Stefan Mero. Bachelor of Economics (Hons) Business School Accounting and Finance

Transcription

1 Identifying macroeconomic determinants of daily equity market returns An Australian study Stefan Mero Bachelor of Economics (Hons) 2016 This thesis is presented for the degree of Master of Philosophy at The University of Western Australia Business School Accounting and Finance

2

3

4

5 Abstract Understanding macroeconomic risk is a fundamental aspect of economic and financial decision-making. More recently, attention has turned to identifying macroeconomic variables as risk factors (Chen, Ross & Roll 1986; Chan, Karceski & Lakonishok 1998; Flannery & Protopapdakis 2002). Most research, to date, has focused on the relationship between macroeconomic data values and stock market prices over long time horizons. This study extends the existing Australian stock market based literature by examining the relationship between macroeconomic news and stock market returns/return volatility at daily a level, in an event study framework. The study covers the period before, during and after the Global Financial Crisis in 2008 to determine whether the effects of news differ during different phases of stock market activity. In the stock market boom leading up to the Crisis, higher than expected overnight cash rate news was found to have a negative impact on stock returns that disappears in the subsequent period of subdued stock market price growth after the Crisis. Macroeconomic fundamentals - such as unemployment, the consumer price index and real gross domestic product - matter only after the onset of the Crisis. Over the whole period, consumer sentiment and real gross domestic product surprises are the only macroeconomic variables to impact stock market volatility.

6 Contents 1 Introduction Theoretical, Empirical and Industry Perspectives Thesis Contribution Thesis Structure Literature Review Theoretical Background Australian Studies Australian Stock Market Returns Australian Stock Market Return Volatility The Australian Stock Market and Efficient Market Hypothesis Foreign Studies Foreign Stock Markets and Macroeconomic Surprises Business Cycles and Macroeconomic Factor relationships with Stock Markets Conclusions from the Literature Hypothesis Unemployment Balance of Trade Retail Sales Producer Price Index Consumer Price Index Real Gross Domestic Product Overnight Cash Rate Consumer Sentiment Methodology Returns Surprises (Unexpected Components of Announcements) Control Variables Returns Estimation Volatility Estimation Data Stock Market Indices Stationarity of Stock Returns August 2016 i

7 5.2 Macroeconomic Surprises Forecasts Announcements Surprises Control Variables Results Continuous Model Results Dummy Variable Based Model Results Continuous Model Results: Pre- and Post-Global Financial Crisis Dummy Variable Based Model Results: Pre- and Post-Global Financial Crisis Robustness Tests Summary and Discussion of Results Conclusion Thesis Contribution Main Results Limitations and Possible Extensions Final Conclusion References Appendices Appendix A Structure of Macroeconomic Announcement Data and Dates Appendix B Model Fitting Fitting ARMA for Stock Market Return Modelling Fitting GARCH/EGARCH for Stock Market Return and Time Varying Volatility Modelling Appendix C Robustness Tests All Ordinaries Index Based Regressions Single Macroeconomic Variable Regressions Alternate Break-Point Regressions August 2016 ii

8 Tables Table 1 Australian Literature Review Summary Table 2 Foreign Literature Review Summary Table 3 ASX 200 Index Total Daily Returns Summary Statistics Table 4 All Ordinaries Index Total Daily Returns Summary Statistics Table 5 Augmented Dickey-Fuller Unit Root Tests - No Drift or Trend Table 6 Money Market Services Consensus Macroeconomic Forecasts Table 7 Other Macroeconomic Forecasts Table 8 Macroeconomic Announcement Values Table 9 Summary Statistics for Macroeconomic Surprises Table 10 Real GDP Growth ADF Test and Akaike Information Criterion Table 11 Consumer Sentiment ADF Tests and Akaike Information Criteria. 100 Table 12 Continuous EGARCH model results based on full period sample Table 13 Dummy variable EGARCH model results based on full period sample Table 14 Continuous EGARCH model results: Pre- and Post-GFC Table 15 Dummy variable EGARCH model results: Pre/Post-GFC Table 16 Summary of Results by Macroeconomic Variable Table 17 Summary of Results Surviving Robustness Tests Table 18 Summary of Results - Returns Table 19 Summary of Results - Return Volatility Table 20 Macroeconomic Announcement Date Structure Table 21 Raw Macroeconomic Announcement Data Series Structure Table 22 AIC - All Ordinaries Total Returns ARMA Regression Table 23 Q-Statistics on ARMA Model Squared Standardised Residuals Table 24 ARCH LM Test on ARMA Squared Residuals Table 25 Continuous Model using All Ordinaries Index based Returns Table 26 Dummy Variable Model using All Ordinaries Index Table 27 Continuous Model using All Ordinaries Index: Pre- and Post-GFC Table 28 Dummy Variable Model using All Ordinaries Index: Pre- and Post- GFC Table 29 Unemployment Dummy Variable based Regression: Post-GFC Table 30 Retail Sales Dummy Variable based Regression: Pre-GFC Table 31 Producer Price Index Continuous Regression Table 32 Consumer Price Index Continuous Regression Table 33 Consumer Price Index Continuous Regression: Post-GFC August 2016 iii

9 Table 34 Consumer Price Index Dummy Variable based Regression: Post- GFC Table 35 Real GDP Dummy Variable based Regression Table 36 Real GDP Dummy based Regression: Pre- and Post-GFC Table 37 Overnight Cash Rate Continuous Regression: Pre-GFC Table 38 Overnight Cash Rate Dummy Variable based Regression: Pre-GFC. 203 Table 39 Consumer Sentiment Index Continuous Regression Table 40 Consumer Sentiment Index Dummy Variable based Regression Table 41 Table 42 Table 43 Consumer Sentiment Index Dummy Variable based Regression: Post- GFC Continuous Model Results using Alternate Breakpoints: Pre- and Post- GFC Dummy Variable Model Results using Alternate Breakpoints: Pre- and Post-GFC August 2016 iv

10 Figures Figure 1 ASX 200 Index Total Daily Returns Figure 2 Australian All Ordinaries Index Total Daily Returns Figure 3 Unemployment Rate Surprises Figure 4 Balance of Trade Surprises Figure 5 Retail Sales Surprises Figure 6 Producer Price Index Surprises Figure 7 Consumer Price Index Surprises Figure 8 Real GDP Surprises Figure 9 Interest Rate Surprises Figure 10 Consumer Sentiment Surprises Figure 11 Brent Crude Oil One-Month Futures Prices and Returns Figure 12 Lagged US Standard and Poor s 500 Index Returns Figure 13 Term Spread - Australian Commonwealth Government Bonds Figure 14 5-Year Australian Corporate Bond Default Spread Figure 15 ASX 200 Index - Sector Composition Figure 16 ASX 200 Index - Total Daily Returns Figure 17 Westpac-Melbourne Institute Consumer Sentiment Index Figure 18 Consumer Sentiment Surprises Figure 19 EGARCH Normal Distribution Quantile Plot August 2016 v

11 Acknowledgements My supervisors, Professor Richard Heaney and Dr. Joey Wenling Yang, spent many hours reading, editing, analysing data and advising me. I am thankful for their efforts, patience and ability to stimulate a creative learning and research environment. My research is a richer tapestry of findings on account of Professor Heaney s ability to encourage creative thinking, and his guidance on research design. Dr Yang has improved my understanding of financial econometrics, academic conventions in drafting research and skills in managing the scope of research. My editor, Eleanor Mulder and my father, Jonn Mero, also spent hours reading, editing and providing useful comments on my drafting. My employer, Greg Watkinson, also provided useful suggestions regarding my written communication of ideas and document structure. These people improved the readability of my thesis immeasurably. Finally, I would like to thank Adam Hearman and Robyn Oliver, for minimising the administrative burden and thereby making my experience as a research student at the university all the more pleasant. 10 August

12 1 Introduction Understanding macroeconomic risk, in order to price individual assets, is a fundamental aspect of economic and financial decision-making. 1 More recently, attention has turned to identifying macroeconomic variables as risk factors (Chen, Ross & Roll 1986; Chan, Karceski & Lakonishok 1998; Flannery & Protopapdakis 2002). A well-functioning financial system should incorporate important macroeconomic information in stock prices and returns quickly and rationally, or in the words of Fama (1970, p.383), the market should be semi-strong form efficient. The returns on an equity market index in an economy with a well-functioning financial system should, therefore, respond quickly to important macroeconomic risk factors because the index is comprised of individual firms. The effect of these risk factors should be identifiable in short run returns. The idea that certain macroeconomic variables are risk factors in stock market returns is well accepted from a theoretical, empirical and industrial perspective. 1.1 Theoretical, Empirical and Industry Perspectives From a theoretical perspective, the share price of individual firms that are used to construct a stock market index, within an economy, are affected by broader economic conditions. This is because economic conditions generally affect the expected future earnings and dividends of individual firms. In addition, expected future earnings and dividends are related to the firm s share price through the required rate of return (Gordon 1962; Campbell & Shiller 1988). The required rate of return itself may also be affected by macroeconomic factors (Ross 1976). This means any variable affecting 1 For example see Markowitz 1952, Treynor, cited in Ross 1976, p.341, Sharpe 1964, Lintner 1965, Fama & French 1992, Black 1972, Ross 1976, Jagannathan & Wang August

13 earnings, dividends, or the required (or expected) future rate of a firm s return, also affects a broad market index made up of firms. Market indices should therefore respond to macroeconomic news (Fama 1981; Schwert 1990). From an empirical perspective, the link between macroeconomic variables and stock returns is supported by evidence from major economies, such as the United States (US), Europe and Japan. In the US, Fama (1981) observed that future real activity, measured by industrial production and real Gross National Product (GNP), eliminated the explanatory power of inflation when included as a variable in a regression used for explaining stock returns. Schwert (1990) confirmed the relationship between growth rates in future production and stock returns, which was discovered by Fama (1981) using 100 years of data. An early application of Arbitrage Pricing Theory (APT) by Chen, Ross & Roll (1986) identified macroeconomic variables and nonequity asset returns as risk factors when explaining equity returns. Industrial production, changes in risk premia, the term structure of the yield curve and also inflation were found to be significant explanatory factors. Cheung and Ng (1998) concluded that future real GNP growth has a significant positive influence on US stock returns. Ratanapakorn & Sharma (2007) showed money supply, industrial production, inflation, the Japanese Yen/US dollar exchange rate and short-term interest rates are positively related to US stock returns, while long-term interest rates are negatively related to stock returns. The relationship between US stock prices and the two variables - industrial production and long-term interest rates - is also found in a later study by Humpe & Macmillan (2009). However, in contrast to Ratanapakorn & Sharma (2007), Humpe & Macmillan (2009, p.118) found a negative relationship between inflation and stock returns. 10 August

14 In Europe, Asprem (1989) examined the relationship between the major stock index and macroeconomic variables in ten different countries. Interest rates, inflation, imports and (perhaps surprisingly) employment were shown as negatively related to stock prices, whereas changes in future industrial production and broad money supply were shown as positively related to changes in stock prices. The results of the cointegrating techniques employed by Cheung and Ng (1998, p.293), for German stock returns, indicated future real GNP growth has a significant positive influence on returns. Following the spectacular rise of the Tokyo Stock Exchange (TSE) leading up to 1990, Mukherjee and Naka (1995) studied the long-term relationship between Japanese macroeconomic variables and TSE index based returns. Their model found, in the long run, local currency depreciation, money supply, industrial production, and short-term interest rates are positively related to stock market returns, while inflation and long-term interest rates are negatively related. Cheung & Ng (1998) reported similar results for Japan, finding lagged money supply and future real GNP are both positively related to Japanese stock market returns. Humpe and Macmillan (2009) also documented Japanese stock returns have a positive relationship with industrial production. Unlike Mukherjee and Naka (1995), Humpe and Macmillan (2009, p.118) found Japanese returns have a negative relationship with the money supply in Japan. Turning to an industrial perspective, Australian financial media coverage on the share market frequently and continuously attributes changes in daily returns to macroeconomic variables, such as employment: Shares on Thursday fell for a fourth straight day, but ended well off the day's lows, thanks to strong jobs data (Cauchi 2015) 10 August

15 Real GDP: Stocks ended near unchanged today after a rally spurred by surprisingly strong growth in the nation's economy (Australian Associated Press 31 October 2003) Inflation and interest rates: The Australian share market is expected to open strongly tomorrow with financial markets awaiting key inflation figures that will offer clues surrounding possible interest rate rises (Carter 2007). This highlights it is generally accepted that certain macroeconomic variables are risk factors in Australian stock returns. In this study, I seek to identify and quantify the effect of such risk factors in Australian stock returns. I will also assess whether their effects on stock returns differ between rapid and more subdued periods of stock price growth, as theory and evidence suggest this may be the case (Shiller 2003; Binswanger 2004). 1.2 Thesis Contribution The most common methods employed to detect macroeconomic risk factors in stock returns are based on APT, present value, or cointegration models, which tend to focus on the relationships between macroeconomic variables and stock market returns over time horizons far longer than one day. Few studies examine the effect of the unexpected component of macroeconomic announcements (news or surprises) on stock market returns over the short run. Most existing research tends to focus on the relationship between announced macroeconomic data values and stock market prices over the long run. The distinction between announced macroeconomic data and news is an important one. Announced macroeconomic data is the reported value for a macroeconomic variable, typically by a statistical agency or other sovereign authority. 10 August

16 Macroeconomic news, however, is the difference between the market participant's expected value of the macroeconomic variable and the announced value. I assume if there is no difference, then no news exists news by definition must be new information, or in other words, a surprise. Assuming stock markets are efficient, and an estimate of the market participant s expected value of a macroeconomic variable reflects broader market expectations, news associated with this variable should quickly cause stock market prices to change if it is a risk factor. Any information that is not new is (by assumption) already reflected in the current market price (Fama 1970, 1991). Those studies that have incorporated news or innovations typically do so within a cointegrating framework, which again, have a focus on long run relationships (Cheung & Ng 1998; Humpe & Macmillan 2009). Additionally, a minority of the macroeconomic risk factor research uses the event study methodology, which focuses on a window of time around an event, such as a macroeconomic announcement, to examine the extent to which macroeconomic news is incorporated in prices. Studies examining the effect of news on short run stock returns, for a fairly comprehensive set of macroeconomic variables, have been carried out for the US, United Kingdom (UK) and also for some European markets (Wasserfallen 1989; Becker, Finnerty and Friedman 1995; Flannery & Protopapadakis 2002). Australian studies examining the effects of macroeconomic variables on returns over short time periods (daily) have, to date, focused on a limited number of macroeconomic variables (Singh 1993; Singh 1995; Brooks et al 1999; Kim & In 2002; Akhtar et al 2011; Hasan & Ratti 2012). Akhtar et al (2011) found a relationship between consumer sentiment and Australian stock market returns, and Hasan and Ratti (2012) found a relationship between oil prices and Australian stock returns. While Kim and In (2002) found Australian stock return volatility was higher on real GDP 10 August

17 announcement days, the relationships between short run stock market returns and fundamental macroeconomic variables, such as unemployment, the consumer price index (CPI) and output, are yet to be established in Australia. 2 My literature review yields eight macroeconomic variables as candidates for examination: (1) unemployment, (2) balance of trade, (3) retail sales, (4) producer price index (PPI), (5) CPI, (6) real Gross Domestic Product (GDP), (7) overnight cash rates and (8) consumer sentiment. I examine the effects of these variables on stock market index returns and volatility using the event study methodology, undertaken within a regression framework. The regression framework is an exponential generalised autoregressive conditional heteroscedasticity (GARCH) specification. To construct macroeconomic surprises, announcement data is sourced from the Australian Bureau of Statistics (ABS), the Reserve Bank of Australia (RBA) and Westpac-Melbourne Institute. The corresponding expected values for the announcements are either sourced from Money Market Services (MMS) Australia or modelled using an autoregressive integrated moving average (ARIMA) model, which is based on prior observations of announcements. A time series of surprises is then constructed for each macroeconomic variable as the difference between the announcements and their corresponding expected value. I employ two different variants of the models used to explain returns. The first variant uses continuous macroeconomic surprise values and attempts to measure the sensitivity of returns/return volatility to the change in magnitude of macroeconomic surprises. Put another way, this variant of the model measures the per cent change in returns/return volatility per one per cent of error in macroeconomic forecasts. The second model assigns a dummy variable to each macroeconomic announcement day, 2 Kim and In (2002, p.578) found some evidence that real GDP news days are positively related to futures returns, but not stock returns based on spot prices. 10 August

18 which is equal to one only if the macroeconomic surprise is not equal to zero and zero otherwise. This captures the average effect of macroeconomic surprises on stock returns. An additional dummy variable is assigned to macroeconomic surprises assumed to be bad news, which captures additional information on whether bad news has a different effect to good news. Good and bad news in this context does not relate to presupposed effects on returns, rather, it is an assumed perception of whether the news is a good or bad sign for the economy. My assumption of what constitutes good and bad surprises follows Kim (2003, p.619) for all variables except cash rates and consumer sentiment, which were not included in his study. For cash rates, I assume the perspective of a leveraged entity, and in this case, higher cash rates are deemed bad news because of higher interest payments and capital costs more generally. For consumer sentiment, I assume the perspective of an entity that relies on sales activity. Higher levels of consumer sentiment mean good news because of higher consumer spending and, therefore, higher sales. The continuous model finds robust relationships for the CPI and consumer sentiment. The dummy variable based model detects robust relationships for unemployment, the CPI, real GDP and the overnight cash rate. These results largely corroborate findings in the reviewed literature that conclude real GDP has, in particular, a strong positive relationship with stock market returns. My study benefits from access to data with a large number of observations falling over a period of prolonged stock market expansion, contraction and subsequent subdued growth following the Global Financial Crisis (GFC) in This enables me to examine whether relationships differ during these phases of stock market activity, as is empirically observed by Binswanger (2004) in foreign markets. 10 August

19 With the exception of the overnight cash rate, all of the significant macroeconomic variables appear to explain stock market returns only in the period following the GFC. Lagged stock returns appear to play a greater role than fundamental macroeconomic factors during the stock market boom leading up to the GFC. This is consistent with Binswanger s (2004, p.248) finding that fundamentals cease to explain stock prices during stock market booms. Amongst the variables with significant coefficients, all news assumed to be good economic news is associated with increased returns, while all news assumed to be bad economic news is associated with decreased returns. Real GDP and the consumer sentiment index significantly influence return volatility both before and after the GFC. Real GDP news exhibits asymmetric effects, which shows good news is more important than bad news. 1.3 Thesis Structure In Chapter 2, I review the relevant literature. I begin by reviewing some of the most relevant theories, such as the APT, the efficient markets hypothesis (EMH) and the event study methodology. An overview of studies on the influence of macroeconomic variables, specific to Australian share market returns and return volatility, follows. Foreign studies, specifically examining the effect of macroeconomic surprises on stock markets, are also reviewed because of the similarity of their research design with my study. I finish by drawing conclusions from the literature that impact my research design. The relationships between macroeconomic variables and stock returns, observed in the literature, are reserved for discussion in Chapter 3. In this chapter, I outline my null hypotheses with respect to the effect of each of the macroeconomic variables' associated surprises on Australian stock returns, along with details of alternative hypotheses. Chapter 4 explains how data is processed and how the econometric 10 August

20 models are used to test for relationships between macroeconomic variables and stock returns. The data sources and their characteristics are one of the most important parts of my study. These are presented and discussed at some length in Chapter 5. Chapter 6 is a discussion of the results vis-à-vis the hypothesis chapter, followed by an overview of robustness tests and a discussion on the results surviving the robustness tests. Conclusions are drawn in Chapter 7, which also highlights some areas for future research. 10 August

21 2 Literature Review My literature review begins by outlining some of the theories most relevant to my study. I note the issues found in the literature that studies these theories, and their implications for this study. Studies that examine the relationship between macroeconomic variables and the stock market in Australia are reviewed with a focus on the aspects relevant to my study. Foreign studies that examine the relationship between macroeconomic variables and stock markets, specifically using macroeconomic surprises, are also discussed in a similar way. Finally, conclusions are drawn from the literature, and the contribution of my thesis is explained. 2.1 Theoretical Background Below are some well-established theories and a description of their relevance to my study. Arbitrage pricing theory proposes the expected return on a particular asset can be explained by risk premia that are associated with a number of macroeconomic risk factors, as well as the risk free rate of return (Ross 1976). Macroeconomic risk factors are identified using a statistical process. If the price of an asset differs from that predicted by the model, an arbitrage opportunity exists, and in an efficient market, this is rapidly taken advantage of. My study has parallels to the multifactor model proposed by Ross (1976) because it identifies potential macroeconomic risk factors and uses them to explain returns. Ross s (1976) focus, however, is on stock market returns in excess of the risk free rate, whereas my study examines the return on the stock market index without deducting the risk free rate. Also, unlike Ross (1976), I examine whether prices react quickly to the unexpected component of macroeconomic risk factor announcements. 10 August

