Return Predictability and State Variables in ConsumptionBased CAPMs: International Perspectives Bin Li Maosen Zhong UQ Business School, the University of Queensland, Brisbane, QLD, 4072, Australia 1 March 2006 _____________________________________________________________________ Corresponding Author. Email Address: m.zhong@business.uq.edu.au (M. Zhong); Fax: (+61) 7 3365 6788. Return Predictability and State Variables in ConsumptionBased CAPMs: International Perspectives Abstract Asset pricing literature contends that aggregate stock returns can be predicted by business cycle state variables. In this study, we compare the return predictability of three state variables in the context of consumption-based capital asset pricing model (CCAPMs): consumption growth, the consumption-wealth ratio, and the consumption in surplus of habit ratio in the eight major equity markets in the world: Australia, Canada, France, Germany, Italy, Japan, the U.K. and the U.S.. We also explore the time-series relationship between expected excess market returns and the conditional risk aversion associated with the world as well as local consumption in these eight countries. Our results reveal that the consumption-wealth ratio has much more predictive power for returns in the international markets than the surplus consumption ratio and consumption growth rate. - JEL Classification: G12; G15 Keywords: Consumption-based CAPM, Consumption-wealth Ratio, Habit Formation, Predictability, Asset Pricing 1. Introduction Recent literature has related the aggregate stock market return predictability to consumption variables that reflect the business cycle. Starting from Rubinstein (1976), Lucas (1978) and Breeden (1979), many authors define equilibrium in the capital markets using consumption variables. Under a number of assumptions, the asset returns should be linearly related to the growth rate in aggregate consumption as long as the parameters of the linear relationship remain constant over time. Despite the theoretical soundness and simplicity, the consumption-based capital asset pricing model (CCAPM) is usually not easy to be explicitly estimated for the highly nonlinear nature and non-separability of consumption utility among periods. The canonical CCAPM does not performed well empirically both on the U.S. data (see for example, Hansen and Singleton 1982; Mankiw and Shapiro 1986; Breeden, Gibbons and Litzenberger 1989) and international data (Wheatley 1988). In response to this, researchers have modified the consumption-based CAPM hoping to account for the time-varying risk aversion left out in the canonical model. Campbell and Cochrane (1999) (hereafter as CC) introduce a habit formation variable – the surplus consumption variable, which is the time-varying consumption in surplus to habit, to modify the optimal choices of consumption over time to explain the cyclical variation in expected returns and volatility. The habit formation of CC is nonlinear, infinite-horizon, slow-moving and external in response to the history of consumption. Investors tend to consume more relative to their habits in good times as a result of low risk aversion and consume less relative to habit in recessions. CC find that due to the fact that risk aversion is inversely related to surplus consumption, a high level of consumption exceeding habit should forecast low expected stock market returns. Li (2001) proposes an alternative habit formation based on a finite-horizon linear model. Li’s (2001) model preserves the inverse relation between expected returns and surplus consumption, and it performs almost as well as CC’s nonlinear habit model for the U.S. data. The modified consumption-based asset pricing model not only can explain U.S. stock market returns, but also has predictive power for industry portfolios of the U.S. market and the international stock market returns as well (See Li and Zhong 2005; Li, Lu and Zhong 2004; Jacobs and Wang 2004). 1 Furthermore, the recent studies by Lettau and Ludvigson (2001a) explain time series variation of US portfolios returns with a log consumption-wealth ratio called cay t , which is constructed as the residual from the shared trend among log consumption (ct), log asset holding (at) and log labor income (yt). Under the assumption of a representative agent’s binding intertemporal budget constraints, they find there is cointegration relationship between log consumption, log asset holding and log labor income, and their constructed log consumption-wealth ratio ( cay t ) is a strong and positive predictor for U.S. stock market excess returns at short and intermediate horizons. cay t can help explain about 9% of one quarter ahead excess market returns, 18% for annual excess returns, and 15% for the 3-year excess returns. The economic intuition for these results is that investors want to maintain a smooth consumption path and insulate future consumption from fluctuations in expected returns. Thus, they increase consumption out of current asset wealth and labor income (allowing consumption to rise above the shared trend) when expected return is high and decrease consumption out of current asset wealth and labor income (allowing consumption to fall below the shared trend) when expected return is low. There are several potential contributions of our study. First, though the performance scrt and cay t have been well documented in the U.S. market, the international evidence of return predictability is quite limited, partly because of the unavailability of reasonably good data to construct scrt and cay t . We fill the gap by providing international evidence of the predictability of the consumption state variables. Second, to our knowledge, little study has compared the relative performance of the predictive power of three state variables in the context of CCAPMs: the consumption growth, scrt , and cay t . We show that the consumption-wealth ratio can predict the variation of excess stock market returns much better than the surplus consumption ratio scrt and the consumption growth in the eight major markets: Australia, Canada, France, Germany, Italy, Japan, the U.K. and the U.S.. Moreover, each of the three state variables explains the return variations that are unexplained by the other. This indicates the two variables represent different states of aggregate markets. Our interpretation is that scrt captures the risk aversion of the representative agent, whereas cay t represents 2 conditioning variable on the consumption risk factor that induces the time-varying risk premium. Both state variables can affect the risk premia of the aggregate stock market but in different dimensions. The rest of the paper is organized as follows: Section 2 presents the theoretical development of the surplus consumption ratio ( scrt ), the consumption-wealth ratio ( cay t ) and the corresponding econometric models for testing. Section 3 presents the empirical comparison results of the international markets using consumption growth, scrt and cay t . Section 4 concludes this paper. 2. The Model This section first presents the theoretical development of the surplus consumption ratio ( scrt ) and the aggregate consumption-wealth ratio ( cay t ) in the context of CCAPM. 