The following section presents results of the predictive regressions. First introducing the methodology in section 4.1, presenting the results from panel regressions in section 4.2, individual findings for countries in section 4.3, an attempt to determine the source of cross country variance in 4.4 and relating the findings for international markets to earlier literature in section 4.5. Finally 4.6 will provide a discussion.
With the average k-period return for country I as dependent variable and several predictors on the right-hand side, including sentiment (sent) and macro variables summed in controls (TS CPI IND SHORT). I estimate panel fixed-effects regressions so that all countries enter regressions jointly. Panel regressions are used to increase the power of the tests and to investigate if there is a significant sentiment-return relationship on average across countries.
Second, I estimate the regression separately for each of the 12 countries in the sample and test for a significant impact of sentiment on future returns across horizons. The tests for significance will be for joint significance of the form: . To test the jointly significant impact at the 1,6,12,24 months horizon. This test has been employed earlier in the literature (Mark, 1995; Ang and Bekeart, 2007 and Schmeling, 2009). This test is more reasonable device to test for predictability than just testing individual horizons, due to the correlated results across horizons. Because of the lack of observations individually when looking at the 24 months horizon in the "crisis period" some caution is needed in interpreting these results. The joint significance test will therefore also be tested with the following limitations.
Finally, I pool countries based on the scores of the dimensions on culture. I create a subset of two panels of countries containing the countries above the median and one below the median and this is done for each of the six dimensions. Using the different panels I run fixed-panel regression of the form (1). To test for significance I employ Wald test of joint significance of the form . Testing the different panels of countries allows for tests that would shed light on the relation between cultural dimensions and sentiment. For example I can test if higher uncertainty avoidance leads to lower impacts of sentiment on returns.
This seemingly simple regressions are however plagued by several econometric problems, so careful implementation of this method requires some corrections. The process of using average returns for the forecast horizon and than running regressions with these overlapping observations leads to strong correlation in the residuals. A second problems lies in the inclusion of persistent independent or predictor variable since this can bias the coefficient estimates as they are predetermined but not strictly exogenous (Stambaugh 1999). More on the biased estimates of slope coefficient caused by persistent repressors is found by Valkanov (2003) or Ferson et al. (2003). Several authors rely on a sort of simulation procedure to account for these problems (Brown and Cliff, 2005, Schmeling, 2009). One could also quantify and adjust for biases using auxiliary regression as done by Amihud and Hurvich (2004). Cambell and Yogo (2006) provide a method for efficient tests of stock return predictability in the presence of near unit-root regressors. Their method is however difficult to extend to multiple regressors and multi-period forecasts. Yet others have switched to using Hodrick (1992) standard errors (Ang and Bekeart, 2007; Lioui and Rangvid, 2007) and do not bias adjust coefficient estimates. Hansen and Hodrick (1980) provide standard errors correcting the overlapping and autocorrelation problems. However this correction does not work well in finite samples as found by Richardson and Stock (1989); Hodrick (1992); or Boudoukh and Richardson (1994). In attempting to prevent this paper to become overly technical I chose to correct the bias using Hodrick (1992) standard errors rather than using a form of simulation. I do however acknowledge the shortcomings of this correction but as existing research uses this method as well and I am confident that this solves the problems in a reliable way.
4.2. Results for panel regressions
To start with the results for fixed-effects panel regressions (tables 9-12). The tables show separately the sentiment coefficients, the significance, the economic magnitude and R-squares. I show these results separately so that all style and size portfolios can be shown in the same table. In this way differences between the types of stocks become clear. First I will discuss results for the overall sentiment-return relation or hypothesis 1, second I will report the size and value effects or hypothesis 2.
