. The two anomalies of interest are the Halloween effect and the January effect. The Halloween effect says that the return in the stock market is higher during the winter months from November-April than the return during the summer months from May-October. The other anomaly, the January effect, says that the return in the stock market is in January higher than in every other month. In this paper, the anomalies are examined at the various REIT- and property type level. Furthermore, the development of the Halloween effect over time is examined. The research is performed on monthly data from the CRSP/Ziman REIT database for the period 1980-2017. Various regression models are used to perform this research. The results show that the January effect is absent in all the REIT- and property type markets. On the other hand, the Halloween effect seems to be present in the mortgage REIT market, but not in the equity REIT market. Furthermore, the effect seems to have weakened over time for the equity REITs, while the effect became more powerful in the mortgage REIT market. The presence of the Halloween effect in a part of the REIT market in the US, should be something for the investor to consider. However, it seems that a trading strategy based on the Halloween effect does not outperform the market.

Keywords: Real estate stocks, Calendar anomalies, January effect, Halloween effect, REIT types

Statement of originality

This document is written by Sven Pijlman 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)

3

1. Introduction

Similarities between the stock market and the property share market have been intriguing both financial professionals and academics for many years. The stock market and the Real Estate Investment Trust (REIT) market became less similar over the last few years, as argued by Chiang and Lee (2010). They implicated that REITs provide investors with real estate exposure in the long run. Another topic that intrigued financial professionals and academics are price anomalies. Therefore, it is interesting to see if price anomalies in the REIT market will behave the same way as in the stock market. In the past, various studies have proven the existence of seasonal price patterns in the stock market. The anomalies of interest for this study are the Halloween effect and the January effect. The Halloween effect says that the return during the winter months are higher than the returns during the summer months (Jacobsen & Bouman, 2002). The winter months are defined as November-April and the summer months as May-October. Possible explanations are mentioned in the existing literature. For example, the timing of the summer vacation (Bouman & Jacobsen, 2002), Seasonal Affective Disorder (Kamstra et al., 2003) and the optimism-cycle (Doeswijk, 2008). The January effect says that the returns during January are on average higher than the returns during the remaining months of the year (Kinney & Rozeff, 1976). Possible explanations for the January effect are, the turn of the tax year (Starks et al., 2006), the presence of inside traders (Seyhun, 1988) and cash receipts at the turn of the year (Ogden, 1990).

After they have been discovered in the stock market, follow-up studies, by Ben-Hamo and Brounen (2009) and Chan and Hui (2015) for example, tested if these anomalies also hold for different real estate shares. The results were ambiguous. Hence, seasonal anomalies in the REIT market still remain a puzzle. If these anomalies really exist in the financial world, this means that investors and portfolio managers could act on these anomalies and use them to secure higher profits. However, when these anomalies are discovered and reported, and investors are starting to act on this knowledge, the effect will disappear (Schwert, 2003). Therefore, this paper will test the calendar anomalies over time. Based on the paper by Schwert (2003), the expectation is that calendar price anomalies will weaken, or disappear over time.

Whether calendar anomalies behave different across various industries has been tested in the stock market, but, according to my knowledge, never at different industry level

(5)

5

for the property share market. This paper will examine the Halloween effect and the January effect at the REIT type and REIT property type level. With the additional knowledge of the existence of seasonal anomalies at different industry levels, investors will be able to predict the market even better and, if the effect is also economical significant it can be used to develop profitable trading strategies based on these price patterns. According to Liang, McIntosh and Tompkins (1991), different types of REITs, show different seasonal price behaviors. Therefore, there is expected to observe differences across the various REIT types.

The presence of the Halloween effect and the January effect will be tested with monthly data from the CRSP/Ziman REIT database from 1980-2017. This database provides information on all REITs traded on the NYSE, the NYSE MKT and the NASDAQ stock exchanges. In this study, linear regression models are used to test for the effect in different types of REITs and properties. This method is based on the paper by Jacobsen and Visaltanachoti (2009) where they did research on the Halloween effect at the industry level in the stock market. To test for the presence of the calendar anomalies, various regression models will be tested using ordinary least square (OLS) regressions. The research starts with a simple model, but additional explanatory variables are included in the process. Eventually, time-period dummy variables are created and included in the regression to examine the development of the Halloween effect over time.

The results show no evidence for higher returns during January compared to the remaining months of the year. So, no evidence is found for the presence of the January effect in the REIT market. This conclusion holds for the entire REIT market, as well as for the various REIT- and property types individually.

The research for the Halloween effect show opposite results. In the sample examined in this study, little evidence is found for the presence of the Halloween effect in the entire sample. Furthermore, the effect tends to be present in the mortgage REIT market and absent in the equity REIT market. For the Health Care- and Retail property type REITs, an opposite effect of the Halloween effect is found. For these REITs the return is significantly lower during the winter months compared to the summer months.

Over time, there is expected that the Halloween effect weakened or vanished away. However, this study shows contradicting results regarding the Halloween effect over time. The results show that for the entire market the effect has weakened over time. The same results are found for the equity REITs and the Health Care, Residential and Retail property

(6)

6

types. The mortgage REITs follow an opposite pattern. In the first period, the returns are significantly lower during the winter months than during the summer months, but over time the Halloween effect entered the mortgage REIT market and became more powerful over the years.

In the end, a REIT portfolio managed based on the knowledge of the Halloween effect is compared to a market portfolio. The results show that such a portfolio tends to earn on average higher returns than the market. However, the higher monthly returns are a result of the higher risk for this strategy than for investing in the market portfolio.

The remainder of this study is organized into four main sections. In Section 2, the existing literature on the Halloween effect, the January effect and the REIT market will be discussed. In Section 3, the data, including descriptive statistics, and the methods used to test for the presence of the Halloween effect and the January effect are clarified. Thereafter, in Section 4, the results of this research are presented. In the fifth and final section, the conclusion is drawn and implications following this study are given, followed by discussion of the research limitations and suggestions for future research.

2. Literature review

In this section the existing literature regarding the Halloween effect and the January effect will be discussed. First the general Halloween effect will be reviewed, followed by a closer look on the relationship between the Halloween effect and the REIT market. Thirdly, the January effect in general will be discussed. This section ends with a review of the literature about the presence of the January effect in the REIT market.

2.1 The Halloween effect

The first written reference to the Halloween effect was observed in the New York Times on 30 May 1964. The exact quote in this edition of the paper was: The Stock Exchange world is in a sort of twilight state at the moment. The potential buyers seem to have “sold in May and gone away”. In the European press a similar saying was addressed: “Sell in May and go away.” This saying implies that the month May indicates the start of a bear market and that investors therefore are better off by selling their stocks and instead hold their cash (Bouman and Jacobsen, 2002). There are two additions to this saying mentioned in the European press, both stating that the investor should go back to the stock market in September.