22 The efficient market hypothesis (EMH) is important to my study because it relates to the speed and degree to which financial market prices incorporate new information or surprises. My study is based on the assumption this happens rapidly in an efficient stock market. If a macroeconomic announcement has an effect on stock returns, then I expect this to be reflected in stock prices on the day of the announcement. The EMH postulates a capital market is efficient if prices always fully reflect all available information (Fama 1970, p.383). To establish the point at which this hypothesis breaks down, Fama (1970) reviewed tests based on three subsets of information: weak form tests based on historical prices; semi-strong form tests based on obviously available (public) information, such as company and economic announcements; and strong-form tests based on privately available information. No important empirical evidence was found to disprove the hypothesis that security prices reflected the first two information sets. However, limited evidence that went against the hypothesis, tested on the strong-form set of information, was found. Semistrong form tests, in particular, are concerned with the speed at which prices adjust to publicly available information. Tests based on company and macroeconomic announcements indicated prices reacted at the time of the announcement. Additionally, there is some evidence to suggest prices moved in anticipation of the announcement, and these movements appeared to be unbiased. My study proceeds on the assumption stock prices incorporate the semi-strong form information set. The cost of getting prices to reflect information is not always zero, and this is explicitly accounted for in Fama s 1991 study. Consequently, prices are hypothesised to reflect information, often to the point where marginal profits, made from acting on 10 August

23 information, are offset by the marginal cost. Additionally, Fama (1991, p.1575) emphasised a test of market efficiency requires an equilibrium asset price determined by a model that itself may be the source of pricing errors. A rejection of the EMH could be a result of a bad pricing model and/or market inefficiency. This was referred to as the joint-hypothesis problem. The weak, semi-strong and strong form tests were replaced by the following three classifications of research identified in the literature: tests for return predictability; event studies; and test for private information. Tests for return predictability and event studies are the most relevant to my study, and so, the most pertinent of Fama s (1991) findings on these tests are outlined below. Aside from testing the EMH, tests for return predictability are important in the formulation of the models used in my study when establishing a relationship between stock returns and macroeconomic announcements. To maximise the possibility of detecting such relationships, the effect of other variables that affect or predict returns needs to be both accounted and controlled for when testing. Tests for return predictability, reviewed in Fama (1991, p.1578), included tests for autocorrelation (that is, the effect of past returns on future returns). The studies found significant positive autocorrelation, and it was more prevalent in those market indices with many small stocks. This point is important when considering whether a stock index with many smaller capitalised stocks should be used in a study of stock returns. This issue is revisited when selecting a stock return index. With respect to market efficiency, however, Fama (1991, p.1609) noted the predictable part of returns was only a small proportion of the variance, and could not warrant a conclusion of substantial market inefficiency. Incidentally, the studies reviewed also showed 10 August

24 differences between return variances for trading and non-trading hours. This is linked to differences in the flow of information collected during trading and non-trading hours. The findings on autocorrelation and differing variances around non-trading hours, such as holidays, indicate these effects should be and are controlled for in my study. Fama s (1991, p.1586) review acknowledged tests based on the volatility of returns are a useful way to show expected returns can vary through time. The tests reviewed, however, could not explain whether this variation is rational (and therefore efficient) or a result of irrational bubbles. Early literature reviewed provided evidence of several cases of return seasonality, in addition to periods such as holidays. These included the day-of-the-week effect, intra-day effects and the January effect. This evidence highlights such effects may also need to be controlled for. 3 With respect to market efficiency, Monday, holiday and end-of-month effects were small compared to the bid-ask spread of the average stock, while the January effect was small in relation to the bid-ask spread of small stocks. Fama s (1991, p.1587) view was these effects are market microstructure anomalies, and so, observation of these effects need not necessarily result in the rejection of the hypothesis of market efficiency. Event studies, reviewed in Fama (1991, p.1601), benefited from the use of high frequency data, which allowed a more precise measurement of the speed at which stock prices respond to a given event. It also substantially assisted the study to overcome the joint-hypothesis problem, particularly when stock price responses were large and concentrated in a few days. This is because the issue of finding an assetpricing model that correctly measures daily returns is not so critical for statistical 3 This is controlled for in my model by the use of holiday dummy variables. See Section August

25 inference when the abnormal returns are very pronounced and expected return variation is small. The typical result from these studies is stock prices appear to adjust within a day of event announcements. This suggests day-to-day changes in stock prices are suitable for my study. This speed of adjustment was viewed to be consistent with the hypothesis of market efficiency. Event studies still do not entirely resolve issues relating to market uncertainty and the joint hypothesis problem. A common finding in the review of event studies is the dispersion of returns increases around information events. Event studies only explain the average variation around events, while the residual variation is unexplained. It is not, therefore, possible to determine whether the remaining increase in variation is a rational reaction to uncertainty about new fundamentals or irrational over/underreaction, and thus, indicative of inefficiency. This suggests using a model that accounts for changes in the variance of returns, around the macroeconomic announcements in this thesis, will result in a more precise test for responses in returns (and thus, market efficiency) to public announcements. 2.2 Australian Studies A review of studies on Australian stock market returns is detailed below. This is followed by a review of studies specifically dealing with stock market return volatility. In light of the importance of the EMH to my thesis, an additional Australian stock market study testing the hypothesis is outlined at the end of this section. A summary of these Australian studies is presented in Table Australian Stock Market Returns Gultekin (1983) tests the relationship between stock returns and inflation in a number of countries including Australia. This study is motivated by the Fisher hypothesis that 10 August

26 the real rate of return on assets is independent of the expected inflation rate in efficient markets (that is, nominal returns on assets will vary one for one with expected inflation). Nominal returns were regressed on expected inflation, which was approximated by the realised inflation rate at the beginning of the return holding period. 4 No significant relationship was found between the two variables in Australia. These findings indicate, in Australia, inflation realised in a past quarter has no effect on the nominal returns realised in the subsequent quarter. Jaffe (1984) tests stock market data for four countries, including Australia, for a week-end effect. Other studies, typically based on US data, had found returns were abnormally high on Friday and abnormally low on Monday. Returns were, therefore, regressed on dummy variables, representing each trading day. The hypothesis that returns were equal on each trading day was rejected. Tuesdays were found to have significantly lower mean returns than all other days. Jaffe (1984, p.4) hypothesises the effect may be a result of the timing difference between the Australian stock exchange and the New York stock exchange. The New York exchange tended to experience its lowest mean returns on Mondays, and by then, Monday trading in Australia had already closed. A regression of the differences in returns, between the Australian market and lagged values of the US market on dayof-the-week dummy variables, found days of the week have unequal effects on the differential, providing evidence to support the hypothesis. This result highlights the importance of lagging US return values when testing for these relationships with the Australian market. 4 For countries other than Australia in this study expected inflation was also estimated using ARIMA models and derived from short-term interest rate data. 10 August

27 Additional tests are carried out in Jaffe s study (1984) to determine whether any part of the day-of-the-week effect found in Australia is independent of those observed in the US market. The results provided evidence that at least part of the day-of-the-week effect is unique to the Australian market. Such effects should thus be controlled for in studies of Australian stock returns. Singh (1993) conducts a study on the response of Australian stock prices to money supply announcements. His treatment of the money supply data is of particular interest for my study because it involves modelling the expected and unexpected component of a macroeconomic announcement. This is in addition to using surveyed Money Market Services (MMS) forecasts. Singh (1993, p.48) used money supply forecasts sourced from MMS Australia for broad money (M3). Multiple ARIMA models were used for forecasting, so only subsequent models incorporated previously unavailable information as it became available with each passing day. This ensured ARIMA forecasts on each day were based only on information available at that time. This avoided biasing the forecasts toward the actual announced outcomes, which would have been the case if the unavailable future data were used to fit an ARIMA model on each day historically. The announcements detailing the money supply s preliminary estimates were sourced from RBA press releases for both narrow money (M1) and M3. Changes in stock price indices are regressed on both expected and unexpected changes in money supply, using ARIMA forecasts in one particular case and MMS forecasts in another. At conventional levels of statistical significance, the results showed no significant relationship between changes in stock prices and money supply changes. The surveybased forecasts appear to be somewhat more reflective of market expectations than ARIMA based forecasts and produced higher absolute values of t-statistics, despite 10 August

28 neither forecast producing statistically significant results at conventional levels (Singh 1993, p.50). There is, therefore, some support for the use of survey forecasts over modelled forecasts, though this is a moot point given the lack of strong statistical support. Singh (1995) examines the role of current account deficit announcements on a number of financial markets, including the Australian stock market. Changes in stock prices were modelled as a function of expected and unexpected announcements, with the expected component being represented by forecasts. Including dummy variables in the model controlled for day-of-the-week effects. Monthly announcements of the current account balances are sourced from the ABS. A survey of the expected value of the current account deficit was sourced from MMS. An ARIMA model was, again, used to model expected and unexpected components of announcements as in Singh (1993). The expected component, based on the MMS survey data, was found to have an insignificant effect on stock returns, while the unexpected component had a negative effect significant at the 10 per cent level. Day-of-the-week effects are found to have no significant effect. When undertaken with ARIMA forecasts as opposed to MMS data, the analysis produced comparable results but reported slightly lower t- values. Again, these findings are consistent with Singh s (1993, p.51) previous suggestion that MMS surveyed forecasts contain more information than ARIMA forecasts (based only on past values). Brooks et al (1999) test the effect of unexpected current account deficit (CAD) and GDP announcements, including revisions, on daily observations of the All Ordinaries share price index. An ARIMA model is used to decompose the initial announcements into their expected and unexpected components, by producing one-step-ahead forecasts representing the expected component, and forecast errors representing the 10 August

29 unexpected component. An Ordinary Least Squares (OLS) model, regressing returns on the unexpected component of announcements and revisions, is used to test for significant effects. The results suggested CAD and GDP announcements and revisions have no significant effect on returns. The results were unchanged when the announcements and revisions were separated into good news (positive sign) and bad news (negative sign) announcements. An additional point to note in Brooks et al (1999, p.199) is the discussion regarding the use of MMS forecast survey data in Australia compared to ARIMA models for forecasting. They advocate the latter on the basis that survey data suffers from the effects of herding behaviour, survival bias and its reliance on median expectations in survey forecasts. They argue the use of the median is inappropriate given there is no reason to expect the marginal investor to hold the median expectation. This view is contrary to that of Singh (1993 and 1995), who supports survey-based data on the basis it has more explanatory power than ARIMA forecasts. Although none of these studies detected any significant relationships, using either MMS or ARIMA forecasts, Singh (1993 & 1995) noted some evidence (outlined above) in favour of the MMS forecasts. In light of Singh s findings, survey data is generally preferred over ARIMA forecasts in this study, although it is important to note, the discussion that occurred in the literature. Where MMS data is not available, I consider ARIMA forecasts the next best option. Kim and In (2002) create a model to explain returns in the Australian stock market. The model uses returns in foreign stock markets, and macroeconomic announcements in Australia and overseas, as explanatory variables. Daily Australian stock returns were based on the ASX All Ordinaries index returns, while the Standard and Poor s (S&P) 500, FTSE 100 and Nikkei 225 based returns are used to represent the US, UK 10 August

30 and Japanese markets respectively. The study used a bivariate Glosten-Jagannathan- Runkle (GJR) GARCH model and two-step estimation procedure. Their model makes specific allowance for asymmetry in stock return relationships with explanatory variables. This is an important aspect of their analysis. Unemployment, the CPI and GDP announcements are used from both the Australian and US economy. Dummy variables for holiday periods are also included. The model indicates that return volatility in Australia is significantly higher on announcement days for Australian real GDP. The model also reports a significant positive relationship between Australian and US/UK returns, and a significant negative relationship between Australian and Japanese stock market returns. Shocks in the UK and Japanese stock market have a significant positive impact on the volatility of the Australian market. US and Australian GDP announcements, as well as holidays, were positively related to volatility. Asymmetry terms in the model were significant, and tests of the model residuals found no remaining sign bias, indicating the model adequately captures the asymmetric effect. Groenewold (2003) tests for a structural break in the relationship between the Australian stock market and real GDP, resulting from financial deregulation in Australia, using a vector autoregression (VAR) framework and plotting impulse response functions (IRFs). Returns are based on the All Ordinaries (non-cumulative) price index. Real GDP, valued at 1999/2000 prices, is used to represent real output. The control variables used include the term spread on government bonds, which is calculated as the difference between 10 year yields on Commonwealth Government bonds and three-month rates on Treasury notes. The default spread is calculated as the difference between five-year yields on Commonwealth Government bonds and State Treasury bonds. 10 August

31 Over the full sample, the VAR model and IRFs found lagged output growth, term spread and default spread had no significant effect on stock returns. In terms of precedence, the VAR model results showed causality ran from stock market returns to output growth (that is, stock market returns from two quarters prior caused positive output growth. This is consistent with the theory that share market returns are a leading indicator of output (Fama 1981; Campbell & Shiller 1988). In the prederegulation period, the VAR model results still found lagged output growth, term spread and default spread had no significant effect on stock returns at the conventional levels of statistical significance. In the post-deregulation period, the VAR results found lagged default spreads and term spreads had a significant negative effect on stock market returns. As found in the results for the whole period, stock market returns from two quarters prior caused positive output growth. The impulse response functions and R-squared values for the VAR models suggest any influence that output has on stock returns has weakened post-deregulation. The implication of these findings is the relationship between the real economy and the share market had, if anything, weakened after opening the economy to international capital flows. In isolation, this study creates an a priori expectation that real GDP announcements affect stock market returns. The findings on both default and term spreads justify the inclusion of these variables as control variables in a model of Australian stock returns. Groenewold (2004) computed fundamental share prices in Australia, based on real GDP, using a structural vector autoregressive (SVAR) model over the period 1959 to He found positive real GDP shocks positively affected stock prices, supporting the theory that the real value of firms is the net present value of expected dividends (Groenewold 2004, p.660). Over the period of relatively subdued stock market price 10 August

32 growth from 1988 to 1993, his computed fundamental share prices indicated stock market prices were not too far from fundamental values. However, they departed substantially from fundamentals in the period prior (from around 1970 to 1987), and from around 1994 to 1999 when stock market price growth was strong. This suggests relationships between macroeconomic fundamentals and stock market returns may differ between strong and subdued periods of stock market price growth. Kim (2003) investigated the effects of US and Japanese macroeconomic news announcements on the stock markets of Australia, the US and Japan. The All Ordinaries index observations for open, high, low and close were used to calculate Australian market returns. Macroeconomic announcements, and surveyed expectations of these announcements, were sourced from MMS International so the unexpected components (or news ) could be estimated by deducting expectations from the announcements. Exponential GARCH (EGARCH) models were estimated using Australian stock returns. Using dummy variables controlled for holidays. The effect of macroeconomic announcements was captured using dummy variables to indicate only those announcements with news content. These were announcements where the announced value was not equal to the surveyed expectation. Asymmetric properties of return and return volatility responses were captured through inclusion of dummy variables to indicate bad news. These were based on the sign of unexpected components in the announcement. Most of the US macroeconomic news announcements had a significant effect on Australian returns. Those with a positive effect on returns included retail sales growth, unemployment, PPI and CPI-based inflation. With respect to balance of trade, GDP growth, retail sales growth, unemployment, PPI and CPI-based inflation, bad news had a significantly negative effect on returns. 10 August

33 All of the US news announcements also had a highly significant effect on Australian return volatility. News announcements that increased volatility included US balance of trade, GDP growth, unemployment and CPI based inflation. Retail sales and PPI news announcements reduced volatility. Bad news, with respect to GDP growth, retail sales growth, unemployment and the PPI, had a negative effect and reduced volatility. Bad news, with respect to balance of trade and CPI, increased volatility. In contrast, few of the Japanese macroeconomic news announcements had a significant effect on returns; only the Japanese CPI and bad unemployment news had a positive effect on Australian returns. For return volatility, however, approximately half of the Japanese announcements had a significant effect. Australian return volatility was positively affected by the Japanese wholesale price index, CPI and bad trade balance news. Trade balance news in general, as well as bad wholesale price index news, had a negative effect on volatility. The study indicates the effects of US news announcements on Australian stock returns are more important than the effects caused by Japanese news announcements. A variable capturing the effect of US announcements (such as US stock returns) will therefore, likely be useful in a model explaining Australian stock returns. Chaudhuri and Smiles (2004) tested the long-term relationship between real Australian stock prices and real macroeconomic variables, including GDP, private consumption, money-supply and oil prices. Their study is similar to mine in many respects, but it is based on vector error correction modelling (VECM) that focuses on relationships over the long run. The All Ordinaries index was used for Australian stock price data, while seasonally adjusted M3 money supply, GDP and private personal consumption expenditure was 10 August

34 sourced from the Organisation for Economic Co-operation and Development (OECD) Main Economic Indicator database. The world oil price index was converted to Australian dollars using the Australian-US dollar exchange rate. Their model of Australian returns also included US, Japanese and New Zealand market returns as explanatory variables. The base error correction model found lagged real GDP growth, private consumption, M3 money supply and oil prices had a highly significant role in explaining real stock price variation over an extended period. Concurrent and lagged US stock price indices are highly significant, and they played a dominant role in explaining long-term real stock price variation. Similar, but much weaker, effects were also found for New Zealand stock price indices, while the Japanese index showed no significant effects. This study further supports the argument that US stock returns are an important variable to include in a model explaining Australian returns; whereas Japanese returns are not. Akhtar et al (2011) examined the effect of consumer sentiment news on Australian stock market returns. The Westpac-Melbourne Institute consumer sentiment index was used to approximate investor sentiment. Returns calculated using the Australian All Ordinaries index were regressed on dummy variables, representing negative and positive changes in the consumer sentiment index and the Morgan Stanley Capital Index (MSCI) (world stock market index), to control for the impact of international factors. Negative changes in the consumer sentiment index had a significant negative effect on returns, while positive news was found to have no effect, confirming a negativity bias in relation to bad news. This finding suggests the consumer sentiment index should be included in my study as a macroeconomic variable of interest. 10 August

35 Hasan & Ratti (2012) studied the relationship between oil price shocks and volatility in Australian stock market returns. Their study was based on sector level returns, as opposed to a broad stock market index or returns. Stock market indices for ten industry sectors in Australia were used to calculate a series of ten different returns. The excess return, over Australian 90-day bank accepted bills, was calculated for each series. Oil price volatility was based on onemonth futures prices of West Texas Intermediate crude oil because they were considered less noisy than spot prices. A GARCH-in-mean specification was used to model the relationship between excess returns and volatility in each industry. Excess returns for each sector were modelled as a function of excess returns on the market overall, as well as excess oil returns and oil return volatility. They found that oil prices were negatively related to overall market returns, and oil return volatility was also negatively related to overall market return volatility. While most sectors showed that an increase in oil returns also meant a decrease in excess returns, the energy and materials sectors excess returns moved in the same direction as those of oil. Increased volatility in oil returns reduced volatility of equity market returns for around half of the sectors (including energy and materials), while significantly increasing equity market return volatility in the financial sector. Hasan & Ratti s (2012) results suggest oil returns are an important variable in explaining Australian stock market returns Australian Stock Market Return Volatility Kearns & Pagan (1993) examined and attempted to explain volatility in Australian stock market index data from 1875 to A variety of models, including EGARCH, were used to model volatility over the period. The EGARCH model was found to explain more of the variation in returns, than either GARCH or the rolling 12-month 10 August

36 standard deviation model, while the rolling 12-month standard deviation model was superior to GARCH in this respect. Tests for sign asymmetry, between positive and negative shocks on volatility, revealed only weak evidence of negative shocks having a greater effect on return volatility. The high values of the lagged coefficients in the volatility models indicated a strong persistence of shocks. This suggests a model of returns/return volatility should account for autocorrelation in volatility. Kearney & Daly (1998) examined the relationship between stock market volatility and a number of macroeconomic variables, including inflation, interest rates, industrial production, the current account balance and money supply. They employed a conditional volatility model based on the absolute values of errors between their returns model and realised returns. Lagged industrial production volatility (over 3 months) reported a negative relationship with stock market return volatility. Volatility in wholesale price inflation, over one month, increased stock return volatility, as did interest rate volatility over the same lag horizon. These findings indicate the absolute values or size of changes in macroeconomic variables, as opposed to the direction of changes, are important in explaining Australian stock market return volatility The Australian Stock Market and Efficient Market Hypothesis Sadique & Silvapulle (2001) tested stock returns in several countries, including Australia, for the presence of long memory in returns. Their study was based on the theory that in efficient markets, arbitrage opportunities are quickly taken advantage of and decrease the correlation between successive returns in the market. Three procedures were used to test for long memory, including rescaled range analysis, the Geweke and Porter-Hudak (GPH) test, and a frequency and time domain score test. These tests found no significant evidence of long memory in the Australian stock 10 August