The results are universally negative on average as theory predicts. The coefficients reported are two times the standard deviation so the impact of sentiment is a rather large shock. Most coefficients become more negative over time. This pattern is consistent with limits of arbitrage hindering the ability of investors on profiting on mispricing that might persist for a significant amount of time (Brown and Cliff, 2005). The inclining returns can also be due to long-horizon regressions (cf. Hong et al.). The increase can result from a bias that mechanically generates significant results over longer horizons. This bias however is corrected by using Hodrick (1992) standard errors. The fact that for some style and size stocks the returns are declining is comforting, if there still is a bias this should not be the case. At the contrary Schmeling (2009) finds diminishing marginal impact which suggests that noise trading effects wash out over time, limits to arbitrage exist in the short run but become weaker in the long run. Overall the results are negative and significant thus supporting hypothesis 1 and finding results in line with earlier evidence. In the "crisis period" some positive coefficients are found, these results further confirm the first hypothesis. Optimism (pessimism) should lead to lower (higher) future returns leading to a negative (positive coefficient).
Based on table 9 or the regression coefficient I find evidence for the value premium. Growth stocks are more heavily influence than value stocks at least for small stocks. For mid and large cap stocks this relation is not present for those sizes value stocks are more influenced than growth stocks. Based on table 11 or significance no differences can be found for the small and midcap stocks. But when looking at large stocks, value stocks are more significant than growth stocks, suggesting that value stocks are more influenced. Summarizing that the effects for value stocks are more pronounced than for growth stocks for mid and large cap stocks and for small cap stocks growth stocks are affected more heavily. Hypothesis 2 finds some support in general because there are differences between growth and value stocks but the evidence is mixed. The size effect is supported by these results as there are visible differences between small and large stocks.
There is a pattern when looking at size. The small or mid cap stocks display more negative returns compared to large stocks. The coefficients find a clear relationship between size and the sentiment effect, with the effects being more pronounced for small stocks. The significance tells the same story, with the results being the most significant for small stocks. These results are in line with the theory and with previous results. Together with the results for the growth and value differences hypothesis 2 is supported in finding significant changes between sizes.
It is interesting to note that there are noticeable differences found when looking at the different time periods. Before the crisis the negative relation is found as predicted since optimism should lead to lower future returns. However during the crisis positive coefficients are found on the 1 and 6 month horizons. When looking at sentiment during the crisis we can classify the sentiment as pessimistic or at least less optimistic. Pessimism should lead to undervaluation and higher future returns so for the short horizons the relationship is caused by pessimism coming from the crisis. At longer horizons the impact turns negative again but lower than at the "normal period" pointing to the return of some optimism about the future after the period of pessimism. The results thus confirm the theory; optimism should lead to overpricing and lower future returns and pessimism to undervaluation and higher future returns. And one can conclude that the crisis definitely dampens optimism.
Results are significant in economical terms when looking at the magnitude of the effect. The magnitude is calculated by taking the coefficient and multiplying the coefficient by the horizon and by the standard deviation of sentiment. Values in table 11 indicate the effect of one standard deviation increase in sentiment on the return over indicated horizon (in percent). For instance, a one standard deviation increase in sentiment of large value stocks is associated with a 3.08 percent decrease over a 12 months period. The average simple return for this portfolio is 12.7% concluding that there is an economically significant reduction. And not as severe, so the result seems plausible.
Table 12 shows the r-squares of the regressions. They tend to be low which is usual for regressions that forecast stock returns. Looking at the change in r-squares sentiment seems to have quite some predicting power especially on the median and long horizons. This change resembles the added predictability when sentiment is added as a predicting variable.
One has to be careful in interpreting the results for the crisis period. Due to a limited number of observations the results must be carefully interpreted and cautiously used to draw conclusions. This is especially the case with the 24 month horizon. In individual countries results some special care will be given to this problem.
4.3. Results for individual countries
This section presents the results for individual countries, table 13 present the different time periods (panels A,B,C), the different styles and size are reported in the same table. The coefficients are mean coefficients across all horizons and the p-value is that of a joint significance Wald test. The tests for significance will be of the form: .
There is quite some heterogeneity across countries, a significant relation between sentiment and returns is found with 8 countries for the aggregate market during the "normal" period in our sample, 2/3 of our sample. In the crisis 6 countries find a significant relation and for the whole period 4 countries find a significant relation. Sentiment seems to be country specific and hypotheses are not supported for all countries. Evidence does not seem to be obviously related to geographical locations, or the size of a country.