(7)

7

Related to this market timing, a similar strategy was reported by Downes and O’Higgins in 1990. This strategy requires an investor to be in the stock market from the 31st_{of October}

until the 30th_{of April and out of the market during the remaining half of the year, therefore}

it is referred to as the Halloween effect. However, Downes and O’Higgins (1990) did not provide a statistical analysis of their results.

The market saying ‘Sell in May and go away’ is often cited in the press, however in the academic literature it was not frequently mentioned. Lewis (1985) mentioned the market saying, but never examined the existence of the Halloween effect. The first extensive academic research on the Halloween effect was by Bouman and Jacobsen in 2002. According to their paper, it is considerable that the Halloween effect is economically significant because a trading strategy based on this anomaly beats a simple buy and hold strategy in most of the countries in their research, making the Halloween effect interesting for practitioners. This is mainly the case because benefits can be obtained by trading only twice a year and therefore the returns will not be affected too much by the transaction costs (Bouman & Jacobsen, 2002). This is in contrast with most other anomalies like the ‘Monday effect’ or ‘the Turn of the Month effect’. Those are difficult to exploit, and it is hard to actually make a profit with these anomalies, because of the high transaction costs associated with strategies based on these anomalies.

A study on a Halloween effect based trading strategy is performed by Haggard and White (2010). The trading rule they examined was: ‘sell stocks in early May, invest in T-bills, and re-invest in stocks on Halloween.’ The results of this study showed that investing on basis of this rule provides an investor with risk-adjusted returns that exceeds the buy and hold equity returns, even after transaction costs are considered. Besides, Haggard and white (2010) concluded that the Halloween effect is robust to the January effect and the consideration of outliers.

Data snooping is often suggested to be an explanation for a calendar anomaly (Sullivan et al., 2001). Data snooping implies that the anomalies are spurious and purely data-driven results, as might be the case for the Monday effect and January effect according to Sullivan et al. (2001). In their paper, a bootstrap procedure was used that explicitly measures the distortions induced by data snooping. However, the data snooping argument does not apply to the Halloween effect because it is based on an addressed market saying

(8)

8

and not just another effect taken out from the range of calendar rules. The number of calendar rules that are based on market sayings seems limited (Bouman & Jacobsen, 2002).

The first to confirm the existence of the Halloween effect were Bouman and Jacobsen in 2002. They concluded that the returns on stocks in 36 out of the 37 countries in their sample, including the US, are higher in the period from November till April than the returns in the months May until October. An annual risk adjusted outperformance between the 1.5% and 8.9%, depending on the country considered, was found for a strategy based on the Halloween effect after transactions costs are taken into account. In this paper, the authors were able to rule out the usual explanations that could cause the Halloween effect, like data mining and risk explanations. In addition, Bouman and Jacobsen (2002) showed that the returns are robust to the January effect and that changes in the interest rate between the summer and winter are not a plausible explanation for the existence of the Halloween effect. Furthermore, some less likely possible explanations were tested in this paper like shifts in interest rates or volumes. None of them could explain why the Halloween effect was present. A possible explanation could be the timing of summer vacations. Bouman and Jacobsen (2002) found a significant relationship between three different proxies for the length and timing of summer vacations and the impact of vacation on trading activity and the Halloween effect.

A similar pattern to the Halloween effect in stock returns is reported by Kamstra et al. (2003). In their paper they examined whether days with less hours of daylight influence the stock market returns. They related this theory to SAD, which stands for Seasonal Affective Disorder where investors will be depressed during days with decreasing hours of daylight. Depression leads to higher risk aversion, based on experimental psychological research (Kamstra et al., 2003). The paper argues that returns in the stock market should become lower during the fall when days are shorter and then become relatively higher when the days will start to lengthen again during the winter months.

Another psychological factor that is related to the Halloween effect are the temperature changes during the year. In a study by Cao and Wei in 2005, they refer to impact of extreme temperatures on human behavior. High temperatures can lead to apathy and aggression, depending on which of these moods is superior, the stock returns will be higher or lower. If aggression is superior this will encourage risk taking and therefore lead to higher stock returns. On the other hand, when apathy dominates, risk avoidance will be

(9)

9

encouraged, and stock returns will be lower. Lower temperatures are associated with more aggressive risk taking and therefore with higher stock market returns (Cao & Wei, 2005). In their study, Cao and Wei (2005) found a significant negative relationship between stock returns and temperature, indicating a difference in the return during summer and winter.

In a paper by Hirshleifer and Shumway (2003) the relationship between daily market index returns and morning sunshine in the city of a country’s leading stock exchange is examined. From a psychologically perspective sunshine affects mood, when the sun shines people will be in a better mood than when it is cloudy (Hirshleifer & Shumway, 2003). Someone in a good mood, because of the sunshine, may attribute this feeling unconsciously to favorable life prospects. Based on their research, Hirshleifer and Shumway (2003) concluded that there is a strong significant correlation between sunshine and stock returns. However, it is not possible to make a profitable strategy based on these findings, because this will go hand in hand with a lot of transactions and therefore high transactions costs (Hirshleifer & Shumway, 2003). The results of this paper are in contrast with the literature discussed so far, because in the summer, with more days of sunshine, the returns must be higher. However, in the paper discussed before by Cao and Wei (2005) is established that the returns during warmer days depends on the superiority of apathy or additional risk taking. It is conceivable that the apathy dominates the effect of being in a good mood and that therefore the stock returns will be lower during the summer months.

The final psychological factor possibly explaining the existence of the Halloween effect is mentioned in a study by (Doeswijk, 2008). He argued that the seasonal difference in the stock market return could be caused by the optimism-cycle. This reflects the optimism bias that most investors are facing. At the end of the year, investors start to look at the next year, but because of the optimism bias, investors usually overstate the economic outlook. Initially, this optimism results in high returns on the stock market, but after a few months the initial optimism becomes hard to obtain and the investors will become more pessimistic resulting in a temporarily dip in the stock market (Doeswijk, 2008). According to research in this paper, it is profitable to base a trading strategy on the Halloween effect.

Hong and Yu (2009) also found evidence that the stock market returns are higher in the winter than in the summer months. They defined the summer as only three months for Northern Hemisphere countries: July, August and September. Just like Bouman and Jacobsen (2002) they linked the lower summer returns to the absence of investors due to

(10)

10

the summer vacation. Hong and Yu (2009) examined this with two measures of vacation activity, namely hotel occupancy rates and airline passengers travel. Their results confirm that the stock market returns are significantly lower in the summer months, because market participants are absent due to summer vacation. On top op of that, Hong and Yu (2009) also found that in countries where the trading activity significantly declined, the mean summer returns were lower. Indicating that apathy, like Cao and Wei (2005) discussed in their study, is actually observed during the summer and lowers the stock market returns.