37 market, indicating the EMH could not be dismissed. This result suggests the Australian stock market is, at least, weak form efficient. An overview of the Australian literature that was reviewed is outlined in Table 1. Table 1 Australian Literature Review Summary Study Observation Period Method Data Frequency Results Gultekin (1983) Relationship between expected inflation and stock market returns January December 1979 OLS Regression International Monetary Fund Australian Stock Market Index and CPI based Inflation Quarterly No significant relationship found between expected inflation and stock market returns Jaffe (1984) Day-of-the-week effect on stock market returns March November 1983 OLS Regression and F-Tests Statex Actuaries Stock Market Index and Composite S&P 500 US stock market index and week day dummy variables Daily Days of the week were found to have unequal effects on returns. Evidence suggested this was partly due to correlation with day-of-the-week effects in the US and partly due to factors unique to the Australian market Kearns & Pagan (1993) Returns asymmetry and persistence of shocks in Australian stock market volatility GARCH, EGARCH, 12 month rolling standard deviation Sydney and Australian Stock Exchange All Ordinaries Index Monthly Weak evidence of sign asymmetry in the effect of stock market shocks on stock price volatility was found. Shocks appeared to be strongly persistent, and return volatility is greater than that in the US, particularly in more recent history Singh (1993) Relationship between money supply and stock returns January June 1987 OLS Regression Statex Actuaries Stock Market Accumulation Index, All Industrials, All Ordinaries, Banks and Finance and Transport indices, MMS M3 money supply forecasts, RBA press releases for M1 and M3, ARIMA forecasts based on RBA press releases Daily No significant relationship was found between stock returns and money supply. Some evidence was found suggesting surveyed expectations were more reflective of market expectations than ARIMA base forecasts Singh (1995) Relationship between current account deficit and stock returns July October 1991 OLS Regression Dividend corrected Statex Actuaries Accumulation index, week day dummy variables, MMS current account balance forecasts, monthly ABS current account balance announcements, ARIMA forecasts based on ABS current account balance announcements Daily Results did not find any significant relationship between stock returns and the current account balance. Evidence was found that suggested surveyed forecasts contain more information than ARIMA based forecasts 10 August

38 Kearney & Daly (1998) Relationship between stock market volatility and inflation, interest rates, industrial production, current account balance and money supply January January 1994 Absolute value conditional volatility and GLS estimated ARCH ASX All Industrial Index, index of industrial production, OECD index of wholesale prices, current account balance, Australian-US Dollar exchange rate, RBA 3- month bank accept bill interest rate, dummy variables representing 1987 stock market crash and monthly seasonal dummy variables Monthly It was shown that conditional stock market volatility was directly related to the conditional volatility of wholesale price inflation and interest rates. Industrial production, current account balance and money supply were indirectly related Brooks et al (1999) The effect of announcements and revisions in GDP and current account balance on stock returns January December 1993 ARIMA/O LS All Ordinaries Index and announcements of current account balance and GDP values, ARIMA forecasts based on current account balance and GDP announcements. Daily Current account balance and GDP news announcements and revisions were found to have no significant effect on returns. The results were unaffected by separating out negative and nonnegative news. Sadique & Silvapulle (2001) Tests for longterm memory in stock market returns January December 1998 Rescaled range analysis, GPH test, frequency and time domain score test Australian aggregate stock price index Weekly No significant long memory of stock market shocks could be found indicating the efficient market hypothesis could not be dismissed Kim & In (2002) Spill over effects from international stock market returns, employment, CPI and GDP announcement days on Australian stock market returns July December 2000 GJR GARCH with two step estimation procedure ASX All Ordinaries Index, S&P 500, FTSE 100, Nikkei 225, Australian and US employment, CPI and GDP announcement day dummy variables Daily A significant positive relationship was found between Australian and US/UK returns while a significant negative relationship was found between Australian and Japanese returns. US and Australian GDP announcements had a positive effect on volatility, as did Australian holidays. Shocks in the UK and US stock market also had a significant positive impact on Australian stock market volatility. The model also found asymmetry terms were significant, indicating negative and positive shocks had a different effect on conditional volatility Groenewold (2003) Relationship between the stock market returns and real output, term spreads, and default spreads pre- and postfinancial deregulation Quarter 1, Quarter 2, 2001 VAR and impulse response functions All Ordinaries Index, Real GDP, term spread between 10 year Commonwealth Government Securities and 3-month Treasury notes, default spread between 5 year Commonwealth Government Securities and 5 year State Government Treasury Bonds Quarterly Pre-deregulation lagged stock returns were found to have a significant effect on themselves. Post-deregulation lagged term spreads and default spreads had a significant negative effect on stock returns. Impulse response functions (IRFs) found stock market return shocks on themselves died out in around 1.5 years. IRFs showed output growth had a weak negative effect on stock returns pre deregulation, but a negligible effect post deregulation 10 August

39 Groenewold (2004) Relationship between the real stock market prices and real output Quarter 4, Quarter 1, 1999 SVAR and impulse response functions All Ordinaries Index divided by the GDP deflator and Real GDP Quarterly The effects of shocks to both output and share prices in the model die out quickly, but more slowly so for share price shocks. A positive output shock had a positive effect on both real output and real share prices. A positive stock market shock initially depressed output but the effect was only temporary. Share prices appeared to be undervalued for most of the 1970s and overvalued for most of the 1990s Kim (2003) Spill over effects of US and Japanese news on the Australian stock market January May 1999 Moving average EGARCH All Ordinaries Index high, low, open and close prices, MMS Surveyed forecasts of US balance of trade, real GDP, retail sales, unemployment rate, PPI and CPI, MMS surveyed forecasts of Japanese trade balance, current account balance, unemployment, money supply, wholesale price index and CPI, dummy variables representing holidays Daily US news announcements that increased volatility included balance of trade, GDP growth, unemployment and CPI inflation. Retail sales and PPI news announcements reduced volatility. Bad news with respect to GDP growth, retail sales growth, unemployment and the PPI were negatively related to volatility Japanese CPI and bad unemployment news was positively related to Australian returns. Australian return volatility was positively related to the wholesale price index, CPI and bad trade balance news. Trade balance news on average and bad wholesale price index news was negatively related to volatility Chaudhuri & Smiles (2004) The effect of aggregate economic activity on the Australian stock market Quarter 1, Quarter 4, 1998 VECM All Ordinaries Index, OECD seasonally adjusted Main Economic Indicators; M3 money supply, GDP, private personal consumption expenditure and world oil price converted using AUD/USD exchange rate Quarterly Lagged Real GDP, real private consumption, real M3 money supply, and real oil prices were found as highly significant in explaining real stock price variation over the long run Akhtar et al (2011) Negatively biased effect of consumer sentiment announcements on Australian stock returns June December 2009 OLS Regression All Ordinaries Index, MSCI World Index, Westpac Melbourne Institute consumer sentiment index, Dummy Variables signifying negative index changes Daily Negative changes in the consumer sentiment index had a significant negative effect on returns while positive news had no effect Hasan & Ratti (2012) The effect of oil price shocks on Australian stock return volatility March December 2010 GARCHin-mean Australian stock market indices for 10 GICS sectors, one-month West Texas Intermediate crude future prices, and excess stock market index returns over 90 day bank accepted bills Daily Oil prices were negatively related to overall market returns and volatility. Most sectors' excess returns were negatively related to those of oil, however energy and materials were positively related. Increased oil return volatility reduced return volatility for around half of the sectors in the study including energy and materials, while financial sector return volatility increased 10 August

40 2.3 Foreign Studies Foreign Stock Markets and Macroeconomic Surprises My literature review found four foreign studies that specifically examined the effect of macroeconomic surprises on stock returns. The foreign literature gives us an additional insight into which methods and data are likely to produce informative results. They also assist in the development of hypotheses, which are detailed in Chapter 3. Wasserfallen (1989) examined the relationship between stock returns and macroeconomic surprises in the UK, Switzerland and West Germany separately. His study used OLS regression with distributed lags and quarterly returns based on a prominent stock market index, present in each country at the time. He used the Frankfurter Allgemeine Zeitung index for West Germany, the Swiss National Bank index for Switzerland and the Financial Times Ordinary index for the UK. Surprises are calculated as the difference between realised values and ARIMA forecasts (residuals), for a number of macroeconomic variables, including real GNP, industrial production, the unemployment rate, consumer prices, money supply, monetary base, real exports, import prices, nominal and real interest rates, real investment, nominal and real wages, and foreign exchange rates. His results indicated the explanatory power of his regressions was very low. For West Germany, unexpected changes in nominal interest rates, and consumer and import prices, have a negative relationship with returns. That is, an unexpected increase (decrease) in these factors is associated with decreased (increased) stock returns. Conversely, unexpected changes in money supply have a significant positive relationship with stock market returns. In Switzerland, unexpected real consumption was the only variable to have a relationship 10 August

41 with returns, and that relationship was negative. The UK study found only unexpected nominal wages had a relationship with real returns that was negative. Becker, Finnerty and Friedman (1995) examined the relationship between macroeconomic surprises and stock market returns in the UK. They employed OLS regression on high frequency (half hourly) returns using the Financial Times Stock Exchange (FTSE) 100 index. Macroeconomic surprises were calculated using MMS surveyed data on the current account, industrial production, money supply (M0), PPI, public sector borrowing requirement, retail price index, retail sales, unemployment, and visible trade to represent expectations. Their results showed that higher than expected visible trade and current account surprises were positively related to stock returns. Heavier than expected government borrowing was negatively related to stock returns. 5 All other variables had no statistically significant effect. Flannery and Protopapadakis (2002) studied the effects of seventeen macroeconomic surprises, and announcement days on stock returns, spanning the 16 years from 1980 in the US. A GARCH model was used to estimate returns as a function of surprises, and the volatility of returns as a function of macroeconomic announcement days. Volatility was of more interest because they thought it likely the impact of macroeconomic news on returns was time varying, in terms of strength and sign. Although an announcement could cause a change in stock returns, it meant the response of returns to a given type of announcement could, at times, be negative and at other times be positive. Volatility captured the magnitude or size of the effect rather than the sign. 5 Becker, Finnerty and Friedman (1995) defined surprises as the actual announced value less the expected value which is opposite to the definition in my paper. 10 August

42 Accordingly, Flannery and Protopapadakis (2002) assessed macroeconomic variables to determine whether they were risk factors in the stock market by also modelling volatility as a measure of risk. Their study used the weighted index of the New York Stock Exchange-American Stock Exchange-National Association of Securities Dealers Automatic Quotation System (NYSE-AMEX-NASDAQ) to calculate returns and base trading volumes. Seventeen macroeconomic variables of interest were examined, including the CPI, PPI, money supply (M1 and M2), employment, unemployment, balance of trade, home sales, housing starts, industrial production, personal income, personal consumption, retail sales, interest rates, consumer credit, construction spending and real GNP. Money supply surprises and announcement days were the only variables to affect both the level and the volatility of returns. Surprises had a negative effect on the level of returns, while announcement day dummies had a positive effect on the volatility of returns. The PPI and CPI surprises negatively affected the level of returns, while the balance of trade, employment, home starts and real GNP announcements days had a positive impact on return volatility. Kim (2003) produced regression results for the effects of US and Japanese macroeconomic news announcements on their own respective stock markets. The Dow Jones Industrial index and Nikkei 225 index were used to represent the stock market returns of the US and Japan respectively. Observations for open, high, low and close index prices were used to calculate market returns. MMS International surveyed expectations of macroeconomic announcements were deducted from the announcements themselves to estimate the unexpected components (or news ). The same EGARCH regression framework for Kim s (2003) Australian stock return model (outlined in Section 2.2.1) was used for estimating the stock returns. 10 August

43 The regression results showed that US returns were negatively related to both good and bad US balance of trade surprises, and positively related to bad retail sales surprises. US return volatility was decreased by good balance of trade news, real GDP news, retail sales news and bad unemployment news announced for the US. Bad balance of trade and PPI news announced for the US increased US return volatility. For Japan, the regression results showed that good Japanese trade balance news, bad money supply news and bad CPI news decreased Japanese returns. Good current account balance news and bad trade balance news increased returns. 6 Japanese return volatility is decreased by good unemployment and CPI news, and bad trade balance news. Good trade balance news, current account balance news, bad unemployment and CPI news and any wholesale price index news increases Japanese return volatility. An overview of some of the first foreign literature on the effect of macroeconomic announcement surprises on stock returns is summarised in Table 2. 6 For Japan, bad news announcements for trade balance are when it is lower than expected. Bad news announcements for money supply are when it is higher than expected. 10 August

44 Table 2 Foreign Literature Review Summary Study Country Observation Period Method Data Frequency Results West Germany Frankfurter Allgemeine Zeitung Index, real GNP, industrial production, unemployment rate, consumer prices, money supply, monetary base, real exports, import prices, nominal interest rate, real interest rate, foreign exchange rates Consumer price, import prices, and nominal interest rate all had a significant negative relationship with returns, while M1 money supply had a significant positive relationship Wasserfallen (1989) Relationship between nominal returns and macroeconomic surprises Switzerland OLS distributed lag regressions and ARIMA forecasts Swiss National Bank Index, real GNP, industrial production, real consumption, real investment, consumer prices, money supply, monetary base, real exports, import prices, nominal interest rate, real interest rate, foreign exchange rates Quarterly Only real consumption had a significant negative relationship with returns United Kingdom Financial Times Ordinary Index, Real GNP, industrial production, consumer prices, nominal wages, real wages, M1, monetary base, real exports, import prices, nominal interest rates, real interest rate, foreign exchange rates Only nominal wages had a significant negative relationship with real returns Becker, Finnerty and Friedman (1995) Relationship between surprises in macroeconomic announcements and UK equity market returns United Kingdom July December 1990 OLS Regression FTSE 100 Index, MMS surveyed expectations and actual announcements from Economic Trends (UK) for current account, industrial production, M0, PPI, public sector borrowing requirement, retail price index, retail sales, unemployment and visible trade Half hourly Higher than expected visible trade and current account balances were positively related to returns. Heavier than expected government was negatively related to returns. All other variables had no statistically significant effect Flannery and Protopapadakis (2002) Relationship between macroeconomic announcement days and surprises in announcements and US equity market returns United States January December 1996 GARCH NYSE-AMEX-NASDAQ market index from Centre for Research in Security Prices, lagged dividend to price ratio, lagged three-month Treasury bill yields, lagged 10 year Treasury bond term premium on three-month bills, lagged default premium between Moody's BAA and AAA seasoned corporate bond indices, surprises based on MMS surveyed expectations and announcements on balance of trade, consumer credit, construction spending, CPI, employment, unemployment, new home sales, housing starts, industrial production, leading indicators, M1, M2, personal consumption, personal income and PPI Daily Higher than expected consumer and producer price indices were negatively related to the level of returns. The announcement days for balance of trade, unemployment/employment, home starts, M1, M2 and real GNP were positively related to the volatility of returns 10 August

45 Kim (2003) Spill over effects of US and Japanese news on the Australian stock market United States Japan January 1991 May 1999 Moving average EGARCH Dow Jones Industrial Index high, low, open and close prices, MMS Surveyed forecasts of US balance of trade, real GDP, retail sales, unemployment rate, PPI and CPI Nikkei 225 Index high, low, open and close prices, MMS surveyed forecasts of Japanese trade balance, current account balance, unemployment, money supply, wholesale price index and CPI, dummy variables representing holidays Daily US returns were negatively related to both good and bad US balance of trade surprises and positively related to bad retail sales surprises. US return volatility was decreased by good balance of trade news, real GDP news, retail sales news and bad unemployment news. Bad balance of trade and PPI news increased US return volatility Good trade balance news, bad money supply news and bad CPI news decreased returns. Good current account balance news and bad trade balance news increased returns. Return volatility is decreased by good unemployment and CPI news, and bad trade balance news. Good trade balance news, current account balance news, bad unemployment and CPI news, and any wholesale price index news increases return volatility 10 August

46 2.3.2 Business Cycles and Macroeconomic Factor relationships with Stock Markets Binswanger (2004) studied the relationship between macroeconomic fundamentals and stock returns in Canada, Japan and an aggregate economy consisting of four European G-7 countries. He found that the fundamental relationship between stock returns and real GDP shown in other research disappeared during the stock market boom of the 1980s. He concluded there is support for the hypothesis that speculative bubbles in the stock market were an international phenomenon affecting major economies during the 1980s and 1990s. Andersen et al (2007) extend the literature by using high-frequency futures market data available on a tick-by-tick basis to quantify the effect of US macroeconomic news announcements on global stock, bond and foreign exchange markets. One of the key findings from this study was that stock markets react differently to macroeconomic news depending on the stage of the business cycle. Over the full sample, including both a period of expansion and contraction, they found almost no significant equity market responses to economic news. Once the sample was split into expansion and contraction periods they found that equity markets responded negatively to positive real economic shocks during expansions while responding positively during contractions. 7 Velinov and Chen (2015) examined whether macroeconomic fundamentals explained stock prices in France, Germany, Italy, Japan, the UK and US with a special emphasis on the period following the GFC. In all countries they found that stock prices fell back 7 Many additional papers studying various aspects of macroeconomic announcement surprises and asset returns can also be found. Some examples include Bomfim (2003), Andersen (2003), Brenner, Pasquariello, and Subrahmanyam (2009) and Rangel (2011). 10 August

47 toward their fundamental values (approximated by industrial production) after the GFC. 2.4 Conclusions from the Literature The literature reviewed for examining the effect of macroeconomic variables on Australian stock returns covers a period from 1947 to There appears to be a dearth of Australian literature dealing with the effect of macroeconomic news or surprises on Australian stock market returns. Kim and In (2002) made an early attempt to detect a relationship between stock returns and announcement days for Australian real GDP, the CPI and unemployment. While they detected a significant relationship between real GDP announcement days and return volatility, no other relationship were established. Additionally, their study was not based on surprises, but only announcement days. More recent studies include Hasan and Ratti (2012), covering the period 2000 to 2010, and Akhtar et al (2011), covering the period 1992 to However, their focus is on the effect of a single macroeconomic variable on Australian stock returns, rather than the effect of a variety of macroeconomic surprises. My thesis aims to provide a contemporary analysis of the effect of a variety of macroeconomic announcements commonly cited in the media, taking advantage of the long high quality data sets now available, such as MMS surveys and the S&P ASX 200 index (discussed further below). I will employ an extensive set of macroeconomic variables that are noted in the literature. While a number of methods have been used to model Australian returns, the EGARCH model in Kim s (2003) model was one of the most successful in terms of explanatory power. This approach has the benefit of modelling both returns and return volatility simultaneously, which is of interest in return studies, as highlighted in the theoretical and Australian literature (Fama 1991; Kearns & Pagan 1993; Kearney & 10 August

48 Daly 1998). Foreign studies support the use of ARCH or GARCH type models. Kearns and Pagan (1993, p.174) found EGARCH had superior explanatory power to GARCH and other specifications used. Their study also presented evidence that volatility persists in Australian returns, and so, GARCH type models that control for these autocorrelation effects should produce more statistically robust results than basic OLS models. This specification thus far appears to be a good candidate for my analysis of returns. Fama (1991, p.1601) noted the event study process, using high frequency data (such as daily observations), allows for a more precise measurement of the speed at which stock prices respond to a given event, such as macroeconomic announcements. The process also assists in overcoming the joint-hypothesis problem. This supports the use of daily data in an event study process for analysing the effect of macroeconomic announcements on Australian stock market returns. The ASX All Ordinaries index is the most utilised Australian stock price index in the literature reviewed, as evidenced by its use in eight of the fifteen studies reviewed above. I note, however, this index has been restructured to include a large number of smaller capitalised firms from 2000 onward. This gives rise to the thin trading issues raised in Fama (1991, p.1579), so my preference is to use the All Ordinaries index only for checking the robustness of results. The S&P ASX 200 index gives less weight to smaller capitalised firms and has maintained a consistent structure since its inception. My study benefits from having access to this data series, over a time frame that falls over a period of prolonged stock market expansion, contraction and subsequent subdued growth following the GFC. This allows me to examine whether relationships are different during these phases of stock market activity because the 10 August

