MSc Finance - Master Thesis
Global Investor Sentiment and
Global Listed REITs
University of Groningen - Faculty of Economics and Business
Author: Luchian Mădălina – Nicoleta Student No: S3841510
Supervisor: dr. Auke Plantinga
Keywords: investor sentiment, return predictability, REIT, real estate investment trusts JEL Classification: G11; G12; G14; G15;
Groningen, Netherlands – January 9th, 2020
1 Table of Contents Table of Contents ... 1 List of Tables ... 2 List of Figures... 2 I. Introduction ... 3
II. Literature Review... 5
2.1. Investor Sentiment ... 5
2.1.1. Theoretical Background of Investor Sentiment ... 5
2.1.2. Measuring Investor Sentiment ... 7
2.2. Real Estate Investment Trusts (REIT) & Investor Sentiment ... 10
III. Data & Methodology ... 12
3.1. Data Collection: Dependent Variables (Country vs Asset Class) ... 13
3.2. Data Collection: Investor Sentiment Proxies... 14
3.3. Orthogonalization ... 18
3.4. Total Country Level Sentiment Index – Principal Component Analysis (PCA) ... 19
3.5. Global Investor Sentiment Index ... 22
3.6. Local Investor Sentiment Index ... 24
3.7. Predictive Regression Analysis ... 26
IV. Results... 28
4.1. Predictive Regression Analysis ... 28
4.2. FF3 Model Inclusion ... 30
V. Conclusion ... 32
2 List of Tables
Table 1: Independent Variables - Summary Statistics ... 14
Table 2: Global Investor Sentiment Proxies - Data Parameters, by Country ... 16
Table 3: Global Investor Sentiment Index - Proxies used for Construction ... 17
Table 4: Summary Statistics for Raw Investor Sentiment Proxies ... 18
Table 5: Summary Statistics for Orthogonalized Investor Sentiment Proxies (Residuals) ... 20
Table 6: Summary Statistics - Total Country Level Investor Sentiment Index ... 22
Table 7: Local Sentiment Index Summary Statistics - All Countries ... 24
Table 8: Predictive Univariate Regression Analysis Results ... 29
Table 9: Predictive Multivariate Regression Analysis Results ... 30
Table 10: Fama - French 3 Factor Model Inclusion: Global Investor Sentiment Index ... 31
Table 11: Fama - French 3 Factor Model Inclusion: Local Investor Sentiment ... 31
List of Figures Figure 1: Total Country Sentiment Index - Factor Loadings - GBP ... 21
Figure 2: Total Country Sentiment Index - Factor Loadings - USA ... 21
Figure 3: Total Country Sentiment Index - Factor Loadings - CND ... 21
Figure 4: Total Country Sentiment Index - Factor Loadings - AUS ... 21
Figure 5: UK - Total Country Level Investor Sentiment ... 22
Figure 6: USA - Total Country Level Investor Sentiment... 22
Figure 7: CND - Total Country Level Investor Index ... 23
Figure 8: AUS - Total Country Level Investor Index ... 23
Figure 9: Global Investor Sentiment Index ... 23
Figure 10: Local Investor Sentiment Index - CND ... 25
Figure 11: Local Investor Sentiment Index - AUS ... 25
Figure 12: Local Investor Sentiment Index - GBP ... 25
3
I. Introduction
The award of a Nobel prize for behavioural economics in 2017 to Richard Thaler and behavioural finance in 2013 to Robert Shiller has resulted in increased attention of academia to the role of investor sentiment in the financial market. The expansionary and recessionary economic cycles that have characterized the last half century has paved the way for academics to use behavioural finance to explain market behaviour in situations where dramatic changes in stock prices occur that defy the traditional explanation provided by a standard finance model. As opposed to assuming that investors are always unemotional and that market prices will equal the present value of expected future cashflows (Baker & Wurgler; 2007), investor sentiment is modelled in the literature to assess how fundamental prices can be changed by the activity of investors that exhibit not fully rational behaviour.
Investor sentiment aims to reflect an investors attitude (generally regarded as pessimism vs optimism) with respect to price development in a market. The relationship between investor sentiment and pricing developments in any asset class is generally regarded from two points of view: (i) a level of optimism (pessimism) in a current period leads to a low (high) return in the following period(s) or (ii) certain asset classes may or may not display higher sensitivity to investment sentiment than others.
Up to the 1990s many studies surrounded the question Does Investor Sentiment affect stock
prices?, while in the period after that, academia focused on the question How is investor sentiment measured?. One of the most notable and widely replicated contributions to measuring investor
sentiment is Baker & Wurgler’s (2006) paper that constructed a sentiment index that affects future stock returns on a cross sectional basis in the US. This index has been widely used in academia (Smith et al., 2016; Neely et al., 2013 among many others) to see if investor sentiment can predict risk premia in equity, as well as other various asset classes like the euro bond market (Beber, Brandt, & Kavajecz; 2009), the US bond market (Nayak, 2010), the gold and metals market (Smales, 2014; Zheng, 2015).
4
related to local sentiment indices in their respective markets. This index has not been as widely used in academia as its earlier counterpart. Many studies focus on local markets – they replicate the methodology to construct local indices - for example, Lam & Hui (2018) study the Hong Kong housing market by constructing a local sentiment index using Baker & Wurgler’s methodology. Therefore, the literature still leaves ample space to examine the connection between investor sentiment measured on a global scale among developed markets and various asset classes. The aim of this paper will be to apply the Baker, Wurgler & Yuan (2012) methodology of constructing a global investor sentiment index and study the relationship w.r.t. real estate as an asset class, focusing exclusively on returns generated by listed real estate investment trusts (REIT) at an aggregate market level.
5
results from the previous predictive analysis are not spurious. I find that the global index does successfully enter the Fama French 3 Factor model, however there is not sufficient evidence to also validate the local index.
II. Literature Review
2.1. Investor Sentiment
2.1.1. Theoretical Background of Investor Sentiment
The standard model, or classical finance theory, finds that in an efficient market, investors are rational and asset prices are determined by fundamental factors. Since the stock market should hold all available information, even if some investors act irrationally and mispricings occur, such opportunities will be quickly taken up by arbitrageurs, therefore the prices will be immediately corrected. Therefore, the prices in the market will be intrinsic values, reflective of the Efficient Market Hypothesis.
Historically, the stock market has been plagued with events that dramatically contradict the definitions of stable markets and true prices above, from the tulip bubble in the Dutch Golden Age, to the Great Crash of 1929, the Black Monday Crash in 1987 and the 2008 Financial Crisis, to name the most prominent ones.
Early studies, going back to the 1980’s have focused on investor sentiment but the role of investor sentiment in return predictability was implicit, due to the lack of strong statistical evidence due to short time series data (where it becomes difficult to distinct a random walk from a bubble) or lack of economic interpretation where there is strong statistical evidence (Baker & Wurgler, 2007). The aim was rather to find if investor sentiment affects stock returns, whereas in the most recent years, the aim of studies is to measure and quantify investor sentiment, in order to better capture its effects on the market.
Researchers in behavioural finance have modified standard models by focusing on two assumptions (De Long et al,1990; Shleifer & Vishny,1997):
(i) Investor sentiment is a particular, individual belief about future cashflows and corresponding risks that cannot be justified by the information available in the market – this is characteristic of irrational investors (one example can be noise traders);
6
Barberis, Shleifer and Vishny (1998) and Daniel, Hirshleifer and Subrahmanyman (1998) propose models that focus on a bottom up approach of measuring sentiment by using biases in investor psychology like overconfidence, conservatism or differences of opinion across investors. Baker & Wurgler (2007) argue that such a bottom approach will rely on traits unique to each individual, which may not yield great results over time if you account for mass psychology. They propose a top down macroeconomic approach, by reducing broader variables to key aggregate sentiment proxies and trace their effect to market level returns. The aim is then not to find if equity returns depend on sentiment, but rather to point out which type of stocks depend on sentiment.