In a study by Jacobsen et al. (2005) the interaction between the Halloween effect and portfolios formed on size, earnings-price ratio, cash-flow-to-price ratio, book-to-market ratio and dividend yield. In all these portfolios there is found evidence for the presence of the Halloween effect. Contrary to the January effect, the Halloween effect seems to be unrelated to the anomalous behavior of portfolios formed on these criteria. The main finding of this paper is that the Halloween effect seems to be unrelated to other well-known anomalies and is a market-wide phenomenon.

Arbitrage is normally a strong argument against the empirically proven Halloween effect. So, history and practice tell us that there is such a thing as the Halloween effect, while the stock market logic and literature tells us that the Halloween effect should not exist (Bouman & Jacobsen, 2002). The puzzle for the Halloween effect is not yet solved, so more research is needed.

Anomalies like the Halloween effect suggests either market inefficiency in the asset market or inadequacies in the underlying asset-pricing model. Most market inefficiencies tend to weaken or disappear after they are reported in published papers (Schwert, 2003). Investments vehicles based on the knowledge about anomalies were implemented by practitioners at the time that these anomalies were discovered. These practitioners aimed at taking advantage of the anomalous behavior of asset prices. Therefore, possibly causing the anomalies to disappear (Schwert, 2003). According to Schwert (2003), this happens because research findings cause the market to become more efficient.

A more recent study on the Halloween effect is done by Degenhardt and Auer in 2018. Their findings show that the Halloween effect still exists in the stock market, enhancing the robustness of the earlier studies. However, their research shows also that the effect has weakened over time, as expected by the theory of Schwert (2003). The weakening of the effect has started when it became public information and the effectiveness and

(11)

11

persuasiveness of trading strategies based on the anomaly are limited (Degenhardt & Auer, 2018).

The differences in the effect has been tested for different industries in the stock market and the conclusion was that the effect was strongly present in the production sectors, but that the effect almost lacked in the consumer consumption related sectors (Jacobsen & Visaltanachoti, 2009). It will be interesting to see if the property share market also contains differences between the various REIT- and property types.

2.2 Literature on the REITs

In a study by Chiang and Lee (2010) is argued that the equity REIT market has become less similar to the common stock market over the last few years. Their research focusses on the long run relationship between the common stock market and the equity REIT market. Their study is split up in two sub-periods, before and after the structural break in the early 1990s because of the recent tax reformations and a boom in IPOs. The two sub-periods are 1978-1993 and 1994-2008. According to Chiang and Lee (2010) the REIT market and the stock market moved similar before this structural break. However, after the early 1990s the REIT market became more like private real estate and less like common stocks. This implicates that REITs provide investors with real estate exposure in the long run. It will be interesting to see if anomalies that exists in the common stock market are also present in the REIT market.

Ben-Hamo and Brounen (2009) were the first to examine the Halloween effect in the property share market. They found statistically and economically significant evidence for the presence of the Halloween effect in the US REIT market. In their study, Ben-Hamo and Brounen (2009) tested a few seasonal anomalies in the ten most prominent property share markets in the world and South-Africa. In all these markets the returns were higher in the winter than in the summer. The effect was statistically significant in five countries including the US. The other countries where there is statistically significant evidence for the Halloween effect in the REIT market are Japan, the UK, Singapore and France (Ben-Hamo & Brounen, 2009). Focusing on the US, they found that the effect was concentrated among the smallest and, to a lesser extent, younger firms. On the other hand, they found hardly any difference between the summer- and winter month returns in larger and older firms.

(12)

Ben-12

Hamo and Brounen (2009) also performed a few tests to find a possible explanation for the presence of the effect but could not find a specific reason.

The research by Ben-Hamo and Brounen (2009) was performed using a traditional regression model. Hui and Chan (2015) used a new approach to test for anomalies in real estate share markets in the US and Asia, namely the Shiryaev-Zhou index with a logistic regression. The advantage of this approach is in the estimator. In contrast to the estimator in the standard regression approach, the Shiryaev-Zhou estimator gives rise to a trading strategy that generally outperforms a simple buy-and-hold strategy. This approach can help investors to increase profits by formulating better trading strategies. For the US they found similar results as Ben-Hamo and Brounen (2009), a significant Halloween effect in the REIT market. However, for some other markets the results differed.

In another study Hui and Chan (2018) extended the Shiryaev-Zhou index approach with the window size to vary, while in their 2015 study they used fixed moving-window size. The practice of this approach is beyond the scope of this study. However, the two main implications of the study by Hui and Chan (2018) are worth mentioning. The first implication says that investors can earn higher profits compared to a buy-and-hold strategy, when they adjust their strategy based on the Halloween effect. The second main implication is that anomalies and trends found in the equity stock market may not be applicable to the property share market.

The statistical significance of the Halloween effect in the real estate world is proven by the papers mentioned before. However, the economic significance is still subject to discussion. While Ben-Hamo and Brounen (2009) found economically significant evidence for the Halloween effect in the REIT market, a study by Hui, Wright and Yam (2014) provided a contradicting result. In their paper the economic significance of various calendar anomalies is tested, the Halloween effect was one of them. The research was performed on 27 real estate security indices from 20 countries and regions around the world. Based on two different approaches, there results showed no economically significant evidence that a trading strategy based on the Halloween effect would outperform a buy-and-hold strategy.

Another paper by Chan, Hui, Wright and Yam (2014), supported the results of Ben-Hamo and Brounen (2009). However, their study did not examine the US property security market, but they used securitized real estate indices in six Asian economies. They found evidence that it is possible for investors to outperform a buy-and-hold strategy even with

(13)

13

transaction costs applied. The differences in findings for economic significance could be due to different testing approaches and/or due to different datasets. This study will contribute to this ongoing discussion.

Pyles (2009) argues that REITs are a unique type of securities, that have a more predictable return pattern and therefore will be a safer investment than equity stocks. With the increased risk aversion of investors when the hours of daylight on a day becomes less (Kamstra et al., 2003), there could be argued that because of SAD the investments in real estate shares during the winter will increase because of their better predictability. In his research, Pyles (2009) found similar results in the REIT market as Kamstra et al. (2003) for the stock market. The returns in the winter months, where the hours of daylight on a day are less, are higher than during the summer months. Only the smallest 40 percent of the REITs show this effect, where the largest 60 percent of the REITs show no significant evidence for the presence the Halloween effect because of SAD. According to Pyles (2009) this is due to the safer nature of the larger securities.