49 literature suggests this may be the case (Binswanger 2004, Andersen et al (2007), Velinov & Chen 2015). Two of the studies examining the Australian stock market found US returns (S&P 500) to have significant effects on Australian returns. Jaffe (1984) found evidence to suggest negative average Monday returns in the US were correlated with the lowest mean returns occurring on Tuesday in Australia, due to time zone differences. Kim and In (2002, p.578) similarly found lagged US stock prices had a positive relationship with Australian prices, affecting both the mean and variance. The most commonly studied macroeconomic variables in the Australian literature include inflation, balance of trade (or more specifically the CAD), interest rates, GDP, employment and oil prices. Each of these variables was included in at least two of the studies. Based on their prominence in the literature, I consider these variables good candidates for inclusion in my model of stock returns, to ensure a comprehensive study of macroeconomic risk factors whilst maintaining a parsimonious model. The interest rate used is the overnight cash rate because its values are announced periodically (rather than being continuously updated), which lends itself well to the study of surprises. Retail sales and the PPI were included in my study on account of good quality surveyed forecasts being available from MMS, and because they had been included in foreign studies using macroeconomic surprises (Becker, Finnerty & Friedman 1995, Flannery & Protopapadakis 2002). Consumer sentiment was included, following the success of Akhtar et al (2011), who found it significant in explaining Australian returns. Singh (1993 & 1995) found evidence to suggest MMS survey data has information content superior to naive ARIMA models of expected values for macroeconomic announcements. In the foreign literature that examined the effect of macroeconomic 10 August

50 surprises, the use of MMS survey forecasts was associated with models more successful in terms of explanatory power than those employing ARIMA forecasts (Kim 2003). Accordingly, I opt to use MMS survey data as forecasts in my modelling, wherever possible, and I use ARIMA forecasts only where MMS data is insufficient. The exception here is money supply, for which MMS data is no longer available, and which reported disappointing results when ARIMA forecasts were used (Singh 1993). For this reason, I exclude money supply as a macroeconomic variable in my study. The findings of Sadique & Silvapulle (2001) suggest the Australian stock market is, at least, weak form efficient, which is an encouraging starting point from which to proceed because my study relies on the Australian stock market being efficient. 10 August

51 3 Hypothesis As outlined in Section 2.3.2: Conclusions from the Literature, the macroeconomic variables examined in my study are unemployment, balance of trade, retail sales, the producer price index, the consumer price index, real GDP, the overnight cash rate and the consumer sentiment index. My null hypothesis for each macroeconomic variable is that the surprise component of related announcements has no effect on aggregate stock returns. As shown in the literature review (Chapter 2), there are quite a number of conflicting theories and empirical results that attempt to explain the relationship between each of the macroeconomic variables in my study and stock market returns. Flannery and Protopapadakis (2002, p.752) raised the possibility of such conflicts, and highlighted that the impact of specific macroeconomic variables might vary with economic conditions. This meant that the effect of macroeconomic risk factors might change depending on the stage of the business cycle. Despite this, they emphasised that economically important surprises should be associated with returns that are abnormally large in absolute value. The models set out in Chapter 4 have been designed to capture any relationships with abnormally large absolute values in returns. My null hypothesis is, therefore, that no macroeconomic variables are economically important. Alternative theories and evidence propose that macroeconomic variables do affect stock market returns but, all too often, explain the relationship as a time invariant effect. Allowing for time varying stock market volatility responses (for each macroeconomic variable) allows a greater chance of detecting economically 10 August

52 important relationships with returns, in the event that the direction of the effect is time varying. The dummy-based model assigns a dummy variable to macroeconomic surprises that are not zero, and a dummy variable to any bad news that depends on the sign of the observed value for the macroeconomic surprise. The surprise is defined in equation (1) below: Surprise E ( Announcement ) Announcement k, t t 1 k, t k, t Where: E ( Announcement ) is the forecast value for period t prior to period t ; and t 1 k, t Announcement kt, are the realised values of each of the k announcements at time t. In my analysis, I also refer to surprises for each macroeconomic variable as good or bad for ease of discussion. This is because the sign of surprises (negative or positive) has a different meaning depending on the variable in question. For example, a negative sign could indicate good news for some variables and bad news for others. I must emphasise good and bad news, in this context, does not relate to presupposed effects on returns, but instead, assumed perceptions of what is good or bad news for the economy. 8 My assumption of what constitutes good and bad surprises follows Kim (2003, p.619) for all variables except cash rates and consumer sentiment, which are not included in his study. For the former, I assume the perspective of a leveraged entity so higher cash rates are bad news (in terms of higher interest payments). For the latter, I assume the perspective of an entity that relies on sales activity so high 8 The words news and surprises are used interchangeably, although strictly speaking news is only the value of surprises that are not equal to zero. In practice, it is rare for surprises to equal exactly zero, as this means expectations forecasted the announcement precisely. 10 August

53 levels of consumer sentiment mean good news (in terms of better buying conditions and sales). To summarise, a positive value of surprises resulting from equation (1), for unemployment, the PPI, CPI and overnight cash rate, is assumed to be good news. This is based on the assumption that when announced values of these variables are low (relative to expectations), it is typically seen as good news. Announced values of those variables that are low (relative to expectations) result in a positive surprise value (according to equation (1)). All negative observations in the set of t surprises, for unemployment, the PPI, CPI and overnight cash rate, are coded with the number one for the bad news dummy and a zero otherwise. I must also emphasise the bad news dummies are not multiplied by the surprise value. A positive value of surprises resulting from equation (1), for retail sales, real GDP, balance of trade and the consumer sentiment index, is interpreted differently. For these variables, a positive value is assumed to be bad news. This is based on the assumption that announced values of those variables that are low (relative to expectations) are typically seen as bad news. Thus, according to equation (1), lower than expected announced values produce a positive surprise value, which is interpreted as bad news. For retail sales, real GDP, balance of trade and the consumer sentiment index, all positive values are coded with the number one for the bad news dummy and a zero otherwise. Again, the bad news dummies are not multiplied by the surprise value. One final complication is that changes in one macroeconomic variable can be a proxy for changes in another macroeconomic variable. For example, in Australia, CPI news can influence expectations of changes in the overnight cash rate due to the inflation targeting objective of monetary policy. There are many possible 10 August

54 permutations of these interrelationships between macroeconomic variables, which may also change depending on the phase of the business cycle. In light of this, I only outline some of the most noted hypotheses on these relationships covered in the literature. 3.1 Unemployment My null hypothesis is that unemployment surprises (or news) have no effect on returns or return volatility. Kim and In (2002, p.578) showed that Australian employment announcement day dummies (as distinct from values) had no significant explanatory power in Australian stock market return regressions over the period 1991 to This finding is directly relevant to my thesis because the Australian employment announcements used by Kim and In (2002, p.574) follow the same release schedule as the unemployment announcements used in my study. In the foreign studies that use the macroeconomic surprise methodology, Kim (2003) assessed the effect of US and Japanese unemployment announcement surprises on stock return levels, and found US unemployment announcements had no effect on US market returns. Such was also the case in Japan; he found no significant effect between Japanese unemployment announcement days and Japanese market returns. Flannery and Protopapadakis (2002, p.766) confirmed Kim s finding that there were no significant effects between US unemployment announcements and the level of US returns over a longer, but earlier, period. Becker, Finnerty and Friedman (1995, p.1206) found UK unemployment announcements had no effect on UK aggregate stock returns. In West Germany, Wasserfallen (1989, p.622) observed no relationship between unemployment values and West German stock market returns. The null hypothesis is set out as follows: H1: No relationship between unemployment news and stock returns/return volatility 10 August

55 The finding that there is no relationship between employment/unemployment announcements and returns could possibly be a result of some studies failing to capture time varying responses to unemployment news. That is, if stock returns do, in fact, change in response to unemployment news, but the sign of the effect changes over time, the effect may be undetectable in return levels but detectable in squared or absolute values of returns. Time varying responses are more likely to be the case if significant relationships are found between news and volatility (or absolute returns). Theories supporting an alternative hypothesis, that unemployment is negatively related to returns, are as follows. Asprem (1989, p.595) initially presumed employment (unemployment) would be positively (negatively) related to real activity, and thus, positively (negatively) correlated with stock returns. Boyd, Hu & Jagannathan (2005, p.650) found the effect of unanticipated increases in unemployment on stock returns is dependent on the state of the economy; that is, whether the economy is expanding or contracting. The reasoning is unemployment news contains information on corporate earnings/dividend growth expectations, meaning unemployment news can be a proxy for growth expectations. They found an unanticipated increase in unemployment often precedes slower growth, particularly during contractions. The subsequent lower growth in corporate cash flows equates to lower stock prices and returns. The alternative hypothesis, based on these theories, is set out below: H1a: Good unemployment news (unexpected decrease in unemployment) increases stock returns and vice versa for bad unemployment news The same authors also recognise another alternative hypothesis that unemployment and stock returns are positively related. Asprem (1989, p.595) reasoned employment (unemployment) may increase (decrease) only in the later stages of a boom period, 10 August

56 and by that time, earnings expectations, and thus stock prices, are starting to decline. That is, stock prices are based on the period ahead, while unemployment relates to the present. Boyd, Hu & Jagannathan (2005, p.650) formed a view that both future interest rate information and information on earnings/dividend growth are implicit in unemployment news. The future interest rate information appears to dominate corporate earnings/dividend growth information during expansion. They reasoned this is because an unanticipated rise in unemployment may signal an expectation that future interest rates will decline in response, and as a result, stock prices (expressed as the present value of corporate cash flows) are higher. This is on account of a lower discount rate. These theories give rise to the alternative hypothesis set out below: H1b: Good unemployment news (unexpected decrease in unemployment) decreases stock returns and vice versa for bad unemployment news In turn, these theories suggest that unemployment news, both good and bad, affects stock return volatility. Kim (2003, p.625) found, in the US, bad unemployment news reduces return volatility. This gives rise to the following hypothesis: H1c: Bad unemployment news (unexpected increase in unemployment) decreases stock return volatility In Japan, Kim (2003, p.626) found evidence to the contrary. Bad unemployment news reduced increased Japanese return volatility, while good unemployment news decreased it. The alternative hypothesis, based on this evidence, is set out below: H1d: Good unemployment news (unexpected decrease in unemployment) decreases stock return volatility, and vice versa for bad unemployment news 10 August

57 3.2 Balance of Trade My null hypothesis is that balance of trade surprises have no effect on returns or return volatility. Brooks et al (1999) and Singh (1995) both examined whether there is any relationship between the Australian current account balance and Australian stock returns, but they found no evidence to suggest one. In the Japanese market, Kim s (2003) study did not show any relationship between Japanese balance of trade surprises and Japanese stock returns. Flannery and Protopapadakis s (2002) study, did not find any evidence that the balance of trade surprises and US returns were related. The null hypothesis is set out as follows: H2: No relationship between balance of trade surprises and stock returns/return volatility Evidence exists to support the alternative hypothesis that balance of trade surprises have a positive relationship with returns, in the sense that bad news reduces returns. In the UK, Becker et al (1995, p.1206) found higher than expected current account balances, which I assume to be good news, in the Australian context, increased returns. In Japan, Kim (2003, p.626) also found good current account surprises increase returns. The alternative hypothesis, based on this evidence, is set out below: H2a: Good balance of trade news (unexpected increase in balance of trade) increases stock returns, and vice versa for bad balance of trade news Other evidence shows the sign of the relationship between balance of trade news (or surprises) and returns can be positive or negative, suggesting an alternative hypothesis that balance of trade news increases return volatility, as opposed to returns. In Kim s (2003, p.624) study US balance of trade surprises, both good and bad, had a negative effect on US returns, highlighting the lack of consistency around the sign of the 10 August

58 relationship between balance of trade surprises and returns. This indicates it may be the absolute value of returns or volatility that is affected, as opposed to returns. More particularly, he also found good balance of trade surprises decreased US return volatility, while bad news increased it. Flannery and Protopapadakis (2002, p.766) found that US balance of trade announcement days, in general, were positively related to US return volatility over an earlier/longer period than examined by Kim (2003). This adds additional support to the hypothesis that balance of trade news affects return volatility. The alternative hypothesis based on this evidence is set out below: H2b: Good balance of trade news (unexpected increase in balance of trade) decreases stock return volatility and vice versa for bad balance of trade news In Japan, Kim (2003, p.626) found that good current account balance surprises (larger than expected) increase return volatility. Assuming this is driven by the balance of trade, the alternative hypothesis, based on this evidence, is as follows: H2c: Good balance of trade news (unexpected increase in balance of trade) increases stock return volatility Kearney and Daly (1998, p.603) found evidence of another alternative hypothesis. They reported a negative relationship between the conditional volatility of the current account deficit (negative balance of trade) and the conditional volatility of stock returns in Australia. The alternative hypothesis, based on this evidence, is set out below: H2d: A negative relationship exists between the size of balance of trade surprises and stock return volatility. 10 August

59 3.3 Retail Sales My null hypothesis for retail sales is that they have no effect on Australian stock market returns. Becker et al s (1995, p.1206) UK study found no relationship between returns and retail sales announcements, although it must be kept in mind, they did not separate the effects of good and bad news. The null hypothesis is set out below: H3: No relationship between retail sales surprises and stock returns/return volatility Economic theory-based alternatives support a relationship between retail sales and stock market returns via the indirect effect of consumption on stock returns through real GDP. As highlighted in Chapter 1, one of the most regular empirical results is that expected and actual output, measured using indicators such as industrial production, real GNP, or GDP, are positively related to stock returns. Retail sales are viewed as an economic indicator, on account of being a measure of consumption (Stock & Watson 1989, p.390). Consumption is a component of the Keynesian model of aggregate expenditure (Keynes 1936). The most basic representation of the model predicts consumption is positively related to income and output. Another possibility is that more emphasis should be placed on future expected real GDP than current real GDP. The permanent income hypothesis emphasises a positive relationship between changes in expectations of future income and changes in consumption spending (Friedman 1957). Changes in consumption, therefore, may be seen as a forecaster of changes in future income and output, and so one would expect consumption to be positively related to measures of output, such as real GDP. In turn, this links consumption to stock returns, through the real GDP/stock market return relationship. The alternative hypothesis, based on this theory, assumes retail sales are 10 August

60 a measure of consumption, which is a component of real GDP. 9 This hypothesis, therefore, assumes retail sales affect stock returns, through the hypothesised positive real GDP/stock return relationship discussed in Section 3.6 below. The alterative hypothesis for the retail sales and stock return relationship based on these theories is outlined as follows: H3a: Good retail sales news (unexpected increase in retail sales) increases stock returns and vice versa for bad retail sales news An opposing alternative is offered by neoclassical economic theory, which characterises output (production) as either being allocated to current consumption or investment, this implies that lower rates of current consumption mean an increased rate of savings, investment, capital accumulation and higher potential future output (Solow 1956, Swan 1956). Additionally, if borrowing supports current consumption, it is future consumption, and thus expenditure and output, that may be adversely impacted through future interest and principal repayments. Under these circumstances, rates of consumption that are considered too high could conceivably be related to lower levels of expected real GDP growth and thus stock returns over the long run. Kim s (2003, p.624) results for the US stock market provide evidence supporting this alternative hypothesis showing that bad (lower than expected) retail sales news announcements increase US returns. An alternative hypothesis based on this theory and evidence is set out below: H3b: Good retail sales news (unexpected increase in retail sales) decreases stock returns and vice versa for bad retail sales news 9 Multicollinearity is not an issue in this study design because real GDP announcements are made on different days to retail sales announcements and also because this study focuses on the differences between expected and actual values of each variable which can move more independently than the levels of the variables themselves. 10 August

61 With respect to stock market volatility, Kim (2003, p.624) also found evidence to support an alternative hypothesis of a relationship with retail sales in the US. Good US retail sales announcements decreased stock return volatility. The alternative hypothesis based on this evidence is as follows: H3c: Good retail sales news (unexpected increase in retail sales) decreases stock return volatility 3.4 Producer Price Index My null hypothesis for the producer price index is that it has no effect on Australian stock market returns. In the US, Kim (2003) found the PPI had no effect on the level of returns in the stock market. Using UK data, Becker et al (1995) could not detect any relationship between PPI announcement surprises and stock returns. Flannery and Protopapdakis (2002) found no significant relationship between the US PPI and US return volatility. The null hypothesis is set out below. H4: No relationship between retail sales surprises and stock returns/return volatility Assuming that the producer and consumer price index are an identical measure of inflation, the alternative hypotheses on the relationship between the consumer price index and stock returns (outlined in Section 3.5) should hold for the producer price index. Tiwari (2012, p.1571) sets out a theory where producer prices are normally set as a mark-up over wage costs that are driven by consumer prices. In turn, consumer prices are set by consumer demand. In these circumstances, consumer price changes should precede producer price changes, and assuming both price indices contain the same information, the stock market should react to consumer prices but not producer prices. 10 August

62 In the same study, however, Tawari (2012, p.1571) highlighted that it is equally plausible for producer price changes to precede consumer price changes in response to cost push shocks such as imported input price shocks. 10 The theory that PPI changes precede CPI changes, combined with the generalised Fisher hypothesis that postulates a one for one relationship between inflation and stock returns (outlined in Section 3.5), gives rise to the alternative hypothesis below: H4a: Good producer price index news (unexpected decrease in the producer price index) decreases stock returns, and vice versa for bad producer price index news Flannery and Protopapadakis (2002, p.766) found evidence to support an additional alternative hypothesis for returns. Using a longer and earlier period than Kim (2003), they showed the US PPI had a negative relationship with the level of returns on the US market. Tiwari s (2012) theory that PPI changes precede CPI changes, combined with the proxy effect hypothesis (outlined in Section 3.5), gives rise to the alternative hypothesis stated as follows: H4b: Good producer price index news (unexpected decrease in the producer price index) increases stock returns, and vice versa for bad producer price index news For return volatility, Kim (2003, p.625) found evidence of an alternative relationship in the US, namely that bad (higher than expected) US PPI news announcements are positively related to US equity market volatility. The alternative hypothesis for volatility, based on this evidence, is set out below: H4c: Bad producer price index news (unexpected increase in the producer price index) increases stock return volatility 10 Tiwari (2012) did, however, find that in Australia consumer price changes precede producer price changes. 10 August

63 3.5 Consumer Price Index My null hypothesis for the consumer price index is that it has no effect on Australian stock market returns. Gultekin (1983) studied the relationship between CPI based expected inflation and Australian stock market returns over the period January 1947 to December No significant relationship was found. Kim and In (2002) analysed the relationship between Australian CPI announcements and returns over a later period than Gultekin (1983), spanning July 1991 to December Again, no significant relationship was found. In Europe, Wasserfallen (1989) found there was no significant relationship between UK/Swiss consumer price surprises and UK/Swiss stock market returns. Kim (2003) found no relationship between US CPI surprises and US stock market returns. Turning to volatility in Australia, Kim and In (2002) found no relationship between CPI announcements and stock return volatility. Overseas, Flannery and Protopapadakis (2002) found no relationship between US CPI announcements and US stock return volatility. Similarly, Kim (2003) found no relationship between US CPI surprises and stock return volatility. The null hypothesis is set out below: H5: No relationship between retail sales surprises and stock returns/return volatility One of the most obvious alternative hypotheses is based on the Fisher Hypothesis. The Fisher Hypothesis characterises interest rates as consisting of a real component, determined by real factors, such as the time horizons of investors, productivity of capital, and expected inflation (Fisher 1930). More generally, the hypothesis predicts nominal expected returns on assets are positively related to expected inflation, as the nominal component will vary one for one with inflation. An alternative hypothesis, based on this theory, is set out as follows: 10 August

64 H5a: Good consumer price index news (unexpected decrease in the consumer price index) decreases stock returns and vice versa for bad consumer price index news Alternatively, Fama (1981, p.545) postulates a negative relationship between inflation and returns may result from a proxy effect. This assumes expected future real output is positively related to stock prices and the demand for money. If a decrease in expected future output is not offset by a decrease in money supply, it results in higher inflation. Inflation, therefore, becomes a proxy for changes in expected future output and stock prices. 11 Evidence, in support of a negative relationship, is presented in Wasserfallen (1989), Flannery and Protopapadakis s (2002) and Kim (2003). Wasserfallen (1989, p.622) found consumer price surprises in West Germany had a negative relationship with West German stock returns over the period 1977 to Flannery and Protopapadakis s (2002, p.766) US study observed a negative relationship between CPI surprises over January 1980 to December For the Japanese stock market, Kim (2003, p.626) found bad (higher than expected) Japanese CPI announcements were negatively related to Japanese returns. The alternative hypothesis, based on this theory and evidence, is stated below: H5b: Good consumer price index news (unexpected decrease in the consumer price index) increases stock returns, and vice versa for bad consumer price index news In Japan, Kim s (2003, p.627) evidence supported an alternative hypothesis for volatility. Good (lower than expected) CPI surprises were observed to have a negative 11 For the case of Australia, it is worth noting since mid-1993, the Reserve Bank of Australia has conducted monetary policy with a particular focus on maintaining inflation within a band of 2 to 3 per cent over the medium term (Reserve Bank of Australia 1999). The implication here is that unexpectedly high levels of inflation may be associated with a tightening of monetary policy or an increase in the overnight cash rate. The hypothesised effects of the overnight cash rate on stock returns are outlined in Section 3.7 below. 10 August