When comparing periods the crisis seems to distort the results. When looking at both the "normal" and "crisis" periods the results are significant for more countries than for the whole period. In most cases large differences between the crisis and the normal period can be detected. These differences probably cause the lack of significance when looking at the whole period. In the crisis period the 1 month horizon causes positive coefficients and the mean coefficients are much lower because of that. Overall the coefficients are negative and significant. Sentiment effects seem to be country-specific and hypothesis 1 is not supported for all countries in the sample. When looking at the "normal" period on aggregate France, Germany, Spain and Italy are affected whereas Australia, Finland and Japan fail to be significant. The U.K. and U.S. cannot be seen as countries particularly prone to sentiment effects.
The size premium does hold well in this sample of individual countries, on average the small stocks are influenced the most; if not usually the mid cap stocks are influenced more than the large stocks. With influence I mean that the coefficients are more negative for small or mid cap stocks. The significance of the regressions provides further evidence for the size relation. In the "normal" period midcap stocks find a significant relationship in 8 countries for growth stocks and 9 for value stocks for large stocks a significant relationship is found for 7 countries for growth stock and at 6 countries for value stocks. Only for Denmark and the U.S. the coefficient for large stocks is more negative than for small stocks while remaining significant.
The evidence for a possible value premium complements the findings of the panel regressions. For individual countries the results are also mixed. For small stocks the effects is more pronounced for growth stocks and for large the value stocks are most affected, finding similar results for individual countries. For Canada, Denmark and Italy growth stocks are more influenced than value stocks. The other countries corroborate the evidence found from the fixed panel regressions, finding the effect to be more pronounced with small growth stock while the being more pronounced for mid value and large value stocks. Both hypothesis 1 and 2 can thus not be accepted for all countries, a comforting results since I can now focus on the determination of these rather large cross-country differences.
The 24 month horizon for the crisis period is somewhat limited due to the number of observation per country so the statistical power and significance of that test is lower than for other horizons. Due to that reason table 14 includes joint significance tests excluding this horizon. This table finds no real differences in the results found earlier but merely reinforces the relations stated before. The results become more reasonable with more relations becoming insignificant so the differences between the normal and the crisis period are smaller.
4.4. Cross-Sectional Analysis
Since it is not clear what drives the differences between countries, not obviously location or size I explore the possibility that the culture of a country drives the magnitude of the effect. It is already documented that individualism and uncertainty avoidance are correlated. Overconfidence and collectivism are said to partially cause the sentiment effect. I pool the countries based on dimensions by Hofstede (2011). Hofstede uses these dimensions basically to index the culture in a country. Using different pooled subsets I create two different panels for each dimension, one above the median and one below. Comparing these two panels gives an indication of the effect that this dimension has on the sentiment-return relation.
The results shown in tables 15 and 16 and clear some questions but cannot answer everything. Since I only pool countries based on the dimensions one it is not clear which dimension is the most important. And other variables not included might drive the effect I assume to come from culture. However most dimensions see a pattern where only one of the subgroups is significant and the significant coefficients are larger than the insignificant. So there seems to be a difference between cultures. The tables 15 and 16 include the dimensions and the periods as well as the horizons. Again for comparability the significance and the coefficient are shown in separate tables. Due to the overall significance at the 24-month horizon the effect may be caused simply by the horizon rather than by the culture. Therefore to explore the culture effects table 17 shows joint significance over the 1,6,12 month horizons.
A quick look at the tables shows large differences. The results for the above and for the below median show large differences at the one month horizon the effects are of opposite sign for some dimensions. When looking at significance the average pattern is that only one of the two panels is significant. Overall the effects are in line with the theory of cultural effects.
Power distance turns out to be somewhat ambiguous as the both panels are significant during the crisis period and during normal periods the above median panel also comes close to being significant. The difference during the crisis can be explained by the corruption perception, a market integrity factor, during a crisis both countries that score high and low on power distance feel that the government has made mistakes thus raising the perception of corruption. And less market integrity should lead to a more pronounced sentiment effect. Countries that score high on individualism are not affected, results are not significant. For collectivistic countries the effect is significant and negative coefficients are found. Differences between periods, "normal or crisis" are small. The more masculine countries are affected whereas results for feminine countries fail to be significant. The coefficients found are negative and become more negative over a longer period of time. Again there are little period differences. Countries that score high on uncertainty avoidance are the only countries that are affected by sentiment finding only significant results for these countries. This result is also found for countries scoring high on long-term orientation and for countries that are more restraint.