2.1 Surplus Consumption ratio Campbell and Cochrane (1999) present CCAPM with external habit formation where a representative agent derives utility from the difference between consumption and a time-varying habit level, and the model can capture much of excess stock returns in the long horizon. We provide a concise introduction of this model. Assume a representative agent take the utility function as Et j (Ct j X t j )1 1 j 0 , 1 (1) where Ct is the per capita real consumption, Xt is the habit level, δ is the subjective discount factor, and γ is the utility curvature parameter. Abel (1990) and Campbell and Cochrane (1999) suggest the level of habit Xt to be external. Xt is also assumed to depend upon a long history of aggregate consumption and follows an infinite-horizon nonlinear formation process (CC 1999). The surplus consumption ratio SCRt , is defined by SCRt Ct X t 1 Ct . (2) 3 SCRt indicates the state of the economy. When SCRt declines, consumption Ct drops to the habit level Xt, and the economy is reaching recession and investor risk aversion rises. On the other side, rising SCRt means that current consumption exceeds Xt, and the economy is expanding, and investor risk aversion decreases. Hereafter, we use uppercase letters to denote variables at their original scale, and lowercase letters to denote the logs of the corresponding uppercase letters. Campbell and Cochrane (1999) suggest a heteroscedastic AR(1) process for the log surplus consumption ratio, scrt log(SCRt ) : scrt 1 (1 ) scr scrt ( scrt ) c ,t 1 , (3) where φ is the habit persistent parameter, and (scrt ) is the sensitivity function, and c ,t 1 is the innovation in consumption growth. Campbell and Cochrane (1999) suggest that we should use a large value of the persistence parameter in order to obtain the predictive power of the log surplus consumption ratio. As in Li and Zhong (2005), we use φ = 0.90 in this paper1. (scrt ) represents the conditional sensitivity of SCRt to Ct, and it is inversely related to the surplus consumption ratio SCRt . When SCRt falls, the sensitivity function (scrt ) and expected excess returns rise. Consumption growth is assumed independently and identically distributed (i.i.d.). Imposing three conditions that the risk-free rate is fixed, habit is predetermined at the steady state, and habit moves non-negatively with contemporaneous consumption elsewhere, we can specify the sensitivity function (scrt ) max{0, (1/ SCR) 1 2( scrt scr ) 1} , (4) where the steady state is SCR c /(1 ) , and c is the standard deviation of the unexpected consumption growth c ,t 1 . The resulting sensitivity function (scrt ) is inversely related to the log surplus consumption ratio, scrt . Following Campbell and Cochrane (1999) and Li, Lu and Zhong (2005), we choose curvature parameter 2 to compute the steady-state value of scrt . Under the external habit formation, (1) implies the intertemporal marginal rate of substitution of the representative investor at time t 1 is Alternative habit persistence values ( 0.85, 0.95 ) do not alter any of the conclusions reached in this study. 1 4 M t 1 ( Ct 1St 1 ) . Ct St (5) Let Ri ,t 1 denote one plus the rate of return on asset i from time t to t 1 , then Ri ,t 1 satisfies the Euler equation of the following form: Et [ M t 1 Ri ,t 1 ] 1 , (6) where Et is the expectation conditional on the information set as of time t. Under the assumption of the jointly lognormally distribution of asset returns and consumption growth, the Euler equation (6) implies that expected excess returns on any asset is 1 Et [rie,t 1 ] it2 [1 ( scrt )]covt (ri ,t 1 , ct 1 ) 2 In (7), 2 . (7) 1 2 it is the Jensen’s alpha, and the risk premiums on asset i are the price of 2 risk, [1 ( scrt )] , times the conditional covariance of the asset’s returns with consumption growth. The price of risk depends on the utility curvature parameter and the sensitivity function (scrt ) . Given that the expected excess returns are inversely related to scrt , (scrt ) is inversely related to the surplus consumption ratio, scrt (Eq. (4)). Thus investors require higher expected returns on assets when consumption falls toward habit. Under the assumption that the conditional covariance of returns with consumption growth is constant, linear approximations to the sensitivity functions (scrt ) imply that the expected excess return on asset i can be written as Et [rte,t 1 ] 1scrt , (8) where scrt is the log surplus consumption ratio and the slope coefficient 1 is constant. Sensitivity function (scrt ) , according to (7), is positively related to the expected excess market returns, and scrt is inversely related to (scrt ) , so the coefficient 1 is expected to be negative. The model for expected one-period returns in (8) can be easily extended to a model 2 for expected returns over multiple periods. Let See Li (2001) for detailed derivation. 5 K r k 1 e i ,t k rie,t 1 rie,t 2 rie,t k denote the cumulative excess stock market returns with continuous compounding over K periods. If scrt are highly persistent, they should be able to predict multi-period returns. We write expected K-period excess returns as K Et [ rie,t k ] K 1K scrt , (9) k 1 where 1K are constant slope coefficients3. Assuming the growth rates of consumption and dividends are i.i.d., expected excess stock returns are determined by a single state variable – the surplus consumption ratio. Campbell and Cochrane (1999) and Li (2001) show that expected excess returns should be negatively related to the state variable because high (low) surplus consumption at the business cycle troughs (peaks) are associated low (high) investor risk aversion, thus reducing (increasing) the required rate of returns. 2.2 Consumption-wealth ratio Here we theoretically derive the aggregate consumption-wealth ratio cay t that provides useful conditioning information for asset returns. Consider a representative investor who invests his wealth in a single asset with a time-varying risky return Rt4. We denote Wt as aggregate wealth including human capital and household asset holding at time t, Ct as consumption and Rt+1 is the net return on the market portfolio of all invested wealth (including financial and non-financial wealth). For a completemarket model where wealth includes human capital as well as financial assets, under the intertemporal budget constraint, the period-to-period aggregate wealth is Wt 1 (1 Rt 1 )(Wt Ct ) . (10) Campbell and Mankiw (1989) derive the log consumption ratio from (10). In order to make the budget constraint function linear, we approximate it by taking first order Taylor expansion (see Campbell and Mankiw 1989 for details), resulting in wt 1 wt 1 wt k rt 1 (1 1/ )(ct wt ) , 3 Similar to Campbell et al. (1997), if coefficients (11) scrt follows an AR(1) process with an autocorrelation , then the law of iterated expectations implies 1K 1 (1 K ) /(1 ) . 4 Individuals, usually in empirical research, are assumed to be aggregated into a single representative agent economy (see Campbell, Lo and MacKinlay (1997), p.304). 6 where the parameter ρ is the average ratio of invested wealth, W-C, to total wealth, W, and k is a constant. The wealth growth rate wt 1 can be arranged in terms of consumption growth rate and the change of log consumption-wealth ratio. Solving it forward recursively, we can get a log-linear version of the infinite-horizon budget constraint ct wt Et i (rt 1 ct 1 ) . (12) i 1 (12) links the log consumption-wealth ratio with future market return and the future consumption growth, suggesting that a higher log consumption-wealth ratio at this period must forecast either higher returns on the market portfolio at future periods or low future consumption growth rates. Moreover, economic intuition indicates that, when investors expected low future returns on assets, they will drop today’s consumption temporarily below the long term relationship among consumption, asset and wealth to secure future higher consumption. Therefore, trend deviation cay t and excess stock returns should have positive covariance. Given that aggregate wealth Wt is the sum of financial asset At and human capital Ht, the log linear approximation of Wt is a convex combination of the loglinear approximation of At and Ht, wt at (1 )ht , (13) where ω equals the average share of asset holdings in total wealth, A/W. If aggregate labor income Yt can describe the non-stationary component of human capital Ht (see Lettau and Ludvigson (2001a) for details), then we can obtain the following relationship between log consumption-aggregate wealth ratio ct at (1 ) yt Et i ra ,t i (1 )rh ,t i ct i (1 ) zt (14) i 1 Assuming all terms on the right-hand side of (14) are stationary, then the left-hand side must also be stationary, implying ct, at, yt must have cointegration relation with a cointegrating vector (1, -ω, ω-1). We now denote cay t as the left side of (14), ct at (1 ) yt . This equation suggests that as long as expected future returns on human capital rh,t+i and consumption growth ∆ct+i are not too volatile, or if they have high correlation with expected future returns on assets, then cay t should help forecast the expected future asset returns. 