Macroeconomic News and Real Estate Returns

Transcription

1 Macroeconomic News and Real Estate Returns Tim A. Kroencke*, Felix Schindler**, and Bertram I. Steininger*** January 15, 2018 Abstract This paper extends recent evidence for the relationship between macroeconomic news announcements and common stock returns for the real estate market in the U.S. and the U.K. Similarly to findings for common stocks, we find economically large announcement returns for real estate of 0.17% in the U.S. and 0.05% in the U.K. These high average daily returns are not explained by the surprise component released on announcement days. We find that the returns of listed real estate in excess of (or orthogonal to) common stocks are positive on macroeconomic announcement days but virtually zero on all other days. JEL-Classification: E44, G1, G12, G32, R30 Keywords: Macroeconomic risk; macroeconomic news; inflation hedging; listed real estate; real estate risk Acknowledgments: The authors are grateful for comments from Terrance M. Clauretie, William M. Doerner, Thomas M. Springer, Michael Stein, Elaine Worzala, and the seminar participants at the ARES Annual Meeting 2015 and at the ERES Annual Meeting We would like to thank the European Public Real Estate Association (EPRA) for their grant in support of this work and for their thoughtful comments. * University of Basel, t.kroencke@unibas.ch ** Steinbeis University Berlin and ZEW Mannheim, schindler@steinbeis-cres.de *** RWTH Aachen University and ZEW Mannheim, steininger@immo.rwth-aachen.de, Tel.: (corresponding author)

2 Macroeconomic News and Real Estate Returns Abstract This paper extends recent evidence for the relationship between macroeconomic news announcements and common stock returns for the real estate market in the U.S. and the U.K. Similarly to findings for common stocks, we find economically large announcement returns for real estate of 0.17% in the U.S. and 0.05% in the U.K. These high average daily returns are not explained by the surprise component released on announcement days. We find that the returns of listed real estate in excess of (or orthogonal to) common stocks are positive on macroeconomic announcement days but virtually zero on all other days. 2

3 1 Introduction This paper studies the reaction of listed real estate returns with respect to a wide range of macroeconomic news announcements. First, we shed light on the question of whether the release of macroeconomic news has an important effect on the pricing of real estate securities, independent of the news component of the announcement. Recently, Savor and Wilson (2013) found that excess returns of common stocks and bonds are higher on days when important macroeconomic news is scheduled to be announced. We extend their results for the listed real estate market. The second question we want to address is that of what kind of macroeconomic news is important for the pricing of real estate securities. Is this news related to broader economic activity (e.g. GDP), labor market conditions, sentiment and confidence in the market, housing market conditions, or to inflation? The real estate markets which we cover are listed real estate vehicles traded in the U.S. and in the U.K. We investigate aggregated indices of listed real estate (split by property foci) on their own and in comparison with the common stock and bond market. The high-frequency nature of the surprise component of macroeconomic data releases limits our sample to listed real estate vehicles. We find that investors receive a large compensation for holding real estate securities on days when macroeconomic news is released. For the U.S., the macroeconomic announcement day return is 0.17%; for the U.K. it is 0.05%. These numbers are economically large when we compare them to the average returns on non-announcement days, which are virtually zero: 0.004% in the U.S. and 0.003% in the U.K. These results are comparable to the findings documented by Savor and Wilson (2013). A limitation of our study is the relatively short sample period that is available (U.S.: ; U.K.: ). As a result, our tests lack precision and even though the announcement day returns that we document are economically relevant, they are not statistically significant for the U.K. We then investigate how real estate securities respond to the surprise component released on such announcement days. In line with the literature (e.g., Gürkaynak et al. (2005b)), we measure macroeconomic news as the difference between the released value of the macroeconomic variable and its expected value (median consensus expert forecasts). We expect that real estate securities will have a positive sensitivity to news related to an expanding economy (Ling and Naranjo (1997), Karolyi and Sanders (1998), Ewing and Payne (2005), Plazzi et al. 3

4 (2008)) and will be positively related to news about inflation (Hartzell et al. (1987), Dokko et al. (1991), Rubens et al. (1989), and Yobaccio et al. (1995)). Moreover, having split the sample into recession and expansion periods, we proceed to study how the sensitivity of real estate returns with respect to the different kinds of macroeconomic news changes over time. This part of the analysis aims at shedding light on several highly debated issues. For example, if real estate securities become less sensitive towards macroeconomic news during booms, this might indicate that other factors which are possibly more behavioral in nature are more important during such times. It follows that such patterns may prove to be helpful in identifying periods of mispricing. Another interesting question that we want to address with this framework is related to the inflation-hedging property of real estate assets. We expect that real estate assets will provide inflation protection in times when the inflation risk is high, but not when there is no risk of soaring price levels. First, in line with Savor and Wilson (2013), we find that adding surprise components of different macroeconomic news releases does not explain the high average returns on announcement days or the even higher average returns that are specific to real estate. Interestingly, we find that real estate returns in the U.S. are, as expected, positively related to news about retail sales (an early indicator of GDP) and news about inflation. We also find evidence that real estate returns are more sensitive to this news in recessions than they are in expansions. However, evidence relating to the other macroeconomic news or to the U.K. is rather inconclusive. Some variables show the expected sign; other variables that do not show the expected sign or results are almost never consistent across the two countries that we study. Overall, the results on the sensitivity of real estate securities are as similarly inconclusive as recent evidence provided for common stocks is (e.g., Savor and Wilson (2013)). One possible explanation might be that consensus expert forecasts do not accurately reflect the expectations priced in security markets (e.g., Savor and Wilson (2013)). Another possible explanation could be that the sensitivity of asset prices is not stable but varies rather over time (e.g., Swanson and Williams (2014)). Nevertheless, we hope that our results will point towards new avenues for future research into the relationship between macroeconomic risk and real estate securities. 4