In Baker & Wurgler (2006;2007) they find that growth firms, high volatility, non – dividend paying, unprofitable, younger stocks, stocks of low capitalization or stocks in financial distress are most subject to sentiment induced variation in returns. This can also be liberally observed in market history – for example, during the Dot.Com bubble in the late 1990’s small start up firms drove the majority of the investors’ excitement.
An investor sentiment index can thus be created to follow characteristics that make the set of companies mentioned above more speculative than other companies in the market. At the other end of the spectrum, there are companies with a long history of earnings, that pay dividends regularly and have tangible assets among other characteristics, which makes them much easier to value and less prone to speculation. Mainly, it is believed that lack of available information (combination of no earnings history and a highly uncertain future) in the market gives rise in difficulty and subjectivity in determining a company’s true value. This paves the way for investors to provide valuations from too high to too low (according to an individual’s beliefs). This type of uncertainty can drive both over - confidence, as well as over - cautiousness.
7
Smith et al. (2016) observe that a growing body of literature suggests that sentiment increases mispricing, and this mispricing is asymmetrical between high & low sentiment periods. High sentiment driven markets are less efficient than low sentiment ones (Stambaugh, Yu, and Yuan, 2012). The authors go on to explain that during high sentiment periods overpricing is more prevalent than underpricing is during low sentiment periods. This is because in high sentiment markets, overpricing cannot be eliminated due to restrictions to short selling, while underpricing can be fully countered by rational investors by holding long positions in securities. Investor sentiment has also been used alongside macroeconomic factors to see if risk premiums can be predicted in equity markets. Neely et al. (2013) find that monthly sentiment changes are positively and contemporraneously correlated with the equity risk premium. Smith et al. (2016) also study the effectiveness of technical indicators in different investor sentiment periods by focusing on strategies employed by hedge fund managers – assumed to be the most skilled and rational investors. It is discovered that in high sentiment periods hedge fund manager employing technical analysis outperform those that do not an also display better timing ability, whereas in low sentiment periods such strategies do not add any value.
2.1.2. Measuring Investor Sentiment
As previously mentioned, the main subject of investor sentiment studies in the past two decades have been focusing on how to quantify and measure investor sentiment. The literature offers a great variety of measurement proxies and methodologies, from proxies that have foundations in investor psychology to proxies that have their origins in economic fundamentals or trading patterns. An investor’s beliefs can also be recorded through surveys, though there is no guarantee that individuals will be honest in their answers, therefore using survey results as a proxy has always been regarded with a certain degree of suspicion. Focusing on company information can also yield misleading results – corporations can alter their structure and alter their fundamentals for many reasons which are not readily obvious.
Investor surveys have been conducted by Robert Shiller since 1989, Qiu and Welch (2006) also use polls for Consumer Confidence Index and show that consumer confidence correlates with small stock returns and returns of firms held disproportionately by retail investors. Investor mood is another subject of scrutiny, with studies trying to relate stock prices to the human emotional spectrum. Examples of this may ring familiar, such as day – of – the – week effect, “Sell in May and Go Away”, or more specifically, a study by Kamstra, Kramer and Levi (2003) who showed that market returns are on average lower in fall and winter, and attributed this to seasonal affective disorder.
8 Table 1: Measurements of Investor Sentiment
Investor Sentiment
Proxy Studies Definition
Retail Investor Trades
Greenwood & Nagel (2008); Barber, Kumar and Lee (2006);
This proxy is used based on the proof that an inexperienced investor is more likely to be subject to sentiment than a professional, sophisticated investor.
Mutual Fund Flows
Brown et al (2002) Overall market sentiment measure based on how fund investors are moving into and out of safe government bonds to risky growth stocks, as mutual fund investors are known to chase high returns.
Dividend Premium
Baker & Wurgler (2004a) Stocks that pay dividends on a relatively stable basis are seem as
safe, bond - like stocks, due to the steady stream of income they
offer.
Closed End Fund Discount
Lee, Shleifer & Thaler (1991)
It is the difference between the net asset value of the fund and the fund’s market price. If retail investors disproportionately hold this type of funds (following the retail investor trade proxy logic) the discount will increase when retail investors are bearish.
Option Implied Volatility
Whaley (2002) The Market Volatility Index (VIX) is often termed the fear index. Since option prices increase when the underlying asset has a higher volatility, the index itself can be a proxy.
Equity Issues over Total New Issues
Baker & Wurgler (2000) This measure focuses on all equity issues not just IPOs and is abroader measure of financing activity. Studies find that high values of equity shares issues are followed by low stock market returns.
Insider Trading
Sehyun (1998) Since corporate executive hold more accurate information than the general public about the actual value of their holdings, personal portfolio decisions may review information about firm mispricings.
9
the first principal component of the six underlying variables (Baker & Wurgler, 2006). However, when focusing on a global investor sentiment index not all information is proportionately available for all countries, therefore only four proxies (Baker, Wurgler & Yuan;2012) will be the subject of this study:
(i) Volatility Premium. This proxy identifies the time when valuations on stocks with high idiosyncratic volatility is low compared to the valuations on stocks with low idiosyncratic volatility. The reasoning for this variable comes from the proof in literature that investor sentiment affects stocks that are difficult to value and arbitrage. Volatile stocks tend to be costlier (wider bid – ask spreads) and are also riskier to trade (association with noise trading risk).
(ii) Turnover. High trading volumes (especially for assets that are overpriced) have been noted in history to precede any economic bubble. Studies find that between 1998 and 2000 (period preceding the Dot.Com crash) volume was much greater in internet related stocks than non - internet related ones (Ofek and Richardson, 2003)., and many other studies from as far back as 1988 (Smith, Suchanek, and Williams; 1988) find evidence that bubbles are associated with high turnover. Subsequent papers also outline that this correlation can be owed to trading fees and short – sale constraints. (iii) IPO First Day Returns. Very distinct and remarkable returns are registered following
initial public offerings. A few examples: in 1998 Computer Literacy Inc. registered a first day return of 99.38%, eSpeed in 1999 registered 99.71% in the US, and other countries showcase the same pattern. In the UK, Cleeve capital registered a return of 93%, in Canada one of the highest first day returns was at 120% and in Australia one of the highest returns was in 2003 at 136%. It is difficult to find fundamental explanations for these numbers and must assume that the incentive is behavioural and thus relates to investor sentiment, especially since evidence suggests that company issues with the highest return were in the highest demand. Ritter (1998) sums up the rationale behind this proxy.
(iv) IPO Volume. The rationale behind this proxy is closely tied to IPO returns, in that in so called hot markets shareholders will see incentives to time their equity offerings with periods when valuations and demands are greatest.
Using Principal Component Analysis in the four variables listed above, we can create a country level investor sentiment index. Afterwards, a global sentiment index can be created by taking once again the first principal component of the four country level indexes.
10
Investor optimism is not contained locally – investors in the US can switch to risky assets in periods of high, positive sentiment, which can include international equities. This can in turn, translate into increase in volume and volatility at a local level in another country. Thus, US sentiment can affect prices in another country, which is what a global sentiment index will capture. Likewise, they outline that one mechanism that facilitates the spread of contagion is international capital flows.
2.2. Real Estate Investment Trusts (REIT) & Investor Sentiment
Historically, the three main asset classes have always been equities (stocks), fixed income (bonds) and cash equivalent (or other money market instruments), but in recent decades other asset classes and sub classes have differentiated themselves from the main three, depending on a number of statistical methodologies that are considered. Generally, an asset class will be expected to behave differently from the risk – return point of view in certain market environments, and among the commonly held asset classes nowadays, we can add foreign currencies, commodities, infrastructure and real estate to the ones previously mentioned. Due to a demonstrated low to moderate correlation with other sectors of the stock and bond markets, real estate has become an essential component for diversifying traditional stock bond portfolios. Long term lease contracts that generate income can be an important source and component of a return from a real estate investment, especially in times where yield is at record low levels. Moreover, real estate also benefits from mega trends such as rapid urbanization, aging populations and development in emerging markets.