7 (14) suggests that cay t has positive linear relation with the excess stock market return. To examine the implication of the single-state-variable model for the predictability of returns on the market portfolio, the regression equation is as follows: Et [rt e,t 1 ] 2 cay t , (15) where Et [ri ,t 1 ] is one-quarter ahead expected excess stock market returns; is a e constant, and 2 is the slope coefficient. Accordingly, the model for expected one-period returns in (15) can be easily extended to a model for expected returns over multiple periods. If cay t are highly persistent, they should be able to predict multi-period returns. We write expected Kperiod excess returns as K Et [ rie,t k ] K 2K cay t , (16) k 1 where 2K are constant slope coefficients. 2.3 The relationship between surplus consumption ratio and consumption-wealth ratio Analogous to Campbell and Cochrane (1999), who shows that the priceconsumption ratio is a function of the surplus consumption ratio, Li (2005) demonstrates that the wealth-consumption ratio dependents upon the surplus consumption ratio. Rather than define the represent investor’s budget constraint as (10), he defines it as Wt 1 Ct 1 Rt 1Wt , (17) Rt 1 (Wt 1 Ct 1 ) / Wt . (18) which is equivalent to Substituting (18) into (6) yields Wt C W Et [ M t 1 t 1 (1 t 1 )] . (19) Ct Ct Ct 1 Equation (19) implies that the wealth-consumption ratio is a function of marginal rate of substitution M t 1 , which is a function of the surplus consumption ratio as specified in equation (5). As the consumption-wealth ratio is the inverse of the wealthconsumption ratio, the consumption-wealth ratio should also depend upon the surplus consumption ratio and the consumption growth rate. Therefore, the consumption- 8 wealth ratio and the surplus consumption ratio should be related in the predictability tests. 3. Data We compile the data of aggregate market returns and macroeconomic data for eight major countries: Australia, Canada, France, Germany, Italy, Japan, the U.K. and the U.S.. As discussed above, the key state variables for return predictability test are the log surplus consumption ratio and the log consumption-wealth ratio. For Australian market, our macroeconomic data include household consumption (nondurable goods and service), after-tax labor income and net household wealth in Australia. Our financial data include All Ordinary stock market returns, and the shortterm bond rate. Appendix A provides detailed descriptions of data source and data construction in Australian market. For Canada, France, Germany, Italy, Japan and the U.K., we obtain the macroeconomic data and financial data from the Datastream International. The consumption data is consumption expenditure from household and non-profit organization serving household. We use household disposable income as the income data. Both the consumption data and income data are quarterly in frequency. However, we are unable to get the household asset holding data at quarterly level. The household wealth data are in the form of the percentages of household income in Datastream. We multiply these percentages with their corresponding household income data to get the household wealth series. They are only available in annual frequency, so we use ARIMA(1,1) model to extrapolate the annual data to the quarterly frequency. These data series are from the first quarter of 1985 to the last quarter of 2004. We also retrieve the consumer price index and the population data from Datastream. Then we use the first quarter of 2000 as the base date, and compute the real per capita household consumption, household income, and household wealth. We take logs of these three variables, and use the dynamic least square estimation to obtain the log consumption-wealth ratio for these six markets (four lags are used in the dynamic least square estimation5). 5 Other lags are also tried, and the results have not been substantially changed. 9 The market indices of six countries are: S&P/TSX Composite for Canadian market, CAC 40 for French market, DAX 30 for German market, Milan COMIT performance index for Italian market, NIKKEI 225 Stock Average for Japanese market, and FTSE 100 for the U.K. market. All of the indices data are from Datastream International. We use Treasury Bill rate to proxy for risk-free rate for these countries except for Japan, for which we use the deposit rate instead because Treasury Bill rate of Japan is not available. The risk-free rates are from the International Financial Statistics. The U.S. consumption-wealth ratio data come from the website of Martin Lettau6, which provides quarterly data from the fourth quarter of 1951 to the fourth quarter of 2004. The variables which are used to construct the U.S. consumptionwealth ratio - consumption, labor income, household net wealth – are log real per capita data. We also use the consumption data to construct the surplus consumption ratio. S&P500 stock market index are from Datastream. The risk-free rate is proxied by the U.S. three-month Treasury bill rate. 4. Empirical Results 4.1 Summary Statistics We construct the surplus consumption ratio and the consumption wealth ratio according to models discussed in section 2. Specifically, for the consumption wealth ratio, we employed augmented Dickey-Fuller (1979) and Philips-Perron (1988) Unit Root test in the series of household consumption, labor income and net household wealth. We find there is only one unit root in each of these three series. Next we move to test whether there is any cointegration relationship among the log per-capita real consumption, wealth, and labor income. To do this, we conduct two kinds of cointegration test: one is Phillips-Ouliaris (1990) residual-based cointegration test, which is to discern whether there is at least one cointegration vector among these three variables; the other test is a more popular one: Johansen’s (1988, 1991) L-Max test and trace statistics test, which will tell us the number of cointegration vectors of the long-term relationship. Appendix B displays the cointegration test results. The 6 Quarterly cay t data for the U.S. can be retrieved from Martin Lettau’s Webpage: http://pages.stern.nyu.edu/~mlettau/. 10 three variables are cointegrated and hence share a common trend in the long run. Therefore, we follow Lettau and Ludvigson (2001a) and construct the consumption wealth ratio using the Dynamic Least Square regression. Table 1 presents the summary statistics of the variables in eight countries: Australia, Canada, France, Germany, Italy, Japan, the U.K., and the U.S. For every country, there are two sub-panels: the correlation matrix and univariate summary statistics. The quarterly variables include log excess return on the market index rt e1 , the surplus consumption ratio scrt , and the consumption-wealth ratio cay t and the consumption growth rate ct 1 . The sample period of each country is given at the top of the each panel. << Insert Table 1 Here>> Sub-panel A in each panel demonstrates that excess return on the stock market is negatively related to the surplus consumption ratio scrt in most markets, but positively correlated with the local consumption-wealth ratio. The sign of the correlation is consistent with the theory we discussed in Section 2. Sub-panel B in each panel presents the first-order autocorrelation of the variables at lag 1. They show that in most countries, excess stock market return and the consumption growth rate exhibit low autocorrelation, but the surplus consumption ratio and the consumption-wealth ratio exhibit high autocorrelation. The high persistence of these two variables may be helpful for forecasting excess market returns over long horizons. 