5 2 Literature review Boyd et al. (2005) find that the news surprise on macroeconomic days explains some variation of asset returns. In contrast, Savor and Wilson (2013) show that stock returns are significantly higher on days when important macroeconomic news (inflation, unemployment, or interest rate) are announced independent of the news component of the announcement. For the period between 1958 and 2009, they find that excess returns for stocks are on average 10.3 basis points higher on such days than on normal days. Flannery and Protopapadakis (2002) analyze the sensitivity of 17 different macroeconomic surprises to stock returns by using a generalized autoregressive conditional heteroscedasticity (GARCH) model. By doing so, they can distinguish between a direct effect of the surprise component as well as a second indirect channel through the volatility of the shock on the returns. Their findings are that changes in the consumer price index (CPI) and the producer price index (PPI) as well as monetary aggregates influence stock returns. The monetary aggregate is the only macroeconomic variable which also influences stock returns through a higher conditional volatility. Conversely, Brenner et al. (2009) could not confirm statistically significant effects on excess returns of stocks, on Treasury bonds, or on corporate bonds when using a GARCH model. Jones et al. (1998) analyze excess returns of Treasury bonds and conclude that PPI and employment news have an effect on mean and volatility of returns. Previous empirical literature finds mixed evidence concerning how asset returns react to macroeconomic news announcements. Black (1997), Black (2009), and Boudoukh et al. (1994) find that the answer to the question of whether monetary policy influences the real economy or not remains unclear, at best. On the other hand, Bernanke and Kuttner (2005) utilize a high frequency and market-based proxy of monetary policy news and are able to show that monetary policy has an economically sizeable effect on the stock market. There is also related research which studies the real estate market. Growing wealth should have a positive influence on the real estate market. As general stocks, listed real estate returns are also dependent on unexpected changes in macroeconomic factors. According to the results of Ling and Naranjo (1997), unexpected changes in macroeconomic factors, such as production and output, induce lower real estate returns. Karolyi and Sanders (1998) compare the predictability of stock, bond, and Real Estate Investment Trust (REIT) returns by using a multi-beta asset pricing model with a broad set of economic variables. They find that these economic variables are highly important for explaining the returns of all these assets and that traditional multi-beta asset pricing models cannot capture the important economic risk premium for REITs. By 5

6 combining firm-specific variables and macroeconomic factors in an asset pricing model, Chen et al. (1998) observe that the unanticipated change in term structure can help to explain return variations in a model that even excludes firm-specific variables. Ewing and Payne (2005) show that shocks to monetary policy, economic growth, and inflation are associated with lower expected returns of REITs. On the other side, a shock to the default risk premium leads to higher returns. Plazzi et al. (2008) analyze the cross-sectional variation of real estate returns and growth in rents for commercial real estate. They find that macroeconomic variables can explain the time-series fluctuation of the returns and they identify the credit channel as a driver for these effects. 3 Empirical Framework The empirical framework of our study aims to estimate whether and how real estate returns are dependent on macroeconomic announcement days and the surprise component of such days. For a preliminary overview, the following equation sums up our regression model: r t = α + γa t + β t S t + θ t C t + ε t. (1) Our approach closely follows Savor and Wilson (2013) and Swanson and Williams (2014) but has three steps: First, we identify the announcement days (A t ) and the surprise component of major macroeconomic announcements (S t ) in the U.S. and the U.K. Second, we estimate the average sensitivity of an announcement day (γ) to returns (r t ) mainly for different real estate returns. Third, we compare the time-invariant sensitivity (β) and the time-varying sensitivity (β t ) during expansion and recession periods of the different returns and determine when and to what extent each series is influenced through macroeconomic news surprises. To control for other factors, we also add a set of control variables (C t ) to our model. 3.1 Asset returns We start by measuring the impact of the macroeconomic announcement day risk and surprise on real estate returns (r t ). In most specifications, the real estate returns are dissected into REIT-specific and sector-specific returns (office, residential, and specialty). Moreover, we construct three different long-short portfolios in order to disentangle the pure effects on real estate. The first portfolio is long in real estate and short in stocks extracting the effects on real estate relative to the stock market. The second portfolio is also long in real estate but 6