Two approaches to real estate investment are now widely recognized – private real estate and real estate investment trusts, or REITs. Private real estate simply refers to the direct ownership of commercial property in categories such as apartments, retail, offices or industrial spaces, akin to private equity. On the other hand, private REITs or listed REITs are a special type of real estate security that allows companies to take advantage of special tax structures that minimizes the tax burden, in return for certain restrictions on their operations. This typically allows them to offer higher dividend yields compared to equities with similar risk profiles.
11
In 2017, HSBC valued the residential and commercial real estate global market to stand somewhere above US$220 trillion. However, since most of the world’s households are homeowners, only a fraction of that is considered an alternative investment possibility. MSCI also values the size of the professionally managed real estate investments to have reached US$8.5 trillion in 2017. Pension funds and similar products are estimated to have one of the largest exposures to real estate globally, therefore the behaviour of the global market in the coming years (especially accounting for their behaviour during economic cycles) may become crucial to documenting possible impacts on dividends and pay outs for passive investors, which are the average, mid to high income individuals and families in society. Listed REITs have historically demonstrated a low correlation with other equity sectors or the bond market, therefore they have been proven to be an efficient diversification method to reduce volatility an mitigate losses. The attractiveness of REITs for active investors is also increased by its income generating capability in an environment with low interest rates, as observed in the past decade. While many characteristics differentiate listed REITs from the common stock market, the two also exhibit many similarities. EY Analysts report that the issue of consistent and reliable pricing has become as challenging in REITs as with other common equities – discounts to net asset value (NAV) have widened in 2017 for REITs compared to previous years and implied spreads between higher and lower quality properties have increased. While sectors like healthcare and industrial spaces continue to provide sizeable premiums, other sectors like office spaces often trade at a 5% - 10% discount. These discounts, coupled with major private equity funds and large institutional investors with capital at their disposal have also generated a wave of public to private transactions in the REIT market in 2017. As such, a look is warranted into how the prices and returns of REITs are fundamentally affected by investor behaviour.
Macroeconomic factors like GDP growth, income levels, population growth, interest rates supply and demand functions (Case & Shiller, 1990; Quigley, 1999) have been previously employed to study how property prices are affected. However, as the 2008 financial crisis has proven there can be substantial differences between expected (rational) and actual (market transactions) prices. Evidence of bubbles and price anomalies can be found prior to the 2008 period in different regions as evidence by Chan, Lee & Woo (2001) or Wong, Hui, Seabrook & Raftery (2005). As with the common equity market, the fundamental reasoning process has always been that large differences in a property’s market price versus the rational price are proof of irrational components factoring into the price. Noise traders can drive the price away from the fundamental value of an asset and as the investment markets is becoming increasingly more diverse and complicated due to the rise of technology and globalization, new angles become available to study price changes through investor sentiment.
12
quantifies news media for residential properties into a different type of sentiment measure to measure impact on housing prices during the 2008 financial crisis. However, Lin, Rahman & Yung (2009) study the effect of investor sentiment on US REITs using the Baker & Wurgler (2006) investor sentiment index and find that when investors are optimistic (pessimistic) REIT returns become higher (lower).
Various papers study REIT returns, such as the time series properties of REIT returns (Devaney; 2001), predictability of returns (Nelling & Gyourko, 1998; Li & Wang, 1995) or the return generating process (Chui et al. 2003a; Payne, 2003; Swanson et al. 2002).
Most of the studies mentioned above focus on local real estate markets (such as the US based REITs versus investor sentiment). The aim of this study will be to use a global sentiment index for developed markets to focus on a global REIT market. As such, the examination can have practical investment implications for one of the largest asset classes in the world, where participants can arguably be found across every type of investors from passive individual investors (savers, individual homeowners, pension holders and pension funds) to the more sophisticated real estate developers, banks and institutional investors.
III. Data & Methodology
The analysis conducted in this study will be a comparative one, assessing how investor sentiment affects returns from different asset classes (stocks vs bonds vs REITs) and secondly how investor sentiment affects returns across countries. The main steps outlined by the methodology section are as follows:
(i) Data Collection: Dependent Variables. Collect dependent variables as market level returns for each asset class (stocks, bonds, REITs) for four different countries (United States of America, United Kingdom, Canada, Australia).
(ii) Data Collection: Investor Sentiment Proxies. Collect data to construct a global investor sentiment index, focusing on four main proxies.
13
sentiment index only from the so - called irrational component, identified by the residuals.
(iv) Total Country Level Sentiment Index. Conduct Principal Component Analysis (PCA) on the raw residuals derived from the investor sentiment proxies to construct a total sentiment index for each individual country, resulting in four country level investor sentiment indexes.
(v) Global Sentiment Index. Conduct Principal Component Analysis (PCA) on the four country level investor sentiment indexes to construct a final global index.
(vi) Local Sentiment Index. The previous total country level sentiment index can be separated in two parts: the global component, which has been extracted into a global sentiment index in the previous step, and a local component, representative of the investor sentiment specific only to a particular country. The local component is extracted by storing the residuals from regressing the total country level sentiment index on the global sentiment index and storing the resulting residuals.
(vii) Predictive Regression Analysis. The resulting Global and Local indexes will be used in univariate and multivariate predictive regression analysis to assess the impact investor sentiment can have on the returns of each of the three asset classes selected for this study, at a country level. As a last check, I also introduce the global and local sentiments in the Fama French 3 Factor Model.
The sections below will outline the data and methodology used in each of the steps mentioned above, as well as any additional required rationale in the procedures.
3.1. Data Collection: Dependent Variables (Country vs Asset Class)
The independent variables used for the analysis and their summary statistics are depicted in Table 2 below. All the stocks collected are the constituents of each country’s largest index (USA – S&P500; UK – FTSE100; Canada – S&P/TSX Composite Index; Australia – ASX). The constituents of each respective index are considered representative of the stock market overall due to the fact that in general, a country’s largest index represents over 80% of the country’s entire market capitalization.
For equities as an asset class, for each country I first gathered daily prices from 31/12/1979 – present, averaged the monthly prices and then calculated monthly returns as:
Ri = 100 x log (pt/pt-1) (1)
14 Table 2: Dependent Variables - Summary Statistics
Correlations
Mean SD MIN MAX Stocks Bonds REIT
United States of America USA
Stocks 1.032 3.846 -17.193 13.110 1.000 Bonds -0.483 7.132 -30.718 26.971 0.172 1.000 REIT 0.137 0.815 -4.407 7.299 0.187 0.160 1.000 United Kingdom GBP Stocks 0.932 4.474 -29.672 13.689 1.000 Bonds -0.765 19.655 -290.472 218.883 -0.012 1.000 REIT 0.022 0.355 -0.315 7.168 0.093 -0.714 1.000 Canada CND Stocks 1.013 8.615 -33.305 35.667 1 Bonds -0.645 7.100 -35.883 28.097 0.075 1 REIT 0.019 0.108 -0.385 0.895 0.100 0.066 1 Australia AUS Stocks 1.122 10.741 -52.296 31.345 1 Bonds -0.647 5.382 -29.911 18.046 0.063 1 REIT 18.587 4.993 9.992 27.574 -0.002 -0.089 1
3.2. Data Collection: Investor Sentiment Proxies
The methodology for estimating the global investor sentiment (“IS”) index along with local components for each country considered in the global index follows the steps set out by Baker, Wurgler & Yuan (2012). I selected four main proxies for investor sentiment from the ones mentioned in the literature review section: volatility premium, market turnover, first day return of IPOs and total volume of IPOs, variables which are outlined in Table 4 below. The data for each proxy has been gathered for each individual country that has been considered in the global index construction. Thus, the number of variables for the index construction will be No of Countries x
No of IS proxies, or 4 x 4 = 16.