2.3 The January effect

Another calendar anomaly of interest for this study was discovered many years before the Halloween effect, namely the January effect. January is part of the winter months in the Halloween effect and could partly explain the Halloween effect. Therefore, January will be one of the control variables in the model used for investigating the Halloween effect, as will be discussed in the methodology section. In this section, the existing literature on the January effect will be discussed.

While previous studies showed that there was no evidence for seasonal patterns in stock returns, Wachtel (1942) presented contradicting evidence. He found that since 1925 the Dow-Jones Industrial Average was outperformed by high-yield stocks in January, starting the literature on the January effect. However, his study lacked statistical significance.

The first to provide statistical significant evidence for the January effect in the US were Kinney and Rozeff in 1976. This effect states that returns in January are higher than returns in the other 11 months of a year. Kinney and Rozeff (1976) used data from the New York Stock Exchange (NYSE) for the period 1904-1974, with the exception of the 1929-1940. Their results show a significant higher return in January than in the remaining months of the year. They argue that tax-loss selling (to claim a capital loss) by investors towards the end of

(14)

14

the year could be a possible explanation for the January effect, although possible explanations are not tested in their paper.

Further research to this seasonal effect was performed by Keim (1983). He found higher abnormal returns in January, providing evidence for the January effect. Furthermore, he concluded that the January effect is most pronounced among small firms. According to Keim (1983), the January premium is for more than fifty percent attributable to abnormal returns in the first week of the trading year, especially on the first trading day. Possible explanations are discussed, but not empirically examined in the paper.

The first to test a possible explanation for the January effect were Gultekin and Gultekin in 1983. They examined the stock market in 18 major industrialized countries and found in almost all of them evidence for the January effect, including the US. Furthermore, they found a close association between observed seasonality and the turn of the tax year in their sample, supporting the tax-loss selling explanation. Because of the nature of the data it is not possible to make definitive statements about the causality of this relation, but it is the first indication of a possible explanation for the January effect (Gultekin & Gultekin, 1983).

This observation was supported by a research of Reinganum (1983). He performed some empirical tests that indicated that the higher returns witnessed at the beginning of January are consistent with the tax-loss selling motive. However, not the entire January effect can be explained by tax-loss selling. Apenbrink, Jones and Lee (1991) also presented evidence for the relationship between tax-loss selling and the January effect. They found that for the stock returns in the Cowless Industrial index no significant January effect was present before 1917, when the personal income taxes were introduced, but thereafter there was.

However, it is no clear case that the January effect is caused by tax-loss selling motives, as other studies found no significant relationship between the two. A study by Jones, Pearce and Wilson (1987) shows that the January effect was present in the stock market long before income taxes had an effective impact on investors. Furthermore, after income taxes were imposed, no significant change was observed in the January effect.

More recent support for tax-loss selling explanation is provided by Starks, Yong and Zheng (2006). In their paper, municipal bond closed-end funds are used as investments to test for the tax-loss selling statement. They argue that these are good candidates to test for this relationship because they are investment products that are almost entirely held by

(15)

15

individual investors who are presumably tax sensitive. Starks et al. (2006) conclude that tax-loss selling activities at the previous year-end can largely explain the January effect.

Another possible explanation for the existence of the January effect is presence of inside traders. Seyhun (1988) presents evidence that corporate insiders in small firms adjust their transactions at year-end to use the price run-up in January to earn higher profits. Insiders are enabled to capture more of the positive abnormal returns in January by becoming net purchases of the stock in December. Some insiders in small firms tend to postpone their stock sales and accelerate their stock purchases in December (Seyhun, 1988). However, additional increased insider purchase of stock does not lead to any price pressure on the prices of small firms in January, as Seyhun (1988) shows in his paper.

Another phenomenon that partly explains the January effect is the fact that the payment system in the US is so that most investors will receive substantial cash receipts at the turn of the year (Ogden, 1990). According to Ogden (1990), these investors will therefore increase their demand for stocks. He also argues that the expected liquid profits, and thus the January effect is negatively related to the stringency of monetary policy. In his study he provides empirical significant evidence for these statements, thus a partial explanation for the January effect, using stock index returns from 1969-1986.

The final possible explanation for the January effect mentioned in the existing literature is the window-dressing hypothesis. This hypothesis states that institutional investors, who are normally evaluated in relation to their peers shortly before the calendar year-end, buy winners and sell losers in order to present respectable year-end portfolio holdings (Sias & Starks, 1997). Mixed results are found in prior researches to support this hypothesis

In line with the paper of Schwert (2003), Gu (2003) found evidence for a weakened January effect. He used 70 years of data on the major US stock indices in his research. Based on this data, he showed that since 1988 the seasonal pattern seems to fade away. Contrary to the earlier study by Keim (1983), Gu (2003) found that the January effect is most pronounced in larger firms, indicating that the January effect may not be related to the size of the firm. Furthermore, he found that the January effect is negatively correlated to the actual and expected real GDP growth, inflation and return of the year. A positive relation was found with volatility. It will be interesting to see if the effect is still present and therefore could partly explain the Halloween effect.

(16)

16

2.4 January effect on the REITs

There are some contradictory conclusions about the presence of the January effect in the real estate share market. The first research to the January effect in the property share market was performed by Colwell and Park (1990). The study was performed on 28 equity REITs and 33 mortgage REITs listed on the AMEX and NYSE during the period from 1964 to 1986. They concluded that there is evidence for the January effect in the real estate related investments. Similar to the general stock market, the January effect tends to be more concentrated in smaller REIT firms and seems to disappear for larger firms. In a paper by Liang, McIntosh and Tompkins (1991) there was reported that the small firm effect, higher returns for smaller firms, was present despite the fact that the smaller REITs did not appear to be more risky than larger REIT firms. Furthermore, they compared equity REITs to mortgage REITs and found a significant higher January return in both markets, although the effect for mortgage REITs exceeds the effect for equity REITs. This article indicates that there are differences in seasonal effects between different types of real estate investments. Similar conclusions were made by Friday and Peterson (1997). Their study examined monthly total return indices from NARIET for the period 1973-1993. On top of their conclusion that there is evidence for the January effect in the property share market, especially for smaller firms, they found that the information related effects are not the main cause for the presence of the January effect in the REIT market. The most likely explanation based on the research by Friday and Peterson (1997) is the tax-loss selling hypothesis.

Contradicting results are found by Hardin, Liano and Huang (2005), who did research on the January effect in REITs by constructing an equally weighted- and a value-weighted indexes for the period after the change in the REIT structure from 1994-2002. They found that the January effect was present for the equal-weighted REIT index, but not for the value-weighted REIT index. So, the puzzle for the January effect is not completely solved, as more research is needed.