7 short in bonds extracting the effects on real estate relative to the bond market. And in the third portfolio, we calculate residual returns, which are orthogonal to the stock and bond market, in order to observe the effects which only influence the real estate market. We convert all returns (r t ) into basis points (1 bp = 0.01% = ). 3.2 Macroeconomic announcements The announcement day regressor (A t ) has a specific value of one on announcement days of macroeconomic variables in our set, and zero on all other days. The macroeconomic variables that we cover are released at a quarterly or monthly frequency. Following Gürkaynak et al. (2005a), we compute the surprise component of each macroeconomic announcement (S t ) as the difference between the realized value of the macroeconomic variable and the forecasted value (the median forecast). The surprise component is standardized by its historical standard deviation such that the beta coefficients (β) are in units of basis points per standard deviation of each macroeconomic surprise. 3.3 Announcement day risk premium To test the macroeconomic announcement day risk premium for different real estate returns, we reduce our regression to the following form: r t = α + γa t + θ t C t + ε t, (2) whereby the returns (r t ) are only explained by the vector of announcement days (A t ) without the surprise component (S t ). Afterwards, we stepwise add a set of control variables (C t ) in order to see whether a potential announcement day premium remains economically large: Model 1 is without any control variables; Model 2 takes the last trading day's average mean and the squared average mean of the market premium into account. The market is represented by a broad country-specific equity index, and the risk-free rate asset by a 30-day cash index. 1 In Model 3, we also add day-of-the-week dummies (Monday-Thursday) to the model. Consequently, we are able to control for any trading-day-specific anomaly regarding the trading volume, cash flows, or any behavioral effect, such as the higher uncertainty over weekends. In 1 For the U.S., we use the Russel 3000 and the S&P UK index for the U.K. to proxy the market returns. The 30-day cash indices are obtained from J.P. Morgan for both countries. 7

8 accordance with the literature (e.g., French (1980)), returns on Mondays tend to underperform returns on Fridays (weekend effect). Flannery and Protopapadakis (1988) conclude that the day-of-the-week variation could be explained by the market discount rate s term premia. Longer term securities will be more exposed to the weekend effect if term premia systematically increase over a weekend. 3.4 Sensitivity towards macroeconomic announcements Finally, we extend our approach by examining whether the announcement day premium is robust after controlling for a vector of unexpected news surprise of each macroeconomic factor (S t ). In the first Model S, the sensitivity of macroeconomic announcement surprise is timeinvariant (β). This relationship can be described by the following equation: r t = α + γa t + βs t + θ t C t + ε t. (3) In the next step, we set the sensitivity time-varying (Model T) over the business cycle. 2 To this end, we add an interaction term for the business cycle (expansion and recession) and the surprises to the model. This specification allows us to identify time-varying sensitivity to macroeconomic news. For example, we can address the inflation-hedging ability of real estate assets during expansion and recessions within this framework. To identify the time-varying sensitivity of macroeconomic announcement surprises, the following model applies: r t = α + γa t + β t S t + θ t C t + ε t. (4) 4 Data 4.1 Returns Our data set is based on listed real estate firms in the U.S. from January 1997 to December 2014 and in the U.K. from January 2005 to December We mainly analyze two groups of return series. First, purely listed real estate returns for the entire real estate market or subsectors (REIT, office, residential, and specialty). Second, listed real estate returns in excess of (orthogonal to) other asset classes (common stocks and bonds). 2 The dates of business cycle expansion and recession periods are obtained from the National Bureau of Economic Research (NBER) for the U.S. and from the Organisation for Economic Co-operation and Development (OECD) for the U.K. 8

9 All asset price data are obtained on a daily frequency from Thomson Reuters Datastream. The daily frequency nature of the surprise component of macroeconomic data releases limits our sample to listed real estate vehicles. For the real estate market, we use aggregated indices of REITs and real estate operating companies (REOCs) of the Datastream index family traded in the U.S. and the U.K. to calculate the daily return. The aggregated indices of listed real estate also cover different property foci (all real estate, REIT, office, residential, and specialty). To compare our results with the stock and bond market and with previous research, we obtain for the U.S. the capitalization-weighted and broad Russel 3000 index and the large cap S&P 500 index, as well as the 10 Year U.S. Government Index, all provided by Datastream. For the U.K., we use S&P United Kingdom as a proxy for the entire market and FTSE 100 for blue chips, and the 10 Year U.K. Government Index for the bond market. For a better interpretation of the small daily returns, we convert all returns into basis points. 4.2 Macroeconomic news Macroeconomic surprises are computed as the forecasted variable minus the announced variables. For the U.S., except for the monetary policy events, both values are obtained from the World Economic Calendar and Economic Indicators (WECO). 3 For the U.K., we use Reuters Poll from Thomson Reuters Eikon as our data source. In the next step, the surprise components of macroeconomic news releases are computed as the difference between the released value and the forecasted value of the macroeconomic variable. The latter value is based on the median of a consensus-based survey among financial market institutions and professional forecasters. Afterwards, the surprise is standardized by its historical standard deviation. In the case of monetary policy data, we directly obtained the announcement day and bank rate decision from the website of the respective monetary committee the FOMC (Federal Open Market Committee) in the U.S. and the MPC (Monetary Policy Committee) in the U.K. As do Bernanke and Kuttner (2005) and Savor and Wilson (2013), we use the price of 30 day federal fund future traded on CME as a price-implied interest rate prediction of the fund rate in the U.S. As long as there is not such a highly traded equivalent for the base rate in the U.K., we rely on the latest quarterly forecast of Oxford Economics. The surprise component of the interest rate is calculated with the same method as are the other macroeconomic variables be- 3 In WECO, Bloomberg comprises historical data on actual and forecasted macroeconomic statistics for the G7 countries. The data can be accessed directly at and originate from Econoday. 9