Based on these simple tests and one can conclude that culture indeed has an effect on sentiment, the effect seems largely driven by it. Most culture dimensions can be explained as overconfidence of collectivism, factors which determine the effect of sentiment according to psychological explanations. Also over-under-reaction, representativeness and conservatism as well as self-attribution are factors used to explain sentiment. Overall the cultural dimensions can be related to one of these values. So a countries culture, explains a countries behavior on aggregate and this behavior leads to sentiment affects.
4.5. Relation of results to earlier studies.
This section relates this papers his findings to earlier documented effects. Due to the limited literature focusing on an international view some caution is advised. Since the sentiment-return relationship is very country-specific. Existing literature focuses on U.S. stock markets so although this can give an indication for the U.S. and for other countries the results do not need to be exactly the same. Schmeling (2009) also has an international focus. Baker, Wurgler and Yuan (2009) find a global and local effect coming from sentiment. The next thing that should be noted is that other evidence results from periods before the crisis, while Baker, Wurgler and Yaun (2009) comment on the crisis their data is not sufficient to make conclusions involving the crisis. When looking at the cultural factors some dimensions have been used to explore the sentiment relation, some have been used to find relations between other economic phenomena and yet others have not been used at all.
Brown and Cliff (2005) provide a natural benchmark, the also provide results for the aggregate stock market, size, growth and value stock separately. They find a two standard deviation movement in their sentiment measure leading to a decline of stock returns of about 1.76% for a 6 month horizon 5.8% decline for a 12 month horizon. These numbers are similar 2.03% and 6.45% declines and thus seem reasonable and not implausible of size. To compare with a more international result Schmeling (2009) finds the declines to be 1.86% and 5.4%. Differences are a likely result of the countries used in this study. Most literature finds a negative relationship when optimistic and positive when pessimistic, consistent with my findings. Due to the length of the time series both with Schmeling (2009) and with Brown and Cliff (2005) it is probable that these time spans would have incurred crisis's as well data runs for example from before the 1987 stock market crash. For that reason when comparing the results there is no distinction made between "normal" and "crisis".
For the effects of value and growth the literature does not provide a consistent picture. Brown and Cliff (2005) find a stronger effect for growth than for value stocks. Baker and Wurgler (2004) show that sentiment effects are similar size for both value and growth stocks. Lemmon and Portnaguina (2006) provide evidence that sentiment effects are significant for value but not for growth stocks. And Kumar and Lee (2006) show that noise traders tend to overweight value stocks relative to growth stocks. Schmeling (2009) finds value stocks to be more affected although both groups are significant. I find evidence consistent with Brown and Cliff (2005) in finding growth stocks to be more effected that value stocks for the small stocks. For mid and large cap stocks results are in line with Baker and Wurgler (2006) finding the relation to be more pronounced with value stocks.
Many earlier papers have also looked at sentiment and the return of small and large stocks (Brown and Cliff, 2005); Lemmon and Portnaguinia, 2006 and Schmeling (2009). My results show that earlier U.S. evidence extends to international markets. Consistent with Schmeling (2009) based on international data, Brown and Cliff (2005) and Lemmon and Portnaguina (2006) based on U.S. data I find small stocks are more affected than large stocks. However my results differ in that there is a significant relation for large stocks which is not found by Schmeling (2009).