4.2 Long-horizon Forecasting Now, we move to examine the predictive power of the surplus consumption ratio scrt, the consumption-wealth ratio cay t and the consumption growth rate ct 1 for the aggregated stock market returns in the long horizons in the world’s eight major equity markets. Table 2 reports the linear regression estimation results using single state variable: the log surplus consumption ratio, scrt, and the log consumption-wealth ratio, cay t , and the consumption growth rate ct 1 as predictive variables. We run the following least square regressions using full sample over the horizons spanning from 11 K 1 to 16. The dependent variable is cumulative K-period excess market returns: r k 1 e t k . Table 2 shows the estimated coefficients of the regression with the state variable at horizons from quarter 1 to quarter 16. It also reports the adjusted R2, and Newey-West (1987) adjusted t-statistics to account for any residual serial correlation. Newey-West adjusted t-statistics are placed in parentheses under the estimated coefficients, and the adjusted R2s of the regressions are placed in the square brackets below the t-statistics. << Insert Table 2 Here>> The log surplus consumption variable, scrt, which is derived from modified Campbell and Cochrane (1999) model. Here, we use the parameter of the curvature of risk aversion γ=2.0, as suggested by Campbell and Cochrane (1999); the habit persistence parameter φ=0.95. In Australia, the first regression shows that scrt can only predict at the very short horizon. At quarter 1, the estimated coefficient is -0.05, Newey-West adjusted tstatistics is -2.51 and scrt can explain 3% of the variation of the excess stock returns. However, scrt hardly has any predictive power of for excess market returns over intermediate and long horizons. The adjusted R2 at intermediate horizon is negligible over long horizons. The estimated coefficient is not significant at the 10% level. Our finding that the estimated coefficient of scrt at short horizons is negative is consistent with the theory. Under the assumption of i.i.d of the growth rates of consumption and dividends, Campbell and Cochrane (1999) and Li (2001) point out that expected excess stock returns are determined by a single state variable – the surplus consumption ratio, and expected excess returns should be inversely related to the state variable because high surplus consumption at the business cycle troughs are associated with low investor risk aversion, thus lowering the required rate of returns. The second regression shows the forecastability of log excess stock market returns over long horizon using cay t as a single state variable. The predictive power of 12 cay t is quite strong in the intermediate horizons (from 1 year to 3 years). The adjusted R2 has jumped monotonically from quarter 1 to quarter 8, partly because the returns are overlapping returns. At quarter 8 (year 2), the predictive effect of cay t on accumulated excess returns is quite large; the point estimate of the coefficient on cay t is about 5.6. Newey-West adjusted t-statistics is 3.3 and cay t can explain 32% of the variation of the accumulated excess stock returns. At short horizons, cay t has little explanatory power, the R 2 is close to zero. cay t ’s low forecasting power at short horizons may be subsumed by its marginal predicting power of surplus consumption ratio at short horizons. As suggested by Eq. (12), a higher log consumption wealth ratio at this period must forecast either higher returns on the market portfolio at future periods or lower future consumption growth rate. cay t almost has no predicting power of consumption growth (result is not reported), thus, a higher log consumption-wealth ratio at this period must forecast higher returns on the market portfolio at future periods. Therefore, consumption trend deviation should covary positively with excess stock returns. As expected, the coefficient of cay t turns out to be positive in all regressions. The third regression examines the performance of the canonical CCAPM, where the consumption growth rate is the only factor. No coefficients are significant at any horizon, therefore the canonical CCAPM performs poorly in explaining Australian equity return. Interestingly, when all three variables enter the multiple regression, the same picture emerges: the significant coefficients enter the same positions as in the single regressions. The adjusted R2s have improved over those in the single regressions. These results indicate that the three state variables are not substitutes of each other in the predictability tests. Rather, they may well represent different aspects of the timevarying risk premia that drive the expected returns. In Canadian market, the surplus consumption ratio scrt has predictive power from quarter 4 to quarter 16, where the t-statistics and the adjusted R 2 rise monotonically. The coefficient of scrt is 0.28 at quarter 16, it is not only statistically significant, but also economically significant. One percentage in the surplus 13 consumption ratio can explain the 28 basis point changes in the stock market returns. The model can explain more than half of the variation of the stock returns. The sign of the coefficient is consistent with the theory, as high surplus consumption would lower the investor’s risk aversion, thus decrease the investor’s required return. In contrast, the consumption-wealth ratio plays no role in predicting Canadian stock market return. Although there are a couple of coefficients are significant for the consumption growth, 2 the explanatory power of the consumption growth rate is poor as the adjusted R s range below 5%. Like Australia, Canadian multiple regressions with all three state variables as regressors preserve their significant coefficients in the same positions as in the single regressions. Contrary to Canada, the surplus consumption ratio has no predictive power of the stock market return, but the consumption-wealth ratio performs quite well in French market. The coefficients are all significant at 5% level except at quarter 16, which has t-stats of 1.80. At quarter 4, the coefficient is 4.87%, which means for one percentage change in the consumption-wealth ratio, there will be 487 basis point 2 changes in the excess stock market return subsequently. The adjusted R reaches above 30%, which means about one third of stock market variation can be explained by the consumption-wealth ratio. In the multiple regressions, the surplus consumption ratio and the consumption growth enter significantly at the long horizons, whereas the consumption-wealth ratio maintain its significance. The sign of the coefficient is consist with the theory. In contrast, the performance of the consumption growth rate is in accordance with the poor empirical performance this ratio documented in previous literature. For German market, the single and multiple regressions results indicate that the surplus consumption ratio is only significant at the longest horizon, but the sign is positive, which is contradictory with the theory. The wrong sign of the coefficient of the surplus consumption ratio probably is due to the poor measurement of the consumption data, which warrants further investigation. The consumption-wealth ratio is statistically and economically significant up to one-year horizon. However the explanatory power the consumption-wealth ratio is quite limited, as the adjusted R 2 is below 10%. As in other countries, the canonical CCAPM perform poorly, the coefficient at the short horizons are negative, which is contrast with the theory. 