10 tween the actual and forecasted values. Our selection of macroeconomic variables is influenced by the availability of the data and previous studies. The 7 news surprises which we cover are: quarterly change of GDP, monthly change of CPI (inflation), monthly unemployment rate, monthly consumer confidence index, index or monthly change of a housing index, monthly change of retail sales index, and the monetary policy target rate. 4 Further details regarding the original publication source, the definition, and release day and time of the macroeconomic variables that we use can be found in the Appendix. 5 Results We provide results for the entire listed real estate market and the sector indices between January 1997 and December 2014 in the U.S. and between January 2005 and December 2014 in the U.K. Additionally, we analyze returns of listed real estate in excess of (or orthogonal to) common stock and bond returns. First, we use stepwise three models to clarify the effects of the announcement day, independent of the news component of the announcement. Second, we test two additional specifications to see whether the surprise component of macroeconomic announcements can explain the variation of listed real estate returns. 5.1 Descriptive statistics Table 1 reports the distribution of the news over trading days during the whole period. In the U.S., most macroeconomic data are released on a Tuesday. According to the literature (e.g., French (1980)), mostly Mondays show significantly lower returns compared to Fridays, so that the returns on announcement days in our sample are not systematically lower. To take a possible Monday effect into account, we add day-of-the-week dummies to our model. Macroeconomic news is released on roughly 24% of the 4,530 trading days. Therefore, only one macroeconomic announcement is released on 21.7% of the days, two announcements on 1.8%, and three announcements on 0.1% of the days. In the U.K., most macroeconomic data are released on a Thursday. Macroeconomic news is released on roughly 27.4% of the 2,524 trading days. Therefore, only one macroeconomic announcement is released on 25.6% of the days, two announcements on 1.7%, and three announcements on 0.1% of the days. Table 1: Distribution over trading days 4 To gauge the importance of each variable from the viewpoint of market participants, we also take the variables' importance into account as measured by how many professionals have subscribed to each variable in Bloomberg. Even if other studies (e.g., Jones et al. (1998)) do find that the surprise component of the producer price index (PPI) helps to explain the return variation of assets, we do not include this variable as an explaining factor as long as the time series of forecasted PPI is too short. 10

11 Panel A: U.S. Panel B: U.K. No. of Macro Day Obs. % Obs. % Events per Day Monday , Tuesday Wednesday Thursday Friday Sum 1, Sum 4, Day Obs. % No. of Macro Events per Day Obs. % Monday , Tuesday Wednesday Thursday Friday Sum Sum 2, The left panels present the distribution of announcements over weekdays. The right panels show how often macroeconomic events are reported per day. The abbreviation Obs. stands for absolute number of observations, % for the relative frequency, and No. for number. The U.S. data between 1997 and 2014 are in Panel A; the U.K. data between in Panel B. Table 2 shows an overview of the descriptive statistics (mean and standard deviation) of the macroeconomic variables for all three types, which we indirectly need in order to calculate the standardized surprise components the actual released value, the median of the forecasted values, and the surprise. The unit we use for each macroeconomic variable corresponds to the original unit in the WECO or the Reuters Poll dataset. On average, the surprises show a tendency to an overestimation: a finding which is well documented in the literature (e.g., for inflation, by Diebold et al. (1997)). Table 2: Descriptive statistics of macroeconomic variables Panel A: U.S. Publication type Variable Actual Median Surprise GDP Mean St. dev Panel B: U.K. Publication type Variable Actual Median Surprise GDP Mean St. dev Infl. Mean St. dev Infl. Mean St. dev Unemp. Mean St. dev Unemp. Mean St. dev Conf. Mean St. dev Conf. Mean St. dev Housing Mean St. dev Housing Mean St. dev RSX Mean St. dev RSX Mean St. dev FED Mean St. dev Mean St. dev This table presents the mean and standard deviation (Std. dev.) of the actual released value, the median of the forecasted value, and the surprise component of the used macroeconomic variables: The acronym GDP stands for the quarterly change of gross domestic product; Infl. for the monthly change of consumer price index (CPI); Unemp. for the monthly unemployment rate; Conf. for the value of monthly consumer confidence index; Housing for a housing index (index value of NAHB s housing market index in the U.S. and monthly change in Nationwide s house price index in the U.K.); RSX for the monthly change in the retail sales index; and FED or BoE for the base rate of the respective central bank (Board of Governors of The Federal Reserve System and Bank of England, respectively). The unit (index, %, or decimal) and whether a growth rate is annualized depends on the original information in WECO or Reuters Poll, respectively. The U.S. data between 1997 and 2014 are in Panel A; the U.K. data between in Panel B. BoE In Table 3, we address the possible issue of multicollinearity of the exogenous variables. Even if the actual macroeconomic variables do co-vary over time in their original announcement 11