The paper by Ben-Hamo and Brounen (2009) discussed before, also described a research on the presence of the January effect in the property share market. Regarding the January effect they found less compelling evidence than for the Halloween effect. Their results indicated that the return in January is on average higher than during the other months in the year, also for the US. However, the difference in monthly returns lacks

(17)

17

statistical significance. Furthermore, based on this paper the effect appears to be most pronounced in small and young firms.

Chan and Hui (2015), also performed a study on the January effect in the REIT market, just as for the Halloween effect they used the Shiryaev-Zhou index approach for the period 1996-2013. Using this method, they did not find statistical evidence for the presence of superior January returns. They mention two possible reasons for their outcomes. Firstly, during turbulent times the January effect may weaken. Secondly, since the effect is reported in the literature, it may have lost its predictive power. This is in line with the statements by Schwert (2003).

Another possible explanation for the contradicting results for the presence of the January effect in the property share market is reported by Lee and Lee (2003), namely the implementation of the new tax regulations in 1994 in the US. This facilitates the increase of institutional investment in the REIT market and according to Lee and Lee (2003), the January premiums decreased with the level of institutional involvement for REITs. This is in line with the tax-loss selling hypothesis. For equity REITs the January premiums declined significantly, while for the mortgage REITs no significant decrease was observed. Trading strategies based on the January effect may work only when institutional investors leave (Lee & Lee, 2003).

From an international perspective, differences between markets can be found. Firstly, Peng (2005) found evidence for the existence of the January effect in the Australian listed property market, thereby providing evidence for market inefficiency. Furthermore, Peng (2005) concluded that anomalies are diminishing and disappearing over time. This is line with the conclusions of Schwert (2003). Suggesting that by reporting anomalies, markets get more efficient.

Ben-Hamo and Brounen (2009) tested the January effect in the 10 most prominent markets and South Africa. Besides the US, only the Netherlands showed superior returns in January, but the statistical significance lacks. Furthermore, they found that the international variation cannot be explained by the size and the maturity of the market. In a study on monthly return irregularities in EPRA/NAREIT indices in 12 European countries by Almudhaf and Hansz (2011) was indicated that only in Switzerland the January effect is present. Five Asian markets were included in the study by Chan and Hui (2015), they only found statistical significant evidence for the presence of the January effect in the Hong Kong property share market.

(18)

18

3. Data and methodology

In the first part of this section, the retrieved data is discussed. In the second part, some descriptive statistics are provided. Finally, the research method is presented, and the models tested in this research are stated and discussed.

3.1 Data

The data used for this research is collected from the WRDS website. From the CRSP/Ziman REIT database monthly values of the following variables are collected for all REITs available in the database: Ticker, REIT type, property type, property sub-type, share code, market capitalization, shares outstanding and total returns. The CRSP/Ziman database contains daily and monthly data on all REITs that have traded on the NYSE, the NYSE MKT and the NASDAQ exchanges. This database has data on REITs from 1980 until 2017, this complete period is used for this research because the development of seasonal anomalies over the years is also of interest for this study. Later on, the full time-period will be broken down into smaller time periods to see if the effect weakens over time. So, the full sample will consist of 37 years of monthly data on all REITs traded on the NYSE, the NYSE MKT and the NASDAQ exchanges.

After the dataset was collected, some modifications were made to make the data appropriate for this study. First, all the observations with another share code than 18 were removed from the dataset. Share code 18 shows the ordinary common shares of REITs, reflecting 68% of the total dataset, as can be seen in Table A1. Another substantial part of the dataset consists of shares of beneficial interest, reflected by share code 48. This type of shares is removed from the dataset because often they are traded infrequently and could therefore distort the results of this research.

Thereafter, some REIT types are removed from the sample. REIT types that are not useful for this research are defined as unknown, unclassified or hybrid REITs. Hybrid REITs do not consist of one particular asset class and therefore they cannot be used to identify differences among asset classes. After this modification, only equity and mortgage REITs are left in the sample.

Then, unknown, unclassified and diversified property types will be removed from the sample. The property types that are of interest of this study are Health Care, Industrial/Office, Lodging/Resorts, Mortgages, Mortgage Backed Securities, Residential,

(19)

19

Retail and Self-Storage. The most common property type in the sample is retail, with 24.1% of all observed property REITs allocated in this category. Finally, observations with unknown values are removed from the dataset as well as some duplicates. This leaves us with a dataset of 41,241 observations before merging it with other datasets.

From the Compustat North-America fundamentals annual database, the data on total assets and total liabilities is collected. These are needed to calculate the book values of the assets. Observations with missing values were removed from the dataset. Finally, from the Kenneth R. French data library, the one-month treasury bill rate is retrieved and is used a proxy for the risk-free rate. The risk-free rate is needed to calculate the excess returns. Furthermore, data on the Small-Minus-Big (SMB) factor, the High-minus-Low (HML) factor, the Momentum (UMD) factor and the excess return on the market will be collected from the Kenneth R. French data library. In this data file there were no missing values and no other reasons to believe that the data will distort the results of this research. Therefore, no observations were removed from this dataset.

After the data collecting process is finished, the multiple datasets are merged together into one data file. The observations in the merged files do not correspond completely, therefore observations that are not in all the datasets were removed. This leaves a total of 22,112 observations for this research. Some additional variables must be created to complete this research. The first one is the dependent variable, the excess return. This is done by subtracting the one-month treasury bill rate from the monthly total return. The next variable created is a dummy for the winter months. This dummy variable will take the value of 1 during the months from November up to and including April and 0 during May until October. The last variable created is again a dummy, this time a dummy indicating the month January, needed to control for the January effect. The dummy variable takes the value of 1 during January and 0 during the remaining months of the year.

3.2 Descriptive statistics

In Table A2 and Table 1 the descriptive statistics are reported for the winter dummy variable and the quantitative variables respectively. The tables show the statistics for all the subsamples used in this research, starting with the entire sample, followed by the subsamples for the different REIT types and at the end, reporting statistics for the different property type samples.

(20)

20

In Table A2 can be seen that observations are almost equally divided between the winter and the summer months. In the complete sample, 50.38% of the observations are reported during the winter months and 49.62% during the summer months. This pattern can be seen in almost all the subsamples except for the Self-Storage sample. In the final column the total observations per sample are reported. As mentioned before, the total number of observations for the entire sample is 22,112. Of these observations 84% are equity REITs and only 16% are mortgage REITs. Considering the property types, Table 1 shows that Retail is most observed with 29.5% of the observations allocated to this category. The lowest number of observations is available for the Self-Storage property type with only 378 observations, this is less than 2% of the entire dataset.