65 effect on Japanese stock return volatility, while bad (higher than expected) CPI announcements had a positive effect. The alternative hypothesis based on this evidence is set out below: H5c: Good consumer price index news (unexpected decrease in the consumer price index) decreases stock return volatility, and vice versa for bad consumer price index news 3.6 Real Gross Domestic Product My null hypothesis for real GDP is that there is no relationship with Australian stock market returns. Over the period January 1989 to December 1993, Brooks et al (1999) found no relationship between GDP news announcements (using ARIMA model based forecasts) and Australian stock returns. Separating negative and positive surprises had no effect on the results. Kim and In s (2002) Australian study confirmed this result using GDP announcements (but without adjustment for the expected component) over a later period January 1991 to December Flannery and Protopapakis (2002) found no relationship between US real GNP announcements and the level of returns. Kim (2003) also investigated the US market, testing the effect of real GDP announcements on the US stock market. Again no relationship was found. The null hypothesis is set out below: H6: No relationship between real GDP surprises and stock returns/return volatility Despite this, there is strong theoretical and empirical support for alternative hypotheses, although this is mainly based on foreign studies. A number of theories are consistent with the view that growth in real GDP (output) is positively related to returns. Increases in output lead to increases in real rates of return 10 August

66 on capital, hence attracting capital investment (Jorgenson 1971). A rational expectations view of the real GDP/stock market price relationship is that expectations of future real output should set current security prices (Fama 1981). Campbell and Shiller (1988) outlined the mechanism through which corporate earnings (which are inextricably linked to output), forecast future dividends, thereby, creating a link between expected output and future dividends. In turn, their expected future dividends are discounted to set current security prices through Gordon s (1962) dividend growth model. These theories establish a positive relationship between expected/actual output (measured using indicators such as industrial production, real GNP, or GDP) and stock returns. This positive relationship is one of the most regularly found empirical results among foreign studies (Asprem 1989, Fama 1981, Schwert 1990, Mukherjee & Naka 1995, Cheung and Ng 1998, Ratanapakorn & Sharma 2007, Humpe & Macmillan 2009). In Australia, Groenewold (2004, p.660) detected a positive relationship between real GDP shocks and Australian stock returns. 12 The alternative hypothesis, based on these theories and evidence, is outlined below: H6a: Good real GDP news (unexpected increase in real GDP growth) increases stock returns, and vice versa for bad real GDP news With respect to stock market return volatility, considerable evidence is found in Australia and overseas to show real GDP announcements affect stock return volatility. Kim and In (2002, p.578) found that Australian return volatility was positively influenced by Australian real GDP announcement days (that is, the announcement itself and not the specifics of the news content). Flannery and Protopapakis s (2002, p.766) study on the US also found real GNP announcement days were positively 12 Groenewold (2004) differs from my study in that he had a focus on long run relationships between output and stock prices as opposed to macroeconomic surprises and daily returns. 10 August

67 related to stock return volatility. This gives rise to an alternative hypothesis that real GDP surprises, of any sign, are positively related to stock return volatility. This hypothesis is as follows: H6b: Real GDP news (unexpected increase or decrease in real GDP growth) increases stock return volatility A second alternative, with respect to volatility, is that good (higher than expected) real GDP surprises reduce stock return volatility. Kim (2003, p.624) observed good real GDP surprises in the US reduce US return volatility, which supports the hypothesis set out below: H6c: Good real GDP news (unexpected increase in real GDP growth) decreases stock return volatility 3.7 Overnight Cash Rate My null hypothesis is that the overnight cash rate has no effect on Australian stock market returns. Wasserfallen (1989) observed Swiss and UK stock market returns showed no significant relationship with nominal or real interest rates. The null hypothesis is outlined as follows: H7: No relationship between overnight cash rate surprises and stock returns/return volatility An alternative hypothesis is that interest rates have a positive relationship with stock returns. Expectations theory, when applied to stock returns, predicts short-term interest rates should have a one for one relationship with stock returns. This is because riskier classes of assets, such as stocks, have a constant premium over risk free assets, such as sovereign bills (Campbell 1987, Fama & Schwert 1977). Constant 10 August

68 returns imply stock prices will adjust to offset changes in the discount rate, stemming from changes in risk-free rate to ensure the premium remains constant. The alternative hypothesis, based on this theory, is set out below: H7a: Good overnight cash rate news (unexpected decrease in overnight cash rate) decreases stock returns and vice versa for bad overnight cash rate news Conversely, Shiller and Beltratti (1992) outlined, in the context of a rational expectations present value model, a rise in the expected discount rate would cause bond prices to fall and bond yields to rise, as their traditionally fixed coupons yield higher returns as a proportion of their price. 13 This makes bonds a more attractive investment vis-à-vis stocks, and so, stock prices need to fall to induce investors to buy stocks. This theory was outlined in the context of long-term bonds, but would apply equally for overnight cash rates if cash rate changes were reflected in longer term bond yields. Empirical support for this alternative is found in Wasserfallen s (1989, p.622) study on the West German market. West German stock market returns showed a significant negative relationship with nominal interest rate surprises. Flannery and Protopadakis s (2002, p.766) results also exhibited a negative relationship between lagged three-month Treasury bill yields and US stock returns. This relationship, however, was based on continuously reported market yields, as opposed to an announcement on interest rate policy. Assuming the relationship between continuously reported interest rates and stock returns holds for interest rate surprises and stock returns, the alternative hypothesis, based on these theories and empirical evidence, is set out below: 13 Discount rates are typically driven by interest rates such as the overnight cash rate which reflect the cost of alternative investment opportunities. 10 August

69 H7b: Good overnight cash rate news (unexpected decrease in overnight cash rate) increases stock returns and vice versa for bad overnight cash rate news Turning to volatility, the results of Kearney and Daly s (1998, p.603) Australian study provided support for an alternative hypothesis, showing Australian stock market volatility is positively related to the conditional volatility of the three-month bank accepted bill rate. I assume volatility in cash rate surprises directly translates into three-month bank accepted bill volatility. The alternative hypothesis, based on this evidence and assumption, is set out below: H7c: A positive relationship exists between the absolute size of overnight cash rate surprises and stock return volatility 3.8 Consumer Sentiment While I refer to consumer sentiment as a macroeconomic variable, it differs from the other variables in that it is a behavioural factor, as opposed to a macro factor (Harvey, Liu & Zhu 2014, p.4). My null hypothesis is that consumer sentiment has no relationship with Australian stock market returns. Although evidence to the contrary exists in Australia (Akhtar et al 2011), I pose this as the null hypothesis for the sake of consistency with the null hypotheses posited for the other macroeconomic variables above. This is also consistent with the formulation of my statistical testing methods, which are designed to detect evidence of a relationship through rejection of the null hypothesis. The null hypothesis, based on this rationale, is set out below: H8: No relationship between consumer sentiment surprises and stock returns/return volatility De Long et al (1990) outlined behavioural theories that support an alternative hypothesis of a positive relationship between consumer sentiment and stock returns. 10 August

70 They argued irrational investors, who trade based on sentiment, induce changes in returns that are both costly and risky for arbitrageurs to force back to fundamental levels. This is because risk stems from the unpredictability of investor sentiment, and arbitrageurs typically have constraints on their investment horizons. For example, an arbitrageur may go out of business waiting for prices to return to fundamentals. Baker and Wurgler (2007) found when investor sentiment is low, the subsequent returns on the stocks of firms that are difficult to value tend to become high (relative to their long-run average). This suggests that low sentiment leads to the stocks of such firms being initially undervalued. In the US, Qiu and Welch (2006) found the consumer confidence index is a proxy for investor sentiment, and that it correlates with the excess rate of return on small firms, thus linking investor sentiment to consumer sentiment. Taken together, these studies suggest consumer sentiment is positively related to returns; however, the studies emphasise this sentiment is linked to irrational beliefs about future corporate cash flows. Another perspective is that consumer sentiment is a forecaster of, and may even cause, changes in consumption expenditure (Carroll, Fuhrer & Wilcox 1994). This could indirectly affect stock returns through changes in output (or real GDP). Other research shows changes in consumer sentiment precede changes in output. Matsusaka and Sbordone (1995) hypothesised that expectations of lower income (low consumer sentiment) lead to lower orders of goods produced to buyer specifications, and which cannot easily be resold without significant loss. As a result, an economy s build-to-order firms experience lower employment, resulting in reduced consumption, incomes and output. As per the discussion in Section 3.3, any effects that consumer sentiment have on output, or real GDP, can result in an indirect effect on stock returns. Given that real GDP is typically observed to be positively related to 10 August

71 returns (see Section 3.6), this theory would suggest that consumer sentiment is positively related to stock returns. Akhtar et al (2011, p.1248) found the Westpac-Melbourne Institute consumer sentiment index announcements were positively related to returns, although it was only the decreases in consumer sentiment that were associated with negative Australian stock market returns; positive announcements had no effect. 14 The alternative hypothesis, based on these theories and evidence, is set out below: H8a: Bad consumer sentiment news (unexpected decrease in consumer sentiment) decreases stock returns With respect to volatility, De Long et al (1990) again provided theoretical support for an alternative hypothesis; that is, investor sentiment may be positively related to stock return volatility. They reasoned irrational investors, trading based on sentiment, induced sustained price movements (in both directions) that are both costly and risky for arbitrageurs to force back to fundamentals. The alternative hypothesis, based on this theory, is outlined below: H8b: A positive relationship exists between consumer sentiment surprises and stock return volatility 14 Note that this is still a positive relationship, albeit asymmetric, because decreases in consumer sentiment are related to decreases in returns meaning the variables move in the same direction hence the correlation is positive. 10 August

72 4 Methodology 4.1 Returns Stock returns based on market indices are calculated using equation (2). R t p t ln x100 pt 1 Where: pt is the closing share market index price on the trading day in question; and p is the closing share market index price on the previous trading day. t Surprises (Unexpected Components of Announcements) The surprise or news series Surprise kt, for each macroeconomic variable is constructed using equation (3). Surprise E ( Announcement ) Announcement k, t t 1 k, t k, t Where: E ( Announcement ) is the forecast value for period t prior to period t ; and t 1 k, t Announcement kt, are the realised values of each announcement at time t. A positive value of the Surprise kt, series for each of the k macroeconomic variables indicates the expected value is high relative to the outcome. The interpretation of whether this is considered good or bad news is not straightforward and is discussed with direct reference to the results (Chapter 6). 10 August

73 The surprise series reflects the unexpected components of the forecasts or news. News by definition is new and unexpected information. Conversely, the expected component of an announcement is assumed not to be news to the market. 4.3 Control Variables US Returns US stock market returns are calculated using equation (4). US t p p US, t 1 US, t ln x 100 Where p US, t is the US stock market index level on day t. Oil Returns Oil returns are calculated using equation (5). Oil t f t ln x 100 ft 1 Where f t is the oil future price index level on day t. Term Spread The term spread on government bonds is calculated using equation (6). Where: TS y y t long, t short, t y long, t is the yield reported by Bloomberg, based on their 10-year government bond index, and expressed as a whole number percentage; and 10 August

74 y short, t is the yield reported by Bloomberg, based on their 5-year government bond index, and expressed as a whole number percentage (details in Section 5.3). Default Spread The default spread on Australian corporate bonds is calculated using equation (7). Where: DS y y t corporate, t government, t y corporate, t is the yield on one of Bloomberg s Australian corporate bond indices, expressed as a whole number percentage (details in Section 5.3); and y government, t is the yield on a Bloomberg government bond index that is of the same tenor as y corporate, t, expressed as a whole number percentage. 10 August

75 4.4 Returns Estimation The form shown in (8) is used to model returns. It is an autoregressive moving average (ARMA) specification that is fitted to observed returns. TS j US p Fri R a a a t c Hol Day a Hol t i, Day i, t i Mon Control j, Control i, t k, Surprise k, t k Unem q ar b i t i i t i t i 1 i 1 CSI a Surprise Where: R t are the daily log percentage returns on the stock market indices; a c is a constant; a Hol is the coefficient on Hol t dummy variables assigned to days after holidays; Fri ai, Day are the coefficients on dummy variables for Monday through to Friday, i Mon but excluding Wednesday; TS a j, Control are the coefficients on each of the control variables: t j US US, Oil, DS t t andts ; t CSI ak, Surprise are the coefficients on the macroeconomic surprise variables; k Unem p a irt i i 1 are the coefficients on the autoregressive lags up to order p; q b i t i i 1 are the coefficients on the moving average terms up to order q; and t are the regression residuals. For the sake of parsimonious presentation p q ar b is abbreviated to M () i t i i t i t i 1 i August

76 Day-of-the-week and holiday variables are used to capture any return effects that may be attributed to different days of the week (Gultekin 1983, Fama 1991). The holiday variable accounts for return effects resulting from information accumulated when the Australian Stock Exchange is closed. Upon opening after a holiday, it is thought this information is factored in. The surprise series coefficient ak, Surprise quantifies the sensitivity of daily returns to each of the k macroeconomic surprises. An additional specification is tested, replacing: CSI k Unem a Surprise k, Surprise k, t with CSI k Unem a CSI Surprise Bad News D k, Surprise k, t a D k, Bad News k, t k Unem. Where: D is a dummy variable that takes the value of one if Surprise kt, does not Surprise kt, equal zero or zero otherwise; and Bad News kt, D is a bad news dummy variable that takes the value of one if the announcement contains bad news, or is otherwise zero. This additional specification is designed to capture the average effect of announcement days containing good news surprises a k, Surprise and the average effect of announcement days containing bad news surprises. The latter is found by adding the marginal effect of bad news ak, Bad News to good news a k, Surprise k, Bad News a. This specification helps to determine if good news has a different effect on returns than bad news. This returns equation is estimated using Eviews 7 econometric software, which fits the specification using the least squares (NLS and ARMA) method. It is estimated alone and then simultaneously (using the autoregressive conditional heteroscedasticity method in Eviews) with the volatility specification outlined below. 10 August

77 4.5 Volatility Estimation This study uses an EGARCH specification to model volatility, and was chosen after carrying out the analysis conducted in Appendix B. As noted in Section 2.3.2: Conclusions from the Literature, Kim s (2003, p.618) EGARCH model was one of the most successful for finding significant relationships between foreign macroeconomic surprises and the Australian stock market. The EGARCH specification is set out in (9). 2 ln( t ) b TS j US Fri Hol b Day Hol t i, Day i, t i Mon CSI b Control Surprise j, Control i, t k, surprise k, t k Unem p r q 2 t j t j j ln( t j) j j j 1 j 1 t j j 1 t j b Where: 2 ln( t ) is the log of the estimated conditional variance or volatility; is a constant; b Hol is the coefficient on Hol t dummy variables assigned to days after holidays; Fri bi, Day are the coefficients on dummy variables for Monday through to Friday, i Tue but excluding Wednesday; TS bj, Control are the coefficients on each of the control variables: t j US and TS ; t US, Oil, DS CSI bk, surprise are the coefficients on the absolute value of surprise variables; k Unem p j are the coefficients on the lagged GARCH effects up to order p; i 1 t t 10 August

78 r j are the coefficients capturing the sign or leverage effects of the j 1 standardised residuals t j t j up to order r = q; q j are the coefficients on absolute value of the standardised residuals j 1 t j t j capturing ARCH effects up to order q; and is the estimated conditional standard deviation at time ( t j) used to t j standardise the regression residuals t. 15 Again, for the sake of parsimonious presentation, p r q 2 t j t j j ln( t j) j j j 1 j 1 t j j 1 t j is abbreviated to () V. The estimated conditional variance, and hence standard deviation, is based on an assumption made regarding the distribution of standardised residuals. This assumption is investigated in Appendix B. Again, holiday, day-of-the-week and control variables are included (see Section 4.4). The coefficient on the absolute value of surprises, CSI k Unem b k surprise captures the sensitivity of return volatility to the absolute size of macroeconomic surprises. As with returns, an additional specification is tested, replacing: CSI k Unem b Surprise k, surprise kt, with CSI k Unem b CSI Surprise Bad News k, Surprise Dk, t bk, Bad NewsDk, t k Unem 15 The volatility equation used here is limited to controlling for the potentially differing effects of negative and positive values of the control variables on volatility. Using the absolute magnitude of movements in control variable returns would be a fruitful addition to the research allowing the absolute size effects of control variables on volatility to be captured. 10 August

79 Where: D is a dummy variable that takes the value of one if Surprise kt, does not Surprise kt, equal zero or zero otherwise; and Bad News kt, D is a bad news dummy variable that takes the value of one if the announcement contains bad news and zero otherwise. This additional specification is designed to capture the average effect of announcement days that contain good news surprises b k, Surprise and the average effect of announcement days containing bad news surprises. The latter is found by adding the marginal effect of bad news bk, Bad News to good news: bk, Surprise bk, Bad News. This specification allows us to determine whether good news has a different effect on return volatility to bad news. 10 August

80 5 Data This chapter sets out the variables of interest for my study and the data used to represent those variables. Stock returns are the independent variable explained by macroeconomic surprises, so here, I detail the data I intend to use, as well as their statistical characteristics. The macroeconomic surprises (as the explanatory variables) are then explained outlining the forecasts and the announcement series used to calculate surprises and statistical characteristics. The chapter finishes with an outline of the control variables, which are used to remove the effects of other factors known to affect Australian stock returns. All variables are observed on a daily basis over the period January 2000 to December Stock Market Indices The ASX All Ordinaries index is the most commonly used stock market index in Australian literature. This index was used in the work of Kearns & Pagan (1993), Brooks et al (1999), Kim & In (2002), Groenewold (2003), Kim (2003), Chaudhuri & Smiles (2004) and Akhtar et al (2011). However, as of 3 April 2000, the All Ordinaries index was restructured by the ASX to reflect a greater proportion of the market and include the 500 largest companies. Prior to this, it reflected only stocks (Worthington 2009, p.46). ASX National Manager of Market Data, John Ying, noted in September 1999: The existing liquidity requirements [on the All Ordinaries index] will be removed as these are far more appropriate to benchmark indices. (Australian Stock Exchange 1999) 10 August

81 The announcement was made in relation to the re-establishment of the All Ordinaries as an index to reflect overall market movements, and also the establishment of new indices, including the ASX 200 as a benchmark index for portfolio performance benchmarking. The All Ordinaries index prior to 2000 is, in fact, more comparable to the ASX 200 index, which was developed post It is important to note the relaxation of liquidity requirements in the All Ordinaries index is likely to lead to thin trading issues in the index, such as delayed price reactions. This could possibly result in the All Ordinaries index being relatively slow to reflect new information because it now includes a large number of small market capitalisation stocks, which can trade infrequently. I have, therefore, opted to use the ASX 200 index in this study. For the sake of robustness, however, I will use both the ASX 200 and All Ordinaries indices to determine if the results are sensitive to the choice of index. Also in this study, I use the total returns (also known as cumulative) indices because this version of the index is conventionally used in stock return studies. However, it is worth noting Groenewold (2003, p.460) found little difference in his results, between those estimated on the cumulative index and those estimate on the non-cumulative index. Daily observations of the S&P ASX 200 and the All Ordinaries total returns index were acquired from Datastream and used as the measure of market returns. 16 Daily percentage returns for both indices were calculated as set out in The indices were not adjusted for dividend effects. The Datastream code for the S&P ASX 200 total returns index is ASX200I(RI). The code for the All Ordinaries total returns index is ASXAORD(RI). 10 August

82 03/01/ /07/ /02/ /09/ /04/ /11/ /06/ /01/ /08/ /03/ /10/ /05/ /11/ /06/ /01/ /08/ /03/ /10/ /05/ /12/ /07/ /01/ /08/ /03/ /10/2013 The ASX 200 Indices give 3652 daily returns observations to work with, spanning the period 4 January 2000 to 31 December The series plotted in Figure 1 appears to exhibit volatility clustering, particularly around 2008 and 2011, which is suggestive of time varying volatility. Figure 1 ASX 200 Index Total Daily Returns per cent The summary statistics for the series in Table 3 indicate daily returns have a significant positive bias, as indicated by the mean of per cent. The minimum daily return of -8.7 per cent occurred on 10 October 2008 with the onset of the GFC. The maximum of 5.6 per cent is smaller by comparison and occurred just a few days after the minimum return on 13 October August

83 Table 3 ASX 200 Index Total Daily Returns Summary Statistics ASX 200 Daily Total Returns (%) Mean Standard Error Median Mode NA Standard Deviation Sample Variance Kurtosis 6 Skewness 0 Range Minimum Maximum 5.63 Count 3542 Like the ASX 200 returns series, 3542 observations were available for the All Ordinaries index from 4 January 2000 to 31 December Visually, the All Ordinaries daily total returns (Figure 2) indicate a very similar pattern to the ASX 200 returns series. Clustering of volatility in the All Ordinaries returns coincides with same time periods as the ASX 200, notably 2008 and Figure 2 Australian All Ordinaries Index Total Daily Returns 10 August