The countries for the global index have been selected based on the fact that they are all developed economies and have a mature (USA) or established (remaining countries) REIT market1. Table 3 presents all eight countries initially considered; they are all developed
1 The status of the REIT regime and stage market has been analysed and reported by EY in Grinis, M., Kaspar, M.,
15
economies with established financial markets that also boast a significant REIT market. With the aim of having a time series as lengthy as possible to create a consistent dataset with few missing observations, only four of the eight countries were finally selected for the index construction. Panel A denotes the four countries selected for the global index: United States of America, United Kingdom, Canada and Australia. Countries in panel B have been excluded for two main reasons:
(i) The number of IPOs registered (in the Eikon Equity Deals Screener) was much too low compared to the selected countries. It cannot be expected for any other country to match the number of IPOs from the US over the selected sample period (1980 – 2019), however for Canada, UK and Australia the number of IPOs leveled in the thousands, while for the countries in panel B it hovered at 600 and below.
(ii) The countries in panel A are the ones with the oldest REIT markets, all registering real estate investment trusts prior to 1983, while the countries in panel B only created the market after 1985, or 2002 in the case of Singapore.
The data for each of the four investor sentiment proxies has been gathered from Thomson Reuters Eikon and Datastream from as far back as December 31st, 1979. The frequency is monthly, and in order to align variables from all countries the starting point for the analysis has been set at 31/01/1985.
The proxies selected for the global index construction have all been previously examined in literature and confirmed as containing an investor sentiment component. The definitions, mnemonics, and formulas used to apply transformations to the raw data have all been detailed in Table 4. Table 5 display summary statistics for the variables themselves, prior to any transformations being applied (log operations or otherwise).
16
Table 3: Global Investor Sentiment Proxies - Data Parameters, by Country
Country Dependent Variables Independent Variables (Index Construction)
REIT Monthly Returns PVOL TURN RIPO NIPO
PA
N
EL
A
United States of America USA
Stock Exchange NYSE,AMEX, NASDAQ S&P500 S&P500 NYSE,AMEX, NASDAQ NYSE,AMEX, NASDAQ Time period 31/12/1985 - 30/09/2019 31/12/1979 - 30/09/2019 31/12/1979 - 30/09/2019 30/01/1985 - 30/09/2019 30/01/1985 - 30/09/2019
Number of stocks (listings) 789 500 500 14,432 14,432
United Kingdom GBP
Stock Exchange ALL Listed Equities FTSE100 FTSE100 London Stock Exchange London Stock Exchange Time period 31/12/1983 - 30/09/2019 31/12/1979 - 30/09/2019 31/12/1979 - 30/09/2019 31/07/1985 - 30/09/2019 31/07/1985 - 30/09/2019
Number of stocks (listings) 1,470 100 100 2,238 2,238
Canada CND
Stock Exchange ALL Listed Equities S&P/TSX Composite S&P/TSX Composite Toronto Stock Exchange Toronto Stock Exchange Time period 31/12/1985 - 30/09/2019 31/12/1979 - 30/09/2019 31/12/1979 - 30/09/2019 31/07/1985 - 30/09/2019 31/07/1985 - 30/09/2019
Number of stocks (listings) 1,566 250 250 1,537 1,537
Australia AUS
Stock Exchange ALL Listed Equities ASX ASX ASX ASX
Time period 31/12/1984 - 30/09/2019 31/12/1979 - 30/09/2019 31/12/1979 - 30/09/2019 31/12/1985 - 30/09/2019 31/12/1985 - 30/09/2019
Number of stocks (listings) 745 200 200 3,578 3,578
PA
N
EL
B
Germany DEU
Stock Exchange ALL Listed Equities DAX(Deutscher Aktienindex) DAX(Deutscher Aktienindex) Various Stock Exchanges Various Stock Exchanges Time period 31/01/1989 - 30/09/2019 31/12/1979 - 30/09/2019 31/12/1979 - 30/09/2019 30/11/1991 - 30/09/2019 30/11/1991 - 30/09/2019
Number of stocks (listings) 1,964 30 30 676 676
Japan JPN
Stock Exchange ALL Listed Equities Nikkei 225 Nikkei 225 Tokyo & Others Tokyo & Others Time period 31/12/1986 - 30/09/2019 31/12/1979 - 30/09/2019 31/12/1979 - 30/09/2019 31/03/1986 - 30/09/2019 31/03/1986 - 30/09/2019
Number of stocks (listings) 599 225 225 237 237
Singapore SNG
Stock Exchange ALL Listed Equities STI (FTSE) STI (FTSE) Singapore Exchange Singapore Exchange Time period 31/07/2002 - 30/09/2019 31/12/1979 - 30/09/2019 31/12/1979 - 30/09/2019 30/06/1987 - 30/09/2019 30/06/1987 - 30/09/2019
Number of stocks (listings) 705 30 30 1,061 1,061
Netherlands NDL
Stock Exchange ALL Listed Equities AEX AEX Amsterdam, Euronext AM Amsterdam, Euronext AM Time period 31/12/1984 - 30/09/2019 31/12/2001 - 30/09/2019 31/12/2001 - 30/09/2019 30/04/1985 - 30/09/2019 30/04/1985 - 30/09/2019
Number of stocks (listings) 400 296 296 296 296
France FRA
Stock Exchange ALL Listed Equities CAC 40 CAC 40 Paris, Euro Paris Paris, Euro Paris Time period 31/12/1984 - 30/09/2019 31/12/1979 - 30/09/2019 31/12/1979 - 30/09/2019 31/12/1985 - 30/09/2019 31/12/1985 - 30/09/2019
17
Table 4: Global Investor Sentiment Index - Proxies used for Construction
Proxy Formula Source Definition
Volatility
Premium PVOL Log(
𝑃/𝐵𝐻𝑉
𝑃/𝐵𝐿𝑉): Datastream
T The ratio of the price – to – book (market to book) ratio of high volatility stocks to the price – to – book ratio of low volatility stocks.
Market
Turnover TURN 𝐿𝑜𝑔 (
𝑇𝑜𝑡𝑎𝑙 𝑉𝑜𝑙𝑢𝑚𝑒𝑡 𝑀𝑎𝑟𝑘𝑒𝑡 𝐶𝑎𝑝𝑖𝑡𝑎𝑙𝑖𝑧𝑎𝑡𝑖𝑜𝑛𝑡−1
) Datastream
Total Market turnover, i.e. the total dollar volume over one month divided
by the market capitalization at the end of the previous month.
IPO First Day
Return RIPO 100 x log (𝑝𝑡/𝑝𝑡−1) Eikon
The return generated from the closing price one day following a company's IPO. Where First day return was not available, first week returns were substituted. Number of
IPOs NIPO 𝑙𝑜𝑔(𝑇𝑜𝑡𝑎𝑙 𝑁𝑜 𝐼𝑃𝑂𝑠/𝑚𝑜𝑛𝑡ℎ) Eikon Total volume of IPOs in one month.
The second proxy refers to market turnover (TURN) which is calculated as the log of the ratio of total dollar volume of individual stocks over one month divided by the market capitalization of said stock in the previous month. A market level aggregate has been obtained by calculating a market capitalization weighted average for all the stocks of each particular country. This proxy has been largely studied in the literature, as outlined in previous chapters, however the main rationale comes from the fact that when short selling becomes costly, investors prone to sentiment exhibition are more likely to trade when they are opitmistic and drive the overall volume up.
The third proxy is derived from initial public offerings (IPO) data – initial first day returns following an IPO. In markets when the investor sentiment is predominantly high, initial returns can increase exponentially. One example is Japan, when in 1999 registered a return of over 130%, while another one can be the maximum value in the summary statistics table below, where the USA registered a 99% return.