12 frequency (e.g., ρ > 0.5 for GDP 5 and Confidence in the U.K.), our daily surprises are not exposed to high correlation values. The maximal correlation in our dataset is between GDP and the retail sales index with a value of around 0.13 in the U.K. This correlation coefficient is the only one which is statistically different from 0 at the 5% level of significance. Table 3: Correlation matrix of macroeconomic variables Panel A: U.S. GDP Surprise 1.00 Infl. Surprise Panel B: U.K. GDP Infl. Unemp. Conf. Housing RSX FED Unemp. Surprise Conf. Surprise Housing Surprise RSX Surprise FED Surprise GDP Surprise 1.00 Infl. Surprise GDP Infl. Unemp. Conf. Housing RSX BoE Unemp. Surprise Conf. Surprise Housing Surprise RSX Surprise -0.13* BoE Surprise This table displays the cross-correlation between different macroeconomic variables. The macroeconomic surprise variables are the same as in Table 2. The U.S. data between 1997 and 2014 are in Panel A; the U.K. data between in Panel B. * indicates a correlation coefficient s significance at the 5% level or better. 5.2 Announcement day returns and effects Univariate test results Figure 1 shows the boxplots of daily returns on announcement days and non-announcement days across the common stock, bond, and real estate market in the U.S. and the U.K. In Table 4, we outline the descriptive statistics (means and standard deviation) of asset returns (common stock, bond, and real estate) and the equal mean test for returns on announcement days and non-announcement days. We focus the discussion on real estate and provide results on common stocks and bonds for completeness. For the U.S. real estate market, the median of daily returns on announcement days is 13.1 bp; the mean is 16.7 bp and significantly different from zero (p-value = 0.006). On non- 5 We assume that the quarterly change in GDP is equally distributed over a quarter. 12

13 announcement days, the median is 5.0 bp; the mean is 0.4 bp and not significantly different from zero (p-value = 0.892). Levene s robust test statistic for the equality of variances reports unequal variances with a highest p-value of for all alternative location estimators (mean, median, and 10% trimmed mean). Using Welch's formula for unequal variances, the mean test shows unequal means with a p-value of For the U.S., we also find significant announcement returns for common stocks and bonds in line with previous research. For the U.K. real estate market, the median of daily returns on announcement days is 6.2 bp; the mean is 5.0 bp and not significantly different from zero (p-value = 0.426). On nonannouncement days, the median is 4.4 bp; the mean is 0.3 bp and not significantly different from zero (p-value = 0.929). Levene s robust test statistic for the equality of variances indicates equal variances with a lowest p-value of for all alternative location estimators (mean, median, and 10% trimmed mean). The mean test indicates equal means because we cannot reject the null hypothesis of equal means with a p-value of We do not find significant and economically important announcement returns for common stocks and bonds in the U.K. In Table 4, we analyze two more indices for common stocks (broad market indices; the Russel 3000 for the U.S. and the S&P UK for the U.K.) and for real estate (country-specific REIT indices). Their test statistics confirm the findings of the other market returns in both countries. These first results indicate that the returns on announcement days are higher for real estate returns. Figure 1: Boxplots of daily returns Panel A: U.S. Panel B: U.K. This figure shows the boxplots for the daily returns (expressed in basis points) of stocks, bonds, and real estate on announcement days (Ann.) and non-announcement days (Non-Ann.) displaying the median, box = quartiles, bars = 1.5 interquartile range; outliers are not displayed. The U.S. data between 1997 and 2014 are in Panel A; the U.K. data between in Panel B. 13

14 Table 4: Descriptive statistics of asset returns Panel A: U.S. Panel B: U.K. Stocks Ann. Non-Ann. Diff. Stocks Ann. Non-Ann. Diff. Russell 3000 Mean Std. dev (0.029) S&P 500 Mean Std. dev (0.033) Bonds Gvt 10 Years Mean Std. dev (0.049) Real Estate All RE Mean Std. dev (0.011) REITs Mean Std. dev (0.028) N 1,071 3,459 UK all Mean Std. dev (0.571) FTSE100 Mean Std. dev (0.572) Bonds Gvt 10 Years Mean Std. dev (0.876) Real Estate All RE Mean Std. dev (0.455) REITs Mean Std. dev (0.420) N 691 1,833 This table presents the summary statistics for different daily stock, bond, and real estate index returns (expressed in basis points) on announcement days (Ann.) and non-announcement days (Non-Ann.). Std. dev. stands for standard deviation of the returns; N stands for the number of observations. The Diff. column displays the difference between the return values on announcement days and non-announcement days. The p-values for the equal mean t-test are in parentheses in the last column. N stands for number of observations. The U.S. data between 1997 and 2014 are in Panel A; the U.K. data between in Panel B Multivariate test results In Table 5, we use ordinary least squares (OLS) regressions in order to test whether a potential announcement day premium remains economically large by controlling for last trading day's average mean and squared average mean of the market premium (Model 2). In Model 3, we add day-of-the-week dummies to control for a possible weekend effect. For comparison, Model 1 reproduces our univariate difference-in-mean test of the previous section. T-statistics are based on Newey-West (1987) standard errors (5 lags). The results of all three models for all asset returns in the U.S., except for Model 3 and REIT, confirm our findings from the univariate tests. The returns on announcement are significantly higher than on nonannouncement days in the U.S. Similarly, we find economically high announcement returns in the U.K.; however, these are not statistically significant. The last trading day's average mean factor is almost statistically significant in every model for each asset class. Interestingly and particularly for real estate returns, some of the day-ofthe-week dummies are also economically large and significant. These days are Monday, Wednesday, and Thursday in the U.S. and Monday and Thursday in the U.K. especially, the Monday-effect in the U.K. is prevalent. The intercept is only statistically significant for bond returns in the U.S. and the U.K., which is in line with other recent studies (e.g., Andersen et al. (2007), Savor and Wilson (2013)). The results for sector-specific real estate returns (office, 14