84 The differences between the All Ordinaries index and the ASX 200 returns only really become apparent in the summary statistics (shown in Table 4). The range and standard deviation of the All Ordinaries returns over the period are marginally smaller than those for the ASX 200, indicating the inclusion of small market capitalised stocks lowers the level and variability of returns over the period. The distribution also has a slightly different shape to the ASX 200 returns, with the mean of sitting further below the median of and a negative skew of one. This suggests a marginally higher probability of lower returns than the ASX 200 index. Table 4 All Ordinaries Index Total Daily Returns Summary Statistics All Ordinaries Daily Returns (%) Mean Standard Error Median Mode NA Standard Deviation Sample Variance Kurtosis 6 Skewness -1 Range Minimum Maximum 5.36 Count Stationarity of Stock Returns The ASX 200 and All Ordinaries return series were tested for stationarity to ensure they were suitable for time series modelling. An examination of Figure 1 and Figure 2 did not indicate any drift or trend in the daily returns. Accordingly, the augmented Dicky-Fuller test without drift or trend was carried out to test for the null hypothesis of a unit root or non-stationarity. 10 August

85 Table 5 Augmented Dickey-Fuller Unit Root Tests - No Drift or Trend Total Returns Series Index test-statistic Critical Value at 5 per cent Standard and Poor s ASX All Ordinaries The results in Table 5 show the absolute value of the test statistics of both series were far below the critical value, thus strongly rejecting the hypothesis of a unit root. Based on the strength of these results, I considered it unnecessary to carry out any additional stationarity tests. 5.2 Macroeconomic Surprises As outlined in 4.2, macroeconomic surprises are the difference between the expected component of the announcement (the forecast) and announcement itself. The equation is reproduced in (10). Surprise E ( Announcement ) Announcement k, t t 1 k, t k, t Where: Et 1 ( Announcement k, t ) is the forecast value for period t immediately prior to period t ; and Announcement kt, are the realised values of each announcement at time t Forecasts Consensus forecasts were obtained from Money Market Services (MMS) International, a former subsidiary of S&P. MMS Asia surveys the forecasts of marketmaking participants in Australia (Haver 2013). The medians of these surveyed forecasts (as first reported) and their corresponding dates are accessed through Haverselect. MMS surveys are prevalent in the literature review (see Singh 1993 & 1995, and Kim 2003, Becker, Finnerty and Friedman 1995 and Flannery and 10 August

86 Protopapadakis 2002) with Singh providing evidence to suggest the survey data has superior information content to that of basic ARIMA forecasts. Unemployment information forecasts are available from 2003, while balance of trade, retail sales, the PPI and CPI are available from The MMS Asia forecasts are outlined in Table 6. Table 6 Forecast Money Market Services Consensus Macroeconomic Forecasts Original Source Format Frequency Unemployment per cent level Monthly Observation Period December December 2013 Missing Forecasts 2 Balance of Trade $ Billion Monthly Retail Sales Producer Price Index per cent change from previous month per cent change from previous quarter Monthly Quarterly June 2005 December 2013 June December 2013 September December Consumer Price Index per cent change from previous quarter Quarterly September September The observation period relates to periods (for example quarter or month) when the macroeconomic variable was under observation. 17 Real GDP, the consumer sentiment index and the overnight cash rate forecasts could not be adequately sourced from MMS. I used other methods to obtain these forecasts, which are outlined in their respective sections. The structure of these forecasts is outlined in Table As distinct from the announcement day date which is when the value for the macroeconomic variable observed during the observation period is released. 10 August

87 Table 7 Real GDP Forecast Other Macroeconomic Forecasts Original Source Format per cent change from last quarter Frequency Quarterly Overnight Cash Rate per cent level Monthly Consumer Sentiment index level Monthly Observation Period March 2000 September 2013 September December 2013 April 2004 December 2013 Missing Forecasts Announcements With respect to macroeconomic announcements, I consider data releases from the Australian Bureau of Statistics (ABS) and Reserve Bank of Australia (RBA) to be the most relevant sources. I view these authorities as the most unbiased source of information available to investors, given their non-commercial objectives. Announcements made by the ABS included balance of goods and services (trade), CPI, PPI, real GDP, unemployment and retail sales. Overnight cash rate announcements were those made by the Reserve Bank of Australia (RBA). The Westpac-Melbourne Institute consumer sentiment index was the only announcement sourced from a non-federal government organisation. I believe the value of the index is sufficiently impartial to commercial interests, on account of the accuracy of the index value itself giving it its commercial value. The announcements are outlined in Table August

88 Table 8 Macroeconomic Announcement Values Announcement Format Frequency Date Range Source Unemployment Per cent level Monthly Balance of Trade $ Billion Monthly Retail Sales Producer Price Index Consumer Price Index per cent change from previous month per cent change from previous quarter per cent change from previous quarter Monthly Quarterly Quarterly January 2004 December 2013 August 2005 December 2013 August 2005 December 2013 October 2005 November 2013 October 2005 October 2013 Australian Bureau of Statistics Australian Bureau of Statistics Australian Bureau of Statistics Australian Bureau of Statistics Australian Bureau of Statistics Real GDP per cent change from last quarter Quarterly January 2000 December 2013 Australian Bureau of Statistics Overnight Cash Rate per cent level Monthly Consumer Sentiment index level Monthly September 2003 December 2013 April 2004 December 2013 Reserve Bank of Australia Westpac Melbourne Institute Accurately pairing the timing of macroeconomic surprises with their associated stock market returns necessitated matching the specific date of the unrevised value to the announcement itself. It was important to use the unrevised value because revised values contained information that was not available on the date when the announcement was first released and, therefore, not reflected in returns. Use of the revised values could have masked the impact of the original unrevised values on the stock market, thus obscuring any underlying relationship that may have existed, and making it undetectable in regression analysis. Unrevised macroeconomic announcement values for each of the ABS announcements were sourced from MMS. The availability of unrevised announcements was the main constraint on the number of observations available for analysis in my study. This is because not all series were available over the entire period from January 2000 to December August

89 5.2.3 Surprises Before applying equation (1), data that was not expressed in unitless measures (such as a percentage level or percentage change) was converted to percentage measures. This was for consistency with stock market returns, which are expressed in percentages. The summary statistics for the resulting surprise series are shown in Table 9. Table 9 Summary Statistics for Macroeconomic Surprises (%) Unemployment Balance of Trade Retail Sales Producer Price Index Consumer Price Index Real GDP Cash Rates Consumer Sentiment Index Mean Standard Error Median Mode NA 0 NA Standard Deviation Range Minimum Maximum Count Each series is explained in detail below. Unemployment The unemployment rate is the percentage of people in the labour force who are unemployed as measured by the ABS monthly labour force survey (Australian Bureau of Statistics 2014a). For example, if the results of the monthly survey show 12 million people are in the labour force, but of these, 708,000 are classified as being unemployed, the unemployment rate would be 5.9 per cent. The raw unemployment 10 August

90 announcements and forecasts are expressed as monthly seasonally adjusted percentage levels. 18 This percentage format was desirable for the purposes of my regression, and no further conversion was required. MMS forecasts were available from December 2003 and paired with unrevised unemployment announcements from this point on. One unemployment announcement was missing (August 2013) and coincided with one of two missing MMS forecasts. These missing values resulted in the loss of two of the 120 observations spanning December 2003 to November Announced values for months after November 2013 were not used because they were announced after December 2013 the limit of my study s observation period. In total, that left me with 118 pairs of observations. The announced values were subtracted from the forecasts to create the unemployment surprise series plotted in Figure The seasonally adjusted series removes the effects of estimated month-to-month seasonal variation in unemployment. 10 August

91 15/01/ /05/ /09/ /01/ /05/ /09/ /01/ /05/ /09/ /01/ /05/ /09/ /01/ /05/ /09/ /01/ /05/ /09/ /01/ /05/ /09/ /01/ /05/ /09/ /01/ /05/ /09/ /01/ /05/ /09/2013 Figure 3 Unemployment Rate Surprises per cent The mean value of the forecast error was 0.05 per cent, with the standard error of 0.02 per cent, indicating a statistically significant upward bias in the forecasts. 19 The implication of these results is that the announced unemployment rate is often lower than expected. This is highlighted in Figure 3 by the large number of values above the zero axis representing overestimates. Balance of Trade The balance of trade measures the net dollar value of goods and services exported against those imported, on a monthly basis. The data from the ABS is expressed in seasonally adjusted billions of dollars (Australian Bureau of Statistics 2014b) This assumes that the series is normally distributed. 20 Seasonally adjusted estimates are derived by estimating the systematic calendar related influences and removing them from the original estimates see 8bd2cdca256f960075c84a!OpenDocument for more details. 10 August

92 The series are converted to percentages for consistency with stock market returns in the regression. The forecast data was converted to percentage changes using equation (11). Where: E ( BOT ) BOT E BOT x t t 1 t t(% t 1) 100 absolute value( BOTt ) Et( BOTt 1) are the balance of trade forecasts in billions of dollars; BOT is the actual balance of trade in billions of dollars; and absolute value( BOT t ) is the absolute value of actual BOT in billions of dollars. The actual balance of trade data was also converted to a percentage change as shown in (12). BOT BOT t 1 t % BOTt 1 x 100 absolute value( BOTt ) The absolute values in the denominator were required to preserve the correct sign of the change. It should be noted that values close to zero in period t might result in very large percentage changes in period t 1. MMS forecasts were only available from June 2005, giving me 103 observations to work with. Two observations were lost because the November and December 2013 observation period announcements occurred in 2014, which is outside the range of my study. The resulting 101 observations are plotted in Figure August

93 02/08/ /12/ /04/ /08/ /11/ /04/ /08/ /12/ /04/ /07/ /12/ /04/ /08/ /12/ /04/ /08/ /12/ /04/ /08/ /12/ /04/ /08/ /12/ /04/ /08/ /12/2013 Figure 4 Balance of Trade Surprises per cent The large spike and dip shown in September 2012 and July 2013 result from the trade balance being unusually close to zero in the month prior. That is, the denominator in equation (11) and (12) was unusually close to zero and as a result dramatically scaled up the expected and actual percentage change calculated by those equations. In order to maintain consistency with the other macroeconomic surprises and avoid manipulation of data which may be interpreted as arbitrary the outliers were left in the data set. 21 The mean forecast error is positive. However, it is very small compared to the mean s standard error, suggesting it is not significantly different from zero. 22 The median forecast error, however, is also greater than zero, which confirms overly optimistic 21 I note that Andersen et al (2007, p.258) implement an alternative data preparation technique in calculating macroeconomic announcement surprises where the surprise is divided by the standard deviation of the surprise component. This technique may possibly mitigate the effect of balance of trade outliers. 22 This assumes a normal distribution. 10 August

94 balance of trade forecasts were common over the period. The two extreme values are of a similar magnitude to each other at 2844 and per cent. Retail Sales Retail sales are the monthly dollar value turnover of retail trade for Australian business. The data reported by the ABS, represent month-to-month percentage changes in dollar values that are seasonally adjusted (Australian Bureau of Statistics 2014c). 23 No conversion to percentage was therefore required. Forecasts were only available for June 2005 onward with one value missing in March 2012, reducing the number of observations available for my study to 102. Additionally, forecasts for November and December 2013 were not announced until after 2013, and hence, are outside the observation period for my study. This further reduced the number of observations to 100. The series are plotted in Figure Estimating the systematic calendar related influences and removing them from the original estimates derive seasonally adjusted estimates. See 8bd2cdca256f960075c84a!OpenDocument for more details. 10 August

95 02/08/ /11/ /03/ /08/ /11/ /04/ /08/ /12/ /04/ /07/ /12/ /04/ /08/ /12/ /03/ /08/ /12/ /03/ /08/ /12/ /04/ /08/ /12/ /04/ /08/ /12/2013 Figure 5 Retail Sales Surprises per cent The series exhibits a higher level of volatility from 2008 to 2010 than for the earlier period. This could be associated with the onset of the global financial crisis. Also, the ABS considered data on retail sales from July to November 2008 as: of limited use for measuring month-to-month estimates because of the increased volatility in these series due to the smaller sample size and the rotation effect of having a different third of the sample reporting each month Consequently, at the time of this study, the ABS did not make seasonally adjusted monthly change retail sales data available over this brief period. The data as it was first announced however, was available through MMS and this is used to construct the forecast errors reported in Figure 5. Statistical precision of the ABS figures is of limited relevance it is the stock market s reaction to these announced figures that is the central concern of this study. 10 August

96 The mean absolute deviation in percentage changes was 0.01, which was not significantly different from zero in light of the mean standard error of This suggests that forecasts are not biased. Additionally, the median is zero, which also tends to indicate unbiasedness. However, the mode is 0.4, suggesting the most common outcome is an overestimate. Producer Price Index The final commodities PPI measures the quarterly change in the price index established by the ABS for products ready to be sold for immediate consumption, capital formation, or export. The quarter-on-quarter seasonally adjusted series is published as a per cent change, and requires no conversion for consistency with the unit of measurement used for stock returns (Australian Bureau of Statistics 2014d). 25 Forecasts were available from the 2005 September quarter onward, reducing the number of observations available to 34. The December 2013 quarter announcement was not released until 2014, reducing the final sample to 33 observations. 26 The series are plotted in Figure This assumes that the series is normally distributed. 25 Estimating the systematic calendar related influences and removing them from the original estimates derive seasonally adjusted estimates. See 8bd2cdca256f960075c84a!OpenDocument for more details. 26 The year 2014 is beyond my study s observation period. 10 August

97 24/10/ /06/ /10/ /04/ /10/ /04/ /10/ /04/ /10/ /04/ /10/ /04/ /10/ /04/ /11/ /05/ /11/2013 Figure 6 Producer Price Index Surprises per cent The mean of 0.02 per cent is not significant. The median and mode confirm the lack of clear evidence of bias with the median exhibiting a slightly positive bias of 0.1, while the mode conversely shows a slightly negative bias of Consumer Price Index The CPI, reported on a quarterly basis, measures the general level of prices for consumer goods and services consumed by Australian households. The index is expressed as a quarter-on-quarter per cent change, or quarterly inflation, and no transformation of the data is required to make it unitless (Australian Bureau of Statistics 2014e). Only 33 CPI forecasts were available from MMS from September 2005 onwards, paring back the number of observations available for analysis from 56 to 33. The surprises based on these observations are plotted in Figure August

98 26/10/ /04/ /10/ /04/ /10/ /04/ /10/ /04/ /10/ /04/ /10/ /04/ /10/ /04/ /10/ /04/ /10/2013 Figure 7 Consumer Price Index Surprises per cent The surprises or forecast errors appear to be fairly symmetrically distributed around zero. This is confirmed by the mean and median returning a value of zero. The most common surprise, however, is negative as shown by the mode of This indicates, if anything, the market tends to underestimate inflation. Real Gross Domestic Product Growth Real GDP growth measures the change in total value of goods and services produced in Australia, holding prices constant from a particular base year. 27 The MMS real GDP consensus forecast series contained 18 forecasts spanning the December quarter 2004 to the September quarter A more comprehensive series of forecasts is available from the RBA spanning March 2000 to August 2011 (Reserve Bank of Australia 2012). However, their more recent forecasts in the Statements on 27 Estimating the systematic calendar related influences and removing them from the original estimates derive seasonally adjusted estimates. See 8bd2cdca256f960075c84a!OpenDocument for more details. 10 August

99 Monetary Policy are available only for June and December, meaning only 26 observations are available from Additionally, in their November 2012 discussion paper, the RBA highlighted their real GDP forecasts have very little explanatory power (Tulip & Wallace 2012, p.30). As a result, I concluded forecasts might not be reliable or frequent enough to produce robust estimates of market expectations of real GDP growth. As an alternative, I sourced the most up-to-date (revised) real GDP data from the ABS, dating back to December 1959, and modelled expectations on a naïve model based on an expanding window of the historical data. This is similar to the approach outlined in Singh 1993 and Augmented Dickey-Fuller unit root tests were carried out on the 216 observations of the seasonally adjusted per cent change series spanning December 1959 to September 2013 (see Table 10). Using the augmented Dickey-Fuller test, the series tested as stationary (when no drift or trend was included), indicating that an ARMA model could be meaningfully fitted. The Akaike Information Criterion (AIC) was used to compare competing models. An AR (1) model resulted in the lowest AIC where only the intercept was statistically significant. This suggests the historical mean of the full information set produces the best forecast, and was used accordingly to produce forecasts. 10 August

100 Table 10 Real GDP Growth ADF Test and Akaike Information Criterion * 5 per cent level of significance ** 1 per cent level of significance *** 0.1 per cent level of significance Augmented Dickey-Fuller Test (no drift or trend) ARMA (p,q) t-statistic Critical Value 10 per cent 5 per cent 1 per cent Akaike Information Criterion (0,0) No Solution (1,0) (0,1) (1,1) (1,2) (2,1) (2,2) (2,0) (0,2) parameter Value standard error Observations 216 t statistic p-value AR(1) Intercept <0.0001*** A forecast for each of the 214 quarters from June 1960 to September 2013 was based on the mean of all of the ABS observations preceding each quarter. 28 Real GDP data was expressed as seasonally adjusted quarter-on-quarter per cent changes, and so, required no conversion to percentages (Australian Bureau of Statistics 2014f). All real GDP figures, as first announced by the ABS, were available from MMS for the entire observation period in the study consisting of 56 quarters. I used the real GDP estimates (explained above) as forecasts. The only constraint paring back the observations for this series was the release date for the December No forecast was made for December 1959 and March 1960, as at least two observations are required to produce an average. 10 August

101 14/06/ /12/ /06/ /12/ /06/ /12/ /06/ /12/ /06/ /12/ /06/ /12/ /06/ /12/ /06/ /12/ /06/ /12/ /06/ /12/ /06/ /12/ /06/ /12/ /06/ /12/ /06/ /12/2013 quarter, which was outside the observation period (after 2013); this resulted in the loss of one observation. The 55 observed surprises are plotted in Figure 8. Figure 8 Real GDP Surprises per cent Real GDP growth surprises appear to exhibit a significant upward bias, particularly after 2007, which is confirmed by summary statistics. The mean of 0.26 is upward biased and has a standard error of The median of 0.26 is similarly upward biased. This is likely a result of a number of quarters of unusually low actual growth deviating from the naïve model s predictions, particularly after Policy Cash Rate The policy cash rate in my study is the overnight cash rate. That is, the interest rate at which financial institutions borrow and lend in the overnight money market. The RBA 29 This assumes a normal distribution. 10 August

102 sets targets for this rate in its implementation of monetary policy, which flows through to other interest rates charged on funds in the Australian economy. I use the market s expectations as overnight cash rate forecasts. The expected overnight cash rate target was derived from the price of 30-day interbank cash rate futures contracts. The data used was a generic, and sourced from Bloomberg using the IB CMDTY ticker over the period August 2003 to December The latest price for the futures contract was used, relating to settlement in the month in which the cash rate announcement was being made. The latest price was the price observed the day immediately prior to the cash rate announcement, which virtually always took place on the first Tuesday of every month. 30 The formula shown in (13) was used to derive the expected value for the cash rate announced the next day. r t 1 x rt n n a b Where: rt 1 is the expected value for the cash rate announced the next day; x is 100 minus the contract price, for that month, prevailing on the close of the day prior to the announcement; r t is the rate prevailing prior to the announcement; nb is the proportion of days in the month before and including the day of the announcement; and n a is the proportion of days in the month after the announcement. 30 Except for January, where no announcement was made. For months where Tuesday was the first day of the month, Bloomberg data had to be manually retrieved to augment the generic 30-day series. This is because the generic series futures prices are always aligned with the month for which the data is being retrieved. 10 August

103 02/09/ /01/ /05/ /09/ /02/ /06/ /10/ /03/ /07/ /12/ /05/ /11/ /03/ /07/ /11/ /03/ /07/ /11/ /03/ /08/ /12/ /05/ /09/ /02/ /06/ /10/ /02/ /07/ /11/2013 Only values for February through to December were reported, as announcements were only scheduled only for these months. This resulted in 112 one-day-ahead expected values, which were used as forecasts. Policy cash rate data was expressed in percentage levels, which is consistent with the ASX 200 returns in the regression equation (in terms of being unitless), and no transformation was required (Reserve Bank of Australia 2013). The forecasts implied on the first Tuesday of every month (except for January), using the methodology outlined above, were paired accordingly with their date and the announced cash rate. The 114 resulting forecast surprises are shown in Figure 9. Figure 9 Interest Rate Surprises per cent The maximum cash rate surprise is 0.5 per cent. This occurred when the cash rate was dropped unexpectedly by this amount in October 2008 with the onset of the GFC. The forecasts do not show any significant evidence of bias with a mean and median of August