18 Table 5: Summary Statistics for Raw Investor Sentiment Proxies
Correlations
Mean SD MIN MAX PVOL TURN RIPO NIPO
United States of America USA
PVOL -1.167 37.175 -554.754 341.146 1.000 TURN 0.168 0.092 0.036 0.512 -0.072 1.000 RIPO 0.297 1.091 -0.268 13.367 -0.005 -0.025 1.000 NIPO 34.795 25.297 0.000 162.000 -0.003 -0.282 0.001 1.000 United Kingdom GBP PVOL 8.061 32.011 -22.309 193.654 1.000 TURN 0.038 0.038 0.000 0.162 0.029 1.000 RIPO 0.470 5.440 -0.990 99.000 -0.021 -0.100 1.000 NIPO 5.374 5.689 0.000 35.000 -0.005 0.132 0.009 1.000 Canada CND PVOL 0.657 3.501 -59.328 32.768 1.000 TURN 0.033 0.030 0.002 0.227 0.035 1.000 RIPO 0.062 0.957 -0.958 19.000 0.014 0.009 1.000 NIPO 3.746 4.161 0.000 23.000 0.031 0.147 0.010 - 1.000 Australia AUS PVOL 0.833 0.725 0.000 3.014 1.000 TURN 0.054 0.022 0.011 0.132 0.228 1.000 RIPO 0.075 0.193 -0.600 1.800 0.151 -0.205 1.000 NIPO 8.472 8.196 0.000 45.000 0.409 0.172 0.206 1.000 3.3. Orthogonalization
Orthogonalization is derived from Baker & Wurgler (2007), which is the first paper to experiment with an investor sentiment index, if only at the US level. It is assumed that each of the proxies collected above will contain economic fundamentals to a certain extent, such as the IPO variables depending in part on the market environment and the available investment opportunities. In order to avoid any macroeconomic influences in the investor sentiment index, the investor sentiment proxies are first individually regressed on a set of macroeconomic indicators2: inflation3 (CPI), unemployment (UNEMP), industrial production (INDPROD), the 2 All data for macroeocnomic indicators have been collected from the Organisation for Economic Co-operation and
Development.
3 Fama & Schwert (1977); Chen, Roll & Ross (1986) study correlation between inflation as an economic indicator
19
short term rate4(STR) and the term premium5 (TERM). From here on out, the resulting residuals
which will be used as variables as referred to as orthogonalized investor sentiment proxies. The main model for the regression to separate economic fundamentals from the sentiment proxies is:
𝑆𝑃𝑐,𝑡 = 𝑎𝑐+ 𝑏1𝐶𝑃𝐼𝑐,𝑡+ 𝑏2𝑈𝑁𝐸𝑀𝑃𝑐,𝑡+ 𝑏3𝐼𝑁𝐷𝑃𝑅𝑂𝐷𝑐,𝑡+ 𝑏4𝑆𝑇𝑅𝑐,𝑡+ 𝑏5𝑇𝐸𝑅𝑀𝑐,𝑡+ ℇ𝑐,𝑡 (2)
Where:
SP = One of the four Sentiment Proxies (PVOL, TURN, RIPO, NIPO) c = country
t = monthly period
Table 6 below presents descriptive summary statistics for the residuals that will be used as proxies for index construction, derived from running the set of regressions described above. Column (e) through (h) represent the correlation among the orthogonalized sentiment proxies. To be noticed is that while previously the correlation among the raw variables in Table 5 was quite low (surrounding 0.2 values and below) the correlation has now jumped to 0.5 and higher among the new variables. As will be discussed in the next section, having high correlation among variables is crucial when employing Principal Component Analysis to extract information from a set of variables. Lastly, column (i) represents the factor loadings of each variable on the country level total sentiment index, to be discussed in detail in the following section.
3.4. Total Country Level Sentiment Index – Principal Component Analysis (PCA)
Once the orthogonalized investor sentiment proxies are obtained as variables, and it is established that there is a high level of correlation among them, Principal Component Analysis can be employed to construct an investor sentiment index for each country in particular.
Principal Component Analysis is a statistical procedure that converts a set of highly correlated observations into a set of values of linearly uncorrelated variables called principal components. Generally, the aim of PCA is to reduce the dimension one is working in, (such as in cases where the dataset consists of tens if not hundreds of variables) and ensure there will be less of a chance of overfitting. The principal components that result from this analysis are completely new uncorrelated variables and they are structured in such a way that the first principal component has the largest variance. In other words, the first principal component will explain as much of the variance in the dataset as possible. It will identify the main co-movements among the entire set of variables, thus also filtering out much of the noise in individual predictors.
4 Fama & Schwert (1977) analyze short term interest rates and the stock market
20
Table 6: Summary Statistics for Orthogonalized Investor Sentiment Proxies (Residuals)
Correlation % Variation
Explained (PCA) Mean SD Min Max PVOL TURN RIPO NIPO SENTCTOTAL
(a) (b) (c) (d) (e) (f) (g) (h) (i) (g) Panel A: USA PVOL 0.053 0.083 -0.076 0.293 1.000 -0.500 73.4% TURN -0.840 0.227 -1.249 -0.526 0.624 1.000 -0.447 RIPO 0.294 0.059 0.119 0.472 -0.521 -0.411 1.000 0.483 NIPO 1.424 0.101 1.084 1.608 -0.777 -0.631 0.871 1.000 0.563 Panel B: GBP PVOL 0.046 0.220 -0.581 0.676 1.000 -0.269 53.1% TURN -0.782 0.570 -1.783 0.388 0.214 1.000 -0.529 RIPO 0.570 0.572 -0.721 1.810 -0.542 -0.642 1.000 0.625 NIPO 0.561 0.245 -0.163 0.927 -0.151 -0.495 0.430 1.000 0.507 Panel C: CND PVOL 0.041 0.041 -0.072 0.150 1.000 0.480 66.7% TURN -1.641 0.332 -2.450 -1.217 0.289 1.000 0.479 RIPO 0.061 0.110 -0.208 0.349 0.811 0.403 1.000 0.507 NIPO 0.419 0.208 -0.260 0.872 0.474 0.856 0.495 1.000 0.532 PANEL D: AUS PVOL 0.070 0.056 -0.069 0.198 1.000 0.474 89.2% TURN -1.305 0.152 -1.630 -1.100 0.846 1.000 0.524 RIPO 0.077 0.065 -0.055 0.175 0.759 0.979 1.000 0.504 NIPO 0.717 0.357 -0.186 1.137 0.790 0.910 0.847 1.000 0.497
By applying this type of analysis to our data set of orthogonalized sentiment proxies we can extract the first principal component for each country to explain as much of the variance in investor sentiment as possible. The first principal component itself will be the total country level sentiment index. The model for constructing the country level sentiment index can be portrayed as in (3):
𝑆𝐸𝑁𝑇𝑐,𝑡𝑇𝑜𝑡𝑎𝑙 = 𝑏1𝑃𝑉𝑂𝐿𝑐,𝑡+ 𝑏2𝑇𝑈𝑅𝑁𝑐,𝑡+ 𝑏3𝑅𝐼𝑃𝑂𝑐,𝑡+ 𝑏4𝑁𝐼𝑃𝑂𝑐,𝑡 (3)
Where: c = country
t = monthly period
21 -0.600 -0.400 -0.200 0.000 0.200 0.400 0.600 0.800
PVOL TURN RIPO NIPO
GBP
Figure 1: Total Country Sentiment Index - Factor Loadings - GBP
The factor loading displayed in equation (3) refers to how much of the individual variable explains the total variation that the principal components accounts for. Figures 1 through to 4 represent the factor loadings for each country level index. By way of interpretation, we can
Figure 2: Total Country Sentiment Index - Factor Loadings - USA
rewrite each of the country level sentiment indexes as:
SENT𝑈𝑆𝐴𝑇𝑂𝑇𝐴𝐿 = − 0.550𝑃𝑉𝑂𝑙 − 0.447 𝑇𝑈𝑅𝑁 + 0.483𝑅𝐼𝑃𝑂 + 0.563𝑁𝐼𝑃𝑂 SENT𝐺𝐵𝑃𝑇𝑂𝑇𝐴𝐿 = − 0.269𝑃𝑉𝑂𝑙 − 0.529 𝑇𝑈𝑅𝑁 + 0.625𝑅𝐼𝑃𝑂 + 0.507𝑁𝐼𝑃𝑂
SENT𝐶𝑁𝐷𝑇𝑂𝑇𝐴𝐿 = 0.480𝑃𝑉𝑂𝑙 + 0.479 𝑇𝑈𝑅𝑁 + 0.507𝑅𝐼𝑃𝑂 + 0.532𝑁𝐼𝑃𝑂 SENT𝐴𝑈𝑆𝑇𝑂𝑇𝐴𝐿 = 0.474𝑃𝑉𝑂𝑙 + 0.524 𝑇𝑈𝑅𝑁 + 0.504𝑅𝐼𝑃𝑂 + 0.497𝑁𝐼𝑃𝑂
Lastly, in the last column of Table 6 above, we can see the percentage of the dataset’s variation that is explained by the first principal component. We can see that for each country, more than 50% of the variation in the investor sentiment proxy dataset is explained through the first
-0.600 -0.400 -0.200 0.000 0.200 0.400 0.600 0.800
PVOL TURN RIPO NIPO
USA
0.450 0.460 0.470 0.480 0.490 0.500 0.510 0.520 0.530 0.540PVOL TURN RIPO NIPO
CND
0.440 0.450 0.460 0.470 0.480 0.490 0.500 0.510 0.520 0.530PVOL TURN RIPO NIPO
AUS
22
principle component, reaching approximately 70% to 89% in three cases. In other words, the total country sentiment index has retained more than 50% (or 89% in the case of Australia) of the dataset’s explanatory power, while removing the noise contained in each variable.