15 residential, and specialty) are similar to those based on the entire real estate market in the U.S. and the U.K. 6 In summary, the announcement day premium remains economically large and statistically significant for all asset returns after controlling for market premium and day-of-the-week dummies in the U.S. For real estate, the announcement day premium is economically large and significant. For the U.K., we find an economically important announcement return for real estate; however, these average returns lack statistical significance at conventional levels. Table 5: Regression analysis of asset returns Panel A: U.S. Stocks Bonds Real Estate Russel 3000 S&P 500 Gvt 10 Years All RE REIT Variable Ann. Day 9.98** 9.67** 9.17** 9.64** 9.32** 9.15** 2.66** 2.67** 3.12** 17.02** 16.57** 12.87* 14.72** 14.28** (2.182) (2.162) (2.001) (2.136) (2.117) (2.032) (1.983) (1.990) (2.218) (2.495) (2.510) (1.921) (2.147) (2.156) (1.600) MKTRF t *** -0.05*** -0.06*** -0.06*** -0.02*** -0.02*** -0.15*** -0.15*** -0.16*** -0.16*** (-2.628) (-2.630) (-3.156) (-3.163) (-3.977) (-3.970) (-3.180) (-3.159) (-3.542) (-3.522) (MKTRF t-1) ** 0.00** 0.00** 0.00** (2.140) (2.137) (2.189) (2.193) (-0.074) (-0.029) (1.427) (1.414) (1.438) (1.426) Monday (-0.095) (0.421) (0.295) (-1.432) (-1.265) Tuesday (0.732) (1.039) (-1.537) (0.160) (0.212) Wednesday (0.392) (0.578) (-1.273) (-1.353) (-1.218) Thursday (0.412) (0.590) (-0.200) (-1.150) (-1.163) Intercept *** 2.11*** 3.05** (0.353) (-0.641) (-0.762) (0.358) (-0.654) (-1.118) (3.148) (3.223) (2.158) (-0.124) (-0.985) (0.626) (0.211) (-0.732) (0.687) N R 2 (%) Panel B: U.K. Stocks Bonds Real Estate S&P UK FTSE 100 Gvt 10 Years All RE REIT Variable Ann. Day (-0.575) (-0.532) (-0.451) (-0.573) (-0.528) (-0.457) (-0.155) (-0.187) (0.473) (0.732) (0.689) (0.250) (0.785) (0.752) (0.410) MKTRF t * -0.04* -0.04* -0.04* * 0.07* (-1.685) (-1.681) (-1.791) (-1.787) (1.547) (1.552) (1.687) (1.710) (1.341) (1.356) (MKTRF t-1) (1.254) (1.236) (1.259) (1.240) (0.358) (0.344) (0.408) (0.388) (0.513) (0.494) Monday *** ** (-0.379) (-0.346) (1.434) (-2.651) (-2.421) Tuesday (0.478) (0.550) (0.213) (-0.116) (-0.028) Wednesday (-0.251) (-0.205) (-1.082) (-0.386) (-0.235) Thursday (-0.617) (-0.566) (-0.630) (-1.298) (-1.399) Intercept ** 1.93* (1.367) (0.536) (0.465) (1.354) (0.503) (0.401) (2.022) (1.654) (0.689) (-0.065) (-0.245) (1.210) (-0.059) (-0.272) (1.080) N R 2 (%) This table presents the results of OLS regressions of daily asset returns (expressed in basis points) in three model specifications. The assets are Russel 3000, S&P 500, 10 Year Datastream U.S. Government, Thomson Reuters Datastream Real Estate, and Thomson Reuters Datastream REIT for the U.S. in Panel A. For the U.K. in Panel B, S&P United Kingdom, FTSE 100, 10 Year Datastream U.K. Government, Thomson Reuters Datastream Real Estate, and Thomson Reuters Datastream REIT. Ann. day is a dummy variable for announcement days of all macroeconomic variables in our dataset. Market premium (MKTRF) is the difference between a country-specific equity index (Russel 3000 and S&P United Kingdom) and a risk-free 6 The corresponding tables are available from the authors upon request. 15

16 rate asset (30-day JPM cash index). Monday-Thursday are day-of-the-week dummies. N stands for number of observations and R 2 for coefficient of determination. t-statistics based on Newey-West (1987) standard errors (with 5 lags) are in parentheses. The U.S. data between 1997 and 2014 are in Panel A; the U.K. data between in Panel B. * 10% significance level, ** 5% significance level, and *** 1% significance level. 5.3 Time-invariant and time-varying sensitivity In the next step, we focus on the real estate market (entire real estate including real estate operating companies, REITs as well as the sectors office, residential, and specialty sectors) and check whether the high average returns can be explained by the macroeconomic surprise components released on announcement days (Model S) and how they interact with the business cycle (Model T). For both models, we also include all control variables of Model 3 in Table 5 (unreported in Table 6). For the U.S., we find that all real estate returns remain economically large on announcement days after controlling for macroeconomic surprises (model S). As expected, news about retail sales (RSX; an early indicator for GDP) has a positive and statistically significant influence; surprises about the unemployment rate influences real estate returns negatively. Allowing the sensitivities to vary in recessions and expansions does not result in a significant reduction of announcement returns. In particular, we find that news about RSX, inflation, and housing are positively related with returns during recession periods. In contrast, GDP, employment, and confidence sentiment are negatively related with returns, but positively during recession periods. During expansion periods, news about inflation and the federal funds rate (FED) are negatively associated with returns. Taking the statistical significance into account, the most important macroeconomic variables in the U.S. are: RSX, inflation, FED, GDP, and unemployment. In the U.K., news about unemployment and housing are statistically significant and positive connected with real estate returns (Model S). Regarding the level of returns on announcement days, the U.K. specialty sector behaves in a similar way to the U.S. real estate market. In Model T, when decomposing the effect of surprise between expansion and recession periods, we find that news about the bank rate (BoE) during expansion, and inflation during recession periods, influences real estate returns negatively. Unemployment during expansion and GDP during recession periods drive real estate returns positively. 16