14 For Italian market, the surplus consumption ratio has explanatory power in the long horizons and the consumption-wealth ratio only plays a significant role at the short horizons, and the consumption growth rate does not have any effect on the stock market return. For Japanese market, none of the coefficient in the surplus consumption ratio or the consumption growth rate is significant. In contrast, the consumption-wealth ratio performs well, economically and statistically significant, and the model can explain almost half of the variation of the stock returns for quarters 8-16. In the U.K., both the surplus consumption ratio and the consumption-wealth ratio are important factors in determining the stock market returns in the long horizons. The two variables have correct signs of the coefficients, whereas the consumption-wealth ratio has wrong signs in the multiple regressions. For the U.S. market, the results are quite similar to what the previous researchers has done. The surplus consumption ratio and the consumption-wealth ratio capture the stock market variation very well. As the regression horizons increase, the adjusted R 2 build up monotonically. They both are significant at every quarter in the regression except the first quarter. One point need mentioning is that the consumption growth rate, though, economically and statistically significant in the first two quarters( high consumption growth are associated with high stock market return), the explanatory power of this model is quite small, less than 5%. And the coefficients are no longer significant in the longer horizons. Interestingly, the consumption growth variable becomes highly significant in the multiple regressions with the surplus consumption ratio and the consumption wealth ratio. In short, the consumption-wealth ratio predicts the future returns quite well in seven out of eight developed markets, and the signs are in consistent with the theory: high surplus consumption-wealth ratio is associated with higher subsequent stock market return. The surplus consumption ratio performs well in the short horizon in some countries (Canada, U.K. and U.S.), and long horizons in other countries (Australia, France, Germany, Italy, and Japan). Consumption growth generally performs badly in the single forecasting regressions. 15 4.3 Relative Explanatory Power of Three State Variable It is worthwhile to investigate the relative explanatory power of the three state variables (the surplus consumption ratio scrt , the consumption-wealth ratio cay t and the consumption growth rate ct 1 ) in explaining the long horizon excess stock market returns in these eight major markets. Table 3 presents the results for explanatory power of cay t relative to the other state variables in the long-horizon regressions for each of the world’s eight major markets. We first run the regression, Rt k a b scrt et k , where Rt k is the K cumulative excess market returns over K period: Rt k rt e k , scrt is the log surplus k 1 consumption ratio, which is derived from modified Campbell and Cochrane (1999) model, where habit persistent 0.90 and investors’ risk aversion 2 , a and b are coefficients, and et k is the regression residual; then we regress et k and the predicted component Rt k et k separately on the consumption-wealth ratio. For the second sub-panel, we run the similar procedure, except that the first step is a multiple regression with both scrt and ct 1 : Rt k a b scrt c ct 1 et k . First consider the regression results in the first sub-panel where the dependent variable is et k . As the coefficients of cay t in most country panel is significant at 5% level, most return variation which has not been explained by scrt can be explained by cay t . When the dependent variable is the predicted component of returns, Rt k et k , the significance of the coefficients of cay t in most country panel has dropped dramatically. When we add ct 1 in the multiple regression, the results do not change drastically. This finding indicates that the consumption-wealth ratio and the surplus consumption ratio are not substitute. Rather, they can explain the different portions of the return variation. Table 4 presents the results for the relative explanatory power of scrt . The table takes the same form as Table 3. In the first sub-panel of each panel, first we run the regression, Rt k a b cay t et k , then we regress et k and Rt k et k separately 16 on the surplus-consumption ratio, scrt . For the second sub-panel, we run the similar procedure, except that in the first step is ct 1 enters the regression. We find the similar pattern with Table 3, which reinforce our understanding that the consumptionwealth ratio and the surplus consumption ratio should be used together rather than alone in explaining the stock market returns. Table 5 displays the results for the relative explanatory power of the consumption growth ct 1 . Except for Canada and U.S., the consumption growth does not explain any return variation that is unexplained by scrt and cay t . Nor does the consumption growth explain the predicted component of returns predicted by scrt and cay t for France, U.K. and the U.S.. This corroborates the earlier observation that the consumption growth generally explains little of the return variation. 5. Conclusion In this study, we use the consumption growth, the log surplus consumption ratio scrt of Campbell and Cochrane’s (1999) and the log consumption wealth variable cay t of Lettau and Ludvigson (2001a) to examine the predictability of stock market returns in the world major eight equity markets. We find that the surplus consumption ratio, scrt , can predict excess stock returns in most countries, but some at short horizons and others at long horizons; on the other hand, the consumptionwealth ratio cay t can predict the variation of excess stock market return in seven out of eight markets (Australia, Canada, France, Germany, Italy, Japan, the U.K. and the U.S.). The estimated coefficient of scrt is negative, which is also in accordance with the theory: because high surplus consumption at the business cycle troughs is associated with low investor risk aversion, thus lowering the expected required return. cay t is positively related to excess stock returns, which is consistent with economic intuition: when investors expect low future returns on assets, they will drop today’s consumption temporarily below the long term relationship among consumption, asset and wealth to secure future higher consumption. Our study compares the relative predictive power of scrt and cay t for excess market returns. One interesting point we find that the consumption-wealth ratio have 17 relatively better performance than the surplus consumption ratio, which is consistent with the findings from Li, Lu and Zhong (2004). The lower predictive power of the surplus consumption ratio may be attributed to serious data measurement errors of aggregate consumption data as discussed in Campbell and Cochrane (1999). Yet, Lettau and Ludvigson (2001b) argue that the construction of the consumption-wealth ratio based on cointegration estimation may have overcome the problems of unobservable variables. Our study enriches the understanding of long-horizon return predictability in the international setting. 18 REFERENCES Abel, A.B., 1990, Asset prices under habit formation and catching up with the Joneses, American Economic Review Papers and Proceedings 80, 38-42. Breeden, D., 1979, An intertemporal asset pricing model with stochastic consumption and investment opportunities, Journal of Financial Economics 7, 265-296. Breeden, Douglas T., Michael R. Gibbons, and Robert H. Litzenberger, 1989, Empirical tests of the consumption-based CAPM, Journal of Finance 44, 231262 Campbell, John Y., and John H. Cochrane, 1999, By force of habit: a consumptionbased explanation of aggregate stock market behavior, Journal of Political Economy 107, 205-251. Campbell, John Y., Andrew W Lo and A. Craig MacKinlay, 1997, The Econometrics of Financial Markets (Princeton University: Princeton, NJ). Campbell, John Y., and Gregory Mankiw, 1989, Consumption, income and interest rates: Reinterpreting the time series evidence; in Olivier J. Blanchard and Stanley Fischer, eds.: NBER Macroeconomics Annual (MIT Press, Cambridge, MA). Dickey, D., and W.A. Fuller, 1979, Distribution of the estimators for time series regressions with a unit root, Journal of the American Statistical Association 74, 427-431. Hansen, Lars Peter, and Kenneth J. Singleton, 1982, Generalized instrumental variables estimation of nonlinear rational expectations models, Econometrica 50, 1269-1288. Harvey, Campbell R., 1989, Time-varying conditional covariance in tests of asset pricing models, Journal of Financial Economics 24, 289-317. Johansen, Soren, 1988, Statistical