3.5. Global Investor Sentiment Index
In the previous section we have created four country level investor sentiment indices. The summary statistics of the four resulting variables can be found in Table 7 below.
Table 7: Summary Statistics - Total Country Level Investor Sentiment Index
Correlations
Mean SD Min Max SENT𝑈𝑆𝐴𝑇𝑂𝑇𝐴𝐿 SENT𝐺𝐵𝑃𝑇𝑂𝑇𝐴𝐿 SENT𝐶𝑁𝐷𝑇𝑂𝑇𝐴𝐿 SENT𝐴𝑈𝑆𝑇𝑂𝑇𝐴𝐿 SENTGLOBAL
% Variation Explained (a) (b) (c ) (d) (e ) (f) (g) (h) (i) (g) SENT𝑈𝑆𝐴𝑇𝑂𝑇𝐴𝐿 7.95 1.713 -5.285 3.708 1 0.645 0.897 SENT𝐺𝐵𝑃𝑇𝑂𝑇𝐴𝐿 -6.22 1.457 -3.931 2.351 -0.1907 1 0.542 SENT𝐶𝑁𝐷𝑇𝑂𝑇𝐴𝐿 -6.51 1.633 -4.588 3.234 -0.1735 0.7765 1 0.510 SENT𝐴𝑈𝑆𝑇𝑂𝑇𝐴𝐿 1.79 1.899 -3.954 2.503 -0.6753 0.7936 0.6899 1 0.583
Once again, we can observe high levels of correlation among the total country level sentiment indexes for each nation, therefore principal component analysis can once again be applied with no restriction. Principal Component Analysis is once again applied to the four country level sentiment indexes, in order to construct a Global Investor Sentiment Index. In column (i) of Table 7 above we can see the factor loadings for each country in SENTGLOBAL. We can see that all
countries contribute almost equally to the global investor sentiment index SENTGLOBAL:
SENTGLOBAL = 0.645 SENT 𝑈𝑆𝐴 𝑇𝑂𝑇𝐴𝐿− 0.542 SENT 𝐺𝐵𝑃 𝑇𝑂𝑇𝐴𝐿+ 0.510 SENT 𝐶𝑁𝐷 𝑇𝑂𝑇𝐴𝐿+ 0.583 SENT 𝐴𝑈𝑆 𝑇𝑂𝑇𝐴𝐿
Lastly, Figures 5 through to 9 graphically represent the total country level and global sentiment indexes.
23
Figure 8: AUS - Total Country Level Investor Index
Analyzing the graphs above, we can see that investor sentiment can be unique to each country’s environment, even if commonalities occur. First of all, all graphs record a sudden drop in sentiment in the period 2008 – 2010, reminiscent of the 2008 Financial Crisis. Likewise, the periods 1995 – 2000 also display steady rises in sentiment and then sudden drops, reflective of the Dot.Com bubble. The USA in particular extends the drop in pessimism past the 2000s, most likely reflective of the 2001 attacks. Australia seems to have the least variation (defined as rises and drops on the index, reflective of optimism vs pessimism) compared to other countries, however another common point is that sentiment drops from the 1980’s to the 1990’s across all countries. The USA seems to have the most unique and frequent movements in sentiment, which is expected seeing as it is by far the largest and most established market. Interpretations on the above sentiment indexes will depend mostly on the interpreter’s historical market knowledge
Figure 7: CND - Total Country Level Investor Index
24
and a large degree of hypothesizing, therefore we will continue the analysis with establishing local indexes for each country.
3.6. Local Investor Sentiment Index
The last step in setting up the investor sentiment indexes is to isolate a local investor sentiment index for each country. This last step is required because the previous total country index will have been made up of two main components:
(i) A global sentiment component which has been previously extracted to create SENTGLOBAL;
(ii) An idiosyncratic sentiment component, which will be unique to each country in particular;
Following our previous logic, we can isolate the idiosyncratic sentiment component and construct a last type of index named SENT𝐶𝐿𝑂𝐶𝐴𝐿, where c = country. This will be done by
regressing each total country index on the global investor sentiment index, using the following model:
𝑆𝐸𝑁𝑇𝑐,𝑡𝑇𝑜𝑡𝑎𝑙 = 𝑎𝑐 + 𝑏1𝑆𝐸𝑁𝑇𝑡𝐺𝑙𝑜𝑏𝑎𝑙+ ℇ𝑐,𝑡 (4)
Where: c = country t = monthly period
The resulting residuals from the regression between each country level total index and the global index will be stored and used as local investor sentiment index - 𝑆𝐸𝑁𝑇𝑐,𝑡𝐿𝑂𝐶𝐴𝐿.
Table 8 below represent the summary statistics for each country local sentiment index, while Figures 10 through to 13 are plot of each local index.
Table 8: Local Sentiment Index Summary Statistics - All Countries
Pearson Correlation P(val)
Mean SD Min Max 𝑆𝐸𝑁𝑇𝑈𝑆𝐴𝐿𝑂𝐶𝐴𝐿 𝑆𝐸𝑁𝑇𝐺𝐵𝑃𝐿𝑂𝐶𝐴𝐿 𝑆𝐸𝑁𝑇𝐶𝑁𝐷𝐿𝑂𝐶𝐴𝐿 𝑆𝐸𝑁𝑇𝐴𝑈𝑆𝐿𝑂𝐶𝐴𝐿 𝑆𝐸𝑁𝑇𝑈𝑆𝐴𝐿𝑂𝐶𝐴𝐿 𝑆𝐸𝑁𝑇𝐺𝐵𝑃𝐿𝑂𝐶𝐴𝐿 𝑆𝐸𝑁𝑇𝐶𝑁𝐷𝐿𝑂𝐶𝐴𝐿 𝑆𝐸𝑁𝑇𝐴𝑈𝑆𝐿𝑂𝐶𝐴𝐿 𝑆𝐸𝑁𝑇𝑈𝑆𝐴𝐿𝑂𝐶𝐴𝐿 1.54 1.410 -4.772 2.701 1.000 0.000
𝑆𝐸𝑁𝑇𝐺𝐵𝑃𝐿𝑂𝐶𝐴𝐿 1.08 0.659 -1.968 1.169 0.734 1.000 0.000 0.000
𝑆𝐸𝑁𝑇𝐶𝑁𝐷𝐿𝑂𝐶𝐴𝐿 1.44 0.892 -2.077 3.238 0.673 0.118 1.000 0.000 0.019 0.000
25
Figure 11: Local Investor Sentiment Index - AUS
Figure 13: Local Investor Sentiment Index - USA
We notice first of all that once the global component has been eliminated from the total country level sentiment index, the graphs look dramatically different, especially for Australia and Canada. While previously, SENT𝐶𝑁𝐷𝑇𝑂𝑇𝐴𝐿 and SENT
𝐴𝑈𝑆𝑇𝑂𝑇𝐴𝐿 displayed steady increases in sentiment
after 2000, now 𝑆𝐸𝑁𝑇𝐶𝑁𝐷𝐿𝑂𝐶𝐴𝐿 and 𝑆𝐸𝑁𝑇
𝐴𝑈𝑆𝐿𝑂𝐶𝐴𝐿 present much higher variation in sentiment. The
commonalities across countries previously noted (Dot.Com bubble and the 2008 Financial Crisis) are still present in the local indexes, however we can notice that the investor sentiment in the USA has been disproportionately affected, with the highest decrease from all four countries (from -2 to -4.77), which can be expected given that this was the crisis’ starting point. Moreover, we can now also notice many similarities in trends in investor sentiment between the UK and USA, suggesting the existence of a great degree of the contagion effect between the two countries. Canada also seems to display the highest level of sentiment in 1995, as opposed to the other three countries where some of the highest levels are displayed in the 1999 – 2000 period attributed to the Dot.Com bubble.