17 Summing up, unemployment, GDP, and BoE are the statistically most important macroeconomic news in the U.K. For housing, RSX, and FED/BoE, the measured effects point in the same direction as for the U.S. However, the influence of macroeconomic surprises regarding GDP, inflation, unemployment, and confidence sentiment are not consistent across the two analyzed countries and are sometimes counterintuitive. Table 6: Regression analysis for macroeconomic surprises Panel A: U.S. All RE REIT Office Residential Speciality Variable Model S Model T Model S Model T Model S Model T Model S Model T Model S Model T Ann. Day 12.96* 14.02** * * * 12.92** (1.936) (2.130) (1.613) (1.809) (1.573) (1.817) (1.395) (1.626) (1.939) (2.068) GDP Surprise (-0.640) (-0.771) (-0.596) (-1.388) (-0.700) Infl. Surprise (0.172) (0.325) (0.321) (0.592) (-0.070) Unemp. Surprise * (-1.389) (-1.333) (-1.323) (-1.790) (-1.121) Conf. Surprise (-0.802) (-0.764) (-0.368) (-0.930) (-0.912) Housing Surprise * (1.326) (1.363) (1.816) (1.322) (1.421) RSX Surprise 63.22*** 64.52*** 76.47*** 69.97*** 52.21*** (2.730) (2.766) (2.830) (2.743) (2.627) FED Surprise (-0.544) (-0.523) (-0.505) (-0.448) (-0.487) GDP x Exp (0.410) (0.241) (0.056) (-0.660) (0.249) GDP x Rec * * ** * (-1.761) (-1.751) (-0.958) (-2.172) (-1.870) Infl. x Exp ** ** ** *** (-2.405) (-2.160) (-2.174) (-1.548) (-2.633) Infl. x Rec (1.075) (1.093) (1.044) (1.093) (1.097) Unemp. x Exp (-0.141) (-0.031) (0.032) (-0.269) (0.197) Unemp. x Rec * * ** ** * (-1.940) (-1.943) (-2.036) (-2.351) (-1.886) Conf. x Exp (0.357) (0.528) (0.550) (0.422) (0.780) Conf. x Rec (-1.001) (-1.006) (-0.544) (-1.186) (-1.276) Housing x Exp (1.181) (1.220) (1.553) (1.011) (1.409) Housing x Rec (0.604) (0.643) (1.135) (0.942) (0.453) RSX x Exp (0.768) (0.804) (0.524) (0.579) (1.091) RSX x Rec *** *** *** *** ** (2.905) (2.945) (3.302) (3.058) (2.576) FED x Exp ** ** * * ** (-2.344) (-2.414) (-1.791) (-1.758) (-2.160) FED x Rec (-0.129) (-0.124) (-0.255) (-0.110) (-0.148) Intercept (0.653) (0.879) (0.716) (0.948) (0.934) (1.144) (0.835) (1.096) (0.098) (0.324) N R 2 (%)