analysis of cointegrating vectors, Journal of Economic Dynamics and Control 12, 231-254. Johansen, Soren, 1991, Estimation and hypothesis testing of cointegration vectors in Gaussian vector autoregressive models, Econometrica 59, 1551-1580. Lettau, Martin, and Sydney Ludvigson, 2001a, Consumption, aggregate wealth, and expected stock returns, The Journal of Finance 56, 815-849. Lettau, Martin, and Sydney Ludvigson, 2001b, Resurrecting the (C)CAPM: a crosssectional test when risk premia are time varying, Journal of Political Economy 109, 1238-1287. Li, Yuming, 2001, Expected returns and habit persistence, Review of Financial Studies 14, 861-899. Li, Yuming, 2005, The Wealth-consumption ratio and the consumption-habit ratio, Journal of Business & Economic Statistics 23, 226-241. Li, Yuming, and Maosen Zhong, 2005, Consumption habit and international stock returns, Journal of Banking & Finance 29, 579-601. Li, Yuming, Weili Lu, and Maosen Zhong, 2004, The predictability of industry portfolio returns, the University of Queensland working paper. Lucas, Robert E., 1978, Asset prices in an exchange economy, Econometrica 76, 1429-1445. Mankiw, N. Gregory, and Matthew D. Shapiro, 1986, Risk and return: Consumption versus market beta, Review of Economics and Statistics 68, 452-459. 19 Newey, Whitney K., and Kenneth D. West, 1987, A simple, positive semidefinite, heteroscedasticity and autocorrelation consistent covariance matrix, Econometrica 55, 703-708. Osterwald-Lenum, Michael, 1992, A note with quartiles of the asymptotic distribution of the maximum likelihood cointegration rank test statistic, Oxford Bulletin of Economics and Statistics 54, 461-472. Phillips, P.C.B., and P. Perron, 1988, Testing for a unit root in time series regressions, Biometrika 65, 335-346. Phillips, Peter, and Sam Ouliaris, 1990, Asymptotic properties of residual based tests for cointegration, Econometrica 58, 165-193. Rubinstein, Mark, 1976, The valuation of uncertain income streams and the pricing of options, Bell Journal of economics and Management Science 7, 407-425. Stock, James H., and Mark Watson, 1993, A simple estimator of cointegrating vectors in higher order integrated systems, Econometrica 6, 783-820. Tan, Alvkin, and Graham Voss, 2003, Consumption and wealth in Australia, The Economic Record 79, 39-56. Wheatley, Simon, 1988, Some tests of international equity integration, Journal of Financial Economics 21, 177-212. 20 Table 1. Summary Statistics for the US Market e Notes: rt 1 is quarterly log excess return on the market index; scrt is the surplus consumption ratio; cay t ct a at y y t , where ct is log consumption, at is the log of household net asset wealth, and yt is log after-tax labor income; ct 1 is the consumption growth rate at time t+1. Autocorrelation coefficient is computed as the first-order autocorrelation. Australia sample period: 1979Q1-2004Q2 e scrt rt 1 cay t ct 1 A: Correlation Matrix e rt 1 1 scrt -0.184 1 cay t 0.114 0.013 1 ct 1 -0.009 -0.185 -0.167 Mean Standard Error Autocorrelation 1 B: Univariate Summary Statistics -0.003 -3.264 -2.197 0.005 0.085 0.279 0.019 0.007 0.050 0.813 0.880 -0.195 Canada sample period: 1985Q1-2004Q4 e scrt rt 1 cay t ct 1 A: Correlation Matrix e rt 1 1 scrt -0.069 1 cay t 0.053 0.559 1 ct 1 0.251 0.059 0.174 Mean Standard Error Autocorrelation 1 B: Univariate Summary Statistics 4.0E-04 -3.541 -1.141 0.004 0.087 0.730 0.014 0.008 -0.029 0.938 0.755 -0.017 France sample period: 1985Q1-2004Q4 e rt 1 scrt cay t ct 1 A: Correlation Matrix e rt 1 1 scrt 0.053 1 cay t 0.258 0.259 1 ct 1 -0.227 -0.241 -0.332 Mean Standard Error Autocorrelation 1 B: Univariate Summary Statistics -0.001 -2.295 -3.132 0.004 0.130 0.266 0.032 0.018 -0.056 0.788 0.759 -0.415 21 Germany sample period: 1985Q1-2004Q4 e scrt rt 1 cay t ct 1 A: Correlation Matrix e rt 1 1 scrt 0.046 1 cay t 0.228 -0.091 1 ct 1 -0.257 -0.086 -0.046 Mean Standard Error Autocorrelation 1 B: Univariate Summary Statistics 0.009 -2.877 -2.800 0.005 0.139 0.635 0.035 0.014 -0.080 0.950 0.841 -0.099 Italy sample period: 1985Q1-2004Q4 e scrt rt 1 cay t ct 1 A: Correlation Matrix e rt 1 1 scrt -0.054 1 cay t 0.299 -0.338 1 ct 1 0.143 0.083 -0.166 Mean Standard Error Autocorrelation 1 B: Univariate Summary Statistics 0.008 -2.213 -25.722 0.001 0.133 0.600 0.236 0.022 0.103 0.882 0.900 0.336 Japan sample period: 1985Q1-2004Q4 e rt 1 scrt cay t ct 1 A: Correlation Matrix e rt 1 1 scrt 0.050 1 cay t 0.143 -0.478 1 ct 1 0.181 0.228 -0.095 Mean Standard Error Autocorrelation 1 B: Univariate Summary Statistics -0.003 -3.597 -4.581 0.004 0.120 0.988 0.070 0.010 0.001 0.984 0.917 -0.200 22 U.K. sample period: 1985Q1-2004Q4 e scrt rt 1 cay t ct 1 A: Correlation Matrix e rt 1 1 scrt -0.012 1 cay t -0.188 0.360 1 ct 1 0.113 0.299 -0.151 Mean Standard Error Autocorrelation 1 B: Univariate Summary Statistics -0.001 -2.761 -1.689 0.009 0.092 0.352 0.013 0.008 -0.049 0.953 0.733 0.330 U.S. sample period: 1964Q1-2004Q4 e rt 1 scrt cay t ct 1 A: Correlation Matrix e rt 1 1 scrt -0.083 1 cay t -0.074 -0.176 1 ct 1 0.263 0.081 -0.275 Mean Standard Error Autocorrelation 1 B: Univariate Summary Statistics 0.008 -5.669 -0.129 0.001 0.084 0.922 0.005 0.001 -0.031 0.957 0.887 0.359 23 Table 2. Long-horizon Forecasting Regressions for the International Stock Markets Notes: This table presents the results from the regression of the long horizon excess market returns in world’s eight major markets: Australia, Canada, France, Germany, Italy, Japan, U.K., and U.S. K denotes the return horizon in quarters, spanning from 1 to 16. The dependent variable is the cumulative excess market returns over K K period: r e tk . In each panel, the three explanatory variables enter the regression independently in the first sub- k 1 panel, which are as the followings: the log surplus consumption ratio, scrt, which is derived from modified Campbell and Cochrane (1999) model, where habit persistent 0.90 and investors’ risk aversion 2 ; the log consumption-wealth ratio cay t ct a at y yt , where ct is log consumption, at is the log of household net asset wealth, and yt is log after-tax labor income, a and y are estimated by dynamic least square method, and scrt and the consumption-wealth ratio cay t are lagged one quarter period; another variable is the consumption growth rate, ct 1 ct 1 ct . The second sub-panel, six lags are used; both the log surplus consumption ratio scrt , cay t and ct 1 enter the regression simultaneously as the explanatory variables. Significant coefficients at the 5% significance level are highlighted in bold face. Newey-West adjusted t-statistics are placed in parenthesis under the estimated coefficients. The adjusted R2 of the regression is placed in the square bracket below the tstatistics. Country Australia Regression 1 2 3 Regression Horizons 1 2 4 8 12 sample period: 1979Q1-2004Q2 Regression with single state variable Regressor scrt cay t ct 1 -0.05 -0.06 -0.03 0.04 0.05 0.11 (-2.51) [0.03] (-1.36) [0.01] (-0.35) [-0.01] (0.30) [-0.01] (0.32) [-0.01] (0.47) [0.00] 0.45 1.32 3.28 5.56 5.03 2.98 (0.87) [0.00] (1.44) [0.05] (2.36) [0.17] (3.32) [0.32] (2.07) [0.20] (1.07) [0.05] -0.10 -2.37 -3.71 -2.73 -3.73 -5.33 (-0.09) [-0.01] 4 scrt cay t ct 1 16 (-1.36) (-1.37) (-1.10) (-1.39) [0.01] [0.01] [0.00] [0.00] Regression with multiple state variables (-1.76) [0.01] -0.05 -0.07 -0.02 0.12 0.17 0.18 (-2.50) (-1.69) (-0.27) (1.04) (1.15) (0.73) 0.39 1.12 3.13 5.91 5.57 3.42 (0.73) (1.19) (2.29) (3.78) (2.66) (1.71) -0.30 -2.23 -2.05 1.67 0.96 -1.64 (-0.26) [0.02] (-1.20) [0.07] (-0.86) [0.16] (0.79) [0.33] (0.47) [0.21] (-0.57) [0.06] 24 Country Canada Regression 1 2 3 4 1 2 12 sample period: 1985Q1-2004Q4 Regression with single state variable scrt cay t ct 1 scrt cay t ct 1 Country France Regression 1 2 3 Regression Horizons 4 8 Regressor -0.02 -0.03 -0.08 -0.13 -0.18 -0.28 (-1.42) [0.00] (-1.64) [0.02] (-2.39) [0.07] (-3.36) [0.17] (-4.42) [0.31] (-7.86) [0.56] 0.33 0.78 1.34 0.37 -0.53 -2.83 (0.43) [-0.01] (0.59) [-0.01] (0.57) [0.00] (0.14) [-0.01] (-0.17) [-0.01] (-0.87) [0.01] 2.62 2.63 5.46 4.38 4.42 1.35 (3.07) [0.05] (1.71) (2.50) (2.02) (1.86) [0.02] [0.06] [0.02] [0.02] Regression with multiple state variables -0.02 -0.03 -0.08 -0.17 -0.25 -0.30 (-1.23) (-1.54) (-2.21) (-3.11) (-5.19) (-6.14) 0.52 1.53 3.22 4.10 4.54 0.93 (0.59) (1.09) (1.33) (1.42) (1.34) (0.32) 2.55 2.36 4.93 4.07 1.35 2.53 (2.87) [0.04] (1.52) [0.03] (2.31) [0.12] (1.65) [0.25] (0.56) [0.41] (0.91) [0.66] Regression Horizons 1 2 4 8 12 sample period: 1985Q1-2004Q4 Regression with single state variable Regressor scrt cay t ct 1 scrt cay t ct 1 (0.43) [-0.01] 16 0.04 -0.03 -0.09 -0.18 -0.35 -0.66 (0.31) [-0.01] (-0.15) [-0.02] (-0.29) [-0.01] (-0.44) [0.00] (-0.70) [0.02] (-1.50) [0.09] 1.14 2.50 4.87 5.95 6.07 4.83 (1.98) [0.05] (3.40) [0.16] (4.67) [0.34] (4.57) [0.32] (4.21) [0.24] (1.80) [0.08] -2.12 -1.03 -0.94 -1.60 -2.38 -4.30 (-1.18) [0.03] 4 16 (-0.72) (-0.65) (-0.83) (-1.18) [-0.01] [-0.01] [-0.01] [-0.01] Regression with multiple state variables (-1.79) [0.00] -0.03 -0.15 -0.40 -0.99 -1.53 -1.43 (-0.28) (-1.05) (-1.69) (-3.88) (-4.44) (-3.36) 0.95 2.80 5.78 7.58 7.03 5.61 (1.45) (3.40) (5.15) (6.00) (3.37) (2.30) -1.55 0.50 1.91 -0.04 -3.76 -5.69 (-0.80) [0.04] (0.36) [0.15] (1.58) [0.42] (-0.03) [0.45] (-2.39) [0.42] (-2.02) [0.27] 25 Country Germany Regression 1 2 3 Regression Horizons 4 8 Regressor 1 2 12 sample period: 1985Q1-2004Q4 Regression with single state variable scrt cay t ct 1 0.00 0.02 0.03 0.07 0.19 0.42 (0.14) [-0.02] (0.40) [-0.01] (0.39) [-0.01] (0.56) [-0.01] (1.03) [0.03] (3.54) [0.20] 0.97 1.71 2.26 3.57 4.27 3.73 (2.58) [0.03] (2.77) [0.07] (2.74) [0.06] (1.66) [0.09] (1.43) [0.08] (1.16) [0.05] -2.40 -3.03 -2.44 -2.43 -3.24 -3.45 (-3.28) [0.05] 4 scrt cay t ct 1 Country Italy Regression 1 2 3 0.04 0.07 0.15 0.33 0.58 (0.35) (0.64) (0.72) (1.22) (1.74) (4.31) 0.94 1.71 2.32 3.81 4.85 4.83 (2.48) (2.87) (3.21) (2.17) (2.30) (2.68) -2.27 -2.73 -1.99 -1.53 -1.57 -0.29 (-3.31) [0.06] (-2.43) [0.10] (-1.20) [0.06] (-0.77) [0.11] (-0.52) [0.18] (-0.08) [0.37] Regression Horizons 1 2 4 8 12 sample period: 1985Q1-2004Q4 Regression with single state variable cay t ct 1 scrt cay t ct 1 (-0.85) [-0.01] 0.01 16 -0.01 -0.02 -0.03 -0.03 -0.17 -0.31 (-0.58) [-0.01] (-0.67) [-0.01] (-0.41) [-0.01] (-0.42) [-0.01] (-1.93) [0.06] (-3.20) [0.18] -0.06 -0.10 -0.17 -0.18 -0.29 -0.48 (-2.25) [0.04] (-2.07) [0.07] (-1.77) [0.09] (-2.07) [0.05] (-3.38) [0.12] (-4.81) [0.29] 0.84 1.09 0.40 0.28 1.16 -0.44 (1.26) [0.01] 4 (-2.75) (-1.49) (-1.01) (-1.01) [0.04] [0.00] [-0.01] [-0.01] Regression with multiple state variables Regressor scrt 16 (1.23) (0.26) (0.10) (0.40) [0.00] [-0.01] [-0.01] [-0.01] Regression with multiple state variables (-0.15) [-0.02] 0.01 0.01 0.02 -0.05 -0.25 -0.39 (0.63) (0.28) (0.32) (-0.55) (-2.99) (-7.14) 0.21 0.33 0.46 0.11 -0.14 0.04 (2.79) (2.55) (1.87) (0.41) (-0.40) (0.08) 1.16 1.61 1.13 0.56 1.49 0.52 (1.79) [0.09] (1.76) [0.11] (0.76) [0.08] (0.20) [-0.03] (0.64) [0.07] (0.29) [0.21] 26 Country Japan Regression 1 2 3 Regression Horizons 4 8 Regressor 1 2 12 sample period: 1985Q1-2004Q4 Regression with single state variable scrt cay t ct 1 0.01 0.01 0.01 0.03 0.05 0.06 (0.40) [-0.01] (0.34) [-0.01] (0.16) [-0.01] (0.29) [-0.01] (0.44) [0.00] (0.43) [-0.01] 0.25 0.58 1.49 2.95 3.96 4.44 (1.40) [0.01] (1.91) [0.04] (3.56) [0.16] (4.69) [0.35] (4.34) [0.48] (3.97) [0.54] 2.07 1.62 3.38 3.51 2.19 3.74 (1.62) [0.02] 4 scrt cay t ct 1 Country U.K. Regression 1 2 3 0.02 0.04 0.09 0.13 0.11 (0.45) (0.83) (0.76) (1.28) (1.61) (1.21) 0.27 0.60 1.39 2.62 3.29 3.38 (2.14) (2.94) (5.47) (8.27) (8.81) (7.53) 1.82 0.94 2.14 0.82 -1.46 0.08 (1.43) [0.03] (0.77) [0.07] (1.19) [0.24] (0.53) [0.48] (-0.99) [0.59] (0.04) [0.59] Regression Horizons 4 8 1 2 12 sample period: 1985Q1-2004Q4 Regression with single state variable cay t ct 1 scrt cay t ct 1 (0.84) [-0.01] 0.01 16 -0.02 -0.07 -0.16 -0.41 -0.62 -0.79 (-0.60) [-0.01] (-1.10) [0.00] (-1.78) [0.05] (-2.74) [0.22] (-2.67) [0.33] (-2.95) [0.45] -1.09 -1.85 -0.86 3.56 8.83 14.05 (-1.56) [0.01] (-1.63) [0.02] (-0.38) [-0.01] (1.15) [0.03] (2.29) [0.15] (2.53) [0.31] 1.41 3.19 1.01 3.08 -4.06 -5.43 (0.81) [0.00] 4 (1.14) (1.28) (0.90) (0.42) [0.00] [0.01] [0.00] [-0.01] Regression with multiple state variables Regressor scrt 16 (1.24) (0.38) (0.82) (-1.75) [0.02] [-0.01] [0.00] [0.00] Regression with multiple state variables (-2.04) [0.00] 0.03 0.03 0.02 -0.15 -0.20 -0.38 (1.06) (0.61) (0.38) (-1.73) (-1.42) (-2.03) -1.71 -2.79 -3.88 -2.27 -2.86 1.60 (-2.23) (-2.70) (-2.01) (-1.12) (-1.07) (0.40) 0.40 1.95 -0.49 4.91 -0.99 1.66 (0.23) [0.02] (0.83) [0.08] (-0.19) [0.07] (1.39) [0.11] (-0.41) [0.12] (0.56) [0.19] 27 Country U.S. Regression 1 2 3 4 Regression Horizons 1 2 4 8 12 sample period: 1965Q1-2004Q4 Regression with single state variable Regressor scrt cay t ct 1 scrt cay t ct 1 16 -0.01 -0.03 -0.06 -0.12 -0.17 -0.21 (-1.96) [0.01] (-2.34) [0.04] (-2.90) [0.10] (-3.34) [0.20] (-3.58) [0.29] (-3.78) [0.38] -0.68 3.52 10.74 24.90 35.85 41.28 (-0.63) [0.00] (1.99) [0.02] (2.79) [0.11] (4.68) [0.33] (6.44) [0.49] (5.69) [0.52] 10.77 15.72 15.07 -1.00 -16.39 -21.14 (3.19) [0.04] (2.73) (1.23) (-0.08) (-0.85) [0.04] [0.02] [-0.01] [0.00] Regression with multiple state variables (-1.02) [0.01] -0.01 -0.02 -0.03 -0.06 -0.08 -0.12 (-1.65) (-1.42) (-1.65) (-1.91) (-2.36) (-3.69) -0.64 3.87 10.59 22.58 31.37 34.54 (-0.55) (1.94) (2.55) (4.16) (6.61) (5.72) 12.25 22.20 29.95 27.68 23.39 28.61 (3.60) [0.05] (3.89) [0.09] (2.60) [0.18] (3.31) [0.38] (1.61) [0.54] (2.47) [0.62] 28 Table 3. Relative Explanatory Power of cayt Notes: This table presents the results of the long horizon regressions in world’s eight major markets: Australia, Canada, France, Germany, Italy, Japan, U.K., and U.S. K denotes the return horizon in quarters, spanning from 1 to 16. In the first sub-panel of each panel, first we run the regression, Rt k a b scrt et k , where Rt k is K the cumulative excess market returns over K period: Rt k rt e k , scrt is the log surplus consumption ratio, k 1 which is derived from modified Campbell and Cochrane (1999) model, where habit persistent 0.90 and investors’ risk aversion 2 , a and b as coefficient, and et k is the regression residual; then we run et k and Rt k et k separately on the consumption-wealth ratio, cay t ct a at y yt , where ct is log consumption, at is the log of household net asset wealth, and yt is log after-tax labor income, a and y are estimated by dynamic least square method, and six lags are used; both the log surplus consumption ratio scrt and the consumption- wealth ratio cay t are lagged one quarter period. For the second sub-panel, we run the similar procedure, except that the first step is Rt k a b scrt c ct 1 et k , where ct 1 ct 1 ct is the consumption growth rate. The results of the second step are reported in the table. Significant coefficients at the 5% significance level are highlighted in bold face. Newey-West adjusted t-statistics are placed in parenthesis under the estimated coefficients. The adjusted R2 of the regression is placed in the square bracket below the t-statistics. Country Australia Regression Horizons 4 8 Regressand 1 2 sample period: 1979Q1-2004Q2 12 16 Rt k a b scrt et k et k Rt k et k 0.41 1.27 3.25 5.64 5.17 3.34 (0.78) [0.00] (1.36) [0.04] (2.32) [0.17] (3.41) [0.33] (2.15) [0.21] (1.23) [0.07] 0.04 0.06 0.03 -0.07 -0.14 -0.35 (0.43) [-0.01] (0.43) [-0.01] (0.50) [0.00] (-1.22) [0.01] (-2.73) [0.05] (-3.24) [0.06] Rt k a b scrt c ct 1 et k et k Rt k et k 0.37 1.07 2.98 5.45 4.88 2.92 (0.72) [0.00] (1.19) [0.03] (2.18) [0.14] (3.34) [0.31] (2.07) [0.19] (1.11) [0.05] 0.08 0.25 0.30 0.12 0.16 0.07 (0.82) [0.00] (1.72) [0.03] (2.11) [0.04] (1.22) [0.00] (1.14) [0.00] (0.30) [-0.01] 29 Country Canada Regression Horizons 4 8 Regressand 1 2 sample period: 1985Q1-2004Q4 12 16 Rt k a b scrt et k et k Rt k et k 0.57 1.25 2.68 3.24 3.31 0.88 (0.77) [0.00] (0.98) [0.01] (1.22) [0.03] (1.29) [0.04] (1.19) [0.04] (0.37) [-0.01] -0.24 -0.47 -1.34 -3.76 -6.14 -9.68 (-5.21) [0.30] (-4.64) [0.30] (-4.03) [0.30] (-3.84) [0.30] (-4.08) [0.29] (-4.43) [0.33] Rt k a b scrt c ct 1 et k et k Rt k et k Country France 0.35 1.02 