In Table 1, the mean for the excess returns, market returns, SMB-factors, HML factors and momentum factors per month are summarized for the entire sample, and for each REIT- and property type individually. Furthermore, the table presents the minimum and the maximum excess return for each sample. Finally, the number of observations per sample is reported. Table 1 shows that in the complete sample the average excess return is 0.95%. In the other subsamples, the mean excess return deviates not much from the complete sample. Only the return for retail is slightly lower and the average return in the Self-Storage sample is substantially higher. High maxima for the excess return are reported. The maximum excess return is not less than 290.27%. This maximum can be found in the Mortgage REITs. The other variables are not exactly equal in all subsamples. However, the differences are not very large.

(21)

21 Table 1: Descriptive statistics.

Entire Sample Equity REITs Mortgage REITs Health-care Industrial/ Office Lodging/ Resorts

Mortgage MBS Residential Retail

Self-Storage Excess return mean (SD) 0.95% (8.92%) 0.95% (8.50%) 0.97% (10.87%) 1.01% (8.28%) 1.00% (7.98%) 0.95% (12.97%) 0.94% (11.13%) 0.98% (8.46%) 1.00% (6.80%) 0.83% (8.55%) 1.78% (6.54%) Excess return Min. -79.83% -79.83%* -67.43% -49.50% -65.68% -79.83% -67.43% -47.43% -53.16% -72.66% -25.15% Excess return max. 290.27% 236.43% 290.27% 55.02% 80.19% 143.40% 290.27% 36.63% 40.39% 236.43% 29.03% Market return mean (SD) 0.73% (4.18%) 0.72% (4.20%) 0.80% (4.00%) 0.70% (4.22%) 0.74% (4.18%) 0.72% (4.07%) 0.76% (4.09%) 0.10% (3.45%) 0.76% (4.23%) 0.67% (4.28%) 0.92% (3.43%) SMB mean (SD) 0.10% (3.16%) 0.10% (3.23%) 0.10% (2.76%) 0.09% (3.19%) 0.10% (3.15%) 0.09% (2.98%) 0.12% (2.81%) 0.01% (2.31%) 0.06% (3.28%) 0.14% (3.38%) 0.06% (2.41%) HML mean (SD) 0.18% (2.95%) 0.20% (3.00%) 0.10% (2.71%) 0.25% (2.71%) 0.15% (2.94%) 0.15% (2.85%) 0.11% (2.77%) 0.01% (2.30%) 0.21% (3.01%) 0.21% (3.10%) 0.35% (2.46%) UMD mean (SD) 0.43% (4.71%) 0.30% (4.39%) 0.30% (4.39%) 0.48% (4.66%) 0.40% (4.73%) 0.31% (4.66%) 0.30% (4.51%) 0.22% (3.62%) 0.51% (4.70%) 0.48% (4.95%) 0.52% (3.43%) Years 1980-2017 1980-2017 1980-2017 1980-2017 1980-2017 1980-2017 1980-2017 1980-2017 1980-2017 1980-2017 1980-2017 #obs. 22,112 18,537 3,539 2,068 4,616 1,455 3,164 530 3,378 6,523 3,164

Note: In this table the summary statistics for the entire sample, and the subsamples based on REIT type and property type, are presented. All REITs listed on the NYSE, NYSE MKT and NASDAQ between 1980 and 2017 are included in the samples. The data is retrieved from the CRSP/Ziman REIT database. The total number of monthly observations, the mean, the standard deviation, the minimum and the maximum per sample are provided for the excess return. For the other variables, the total number of monthly observations, the mean and the standard deviation is presented in this table. The excess return is the raw return minus the one-month t-bill rate used as a proxy for the risk-free rate. The data for the SMB, HML and UMD (momentum) factor is retrieved from the Kennet R. French database. The information is given in percentages.

(22)

22

3.3 Methodology

The first test performed in this study, is simply to check if the effect is still present in the REIT market. After this test is completed, the different property and REIT types will be examined, to see if there are any differences among them. For these tests, usual regression techniques are used. A simple model with only a dummy for the Halloween effect on the right-hand side is used in this first regression. The tests in this study are performed with ordinary least squares (OLS) regression. The first model is presented below and is based on the paper by Bouman and Jacobsen (2002) where they provided the first statistical evidence for the presence of the Halloween effect.

(1) − = + +

The dependent variable is the monthly excess return for REIT i, in industry j and in year t and is reflected by the return for the REIT stock minus the risk-free rate. This is done to check if there is evidence for the Halloween effect for the different property types. The μ is the constant and reflect the returns during the summer, is the variable of interest that takes a value of 1 during the winter months, from November till April, and 0 otherwise. The returns during the winter months, from May till October, exists of the constant term plus the coefficient of the Halloween dummy. Finally, reflects the usual error term.

To test for the presence of the Halloween effect in the REIT market and the different types of REITs and properties, there is tested if the coefficient for the Halloween dummy is significantly different from zero. When this coefficient is positive and significant, this is the first prove for the presence of the Halloween effect in the specific REIT market. For these regressions, robust standard errors are used. The significance of these effects will be tested with t-tests at the significance levels of 10%, 5% and 1%. This will be done for all the remaining regressions in this study. In essence, this equation reflects a simple mean test, where there is examined if the mean return in the November-April period is significantly higher than the mean return in May-October. An advantage of using this regression model is that other variables can easily by include, as is done later on in this paper.

(23)

23

As described in the literature review, the January effect shows that returns in the stock market tend to be higher in January than in other months. As also discussed in the literature review, there is still discussion going on whether this effect is present in the property share market or not. Some might argue that the Halloween effect is the January effect in disguise. Bouman and Jacobsen (2002) and Jacobsen and Visaltanachoti (2009) both include a January dummy in their regression model to control for the January effect. It will be interesting to see if in the data used for this study a January effect appears. To examine the January effect the regression model below will be tested. This model is based on the paper by Ben-Hamo and Brounen (2009).

(2) − = µ + + + … + + +

+

In this model the dependent variable is the monthly excess return, just as in the first model and µ measures the January effect. The variables Feb until Dec are dummies for the remaining months of the year, where Feb is a dummy for the month February and takes the value of 1 in February and 0 otherwise. This holds for all the remaining months in the equation. is included to capture potential spill-over effects across consecutive trading months and is used to control for bias related to serial autocorrelation in the data. The last term, , is the error term. To test for the January effect the statistical differences between January and the other months are examined. If the differences are all negative and significant, there is enough evidence for the January effect. These tests are performed on the total REIT market and on all the REIT and property types individually.