18 Panel B: U.K. All RE REIT Office Residential Speciality Variable Model S Model T Model S Model T Model S Model T Model S Model T Model S Model T Ann. Day (0.278) (0.904) (0.445) (1.063) (0.509) (1.117) (0.212) (0.785) (1.240) (1.587) GDP Surprise *** (1.467) (1.490) (1.262) (1.583) (2.863) Infl. Surprise (-1.281) (-1.366) (-1.445) (-1.256) (-1.625) Unemp. Surprise 29.80*** 30.89*** 39.01*** 25.06*** 3.66 (10.546) (10.592) (11.619) (8.690) (1.196) Conf. Surprise (1.204) (1.163) (1.458) (0.810) (0.789) Housing Surprise * * (1.592) (1.658) (1.408) (1.895) (1.591) RSX Surprise (0.222) (0.206) (-0.138) (0.409) (0.142) BoE Surprise *** (-0.564) (-0.478) (-0.348) (-0.378) (-2.581) GDP x Exp *** (-0.227) (-0.030) (-0.282) (-0.018) (3.325) GDP x Rec ** ** ** *** (2.389) (2.530) (1.969) (2.977) (1.571) Infl. x Exp (0.806) (0.665) (0.615) (0.569) (0.118) Infl. x Rec * * * *** (-1.656) (-1.697) (-1.681) (-1.621) (-3.808) Unemp. x Exp *** 31.24*** 39.84*** 24.97*** 4.65** (14.857) (15.002) (17.319) (11.396) (2.244) Unemp. x Rec (-0.201) (-0.185) (-0.528) (0.082) (-0.819) Conf. x Exp (0.891) (0.885) (1.116) (0.683) (0.311) Conf. x Rec (0.781) (0.729) (0.930) (0.411) (0.923) Housing x Exp * (1.188) (1.330) (1.030) (1.503) (1.706) Housing x Rec (1.361) (1.368) (1.213) (1.526) (1.248) RSX x Exp (0.970) (0.965) (0.806) (0.952) (0.902) RSX x Rec (0.463) (0.448) (-0.079) (0.803) (-0.313) BoE x Exp ** ** *** * (-2.354) (-2.468) (-2.898) (-1.936) (-0.962) BoE x Rec ** (-0.262) (-0.180) (0.035) (-0.171) (-2.344) Intercept (1.313) (1.453) (1.183) (1.315) (1.302) (1.403) (1.090) (1.243) (1.264) (1.314) N R 2 (%) This table presents the results of OLS regressions of various daily real estate returns (expressed in basis points) in two model specifications. Model S focuses on the surprise of macroeconomic variables on announcement days, Model T also on the impact of the business cycle (Exp. for expansion and Rec. for recession). The assets are obtained from Thomson Reuters Datastream. Ann. day is a dummy variable for announcement days of all macroeconomic variables in our dataset. The other control variables from Table 5 are included but not reported. The macroeconomic surprise variables are the same as in Table 2. N stands for number of observations and R 2 for coefficient of determination. t-statistics based on Newey-West (1987) standard errors (with 5 lags) are in parentheses. The U.S. data between 1997 and 2014 are in Panel A; the U.K. data between in Panel B. * 10% significance level, ** 5% significance level, and *** 1% significance level. 18

19 5.4 Long-short and orthogonal portfolios To disentangle the pure real estate impacts, we also construct three different long-short portfolios: The first portfolio is long in real estate and short in stocks so that we isolate the real estate element relative to the stock market. The second portfolio is also long in real estate but short in bonds so that we isolate the real estate element relative to the bond market. And in the third portfolio, which is long in real estate and short in stocks and bonds, we calculate residual returns which are orthogonal to the stock and bond market. We start with multivariate tests (Models 1-3) in analogy to Table 5. 7 The results of all three portfolios confirm high average returns on announcement days. The last trading day's average mean factor is almost in every model and portfolio statistically significant. Relative to the stock market, the day-of-the-week dummies are statistically significant and negative for Monday, Wednesday, and Thursday in the U.S. and for Monday in the U.K. In Table 7, we present the results in relation to the surprise (Model S) and the business cycle (Model T) for all three portfolios. The announcement day premium remains economically large (except for U.K. real estate minus bonds). For all long-short portfolios, the RSX surprise has a positive effect on the pure real estate return in the U.S. This holds true for inflation for real estate minus common stocks and to the orthogonal portfolio. In the U.K., surprises about unemployment, inflation, and housing are statistically important for all three portfolios. Once again, results on the sensitivity of real estate securities are rather inconclusive. Some variables show the expected sign; other variables do not show the expected sign, and results are frequently not consistent across the two countries that we study. Overall, we find that real estate returns in excess of (orthogonal to) common stocks are completely earned on announcement days and are virtually zero on all other days. These results hold for the U.S. and the U.K., and cannot be explained by the surprise component of macroeconomic news announcements. 7 This table is not reported but is available from the authors upon request. 19

20 Table 7: Regression analysis for long-short portfolios Panel A: U.S. Real Estate-Stocks Real Estate-Bonds Real Estate Ortho. Variable Model S Model T Model S Model T Model S Model T Ann. Day (0.774) (0.930) (1.419) (1.616) (0.755) (0.914) GDP Surprise (-0.534) (-0.517) (-0.522) Infl. Surprise 20.66* * (1.686) (0.246) (1.668) Unemp. Surprise * (-1.275) (-1.712) (-1.339) Conf. Surprise (-1.263) (-0.773) (-1.251) Housing Surprise (1.077) (1.445) (1.109) RSX Surprise *** 23.90* (1.632) (2.846) (1.703) FED Surprise (-0.248) (-0.517) (-0.255) GDP x Exp (0.127) (0.492) (0.144) GDP x Rec ** * ** (-2.000) (-1.660) (-1.991) Infl. x Exp ** (1.244) (-2.283) (1.211) Infl. x Rec (1.029) (1.100) (1.037) Unemp. x Exp (-0.439) (-0.306) (-0.455) Unemp. x Rec * ** * (-1.683) (-2.307) (-1.766) Conf. x Exp (-0.802) (0.693) (-0.759) Conf. x Rec (-1.044) (-1.097) (-1.063) Housing x Exp (0.190) (1.354) (0.251) Housing x Rec (1.587) (0.626) (1.560) RSX x Exp (-0.503) (1.451) (-0.397) RSX x Rec ** *** 83.52** (1.997) (2.781) (2.019) FED x Exp ** (-0.252) (-2.431) (-0.335) FED x Rec (-0.238) (-0.045) (-0.222) Intercept (1.491) (1.575) (0.086) (0.316) (1.333) (1.422) N R 2 (%)