Figure 10: Local Investor Sentiment Index - CND
26
3.7. Predictive Regression Analysis
In order to directly asses the predictive power of the constructed global sentiment index and local sentiment indexes, the conventional framework for in – sample analysis is employed in the form univariate and multivariate predictive regressions. The model for assessing the global sentiment index has been set as follows:
𝑅𝑖,𝑐,𝑡+1 = 𝑎𝑖,𝑐+ β1𝑆𝐸𝑁𝑇𝑐,𝑡𝐺𝐿𝑂𝐵𝐴𝐿 + ℇ
𝑡+1 (5)
Where:
R = value weighted return of stocks that constitute the largest index in the respective country, 10Y government bond or aggregate REIT return in excess of the risk - free rate;
i = asset class (stock return, bond return or REIT return) c = country
t = monthly period
𝑆𝐸𝑁𝑇𝑐,𝑡𝐺𝐿𝑂𝐵𝐴𝐿 = Global Investor Sentiment Index ℇ𝑡+1 = mean disturbance term
Moreover, in order to see if the local index, representative of country specific investor
sentiment brings any addition to the model we also employ the following multivariate model: 𝑅𝑖,𝑐,𝑡+1 = 𝑎𝑖,𝑐+ β1𝑆𝐸𝑁𝑇𝑐,𝑡𝐺𝐿𝑂𝐵𝐴𝐿+ β2𝑆𝐸𝑁𝑇𝑐,𝑡𝐿𝑂𝐶𝐴𝐿+ ℇ𝑡+1 (6) Where:
R = value weighted return of stocks that constitute the largest index in the respective country,10Y government bond or aggregate REIT return in excess of the risk - free rate;
i = asset class (stock return, bond return or REIT return) c = country
t = monthly period
𝑆𝐸𝑁𝑇𝑐,𝑡𝐺𝐿𝑂𝐵𝐴𝐿 = Global Investor Sentiment Index
𝑆𝐸𝑁𝑇𝑐,𝑡𝐿𝑂𝐶𝐴𝐿 = Local Investor Sentiment Index ℇ𝑡+1 = mean disturbance term
Under the null hypothesis of no predictability, βi = 0. Baker, Wurgler & Yuan (2012) suggest that
investor sentiment is a contrarian predictor, therefore the alternative hypothesis should be one sided. However, they recommend running a two sided alternative to follow convention when testing relatively new or unknown predictors. Therefore, for the purpose of this exercise the null and alternative hypotheses are:
27
Lastly, we test if the global sentiment and local sentiment indexes have significant effect in a conditional asset pricing model. We introduce each index separately into the Fama - French 3 Factor Model:
𝑅𝑖,𝑐,𝑡+1− 𝑅𝑓,𝑡+1 = 𝑎𝑖,𝑐+ β𝑖,𝑐,𝑡𝑀𝐾𝑇+ β𝑖,𝑐,𝑡 𝑆𝑀𝐵 + + β𝑖,𝑐,𝑡 𝐻𝑀𝐿 + β𝑖,𝑐,𝑡 𝑆𝐸𝑁𝑇𝑐,𝑡𝐺𝐿𝑂𝐵𝐴𝐿+ ℇ𝑡+1(7)
𝑅𝑖,𝑐,𝑡+1− 𝑅𝑓,𝑡+1= 𝑎𝑖,𝑐+ β𝑖,𝑐,𝑡𝑀𝐾𝑇+ β𝑖,𝑐,𝑡 𝑆𝑀𝐵 + + β𝑖,𝑐,𝑡 𝐻𝑀𝐿 + β𝑖,𝑐,𝑡 𝑆𝐸𝑁𝑇𝑐,𝑡𝐿𝑂𝐶𝐴𝐿+ ℇ𝑡+1 (8)
Where:
𝑅𝑖,𝑐,𝑡+1 = value weighted return of stocks that constitute the largest index in the respective country or aggregate
REIT return;
𝑅𝑓,𝑡+1= risk free rate;
i = asset class (stock return, bond return or REIT return) c = country
t = monthly period
𝑆𝐸𝑁𝑇𝑐,𝑡𝐺𝐿𝑂𝐵𝐴𝐿 = Global Investor Sentiment Index
𝑆𝐸𝑁𝑇𝑐,𝑡𝐿𝑂𝐶𝐴𝐿 = Local Investor Sentiment Index ℇ𝑡+1 = mean disturbance term
For each of the stock or REIT returns in the respective countries, we test the joint null and alternative hypotheses for global sentiment:
𝐻𝑂 : β𝑐𝑀𝐾𝑇 = β 𝑐 𝑆𝑀𝐵 = β 𝑐 𝐻𝑀𝐿= β 𝑐 𝑆𝐸𝑁𝑇𝐺 = 0 𝐻𝐴 : : β𝑐𝑀𝐾𝑇 = β𝑐𝑆𝑀𝐵 = β𝑐𝐻𝑀𝐿 = β 𝑐𝑆𝐸𝑁𝑇𝐺 ≠ 0
And local sentiment:
𝐻𝑂 : β𝑐𝑀𝐾𝑇 = β𝑐𝑆𝑀𝐵 = β𝑐𝐻𝑀𝐿 = β 𝑐𝑆𝐸𝑁𝑇𝐿 = 0 𝐻𝐴 : : β𝑐𝑀𝐾𝑇 = β 𝑐 𝑆𝑀𝐵 = β 𝑐 𝐻𝑀𝐿= β 𝑐 𝑆𝐸𝑁𝑇𝐿 ≠ 0
28
economic relevance in terms of asset pricing increases the likelihood that the significant predictive ability of the predictors was not a spurious result.
IV. Results
4.1. Predictive Regression Analysis
Table 9 below showcases the results from the univariate regression analysis with 𝑆𝐸𝑁𝑇𝑐,𝑡𝐺𝐿𝑂𝐵𝐴𝐿as
the predictor in Panel A and 𝑆𝐸𝑁𝑇𝑐,𝑡𝐿𝑂𝐶𝐴𝐿as a predictive variable in Panel B. Results indicate that
the global investor sentiment index is a statistically significant predictor for REIT market level returns in the US, Canada and Australia. Following PCA convention, the sentiment indices (local and global) are standardized, therefore we can interpret the results as for one standard deviation increase in investor sentiment, there will be a 0.046% increase in REIT market level returns in the US, a 0.008% increase in Canada and a 2.578% increase in Australia. In other words, returns for REITs increase during periods of high sentiment (or optimism being displayed by investors). Looking across countries in REITs as an asset class we can see that the Australian REIT market is the most prone to investor sentiment driven changes in returns, with the highest registered increase. This can be explained by the fact that the Australia is a developed economy with a well - established stock market with one of the largest capitalizations in the world, but the REIT market is not yet mature. Even though the US and Canada also exhibit statistically significant increases in return during periods of high sentiment, the level of significance is only of 10%, and the increase is small at the single digit bps level.