Some evidence is in line for individual countries as well, mainly the U.S. data which finds a significant effect. But also for Australia, Jackson (2004) finds no significant effect for noise trader induced returns, my results also are insignificant. And Schmeling (2006) finds evidence for a significant impact in Germany. Extending the comparison to Schmeling (2009) he includes some of the same individual countries as I do. He finds strong relationships with Italy, Germany and Japan, where I find them for Germany and Italy but not for Japan. The effects of the U.K., U.S, Australia and New Zealand are consistent. Finland and Denmark show some differences mainly when looking at size and style. The differences are likely to come from the differences in portfolios were Schmeling uses only value and growth stocks, I use portfolios based on style and size differences. A second difference may come from the use of MSCI indexes were Schmeling uses returns found by Kenneth French.
In looking at culture as an explanation of differences the results are consistent for at least two dimensions. UAI and IND have been employed before. Schmeling (2009) finds the same results. Although not looking for the sentiment return relation but if sentiment affects stock crisis Zouaoui, Nouyrigat and Beer (2011) find that UAI and IND affect the sentiment relation in the same way is my results find. They also give some indication if the sentiment effect is distorted by the crisis. They find investor sentiment to have an effect on the occurrence of a stock market crisis. Chui, Titman and Wei find the effects to be more pronounced for countries prone to herd-like behavior and overreaction. Market integrity is found to be significant by La porta et al. (1998), Chui, Titman and Wei (2008) and by Schmeling (2009). Although this paper does not use a direct measure of market integrity some of the cultural dimensions (PDI, IVR) can be related to weaker or stronger market integrity, the results corroborate earlier finding that in countries with less market integrity the sentiment effects are more pronounced.
Relating the other dimensions is harder since to my knowledge these dimensions have not been used to research the sentiment-return relationship up to this point. They are however found to be significant in earlier literature exploring different economic phenomena. Chui and Kwok (2008, 2009) find the inclusion of cultural factors to increase the predictive ability of the regression model on life insurance. de Jong and Semenov (2002) find lower levels of uncertainty avoidance and higher levels of masculinity leading to more developed stock markets. Park and Lemaire (2011) find Long-term orientation to be significant in a way that countries that are long-term orientated will want more insurance. They also find this to be true for countries that exhibit low Power Distance and high Individualism and Uncertainty Avoidance scores. It is hard to directly relate these earlier results but the overall picture is the same. The inclusion of cultural dimension increases the predictability and Hofstede's dimensions are significant, the scores can well be used to explain cross-country differences. In earlier research significant effects are found if a country scores high or low on a dimension but not for both.
On the negative relationship on average is very consistent with earlier evidence as goes for the positive relationship when sentiment turns out to be pessimistic. The panel regressions are consistent with Baker and Wurgler (2006) finding significant effects for both groups. For small stocks consistent with Brown and Cliff (2005) and for large and mid cap stocks is resembles the results found by Lemmon and Portnaguinia (2006) and Schmeling (2009). Overall the most resembles is found in comparing results with Schmeling (2009). Due to the mixed consistency with the U.S. found evidence and the resemblances with the internationally orientated research, my results do not exclusively support earlier findings in the U.S. literature. The size premium is very consistent and only differs in finding some significance for large stocks as well. The idea that countries more prone to herd-like behavior and overreaction are also consistent with both my results and with earlier results. Findings on different dimensions are consistent in finding similar relations found compared to literature using the dimensions but looking at different financial and economic phenomena. Overall the fact that there is a significant cultural effect is consistent with existing literature. The differences are presumably based on the use of different proxies, different stock markets, the use of different macro risk factors as controls and of course different countries.
4.6. Discussion
This paper gives a first indication that culture influences financial phenomena. Based on Hofstede his dimensions one can conclude that for the sentiment-returns relation culture is significant in determining the strength of the effect of noise trading on returns. Since it is known that Hofstede his dimensions have several limitations or disadvantages further research is needed in this field focusing on other dimensions and a more complex measure of culture. For a first exploratory research to the possibility of culture affecting the sentiment-return relation hofstede his dimensions prove to be a usefull metric of culture.
Furthermore this research uses consumer confidence as a proxy for investor sentiment. To check for robustness other proxies for investor sentiment could be tested on the effects culture as well. For replicability MSCI indices are used, again as a test of robustness of the results different indices can be used to see if this produces on general similar results.