104 Consumer Sentiment The Westpac-Melbourne Institute consumer sentiment index reflects consumers evaluations of their household finances over the past and coming year, expectations of economic conditions over the coming years, and buying conditions for major household items. MMS forecasts of the index contained an insufficient number of observations to create an adequate number of forecast errors for analysis. Augmented Dickey-Fuller tests on the series showed the announcement series was non-stationary in levels, but stationary in first differences (see Table 11). This indicates that the consumer sentiment index can be modelled as an ARIMA process to produce forecasts. The Akaike Information Criteria on a number of specifications, modelled over the entire data set from October 1974 to June 2013, suggested that an ARIMA (1,1,2) without an intercept was the most parsimonious fit. An expanding window historical data set was used to estimate ARIMA (1,1,2) specifications and produce one-step ahead forecasts. The expanding window used all historical consumer sentiment index values dating back to October 1974, before the month in which the forecast was made. This data was used to estimate the ARIMA (2,1,1) specification, which in turn, was used to produce the one period ahead forecast. For the subsequent month, the process was repeated, expanding the historical data set by one more observation. This process was repeated for all months from January 2000 to June 2013 producing 168 forecasts. 10 August

105 Table 11 Consumer Sentiment ADF Tests and Akaike Information Criteria * 5 per cent level of significance ** 1 per cent level of significance *** 0.1 per cent level of significance Augmented Dickey-Fuller Test (no drift or trend) ARMA (p,q) t-statistic Critical Value 10 per cent 5 per cent 1 per cent Akaike Information Criterion (0,0) No Solution (1,0) (0,1) (1,1) (1,2) (2,1) (2,2) (2,0) (0,2) parameter Value standard error Observations 216 t statistic p-value AR(1) Intercept < The data was expressed as a seasonally adjusted index and required conversion to percentages using the equation (14). Where: E ( CSI ) CSI E CSI x t t 1 t t(% t 1) 100 CSIt E ( CSI ) are the modelled forecasts explained above; and t t 1 CSI t are the seasonally adjusted consumer sentiment index actual figures. Et(% CSIt 1) can be interpreted as the expected percentage change on the current level of the CSI. 10 August

106 07/04/ /09/ /02/ /07/ /12/ /05/ /10/ /03/ /08/ /01/ /06/ /11/ /04/ /09/ /02/ /07/ /12/ /05/ /10/ /03/ /08/ /01/ /06/ /11/2013 The actuals data was also converted to percentage changes as shown in (15). CSI CSI t 1 t % CSIt 1 x 100 CSIt Equation (14) was then deducted from equation (13) to create the times series of consumer sentiment index surprises. On account of release dates only being available from MMS from April 2004, the number of useful observations was reduced from 168 to 117. The sample was further reduced to 108 on account of nine missing release dates within the set of available dates. 31 That is, some months in the MMS data set did not report the day on which the announcement was made, and so, it is assumed no announcement occurred. The forecast errors/surprises are plotted in Figure 10. Figure 10 Consumer Sentiment Surprises per cent The missing release dates were for the months July, August, November and December 2004, April, May and November 2005, July 2009 and February August

107 All the data observations were plotted, except for the nine missing dates. Casual observation of the plot does not reveal any bias. Additionally, the forecast over- and underestimates do not appear to increase in magnitude with the onset of the GFC. The mean forecast error is -0.22, which is insignificant. 5.3 Control Variables Crude Oil A number of oil price benchmarks are available, with Brent and West Texas Intermediate (WTI) as the two most notable. Brent accounts for around two thirds of global physical trade in oil, despite only accounting for one per cent of crude oil production (Dunn & Holloway 2012, p.68). WTI has a strong US focus and recent developments in the oil and gas market resulted in a preference for Brent as an indicator of international oil prices. Thus, Brent was selected as the index for oil prices in the analysis. I found that some stock returns studies use spot prices for oil as an explanatory variable, while others use closest to maturity futures prices. Sardosky (2001), Boyer and Filion (2007), and Hasan and Ratti (2012) used one-month futures prices due to spot prices being more affected by temporary, random events that introduce noise into the analysis. I assume market participants place more weight on futures prices, which is consistent with the recent literature. Accordingly, one-month Brent future contract prices were sourced from Bloomberg. Missing values were replaced with the last known price. 32 The Australian Dollar contract price per barrel and returns are shown in Figure The Bloomberg ticker for the Brent futures index used is CO1 Cmdty. 10 August

108 Figure 11 Brent Crude Oil One-Month Futures Prices and Returns per cent 15 AUD price US Stock Market Index I included the lagged daily US S&P 500 Index returns to control for international effects on the Australian stock market. The closing index value was acquired from Datastream and converted to Australian returns using the closing US-Australian dollar exchange rate on each day. 33 Returns were calculated using equation (7) and plotted in Figure The S&P 500 index code used in data stream was S&PCOMP(RI) divided by AUUSDSP observations for each day. 10 August

109 Figure 12 Lagged US Standard and Poor s 500 Index Returns The returns exhibit volatility clustering in similar periods to those evident in the Australian return series in Figure 1 and Figure 2. This suggests some commonality may exist and inclusion of this variable is appropriate for this study. Day of Week and Holidays Day-of-the-week dummy variables and holiday effect dummy variables are commonly used as control variables in return studies (see Jaffe 1984, Singh 1993 and Kim 2003). Australian Stock Exchange holiday dates were sourced from Bloomberg. Dates that correspond to days with no price changes were deleted from my study, and the following day, assigned a dummy variable to control for opening price reactions to information accruing over the holiday period. All days of the week (except for Wednesday) were assigned their own series and coded with a dummy variable taking the value of one if it was that particular day-of- 10 August

110 the-week, or zero otherwise. This established Wednesday as the base case, against which all other days were compared. Term Spreads Flannery and Protopapadakis (2002), and Groenewold (2003) included term and default spreads in their analysis as control variables. The term spread is high on bonds (upward sloping yield curve) during economic downturns, when future conditions are expected to improve, also signalling high-expected returns (Fama 1991, p.1585). Harvey (1989, p.39) explained the role of the term spread in predicting economic growth. He reasoned that an investor s marginal value of a dollar is high during recessions (due to lower consumption) than in affluent times when consumption is high. Foreseeing this, rational investors sell short-term bonds and buy long-term bonds as insurance against an expected down turn. Holding all other factors constant, this raises the yield on short-term bonds (through reduced price) and depresses the yield on long-term bonds through long-term bonds prices increasing (Harvey 1989). Stock returns are linked to a firm s earnings, and thus, real economic growth (Gordon 1962; Fama 1981; Campbell & Shiller 1988; Schwert 1990). The link between term spreads and economic growth indirectly suggests term spreads are a gauge of expected returns from equities. Flannery and Protopapadakis (2002, p.760) used the Treasury term structure premium, measured as the difference in yield to maturity between ten-year Treasury bonds and three-month Treasury bills. Groenewold (2003, p.460) used the term spread between the rates on ten-year Government Bond and three-month Treasury notes. Three-month Commonwealth Government Treasury Bond data sourced from Bloomberg contained many missing observations between 2000 and 2013, so I used 10 August

111 one-year bonds to replicate the term spread used by Groenewold for the Australian market. The spreads between ten- and one-year Government bonds are shown in Figure The term spread becomes negative around the year 2000 Dot-Com bubble and in the years leading up to the 2008 Global Financial Crisis. This indicates the yields on shorter-term bonds were becoming large in comparison to longer-term bonds at the onset of economic turmoil, which is consistent with Harvey s theory outlined above. Shortly after each of these crises, the spread rapidly becomes positive, which is also consistent with the view that economic growth and, thus, future returns are expected to improve. 34 The Bloomberg tickers used for the one and ten year Government bonds are GACGB1 Index and GACGB10 Index respectively. 10 August

112 03/01/ /10/ /07/ /04/ /01/ /11/ /08/ /05/ /02/ /11/ /09/ /06/ /03/ /12/ /09/ /07/ /04/ /01/ /10/2013 Figure 13 Term Spread - Australian Commonwealth Government Bonds per cent Default Spreads Fama (1991, p.1585) outlined that during economic downturns default spreads are high on bonds, while very low stock prices result in relatively high dividend yields. This implies high-expected returns on bonds and stocks. That is, during persistent downturns, an investor requires higher compensation if risking depressed levels of wealth. Flannery and Protopapadakis (2002, p.760) calculated a default premium, measured as the difference in yield between Moody s BAA and AAA seasoned corporate bond indices. Groenewold (2003, p.460) measured the default spread between the rates for fiveyear Government bonds and five-year New South Wales Treasury Bonds. He noted that bonds with a greater quality difference would have been preferred, but was restricted to this pair due to data availability. 10 August

113 For my research, the longest continuous series of Australian corporate bond yields available at the time was the Bloomberg 5-year AA fair value curve index. While other Australian corporate bond series are available from Datastream and UBS, they do not hold a credit rating or term to maturity constant. This means the default spread calculated on these bonds is contaminated with term and credit rating variations that do not reflect the pricing of a given default category. Bloomberg had a variety of other Australian corporate bond indices, including a BBB band, which would have been preferable due to the greater premium on these bonds. However, the AA 5-year index was the only series available with continuous observations from January The Bloomberg 5-year Australian Commonwealth Government bond index series was deducted from the AA fair value curve series to derive a default spread. 35 Anomalous data points were replaced with the preceding day s value. The resulting series is shown in Figure The Bloomberg tickers used for the five year Government bond index and AA 5 year fair value curve are GACGB5 Index and C3585Y Index respectively. 10 August

114 03/01/ /06/ /11/ /04/ /10/ /03/ /08/ /02/ /07/ /12/ /05/ /11/ /04/ /09/ /03/ /08/ /01/ /07/ /12/ /05/ /10/ /04/ /09/ /02/ /08/ /01/ /06/ /11/ /05/ /10/ /03/ /09/2013 Figure 14 5-Year Australian Corporate Bond Default Spread per cent The default spread is relatively stable leading up to the GFC in 2008, fluctuating within a band of 0.5 to one per cent. The peaks in the series appear to roughly align with the clustering of increased volatility displayed in Figure 1 and Figure 2. This is suggestive of a shared relationship between stock market risk and bond returns. 10 August

115 6 Results The models outlined in Chapter 4 were estimated using ASX 200 daily returns from 26 October 2005 to 31 December This period was chosen because the data for all of the macroeconomic surprises was available between those dates. To reiterate Chapter 4, two variants of the model were estimated: one using values/absolute values of macroeconomic surprises (continuous model), and the other using dummy variables for macroeconomic announcement days separated into good and bad news days (dummy variable based model). The continuous models dependent variables were the forecast errors expressed as a percentage. The continuous model allowed me to capture the percentage change in stock returns and stock return volatility per one per cent of error in macroeconomic forecasts. This provided information on the sensitivity of stock returns and return volatility to macroeconomic surprises. The details of model fitting are outlined in Appendix B. 6.1 Continuous Model Results In the mean equation in Table 12, surprises can be negative or positive as they are all based on the equation (16): Surprise E ( Announcement ) Announcement k, t t 1 k, t k, t In the variance equation, absolute values of surprises are used, so both good and bad news is positive and, therefore, summarised into surprises more generally. A positive coefficient for the macroeconomic variables in the variance equation indicates surprises in general increase volatility, while a negative coefficient on the variables indicates surprises, in general, decrease volatility. 10 August

116 The coefficients represent the effect on returns, or return volatility expressed as whole number percentages. For example, a coefficient of 0.5 would represent a 50 basis point or a one half a per cent increase in return on a given day. Table 12 Continuous EGARCH model results based on full period sample * 5 per cent level of significance ** 1 per cent level of significance *** 0.1 per cent level of significance tests are two sided based on a null hypothesis of zero ASX 200 Daily Total Returns: 26 October December 2013 Mean Equation Fri TS CSI Rt M( ) a Hol a Day + a Control a Surprise t Hol t i, Day i, t j, Control i, t k, Surprise k, t i Mon j US k Unem Variable Coefficient p-value Intercept ** ahol Holiday *** aday Monday aday Tuesday aday Thursday aday Friday acontrol US Returns (Lagged) *** acontrol Oil Returns (Lagged) ** acontrol Term Spread acontrol Default Spread ** asurprise Unemployment asurprise Balance of Trade asurprise Retail Sales asurprise Producer Price Index asurprise Consumer Price Index ** asurprise Real Gross Domestic Product asurprise Overnight Cash Rate asurprise Consumer Sentiment Index August

117 Variance Equation Fri TS CSI 2 ln( ) V ( ) b Hol t Hol t bi, Day Day i, t bj, Control Control Surprise i, t bk, surprise k, t i Mon j US k Unem Variable Coefficient p-value Intercept ARCH (1) term *** Asymmetry term *** GARCH (1) term *** bhol Holiday bday Monday bday Tuesday ** bday Thursday bday Friday bcontrol US Returns (Lagged) *** bcontrol Oil Returns (Lagged) bcontrol Term Spread bcontrol Default Spread bsurprise Unemployment bsurprise Balance of Trade bsurprise Retail Sales bsurprise Producer Price Index ** bsurprise Consumer Price Index bsurprise Real Gross Domestic Product bsurprise Overnight Cash Rate bsurprise Consumer Sentiment Index ** Included observations 2070 Adjusted R-squared Log likelihood Akaike Information criterion Diagnostics Q(20) [p-value] [0.8020] Q 2 (20) [p-value] [0.9340] 10 August

118 ARCH LM Test F-Statistic [p-value] [0.9119] Engle-Ng Joint Sign Bias Test F-Statistic [p-value] [0.2997] In Table 12 (and the subsequent tables of results for the models that follow), the p- values are based on Bollerslev-Wooldridge robust standard errors. The p-values given are based on a two sided test and a null hypothesis of zero consistent with the null hypotheses outlined in Chapter 3. Two sided tests have been used because they allow simultaneous testing for negative and positive effects outlined in the alternative hypotheses in Chapter 3, yet are still more conservative than one sided tests as they require greater deviation from zero to detect significance. I mainly discuss variables significant at the five per cent level for the sake of brevity. To start with, I briefly discuss the effect of control variables. For returns (shown in the mean equation), the results indicate holidays, US returns and Brent crude oil futures returns have a positive relationship with Australian stock returns. The result for US returns is consistent with Kim and In (2002). The relationship between oil prices and Australian stock market returns are the opposite of that found by Hasan and Ratti (2012). The default spread shows a negative relationship with returns, which suggests stock price changes are related to bond price changes. As corporate bond prices fall, the yields that reflect the coupon payment, as a proportion of the price, increase which results in an increased default spread. The negative relationship with stock prices reported in Table 12, therefore, shows that bond and stock prices fall together. This is consistent with the findings of Fama (1991) who observed that increased default spreads are related to decreased stock prices during economic downturns. ASX 200 return volatility shows evidence of a negative relationship with Tuesdays and US returns. This is consistent with Jaffe (1984) who found that day of the week effects were unequal. While these results are interesting and reasonable, they 10 August

119 are not the main focus of my research and so I move on to discuss the effect of macroeconomic surprises. For returns, only the CPI reports a significant relationship. Good (bad) CPI surprises report a positive (negative) relationship with returns. That is, stock returns increase by basis points for every one per cent that the CPI is lower than expected. 36. This rejects my null hypothesis of no relationship between CPI surprises and stock returns. Instead, it appears to support the alternative proxy effect hypothesis (H5b) where inflation acts as a proxy for changes in expected future output (which is typically positively related to stock market returns) and stock prices. For volatility, the PPI relates to increased ASX 200 return volatility, reporting a basis points increase in returns for every one per cent surprise (or forecast error) of any sign. This rejects my null hypothesis of no effect and means greater PPI surprises are associated with greater stock return volatility. This is consistent with alternative hypothesis H4c and hence the findings of Kim (2003, p.625) in the US market. 37 The CPI, however, was not found to be significant. These results are in contrast to Tiwari (2012) who found that CPI changes precede PPI changes. If the CPI is a forecaster of the PPI, one would expect only information in the CPI to be of importance to the stock market and, therefore, the CPI to be significant instead of the PPI. This result is examined further for robustness in Section 6.5. Consumer sentiment index surprises (of any sign) significantly increase volatility, rejecting my null hypothesis of no relationship. For every one per cent that consumer sentiment differs from that which was expected, volatility increases by 3.56 basis points. This supports alternative 36 As per equation (1), lower actuals against expectations result in a positive macroeconomic surprise. 37 It should be noted that the test conducted here cannot isolate the effect of good from bad news on volatility. Alternative hypothesis H4c specifies that only bad PPI news increases return volatility. Despite this, the results presented here are not inconsistent with this alternative hypothesis. The model specification in section 6.2 allows for a more specific test of alternative hypothesis H4c. 10 August

120 hypothesis H8b based on De Long et al s (1990) hypothesis that the presence of irrational investors, trading based on sentiment, increases volatility in excess of that justified by fundamentals. This assumes consumer sentiment is a reasonable proxy for investor sentiment (Qiu and Welch 2006, Akhtar et al 2011). 6.2 Dummy Variable Based Model Results The dummy based model captures the average change in stock returns and stock return volatility in response to good or bad macroeconomic surprises. The definition of good and bad surprises is outlined in Chapter 3. The results for the full period are shown in Table August

121 Table 13 Dummy variable EGARCH model results based on full period sample * 5 per cent level of significance ** 1 per cent level of significance *** 0.1 per cent level of significance tests are two sided based on a null hypothesis of zero ASX 200 Daily Total Returns: 26 October December 2013 Mean Equation Fri Rt M( ) a Hol a Day + a Control CSI k Unem a t Hol t i, Day i, t j, Control i, t i Mon j US CSI Surprise Bad News D k, Surprise k, t a D k, Bad News k, t k Unem Variable Coefficient p-value AR (1) ** ahol Holiday *** aday Monday aday Tuesday aday Thursday aday Friday acontrol US Returns (Lagged) *** acontrol Oil Returns (Lagged) ** acontrol Term Spread acontrol Default Spread asurprise Unemployment asurprise Balance of Trade asurprise Retail Sales asurprise Producer Price Index asurprise Consumer Price Index asurprise Real Gross Domestic Product ** asurprise Overnight Cash Rate asurprise Consumer Sentiment Index Bad News Announcements abad News Unemployment abad News Balance of Trade TS 10 August

122 abad News Retail Sales abad News Producer Price Index abad News Consumer Price Index abad News Real Gross Domestic Product abad News Overnight Cash Rate abad News Consumer Sentiment Index ln( ) Variance Equation Fri TS 2 V ( ) b Hol t Hol t bi, Day Day i, t bj, Control Controli, t i Mon j US CSI k Unem b CSI Surprise Bad News k, SurpriseDk, t bk, Bad NewsDk, t k Unem Variable Coefficient p-value Intercept ARCH *** Asymmetry term *** GARCH *** bhol Holiday bday Monday bday Tuesday ** bday Thursday bday Friday ** bcontrol US Returns (Lagged) *** bcontrol Oil Returns (Lagged) bcontrol Term Spread bcontrol Default Spread b Unemployment Surprise bsurprise Balance of Trade bsurprise Retail Sales bsurprise Producer Price Index bsurprise Consumer Price Index bsurprise Real Gross Domestic Product *** 10 August

123 bsurprise Overnight Cash Rate bsurprise Consumer Sentiment Index Bad News Announcements bbad News Unemployment bbad News Balance of Trade bbad News Retail Sales bbad News Producer Price Index bbad News Consumer Price Index bbad News Real Gross Domestic Product *** bbad News Overnight Cash Rate bbad News Consumer Sentiment Index ** Included observations 2070 Adjusted R-squared Log likelihood Akaike Information criterion Diagnostics Q(20) [0.9170] Q 2 (20) [0.8900] ARCH LM Test F-Statistic [p-value] [0.9011] Engle-Ng Joint Sign Bias Test F-Statistic [p-value] [0.3736] The results for the mean equation indicate that holidays, US returns and Brent crude oil futures returns have a positive relationship with ASX 200 returns. This is consistent with the mean equation results for the continuous model in Table 12. With respect to the eight macroeconomic variables, returns respond positively to real GDP surprises, showing an asymmetric response that increases returns by basis points in response to good news, but exhibiting no significant response to bad news. This rejects my null hypothesis of no relationship between real GDP surprises and stock returns, indicating good real GDP news is related to increased returns. The response to good news is consistent with alternative hypothesis H6a and hence the 10 August