29 Table 9: Predictive Univariate Regression Analysis Results
Panel A: 𝑆𝐸𝑁𝑇𝑐,𝑡𝐺𝐿𝑂𝐵𝐴𝐿 Panel B: 𝑆𝐸𝑁𝑇𝑐,𝑡𝐿𝑂𝐶𝐴𝐿
Equity Bonds REIT Equity Bonds REIT
β β β β β β USA 0.013 0.015 *0.046 0.247 0.031 -0.002 (0.123) (0.195) (0.025) (0.143) (0.227) (0.029) GBP -0.157 **-0.357 -0.011 0.186 0.603 -0.002 (0.131) (0.574) (0.011) (0.328) (1.438) (0.012) CND **-0.157 -0.357 **0.008 -0.143 0.266 **-0.009 (0.131) (0.577) (0.003) (0.242) (1.063) (0.006) AUS -0.157 -0.357 ***2.578 ***-0.028 -2.474 ***1.395 (0.131) (0.577) (0.071) (0.418) (1.829) (0.462)
Standard errors in parantheses. Siginificance: ***(1%); **(5%) and *(10%)
Examining the results from the local sentiment index, we observe the same trends we have observed using the global index. First of all, there are more statistically significant results across REITs as an asset class than the other two. Secondly, government bonds as an asset class display no significant results which is in line with the literature. Finally, Australia exhibits more return predictability than the other countries in the sample, with market level returns at a 1% significance across both equities and REITs. It is well worth to note that for equity as an asset class the local investor sentiment is a contrarian predictor, while for REITs the coefficient is positive. In other words, REIT returns increase in high sentiment periods, while equity returns will have a tendency to decrease, though with only 0.028% on average. This can be considered an argument for REIT to remain established as a separate asset class, even though they are essentially listed equities, as they exhibit different behaviour compared to conventional stocks. However, this phenomenon is only exhibited in Australia. We can also see that the Canadian REIT market displays significant results at a 5% level, with market level returns decreasing with -0.009 for a one standard deviation increase in investor sentiment. Therefore, a broader assumption cannot be made about REITs as an asset class. This predictive power of local investor sentiment in Australia can also be justified through the fact that the US is the only mature REIT market, and as previously established in investor sentiment literature, young companies with little history are prone to speculation, over confidence and over valuations which can drive up returns in periods of high sentiment.
30
Canada and the UK, and at a 1% level for Australia. For all three countries, for a one standard deviation increase in investor sentiment market level stock returns will decrease on average with 0.157%. Moreover, global investor sentiment is also significant across REITs as an asset class for US, Canada and once again Australia. The local investor sentiment index remains significant only for REITs and only for Canada and Australia, as previously established. The US is the only country which does not exhibit significance either at a local or global level, except for REITs where there is a result significant at only a 10%, therefore an encompassing assumption can be made that a multivariate model containing both global and local investor sentiment indexes is not effective in predicting US market level returns across any asset class.
However, significance levels for equities and REITs increase when global and investor sentiment indexes are being combined in a multivariate model. Therefore, we can conclude that the model does have predictive power at a market level for equities and REITs in the UK, Canada and Australia.
Table 10: Predictive Multivariate Regression Analysis Results Investor Sentiment Index𝑆𝐸𝑁𝑇𝑐,𝑡𝐺𝐿𝑂𝐵𝐴𝐿; 𝑆𝐸𝑁𝑇𝑐,𝑡𝐿𝑂𝐶𝐴𝐿
Equity Bonds REIT
β
GLOBAL βLOCAL βGLOBAL βLOCAL βGLOBAL βLOCAL
USA 0.013 0 .165 0 .015 0.031 *0.046 -0.029 (0.123) (0.143) (0.195) (0.228) (0.025) (0 .029) GBP **-0.157 0.186 -0.357 0.603 -0.011 -0.002 (0.131) (0.328) (0.578) (1.439) (0.011) (0.012) CND **-0.157 -0.143 -0.356 0.266 **0.008 *-0.009 (0.131) (0.265) (0.578) (1.064) (0.003) (0.006) AUS ***-0.157 -0.028 -0.357 -2.474 ***2.578 ***1.395 (0.131) (0.418) (0.576) (1.830) (0.067) (0.215)
Standard errors in parantheses. Siginificance: ***(1%); **(5%) and *(10%) 4.2. FF3 Model Inclusion
As a last step the global and local indexes have been introduced in a conditional Fama – French 3 Factor (FF3) model to assess market level returns for equities and REITs. Table 11 below displays the results from the model being applied with the global investor sentiment index included. Table 12 displays the results from the FF3 model with the local investor sentiment index included.
31
Therefore, we can feel assured that the global investor sentiment index does not suffer from data mining and the results are not spurious, specifically in the case of the REIT market in Australia. Likewise, in the US we also find all factors are significant for the REIT market, but at a lower 10% level.
Table 11: Fama - French 3 Factor Model Inclusion: Global Investor Sentiment Index
Asset Class: Equity Asset Class: REIT
MKT SMB HML 𝑆𝐸𝑁𝑇𝑐,𝑡𝐺𝐿𝑂𝐵𝐴𝐿 MKT SMB HML 𝑆𝐸𝑁𝑇𝑐,𝑡𝐺𝐿𝑂𝐵𝐴𝐿 USA -0.073 -0.013 -0.002 0.014 *0.161 0.016 *0.026 *0.045 (0.047) (0.068) (0.071) (0.123) (0.009) (0.013) (0.014) (0.025) GBP -0.050 -0.039 -0.055 **-0.155 0.004 0.001 0.008 -0.011 (0.059) (0.073) (0.078) (0.132) (0.004) (0.006) (0.006) (0.011) CND -0.057 -0.039 -0.055 **-0.155 **0.002 0.061 0.001 **0.008 (0.051) (0.073) (0.078) (0.132) (0.012) (0.001) (0.007) (0.003) AUS -0.057 -0.039 -0.055 ***-0.154 *0.049 **-0.079 **-0.090 ***2.585 (0.051) (0.073) (0.078) (0.128) (0.027) (0.039) (0.042) (0.071)
Standard errors in parantheses. Siginificance: ***(1%); **(5%) and *(10%)
When examining the inclusion of a local investor sentiment index in the FF3 model we first notice that there are few results thata re statistically significant. Only market premium is significant for Canada across equities, and the local sentiment index is significant just for Asutralian equities. Looking at reits, we only find three statistically significant results, market premium for canadian REITs, high minus low for USA REITs and local snetiment for Australia. The results are reflective of the previous models, however with such few significant results we cannot make an overarching assumption that the results of the local snetiment index are not at least in part driven by data mining, therefore the local investor sentiment index constructed in this study cannot be fully validated as having significant predictive ability for any particular asset class or country.
Table 12: Fama - French 3 Factor Model Inclusion: Local Investor Sentiment
Asset Class: Equity Asset Class: REIT
MKT SMB HML 𝑆𝐸𝑁𝑇𝑐,𝑡𝐿𝑂𝐶𝐴𝐿 MKT SMB HML 𝑆𝐸𝑁𝑇𝑐,𝑡𝐿𝑂𝐶𝐴𝐿 USA -0.071 -0.016 -0.006 0.015 0.015 0.018 *0.026 -0.003 (0.047) (0.069) (0.073) (0.143) (0.011) (0.014) (0.014) (0.029) GBP -0.048 -0.045 -0.054 -0.171 0.004 0.001 0.008 -0.031 (0.051) (0.073) (0.078) (0.132) (0.004) (0.006) (0.006) (0.027) CND **-0.051 -0.042 -0.051 -0.132 **0.002 0.061 0.002 -0.009 (0.051) (0.073) (0.073) (0.245) (0.013) (0.001) (0.002) (0.006) AUS -0.047 -0.045 -0.054 ***-0.044 0.026 0.021 -0.086 ***1.361 (0.051) (0.073) (0.078) (0.421) (0.056) (0.081) (0.087) (0.464)