The main focus was to explore cultural effects and for that reason some macro risk factors may still be included, a more thorough research might include dividend yield, the default spread, unemployment rate or consumption rate. I therefore acknowledge that it might be missing some important macro risk factor but I feel my set of control variables is a reasonable effort to correct for this problem.
Finally the used method to correct for econometrical problems might be less technical; it may also fail to correct the bias fully. Further research might thus opt to use a more technical form to correct this problems and rely on the use of estimation methods to bias correct the coefficients and standard errors.
Short of these limitations the main contribution still stands, finding culture to be a cross-sectional determinant of the sentiment-return relation. To my knowledge no existing literature covers the 6 dimensions of Hofstede to define culture when testing the sentiment-return relation. This paper thus looks upon the relation both internationally and culturally, a view that is uncommon in literature and a view that thus needs extra attention. I therefore suggest more research is done to explain financial phenomena via the cultural way.
5. Conclusion
I investigate the relation between investor sentiment and future stock return for 12 developed countries and find that consumer confidence as a proxy for investor sentiment is a significant predictor of expected returns on average across countries. The largest shift in magnitude is when going from 1 month to 6 months. Overall the effects incline when looking at mid and large cap stocks and decline over time when looking at small cap stocks. The effect is more pronounced for small growth stocks and for mid and large value stocks. However the predictive power of sentiment varies across countries and sentiment does not contain predictive power for several countries in the sample.
In order to investigate the cross country differences I look at possible determinates of the strength of the sentiment-return relation and find that the influence of noise traders on markets varies cross-sectionally in a way that is economically intuitive. The effect is more pronounced for countries that have less efficient regulatory institutions or less market integrity, countries that are more prone to herd-like behavior or countries that score on cultural dimensions related to the values overreaction, self-attribution, conservatism, representativeness and collectivism. The effects are more pronounced for countries with high scores on UAI, MAS, ITWOS, low on IND and IVR and for PDI the effect is ambiguous.
One cannot transfer evidence from the U.S. to other markets and presume that irrational noise traders move stock markets in general. Moreover cultural factors as well as institutional quality are strong determinants of the sentiment-return relation. High quality markets institutions seem to alleviate effects of noise trading. Culture is however not easily changeable so that sentiment effects should remain a persistent phenomenon in countries. The sentiment-returns relation turns out to be very country specific and to be partially determined by cultural differences.
7. Appendices
7.1. Appendix I. Different Proxies for investor sentiment
Existing studies use different measures for (unobserved) sentiment ranging from direct measures, to indirect measures and to self-constructed indices. For example closed-end fund discounts (CEFD) are used by Lee, Shleifer and Thaler (1991), by Swaminathan (1996) or by Neal and Wheatley (1998). While Lee, Shleifer and Thaler (1991) find contemporaneously correlation of closed end fund discount with small stock returns, Chen, Kan and Miller (1993) find evidence rejecting this. Swaminathan (1996) finds CEFD forecasts the size premium and that information in discounts is related to expectations of future earnings growth and inflation. Neal and Wheatley (1998) also find evidence to support the size premium based on research using CEFD as a proxy for sentiment. Baker and Wurgler (2006) show that the returns on equity are prone to speculation and are difficult to arbitrage making their prices sensitive for changes in sentiment with differences in style and size (small, growth) investment characteristics. They find low periods of sentiment are followed by high returns on small, young, unprofitable and dividend-nonpaying stocks using a constructed proxy based on CEFD and several other market-based variables including the number of IPO's, turnover, ect. These studies use indirect that are made up of time series of macroeconomic and financial variables and therefore might not exclusively represent investors' sentiment. The use of direct measures or surveys is evenly wide spread. Charoenroek (2006) finds that changes in consumer confidence help to forecast aggregate market returns in the United States. Lemmon and Portniaguina (2006) use consumer confidence as a proxy and find evidence supporting the size premium but not supportive of any variance in value or momentum premiums. Brown and Cliff (2004) use a survey of the American Association of Individual Investors and do not find any evidence that fund discounts reflect