124 theories of Jorgenson (1971), Fama (1981) and Campbell and Shiller (1988) who offer various explanations for a positive relationship between stock returns and expected future output growth. The variance equation shows Tuesdays and US returns are related to lower volatility in ASX 200 returns, which is consistent with the results in Table 12 discussed above. The dummy based model picks up an additional day-of-the-week effect for Fridays, which is also related to lower return volatility. Good real GDP surprises appear to reduce volatility by basis points, with an asymmetric bad news response that increases volatility by 6.73 basis points ( basis points). This rejects my null hypothesis of no relationship between real GDP surprises and stock return volatility. It indicates good real GDP news is associated with decreased return volatility, while bad real GDP news is associated with increased return volatility and that the effect of bad real GDP news is slightly stronger than good news. This also suggests better than expected growth prospects have a calming influence on the stock market, while worse than expected growth has the opposite effect. The good news effect on volatility is consistent with alternative hypothesis H6c and hence Kim s (2003, p.624) observations in the US. Consumer sentiment index surprises have an asymmetric response to bad news, decreasing volatility by basis points and rejecting my null hypothesis of no relationship between the consumer sentiment index and stock market volatility. This asymmetric bad news effect is not intuitively convincing because one would typically associate bad consumer sentiment with deteriorating business conditions and increased uncertainty, both of which would be expected to result in increased volatility. The finding also contradicts Akhtar et al (2011) who found bad consumer sentiment news decreases returns. Under these circumstances, bad consumer 10 August

125 sentiment news would be more likely to increase (rather than decrease) the volatility of returns. This result is re-examined in Section Continuous Model Results: Pre- and Post-Global Financial Crisis To determine whether the effects differ before and after the onset of the Global Financial Crisis in 2008, both variants of the model were estimated before and after (and including) 10 October Lim, Durand and Yang (2014, p.171) observed the crises encountered over the period in my study climaxed during October In Australian markets, the largest fall in returns was on 10 October, dropping 8.70 per cent. Two, as opposed to three or more, sub-periods were chosen using this date as the break point to maximise the number of sub-period observations. The validity of the results under this structure is tested using alternate dates and three sub-periods outlined in Section 6.5. The results for the continuous regression pre- and post-gfc are shown in Table 14. Table 14 Continuous EGARCH model results: Pre- and Post-GFC * 5 per cent level of significance ** 1 per cent level of significance *** 0.1 per cent level of significance tests are two sided based on a null hypothesis of zero ASX 200 Daily Total Returns Mean Equation Fri TS CSI Rt M( ) a Hol a Day + a Control a Surprise t Hol t i, Day i, t j, Control i, t k, Surprise k, t i Mon j US k Unem Variable Coefficient (pre-gfc) p-value Coefficient (post-gfc) p-value AR(1) *** - - ahol Holiday *** ** aday Monday *** aday Tuesday aday Thursday *** aday Friday August

126 acontrol US Returns (Lagged) *** *** acontrol Oil Returns (Lagged) ** acontrol Term Spread acontrol Default Spread ** asurprise Unemployment asurprise Balance of Trade asurprise Retail Sales asurprise Producer Price Index asurprise Consumer Price Index *** asurprise Real Gross Domestic Product asurprise Overnight Cash Rate *** asurprise Consumer Sentiment Index Variance Equation Fri TS CSI 2 ln( ) V ( ) b Hol t Hol t bi, Day Day i, t bj, Control Control Surprise i, t bk, surprise k, t i Mon j US k Unem Variable Coefficient (pre-gfc) p-value Coefficient (post-gfc) p-value Intercept ARCH (1) term ** *** Asymmetry term *** *** GARCH (1) term *** *** bhol Holiday bday Monday bday Tuesday bday Thursday bday Friday bcontrol US Returns (Lagged) *** *** bcontrol Oil Returns (Lagged) bcontrol Term Spread ** bcontrol Default Spread ** bsurprise Unemployment August

127 bsurprise Balance of Trade bsurprise Retail Sales bsurprise Producer Price Index bsurprise Consumer Price Index bsurprise Real Gross Domestic Product bsurprise Overnight Cash Rate bsurprise Consumer Sentiment Index pre-gfc post-gfc Included observations Adjusted R-squared Log likelihood Akaike Information criterion Diagnostics pre-gfc post-gfc Q(20) [p-value] [0.8790] [0.9920] Q 2 (20) [p-value] [0.8700] [0.2800] ARCH LM Test F-Statistic [p-value] [0.8907] [0.3853] Engle-Ng Joint Sign Bias Test F-Statistic [p-value] [0.9719] [0.6740] Prior to the GFC, holidays, Mondays, Thursdays and US returns were positively related to returns in the mean equation. The default spread shows a negative relationship. These results are similar to those for the full period s continuous regression. Post-GFC, the effect of US returns and holidays on the ASX 200 returns remains significant and of a consistent sign. However, they both have a diminished effect. Day-of-the-week and default spread effects on ASX 200 returns become insignificant after the onset of the crisis, while the effect of oil futures returns becomes significant and positive. With respect to model fitting, the inclusion of an autoregressive lag no longer results in the most parsimonious fit. These results tend to suggest, after the 10 August

128 stock market euphoria leading up to 2008, there is an increased role for fundamentals. 38 The overnight cash rate is the only macroeconomic announcement to show a relationship with returns prior to the GFC, strongly increasing returns by per cent for every one per cent lower the cash rate turned out to be (compared to expectations). That is, good (bad) cash rate news increases (decreases) returns. This result rejects my null hypothesis of no relationship and supports alternative hypothesis H7b which is underpinned by Shiller and Beltratti s (1992) hypothesis that investors substitute between dividend paying and interest bearing instruments when the discount/interest rate changes. The effect becomes insignificant in the post-gfc period. This result is in contrast to results in the whole period s regressions in Table 12 and Table 13 above. Both tables did not detect any significant relationship between cash rates and ASX 200 returns. Reflecting back on Figure 9 in Section 5.2.3, the overnight cash rate fell significantly more than expected on 7 October 2008, which marginally falls within my definition of the pre-crisis period. This could explain the cash rate s significance exclusively in the pre-gfc period. The robustness of this result is tested in Section 6.5. The coefficient on the CPI surprises becomes significantly positive only after the GFC. That is, ASX 200 returns increase by per cent for every one per cent the CPI decreases (compared to expectations). Based on this, it appears good (bad) CPI news increases (decreases) returns, thus rejecting my hypothesis of no relationship. The results are consistent with the sign on the CPI in the whole period regression results in Table 12, but suggest the relationship, detected between CPI and ASX 200 returns over the whole period, stems from the period following the onset of the GFC. 38 The term fundamentals is used here in the same way that Harvey, Liu and Zhu (2014) classify macro factors. 10 August

129 This is perhaps due to increased importance placed on fundamentals thereafter. Fama s (1981) proxy effect hypothesis, where inflation becomes a proxy for changes in expected future output and stock prices (alternative hypothesis H5b in Section 3.5), is supported by this result. 39 The variance equation shows positive US returns reduce return volatility both before and after the GFC. Prior to the GFC, increased default spreads are related to increased stock market volatility, whereas there is no significant default spread effect on returns reported in the period following the crisis. The term spread coefficient shows no relationship with ASX 200 return volatility in the pre-gfc period, but reports a significant positive relationship thereafter. This appears to be counterintuitive because falling term spreads are associated with an increased risk of recession (Harvey 1989), and so, one should expect a falling term spread to be associated with increased stock market risk or volatility. Under these circumstances, a negative relationship between the term spread and stock market volatility should be observed - not the positive one reported. The tests in Section 6.5 assess whether this result is robust. The variance equation reports no significant relationship with any of the macroeconomic surprises. This is in contrast with the results, for the whole period regression in Table 12, that report a positive relationship between the PPI and consumer sentiment index news (of any sign), and return volatility. The finding that both of these variables are not significant in the sub-periods, pre- and post-gfc, is possibly a result of the smaller sample sizes within these periods (compared to the whole period). 39 As noted in the introduction, output is regularly found to be positively related to returns throughout in the literature. 10 August

130 6.4 Dummy Variable Based Model Results: Pre- and Post-Global Financial Crisis The results for the dummy variable variant of the model are shown in Table 15. Table 15 Dummy variable EGARCH model results: Pre/Post-GFC * 5 per cent level of significance ** 1 per cent level of significance *** 0.1 per cent level of significance tests are two sided based on a null hypothesis of zero ASX 200 Daily Total Returns Mean Equation Fri Rt M( ) a Hol a Day + a Control CSI k Unem a t Hol t i, Day i, t j, Control i, t i Mon j US CSI Surprise Bad News D k, Surprise k, t a D k, Bad News k, t k Unem TS Variable Coefficient (pre-gfc) p-value Coefficient (post-gfc) p-value AR (1) *** - - ahol Holiday ** ** aday Monday *** aday Tuesday aday Thursday *** aday Friday acontrol US Returns (Lagged) *** *** acontrol Oil Returns (Lagged) ** acontrol Term Spread acontrol Default Spread *** asurprise Unemployment ** asurprise Balance of Trade asurprise Retail Sales asurprise Producer Price Index asurprise Consumer Price Index asurprise Real Gross Domestic Product *** 10 August

131 asurprise Overnight Cash Rate asurprise Consumer Sentiment Index Bad News Announcements abad News Unemployment abad News Balance of Trade abad News Retail Sales abad News Producer Price Index abad News Consumer Price Index *** abad News Real Gross Domestic Product *** abad News Overnight Cash Rate ** abad News Consumer Sentiment Index Variable ln( ) Variance Equation Fri TS 2 V ( ) b Hol t Hol t bi, Day Day i, t bj, Control Controli, t i Mon j US CSI k Unem b CSI Surprise Bad News k, SurpriseDk, t bk, Bad NewsDk, t k Unem Coefficient (pre-gfc) p-value Coefficient (post-gfc) Intercept p-value ARCH *** *** Asymmetry term *** *** GARCH *** *** bhol Holiday bday Monday bday Tuesday ** bday Thursday *** bday Friday *** bcontrol US Returns (Lagged) ** *** bcontrol Oil Returns (Lagged) bcontrol Term Spread bcontrol Default Spread *** b Unemployment Surprise 10 August

132 bsurprise Balance of Trade bsurprise Retail Sales bsurprise Producer Price Index bsurprise Consumer Price Index bsurprise Real Gross Domestic Product *** ** bsurprise Overnight Cash Rate bsurprise Consumer Sentiment Index Bad News Announcements bbad News Unemployment bbad News Balance of Trade bbad News Retail Sales ** bbad News Producer Price Index bbad News Consumer Price Index bbad News Real Gross Domestic Product *** bbad News Overnight Cash Rate ** bbad News Consumer Sentiment Index ** pre-gfc post-gfc Included observations Adjusted R-squared Log likelihood Akaike Information criterion Diagnostics pre-gfc post-gfc Q(20) [p-value] [0.8370] [0.9950] Q 2 (20) [p-value] [0.8280] [0.3860] ARCH LM Test F-Statistic [p-value] [0.8514] [0.3937] Engle-Ng Joint Sign Bias Test F-Statistic [p-value] [0.9033] [0.5394] For control variables, the dummy variable specification of the model in Table 15 indicates the same relationship with ASX 200 returns as the continuous model in Table 14. Returns respond positively to holidays, and positively to US returns in both 10 August

133 the pre- and post-gfc sub-periods. Day-of-the-week effects on returns are positive for Monday and Thursday during the pre-gfc period, but thereafter, report no significant effects. The default spread reports a negative relationship with returns in the period preceding the crisis, but reports no relationship after its onset. Post-GFC, oil returns exhibit a positive relationship with stock returns, while no effects were found in the period prior to the GFC. The autoregressive lags no longer have a role in fitting a parsimonious model. As with the continuous model, it appears that after the GFC, fundamentals play a greater part in explaining stock returns. Turning to the eight macroeconomic variables, good unemployment surprises are positively related to ASX 200 returns, causing them to increase by basis points on average. That is, good unemployment news is associated with an increase in returns in the post-gfc period. The null hypothesis of no relationship is rejected. The result supports alternative hypothesis H1a which is explained by Boyd, Hu & Jagannathan (2005, p.650). They highlighted unemployment news may be a proxy of growth expectations. This is because unanticipated decreases in the unemployment rate may signal faster future output growth. Higher output growth typically equates to higher growth in corporate cash flows, and higher stock prices and returns. My results support this hypothesis, but only in the post-gfc period. Good real GDP surprises are also positively related to ASX 200 returns in the post crisis period, causing them to increase, on average, by per cent. During the same period, bad real GDP news has a negative effect on returns ( per cent on average), at the margin that more than offsets the good news effects. On average, this results in a negative; effect of basis points ( basis points) on returns. 10 August

134 These results indicate that post-gfc, good real GDP news is associated with increased returns, rejecting my null hypothesis of no relationship and supporting alternative hypothesis H6a; that stock returns are related to output growth expectations (Jorgenson 1971, Fama 1981). This is one of the most fundamental and regular empirical findings in the literature. One of the more interesting things about this result is that this fundamental relationship is found only post-gfc. The sign of the effects are still consistent with the dummy variable based full period s regression, which is an encouraging sign of robustness. In the post-crisis period, bad CPI news has an asymmetric negative relationship with returns, causing them to decrease by basis points on average, while good CPI news reports no significant effect. This finding rejects the null hypothesis of no relationship and, again, supports Fama s (1981) proxy effect hypothesis (outlined in alternative hypothesis H5b). Unexpected increases in inflation (bad CPI news) may be a sign of a deteriorating outlook for future real output and stock returns. Overnight cash rate news has an asymmetric effect, only evident prior to the crisis when the cash rate tended to be rising. Bad news is associated with a decrease in returns of basis points on average, rejecting the hypothesis that the overnight cash rate has no relationship with stock prices. These results are consistent with those in the continuous model in Table 14; however, this regression provides additional information that indicates the cash rate relationship in the continuous model is an asymmetric one. The results support alternative hypothesis H7b which reasons that investors substitute from dividend-paying to interest-bearing instruments, when the discount/interest rate increases (Shiller and Beltratti 1992), and indicate that investors are particularly sensitive to interest rate increases in the lead up to the GFC. 10 August

135 The volatility equation in the dummy based model shows that, Tuesdays, Thursdays and Fridays were associated with reduced volatility prior to the crisis. The effect is not persistent and disappears after the onset of the GFC. US returns are negatively related to volatility over the whole period. This result is remarkably consistent throughout out all of the modelling. It appears that stock market risk in Australia is strongly linked to the performance of the US economy, regardless of Australian economic conditions. The default spread is positively related to volatility but only prior to the GFC. The same result was found in the continuous model and suggests, leading up to the GFC, increased default risk in the debt market is an indicator of increased risk in the equity market. With respect to the eight macroeconomic variables, prior to the GFC, good real GDP news is associated with decreased volatility ( per cent on average), while the marginal effect of bad real GDP news ( per cent on average) more than offsets this. This means that bad news, overall, is associated with an increase in volatility of basis points ( basis points) on average. Only the relationships with good news persist after the onset of the GFC, and are associated with return volatility being reduced by basis points on average. Both results reject the null hypothesis of no effect and highlight that real GDP announcements have a significant asymmetric relationship across the whole period and are consistent with alternative hypothesis H6c. 40 The effect of bad news on return volatility, however, is limited to the period prior to the GFC. These results appear sensible, with the good news effect being supported by Kim s (2003, p.624) US findings. It can be reasoned that the bad news (lower than expected real GDP) may be seen as a sign of deteriorating business 40 Hypothesis H 6c only specifies an effect based on good news, however the additional bad news effect observed here is consistent with this hypothesis in terms of an inverse relationship existing between the sign on real GDP surprises and the size of returns volatility. That is, higher than expected (good) real GDP reduces return volatility and vice versa for lower than expected (bad) real GDP. 10 August

136 conditions and, therefore, increased uncertainty, risk and heightened volatility in financial markets. The opposite appears to be the case for good news, or higher than expected GDP. Prior to the crisis, bad retail sales have an asymmetric relationship with return volatility, which decreases on average by basis points on bad retail sales news days. This finding rejects the null hypothesis of no relationship with return volatility and is not consistent with any of the alternative hypotheses postulated. It supports the possibility that bad, or lower than expected, retail sales news is in fact good news for the economy and calms the market, reducing volatility. This finding is more closely examined in Section 6.5 where robustness tests are carried out. Cash rate news also reports an asymmetric relationship with return volatility prior to the GFC, with volatility decreasing on average by basis points on days where bad cash rate news is released. This rejects the hypothesis of no relationship between interest rates and return volatility. There is no intuitive reason why an unexpected increase in cash rates would dampen market volatility and the result is inconsistent with the alternative hypotheses postulated. As previously discussed, unexpected increases in interest rates are defined as bad news. The results are, therefore, showing an unexpected increase in the cash rate is associated with decreased volatility. This, if anything, is opposite to what one would expect. The result is tested for robustness in Section 6.5. In the period following the onset of the GFC, bad consumer sentiment news has an asymmetric relationship with return volatility, and it is also associated with a basis point decrease in the volatility of returns on average. This result is similar to that found in the whole period dummy variable based regression shown in Table 13. As discussed, and in relation to those results, the asymmetric bad news effect is not 10 August

137 intuitively convincing, and it also contradicts the findings of Akhtar et al (2011). Again, this result is examined more closely for robustness in Section Robustness Tests The fourteen results (including the split into pre- and post-gfc regressions) that are found to be significant are summarised in Table 16. Table 16 Summary of Results by Macroeconomic Variable Variable Continuous Model Dummy Variable based Model Returns Unemployment not significant post-gfc Retail Sales not significant not significant Producer Price Index not significant not significant Consumer Price Index full period and post- GFC post-gfc Real GDP not significant full period and post-gfc Cash Rate pre-gfc pre-gfc Consumer Sentiment Index not significant not significant Return Volatility Unemployment not significant not significant Retail Sales not significant pre-gfc Producer Price Index full period not significant Consumer Price Index not significant not significant Real GDP not significant full period, pre- and post-gfc Cash Rate not significant pre-gfc Consumer Sentiment Index full period full period and post-gfc To ensure the robustness of these results, three separate robustness tests have been carried out. Firstly, all of the regression models were re-estimated using the All Ordinaries index based total returns (instead of the ASX 200). Secondly, all significant macroeconomic surprises were cross-checked by re-estimating an ASX 200 returns based regression, on each macroeconomic surprise series, in isolation of the others. Lastly, given that a large number of the findings were significant, as a result of splitting the sample into pre- and post-october 2008 sub-periods, an alternative choice of sub-periods was used to check if the results were robust to the 10 August

138 choice of break point. The details and results of these robustness tests are outlined in Appendix C. The seven main results (including the split into pre- and post-gfc regressions) that survived the robustness tests are summarised in Table Table 17 Summary of Results Surviving Robustness Tests Variable Continuous Model Dummy Variable based Model Returns Unemployment not significant post-gfc Consumer Price Index post- GFC post-gfc Real GDP not significant post-gfc Cash Rate not significant pre-gfc Return Volatility Real GDP not significant full period, pre- and post-gfc 42 Consumer Sentiment Index full period not significant 6.6 Summary and Discussion of Results The overnight cash rate is special because it is the only variable that has a robust relationship with stock market returns prior to the GFC. The following theories and evidence offer an insight into why this may be. Flannery and James (1984) hypothesise the effect of nominal interest rate changes is related to a firm s maturity composition of nominal contracts. They found that interest rates were significantly related to the stock price of deposit taking institutions, and that the sensitivity of the relationship was related to the extent of the maturity mismatch between assets and liabilities. In the Australian context, this hypothesis is particularly relevant because 41 With respect to control variables, the positive term spread coefficient observed in the post-gfc period using the continuous model in Section 6.3 was not robust to using an alternative choice of sub-periods. 42 The effects of bad real GDP news on return volatility did not survive the test using an alternative choice sub-periods. 10 August

139 47.6 per cent of the S&P ASX 200 is comprised of financial institutions (as shown in Figure 15). Figure 15 ASX 200 Index - Sector Composition Note. From Standard and Poor s indices S&P/ASX 200 sector breakdown, July 2015 (S&P 2015) Faff and Howard (1999) studied the relationship between long-term interest rates and large Australian bank stock returns, and found a negative relationship during the period of rapidly rising stock prices between 1978 and During the period of relatively subdued stock market growth, between November 1987 and December 1992, no significant relationship was found. Over the period January 1992 to January 2007, Jain, Narayan and Thompson (2011, p.971) found short-term interest rates are negatively related to the largest four banks stock returns. 43 These studies taken with the results of my own study, suggest financial institutions, (specifically large banks) stock returns may be sensitive to interest rate changes, mainly during periods of rapidly rising stock market prices. They also suggest returns on these stocks are negatively related. This is possibly a result of exacerbated maturity 43 ANZ, Commonwealth Bank of Australian, National Australia Bank and Westpac Banking Corporation. 10 August