After the January effect is tested, the focus is shifted back to the Halloween effect. The equation of the main model is presented below and is based on the paper by Jacobsen and Visaltanachoti (2009) where they examined the differences between industries in the regular stock market. The model is:

(3) − = + + + − +

The dependent variable is again the monthly excess return for REIT i, in industry j and in year t. The Halloween dummy takes in this model takes the value of 1 in the months

(24)

24

November-April, except for January when the dummy takes a value of 0 just as in the summer months. An additional January component is included to control for the fact that the Halloween effect is simply the January effect in disguise and takes the value of 1 in January and 0 otherwise. The last term, − , is to correct for general movements in the market. This test will be done for all the different REIT types.

The fourth regression equation of this study will be used to check for the robustness of the methods used before and is equal to the model presented by Jacobsen and Visaltanachoti (2009). They used the Fama- and French three factor model to test for the robustness of the results. Therefore, these Fama- and French factors will be included in the model. The three Fama- and French factors are the excess return on the market, the small minus big (SMB) factor and the high minus low (HML) factor. Furthermore, a momentum factor will be included to control for the different investment styles of investors. The control variable for January will be left out in this regression because the SMB and HML factors are partially driven by the January effect. This leaves the following regression model:

(4) − = + + − + + +

+

Just as in the equations before, the monthly excess return is the dependent variable and is the constant. The Halloween dummy takes the same values as in the first model, this means a value of 1 during the winter months from November-April, and 0 during May-October. The three factors of the Fama- and French model are in the equation presented by

− , and respectively. The momentum factor is included by the variable .

In the final model of this paper, the differences over time are examined. To test for the differences over time, interaction terms are implemented in model (3). The complete time-period 1980-2017 will be broken down into three sub-periods. The first one will contain the years 1980-1994, the second 1995-2005 and the final sub-period ranges from 2006-2017. For these three sub-periods dummy variables will be created and included in the model. The dummies for the sub-periods will also be added to the model as interaction terms with the Halloween dummy. The advantage of using interaction terms over the use of

(25)

25

multiple regressions for the three time-periods is that, with the use of interaction terms, it is possible to test for significance of the changes over time. The final model is presented below.

(5) − = + + + − + 2 +

3 + 2 ∗ + 3 ∗ +

This equation is equal to model (3), only dummy variables for the sub-periods, and interaction terms between the Halloween effect and the time sub-periods are included. To avoid multicollinearity in the model, one of the 3 categories must be left out of the equation. In this model, the first sub-period is left out because the remaining interaction terms then reflect the change over time. The dummy variables for the sub-periods show the differences in excess return over time. The interaction terms are the variables of interest in this model, the coefficients show the change in the magnitude of the Halloween effect in the two subsequent periods compared to the first period from 1980-1994. If the coefficient for the variables is significant and lower than zero, then the effect weakened or dispersed over time, as would be expected based on the paper by Schwert (2003).

In order to test for the differences between property types and REIT types, the models and tests described above are all performed on the total REIT market and the different REIT- and property types. In all the regressions, clustered standard errors are used. These are used because they control for serial correlation and heteroscedasticity. The standard errors are clustered on the REIT ID. In the next section, the results of this research are presented.

4. Results

In this section the results of the research are presented. In the first part, the results for the different regressions are discussed. Secondly, the development of the Halloween effect over time will be discussed for all the different samples. In the final part of Section 4 is tested if it is possible to outperform the market with a sector rotation strategy based on the Halloween effect.

(26)

26

4.1 Regression results

This section contains the results of the multiple regressions performed in this study. First, the regression output of model (1) is presented, the output is given in Table 2. In this model, only the differences in returns between the winter and the summer months are observed, with no additional control variables in the equation. In fact, this model reflects a simple mean test. The first column shows the returns during the summer for every REIT or property type. The second column presents the additional winter return, this is the return during the winter in excess of the return for the summer period. In the last column, the R-squared statistic is noted.

Table 2: Regression results model (1), Halloween effect.

Summer Return Additional winter return R-squared Entire Sample 0.44%*** (0.001) 1.04%*** (0.001) 0.0034 REIT Types Equity 0.55%*** (0.001) 0.81%*** (0.001) 0.0023 Mortgage -0.13% (0.002) 2.23%*** (0.004) 0.0105 Property Types Health Care 1.02%*** (0.002) -0.02% (0.003) 0.0000 Industrial/Office 0.53%*** (0.001) 0.93%*** (0.002) 0.0034 Lodging/Resorts -0.06% (0.002) 2.04%*** (0.004) 0.0062 Mortgage -0.15% (0.002) 2.20%*** (0.004) 0.0098 MBS -0.17% (0.004) 2.33%*** (0.005) 0.0191 Residential 0.67%*** (0.001) 0.65%*** (0.002) 0.0023 Retail 0.45%*** (0.001) 0.78%*** (0.002) 0.0021 Self-Storage 1.51%*** (0.003) 0.54% (0.004) 0.0017

Note: The coefficients are returns in percentage-points. Clustered standard errors in parentheses. *, **, *** significant at the 10%, 5% and the 1% significance level, respectively. In the final column the -statistic is noted.: Fictitious d for illustration purposes only

(27)

27

The first thing to notice in Table 2 is that the monthly average return for the entire sample is positive during the winter months, as well as during the summer months. If the average return is observed per REIT or property type, there can be seen that the return in the winter is for none of the samples negative, and only for Health Care properties the return is on average lower during the winter months than during the summer. The average summer return is only negative for the Mortgage REITs and the Lodging/Resorts property type. However, for none of the REIT or property types, the average return is statistically significantly lower than zero, in both the summer and the winter period.

The yearly mean return is 11.26%. On average, 8.84% of this return is earned during the winter months, while the remaining 2.42% are returns earned during the summer months. Based on the first regression, the average monthly return for the winter period was on average 1.04%-points higher compared to monthly returns during the summer months for the entire sample in the period 1980-2017. From an economical perspective, this is a substantial difference in favor of the monthly winter returns, regarding that the yearly mean return is 11.26%. As can be seen in Table 2, this difference is statistically significant at the 1% level, indicating that the Halloween effect is present in the REIT market for the period 1980-2017. With this knowledge, the study can continue to examine the differences between the various REIT and property types.

For both REIT types, the winter return is significantly higher than the return during the summer, at the 1% significance level. For the Mortgage REITs the monthly return in the months November-April is 2.23%-points higher than the monthly return in months from May until October. The difference in Equity REITs in ‘only’ 0.81%-points. However, for Equity REITs the return during the summer months is significantly higher than zero, this is not the case for Mortgage REITs, where an insignificant negative return is observed during the summer. These findings support the conclusion of Ben-Hamo and Brounen (2009), who concluded that the Halloween effect was present in the equity market.