investor sentiment. Menkoff and Rebitzky (2008) also use survey data. Qui and Welch (2006) report only weak correlation between CEFD and consumer confidence and that only the consumer confidence measures are correlated with a measure of investor sentiment derived from UBS/Gallup. Making the consumer confidence measure as a proxy for sentiment the better choice. They also found consumer confidence to yield a robust contemporaneous correlation with the size premium. Fisher and Statman (2003) report positive correlations between measures of consumer confidence as a direct measure of investor sentiment compiled by the American Association of Individual Investors. Doms and Morin (2004) find, after controlling for macro economic factors, that consumer confidence is responding to the tone and volume of economic news rather than economic content. This implies a presence of an irrational factor in consumer confidence. Qui Welch (2006); Lemmon and Portniaguina (2006) and Ho and Hung (2009) present several additional argument to support the proxy for sentiment. Participation of individual households in financial markets has increased over the recent years, suggesting that consumer confidence measures may be useful in measuring how individual investors think and feel about the economy and financial markets. For the U.S. the consumer confidence index lines up with anecdotal evidence of changes in sentiment. The consumer confidence index turns out to be very highly correlated with changes in stock prices although consumers participating in the survey are not asked directly for their views on security prices. Shleifer (2000) states that because the consumer confidence index captures individual beliefs , it reflects the philosophy of behavioral finance by including opinions of imperfect people who have social, cognitive and emotional biases. An critique might be that consumer confidence does not proxy for investor sentiment but only captures some relevant macro information about time-varying risk premia that is not controlled for by other macro risk factors included in the regressions (industrial production, Term spread, Inflation and Short interest rate). Earlier evidence has already pointed out consumer confidence contains an irrational element and that it is well suited as a device for tracking trader noise sentiment (Doms and Morin, 2004). Schmeling (2009) finds that sentiment remains statistically and economically significant as a predictor while expected business conditions show no significant forecasting power. Therefore it seems unlikely that consumer confidence is just a simple business cycle proxy which is not driven to insignificance by other control variables.
7.2. Appendix II. Details on consumer confidence surveys
In the data discussion chapter it is said that the consumer surveys are selected by way of similarities. This appendix provides details on how the consumer surveys are carried out in different countries. The objective of this appendix is to highlight similarities and discus possible differences across countries. Internationally there is an standardized set of questions for surveying consumer confidence. This standard comes close to the survey of the university of Michigan. And all surveys of developed countries make use of this standardized countries to ensure international comparability. The questions asked regard both the past and future financial situation of the household, the past and future economic situation more generally and about purchases of durable goods. These questions are the most important in every survey, although most surveys do ask additional questions these are often similar across the countries used in this research. For example the survey of the European commission asks questions regarding the expected employment situation and the CPI developments and the Australian survey asks for long-term expectations. But given that the core questions are extremely similar across the countries included, one might expect that consumer confidence indices are comparable internationally. One difference in the consumer confidence indices is with respect to the seasonal adjustments. All countries do seasonally adjust but some differences appear in the way they adjust for the seasons. For example the European commission uses "Dainties" while Japan adjusts based on "X-11". There is no evidence in our results that using different procedures may affect econometric estimates quantitatively. The second difference is the forecast horizon. While most surveys ask questions about future developments the horizons differ. Most surveys (European) ask for one year horizons, Australia includes some additional questions over 5 year horizons and the Michigan surveys asks for one year horizons focusing on financial household situations but 5 year horizon regarding economic situations. Thus forecast horizons differ somewhat between countries. Finally, the numbers of participants differ by countries. Most surveys are based on more than 1000 households but there are large differences between large countries and small countries for example Denmark has 1500 participants and France 3300. Outliers are the U.S. with only 500 participants and Japan having more than 5000 participants.
7.3. Appendix III. MSCI indices construction methodology
The MSCI indices are based on Global Investable Market Indices ("GIMI") methodology. This methodology has several advantages since is aims to provide coverage of the relevant investable opportunity set with non-overlapping size and style segmentation. Second, there is a strong emphasis on investability and replicability of the indices through the use of size and liquidity screens. The size segmentation aims to balance the objectives of global size integrity and country diversification. There is also a balance between index stability and timely reflection of changes in the opportunity set. Finally a complete and consistent index family with Standard, Large Cap, Mip Cap, Small Cap and Investable Market Indices provides a range of styles and sizes. Additionally the method uses minimum free float requirements for eligibility and free float-adjusted capitalization weighting to appropriately reflect the size of each investment opportunity and facilitate the replicability of the indices as well as timely and consistent treatment of corporate events and synchronized rebalancings globally. This leads to international consistent indices that can be compared. In certain cases, where there are no qualifying securities, it is possible for MSCI Indices to be empty following a security deletion or GICS change. If an index becomes empty it would be dynamically discontinued or 'ruptured'. This is the case for the Large Growth index in New Zealand. It is then possible for the index to be re-started once a new security qualifies for the index, and this index level would be rebased to an appropriate level at that time. MSCI global value and growth indices categorize value and growth securities using clear and consistent sets of attributes and a rigorous methodological frame work. Style characteristics are defined using 8 historical and forward looking variables. Each security in an underlying MSCI index is given an overall style characteristic derived from its value and growth scores and placed in one of the two categories. The adjusted market capitalization of each constituent of the underlying index is fully represented in the combination of the value and growth index with no double counting. The value attribute is defined using Book value to price ratio, 12 month forward earnings to price ratio and dividend yield. For the growth classification long-term forward looking earnings per share (EPS) growth rate, short-term forward EPS growth rate, current internal growth rate, long-term historical EPS growth trend and long-term historical sales per share growth trend are used. The large cap indices cover all investable large cap securities with a market capitalization of the above average targeting 70% of each market's free-float adjusted market capitalization. Mid cap indices cover the mid 15% and small cap indices cover all investable small cap securities with a market capitalization below that of the companies in the MSCI standard indices, targeting approximately 14% of each market's free float adjusted market capitalization. While MSCI index covers all investable securities.
7.4. Appendix IV. Discussion of Hofstede's his dimensions
This appendix will discuss the main critiques that the work has suffered during its existence. Hofstede's work on culture is the most widely cited in existence (Bond 2002; Hofstede 1997). Such a widely used body of work does however not escape criticism. Arguments supporting and criticizing his work will be briefly explained below. McSweeny (2002) is concerned that the four or later 6 dimensions offered by hofstede imply a simplistic view of culture and other dimensions should be considered to explain culture more wholly. Hofstede (2002) acknowledges this shortcoming but argues that the major dimension are the indentified but not necessarily all dimensions that differentiate cultures. He is also open to addition of other dimensions, ITWOS by Bond (1991) and IVR by Minkov (2010). Williamson (2002) suggests that although Hofstede's dimensions may seem simplistic in number and in their bi-polarity, they offer a method in which quantitative analyses can be pursued. Culture as said before is hard to test and operationilize, in order to simplify the operationlization and to allow at least some aspects of culture to be more easily applied researchers suggest the use of cultural indices. Further hofstede his first assumption (1980) is that organizational, occupational and national culture are independent of each other thus it is ignoring within-country heterogeneity. Hofstede (2002) refutes this by stating that he measured differences between national cultures. He points out that national identities are the only means available in identifying and measuring cultural differences. Often a point of criticize is the relative age of his data. So results come from old data. This he combats by the fact that the relative scores of his dimensions have been proven to be quite stable over decades. The forces that cause cultures to shift tend to be global or continent-wide. Meaning they affect many countries at the same time, so that if culture shifts is shifts in the same direction for these countries and their relative positions remain the same. Also culture is by definition a very slow changing factor, culture does not change overnight. Smith and bond (1999) conclude that large-scale studies published after hofstede's (1980a) work (Chinise Culture Connection, 1987; Schwarts, 1992,1994; Trompenaars, 1993) have sustained and amplified Hofstede his conclusions. Ng, Lee and Soutar (2007) find hofstedes's and Schwartz's value frameworks to be congruent in some ways. Researchers have used Hofstede his framework successfully to select countries that are culturally different in order to increase variance. Hofstede's values are found to be relevant for additional cross-cultural research by Kirkman, Lowe and Gibson (2006).