At the property type level, all the property markets, except the Health Care and the Self-Storage properties, show evidence for the presence of the Halloween effect. The monthly returns during the winter months in these markets are significantly higher at the 1% significance level. The highest additional winter return is observed in the Mortgage Backed Securities (MBS) market. The monthly return in this market was during the winter on average 2.33%-points higher than during the summer. The observed R-squares in these

(28)

28

regressions are quite low, however that is not unusual for these type of models, where seasonal patterns are examined.

The fact is that an investment strategy based on the Halloween anomaly will induce only limited transaction costs. The transaction costs are limited, due to the fact that this investment strategy only requires two transaction moments each year (Ben-Hamo and Brounen, 2009). Based on the results presented in this section it seems that the Halloween effect offers interesting arbitrage opportunities. However, because of the positive summer returns for the most REIT- and property types, leaving the market during the summer requires a cost of capital that will be hard to obtain at current interest rates.

In Table 3, the results for model (3) are presented. In this model a dummy variable for January is included, as well as a term to correct for general market movements. These terms are included to check if the Halloween effect is not just the January effect in disguise. The coefficient for the control variable for the general market movements can be compared to the beta of the REIT or property type. The table is equal to Table 2, only the additional variables are included.

Again, the average monthly winter and summer returns are positive for the entire sample. Just as in the first regression, the summer returns are negative only for the Mortgage REITs and for the Lodging/Resorts properties. However, in the first regression they were never significantly negative and this time the summer return for these property types are significantly lower than zero at the 5% significance level.

In the entire sample, there is still evidence for the presence of a Halloween effect. However, the effect tends to lose its power. In the first model, the monthly winter return was on average 1.04%-points higher than the monthly summer return, while the difference with this model is only 0.29%-points in favor of the winter months. Although the coefficient is lower, it is still significantly higher than zero at the 1% significance level, providing evidence for the presence of the Halloween effect in the REIT market. These results support the findings of Ben-Hamo and Brounen (2009).

With the first simple model, there was evidence for the presence of the Halloween effect in both the equity and mortgage REITs. However, with the model correcting for the January effect and the general market movements, the effect only seems to appear in the mortgage REIT market. The average monthly return in the winter for the equity REITs is still 0.13%-points higher than the average monthly return during the summer, but the difference

(29)

29

is no longer statistical significant. The power of the effect also decreased in the mortgage REIT market. Nevertheless, the monthly returns are on average still 1.12%-points higher during the winter months than during the summer. The difference for mortgage REITs is significant at the 1% level.

Table 3: Regression results model (3), Halloween effect controlling for January. Summer Return Additional winter return January General Market R-squared Entire Sample 0.21%*** (0.001) 0.29%*** (0.001) 1.01%*** (0.002) 0.74*** (0.042) 0.1207 REIT Types Equity 0.32%*** (0.001) 0.13% (0.001) 0.54%*** (0.002) 0.73*** (0.041) 0.1297 Mortgage -0.43%** (0.002) 1.12%*** (0.002) 3.54%*** (0.005) 0.82*** (0.157) 0.0974 Property Types Health Care 0.84%*** (0.002) -0.91%** (0.003) 1.60%** (0.007) 0.59*** (0.073) 0.0923 Industrial/Office 0.29%*** (0.001) 0.20% (0.001) 0.42% (0.004) 0.79*** (0.068) 0.1707 Lodging/Resorts -0.57%** (0.002) 0.49% (0.004) 2.39%** (0.009) 1.55*** (0.205) 0.2370 Mortgage -0.38%** (0.002) 1.04%*** (0.003) 3.53%*** (0.005) 0.79*** (0.163) 0.0916 MBS -1.09%** (0.004) 1.44%*** (0.002) 4.09% (0.025) 1.14*** (0.284) 0.2246 Residential 0.47%*** (0.001) 0.22% (0.002) -0.09% (0.004) 0.57*** (0.047) 0.1274 Retail 0.30%** (0.001) 0.22% (0.002) 0.23% (0.003) 0.63*** (0.063) 0.1012 Self-Storage 0.84%** (0.003) 0.03% (0.001) 2.00%* (0.08) 0.82** (0.235) 0.1881 Note: The coefficients are returns in percentage-points, the coefficient for the general market reflects the Beta of the particular REIT market. Clustered standard errors in parentheses. *, **, *** significant at the 10%, 5% and the 1% significance level, respectively. In the final column the -statistic is noted. Source: Fictitious data,

for illustration purposes only

If the results per property type of this model are compared with the findings of the first regression, quite substantial differences are observed. For Industrial/Office, Lodging/Resorts, Residential and Retail, there is no longer evidence that the monthly returns during the winter months are significantly higher than the monthly returns during the summer months. The Halloween effect is only observed in the Mortgage and the MBS market. An opposite effect seems to be present in the Health Care property market. The monthly return in this market is on average 0.91%-points lower during the winter months

(30)

30

than during the summer. This opposite effect is significant at the 5% level. These findings are in line with statements made by Liang, McIntosh and Tompkins (1991) and Lee and Lee (2003), who stated that differences can be observed in seasonal effects between different types of real estate investments. The observed R-squares are slightly higher than with the first regression, which makes sense. Fictitious data, for illustration purposes only

In the fourth model of this study, the Fama- and French factors and a momentum factor are included in the regression that tests for the presence of the Halloween effect. This model is used to test for the robustness of the results. Overall, the results are quite similar to the results of the main model (3), so the results tend to be robust. The results of model (4) are presented in Table 4. Compared to previous tables in this paper, additional columns are created for the new factors included in the model. The dummy for January is removed from the equation because the Fama- and French factors are partly driven by the January effect.

This time the average returns are not positive for all the REIT and property types during the winter months. For Lodging/Resorts, the average monthly return is negative in both the winter and in the summer, with the return in the winter slightly higher with 0.05%-points per month. However, this difference is insignificant. MBS is again the only type that has a significant negative return during the summer months.

In contrast to the models before, with the Fama- and French factors, and the momentum factor included, there is no statistical evidence for the presence of the Halloween effect in the entire sample. The opposite tends to be the case, according to this model the return in the winter is 0.09%-points lower per month than during the summer. However, the coefficient lacks significance. These findings are in contrast with the findings of Ben-Hamo and Brounen (2009). However, they did not include the Fama and French factors or the moment factor in their models.

For both REIT types, the results based on this regression are in opposite directions, as can be seen in Table 4. This in line with the findings by Liang, McIntosh and Tompkins (1991) and Lee and Lee (2003). In the mortgage REIT market, the Halloween effect tends to be present. On average, the monthly return is 1.19%-points higher in the winter months than in the summer months. In the equity REIT market, a contradicting result is found, namely a significantly lower return of 0.33%-points per month during the winter months

Seasonal anomalies in REIT returns : the Halloween effect and the January effect