2.15 2.75 3.04 0.59 (0.48) [-0.01] (0.83) [0.00] (1.02) [0.02] (1.14) [0.03] (1.11) [0.03] (0.25) [-0.01] -0.01 -0.24 -0.82 -3.27 -5.87 -9.39 (-0.08) [-0.01] (-1.31) [0.00] (-1.68) [0.02] (-3.10) [0.18] (-3.83) [0.26] (-4.24) [0.31] 12 16 Regression Horizons 4 8 Regressand 1 2 sample period: 1985Q1-2004Q4 Rt k a b scrt et k et k Rt k et k 1.08 2.55 5.14 7.07 7.03 5.81 (1.93) [0.04] (3.43) [0.16] (4.93) [0.39] (6.28) [0.38] (3.81) [0.24] (2.18) [0.11] 0.06 -0.05 -0.28 -1.00 -2.20 -2.62 (1.33) [0.05] (-1.17) [0.05] (-1.01) [0.05] (-0.98) [0.05] (-1.34) [0.07] (-1.58) [0.13] Rt k a b scrt c ct 1 et k et k Rt k et k 0.81 2.39 4.94 6.48 5.70 4.20 (1.48) [0.02] (3.15) [0.14] (4.61) [0.36] (6.30) [0.33] (3.84) [0.17] (1.73) [0.06] 0.33 0.10 -0.07 -0.41 -0.87 -1.01 (2.19) [0.09] (1.05) [0.02] (-0.24) [-0.01] (-0.37) [-0.01] (-0.46) [-0.01] (-0.56) [-0.01] 30 Country Germany Regression Horizons 4 8 Regressand 1 2 sample period: 1985Q1-2004Q4 12 16 Rt k a b scrt et k et k Rt k et k 0.98 1.76 2.35 3.81 4.82 4.76 (2.56) [0.04] (2.75) [0.08] (2.88) [0.07] (1.99) [0.11] (2.01) [0.12] (2.26) [0.14] -0.02 -0.05 -0.09 -0.23 -0.55 -1.03 (-0.66) [-0.01] (-0.57) [-0.01] (-0.48) [-0.01] (-0.55) [-0.01] (-0.64) [-0.01] (-0.71) [-0.01] Rt k a b scrt c ct 1 et k et k Rt k et k Country Italy 0.93 1.69 2.29 3.75 4.75 4.73 (2.48) [0.03] (2.84) [0.08] (2.98) [0.07] (1.98) [0.11] (2.01) [0.12] (2.26) [0.14] 0.04 0.02 -0.03 -0.18 -0.49 -1.00 (0.23) [-0.02] (0.08) [-0.02] (-0.13) [-0.02] (-0.40) [-0.01] (-0.55) [-0.01] (-0.68) [-0.01] 12 16 Regression Horizons 4 8 Regressand 1 2 sample period: 1985Q1-2004Q4 Rt k a b scrt et k et k Rt k et k 0.17 0.27 0.40 0.09 -0.14 0.03 (2.41) [0.07] (2.07) [0.08] (1.56) [0.08] (0.31) [-0.01] (-0.37) [-0.01] (0.05) [-0.02] 0.01 0.02 0.03 0.05 0.21 0.36 (3.00) [0.10] (2.54) [0.10] (2.10) [0.10] (1.79) [0.10] (1.68) [0.10] (1.62) [0.09] Rt k a b scrt c ct 1 et k et k Rt k et k 0.18 0.28 0.40 0.09 -0.12 0.04 (2.69) [0.08] (2.27) [0.09] (1.60) [0.08] (0.33) [-0.01] (-0.31) [-0.01] (0.06) [-0.02] 0.00 0.01 0.03 0.05 0.19 0.35 (-0.18) [-0.01] (0.78) [-0.01] (1.88) [0.04] (1.75) [0.07] (1.66) [0.07] (1.61) [0.08] 31 Country Japan Regression Horizons 4 8 Regressand 1 2 sample period: 1985Q1-2004Q4 12 16 Rt k a b scrt et k et k Rt k et k 0.30 0.65 1.57 3.15 4.24 4.68 (1.65) [0.01] (2.21) [0.06] (3.91) [0.18] (5.82) [0.41] (5.58) [0.56] (4.88) [0.61] -0.04 -0.08 -0.08 -0.20 -0.29 -0.24 (-2.80) [0.22] (-2.35) [0.22] (-1.93) [0.22] (-1.59) [0.20] (-1.44) [0.17] (-1.35) [0.12] Rt k a b scrt c ct 1 et k et k Rt k et k Country U.K. 0.29 0.65 1.56 3.14 4.24 4.67 (1.69) [0.01] (2.23) [0.06] (3.95) [0.18] (5.89) [0.41] (5.60) [0.56] (4.98) [0.61] -0.04 -0.07 -0.07 -0.19 -0.28 -0.23 (-1.04) [0.00] (-1.69) [0.06] (-0.97) [0.00] (-1.26) [0.08] (-1.33) [0.15] (-1.10) [0.06] 12 16 Regression Horizons 4 8 Regressand 1 2 sample period: 1985Q1-2004Q4 Rt k a b scrt et k et k Rt k et k -1.40 -2.37 -1.80 1.68 5.96 11.02 (-1.99) [0.02] (-2.19) [0.04] (-0.90) [0.01] (0.69) [0.00] (3.09) [0.06] (3.23) [0.20] -0.03 -0.25 -0.75 -2.17 -3.90 -5.26 (-2.52) [0.12] (-2.14) [0.12] (-1.79) [0.12] (-1.59) [0.12] (-1.56) [0.11] (-1.49) [0.10] Rt k a b scrt c ct 1 et k et k Rt k et k -1.14 -1.76 -1.45 2.76 6.11 11.28 (-1.64) [0.01] (-1.67) [0.02] (-0.72) [0.00] (1.15) [0.02] (3.18) [0.06] (3.37) [0.21] -0.28 -0.86 -1.10 -3.25 -4.05 -5.52 (-3.05) [0.08] (-3.46) [0.12] (-2.47) [0.19] (-2.19) [0.19] (-1.60) [0.12] (-1.53) [0.11] 32 Country U.S. Regression Horizons 4 8 Regressand 1 2 sample period: 1965Q1-2004Q4 12 16 Rt k a b scrt et k et k Rt k et k -1.26 2.16 7.60 18.44 26.56 29.23 (-1.21) [0.00] (1.25) [0.00] (2.00) [0.05] (3.61) [0.21] (5.06) [0.36] (4.70) [0.40] 0.58 1.36 3.14 6.46 9.30 12.05 (4.77) [0.12] (4.07) [0.12] (3.38) [0.12] (2.78) [0.12] (2.53) [0.12] (2.42) [0.12] Rt k a b scrt c ct 1 et k et k Rt k et k -0.48 3.37 9.01 19.23 26.73 29.50 (-0.51) [-0.01] (2.18) [0.02] (2.66) [0.08] (3.96) [0.24] (5.13) [0.36] (4.80) [0.41] -0.20 0.15 1.73 5.68 9.13 11.78 (-0.48) [0.00] (0.19) [-0.01] (1.32) [0.02] (2.33) [0.09] (2.48) [0.12] (2.36) [0.12] 33 Table 4. Relative Explanatory Power of scrt Notes: This table presents the results of the long horizon regressions in world’s eight major markets: Australia, Canada, France, Germany, Italy, Japan, U.K., and U.S. K denotes the return horizon in quarters, spanning from 1 to 16. In the first sub-panel of each panel, first we run the regression, Rt k a b cay t et k , where Rt k is K Rt k rt e k , cay t ct a at y yt is the consumption- the cumulative excess market returns over K period: k 1 wealth ratio, where ct is log consumption, at is the log of household net asset wealth, and yt is log after-tax labor income, a and y are estimated by dynamic least square method, and six lags are used, a and b are slope coefficients, and et k is the regression residual; then we run et k and Rt k et k separately on scrt , the log surplus consumption ratio, which is derived from modified Campbell and Cochrane (1999) model, where habit persistent 0.90 and investors’ risk aversion 2 ; both the log surplus consumption ratio scrt and the consumption-wealth ratio cay t are lagged one quarter period. For the second sub-panel, we run the similar procedure, except that the first step is Rt k a b cay t c ct 1 et k , where ct 1 ct 1 ct is the consumption growth rate. The results of the second step are reported in the table. Significant coefficients at the 5% significance level are highlighted in bold face. Newey-West adjusted t-statistics are placed in parenthesis under the estimated coefficients. The adjusted R2 of the regression is placed in the square bracket below the tstatistics. Country Australia Regressand 1 2 sample period: 1979Q1-2004Q2 Regression Horizons 4 8 12 16 Rt k a b cay t et k et k Rt k et k -0.06 -0.06 -0.02 0.11 0.14 0.17 (-2.49) [0.02] (-1.27) [0.01] (-0.21) [-0.01] (0.67) [0.01] (0.68) [0.01] (0.58) [0.01] 4.6E-04 -6.2E-05 -0.01 -0.05 -0.07 -0.01 (0.09) [-0.01] (0.00) [-0.01] (-0.17) [-0.01] (-0.86) [0.02] (-1.88) [0.07] (-2.14) [0.10] Rt k a b cay t c ct 1 et k et k Rt k et k -0.06 -0.07 -0.03 0.10 0.12 0.12 (-2.47) [0.02] (-1.47) [0.02] (-0.33) [-0.01] (0.64) [0.00] (0.59) [0.00] (0.41) [0.00] 0.00 0.01 0.01 -0.05 -0.05 0.04 (-0.02) [-0.01] (0.62) [0.00] (0.16) [-0.01] (-0.79) [0.01] (-1.39) [0.02] (2.17) [0.04] 34 Country Canada Regressand 1 2 sample period: 1985Q1-2004Q4 Regression Horizons 4 8 12 16 Rt k a b cay t et k et k Rt k et k -0.01 -0.02 -0.06 -0.12 -0.17 -0.20 (-1.11) [0.00] (-1.26) [0.01] (-1.84) [0.05] (-2.61) [0.16] (-4.35) [0.28] (-5.62) [0.38] 0.00 0.01 0.01 -0.01 -0.03 -0.09 (5.84) [0.30] (5.42) [0.30] (4.97) [0.30] (-5.50) [0.30] (-5.93) [0.29] (-8.19) [0.33] Rt k a b cay t c ct 1 et k et k Rt k et k Country France -0.01 -0.02 -0.06 -0.12 -0.17 -0.20 (-1.04) [0.00] (-1.22) [0.01] (-1.83) [0.05] (-2.61) [0.16] (-4.37) [0.28] (-5.63) [0.38] 0.00 0.01 0.01 -0.01 -0.03 -0.09 (0.97) [-0.01] (2.47) [0.04] (1.76) [0.02] (-1.71) [0.01] (-6.56) [0.25] (-8.51) [0.32] 12 16 Regressand 1 2 sample period: 1985Q1-2004Q4 Regression Horizons 4 8 Rt k a b cay t et k et k Rt k et k -0.01 -0.14 -0.39 -0.92 -1.35 -1.18 (-0.09) [-0.02] (-0.98) [0.01] (-1.52) [0.11] (-3.31) [0.26] (-4.96) [0.33] (-3.34) [0.21] 0.05 0.11 0.21 0.27 0.22 0.18 (1.47) [0.05] (1.30) [0.05] (1.15) [0.05] (1.17) [0.05] (1.59) [0.07] (2.34) [0.13] Rt k a b cay t c ct 1 et k et k Rt k et k -0.03 -0.13 -0.36 -0.90 -1.36 -1.21 (-0.27) [-0.02] (-0.91) [0.00] (-1.38) [0.09] (-3.12) [0.25] (-5.03) [0.34] (-3.51) [0.22] 0.07 0.10 0.18 0.24 0.23 0.20 (1.66) [0.08] (1.17) [0.04] (0.94) [0.03] (1.03) [0.04] (1.68) [0.08] (3.03) [0.12] 35 Country Germany Regressand 1 2 sample period: 1985Q1-2004Q4 Regression Horizons 4 8 12 16 Rt k a b cay t et k et k Rt k et k 0.02 0.04 0.07 0.15 0.33 0.57 (0.51) [-0.01] (0.76) [0.00] (0.78) [0.01] (1.21) [0.04] (1.73) [0.12] (4.15) [0.35] -0.01 -0.01 -0.01 -0.02 -0.04 -0.03 (-0.62) [-0.01] (-0.53) [-0.01] (-0.45) [-0.01] (-0.50) [-0.01] (-0.58) [-0.01] (-0.65) [-0.01] Rt k a b cay t c ct 1 et k et k Rt k et k Country Italy 0.01 0.04 0.06 0.15 0.32 0.56 (0.35) [-0.02] (0.64) [0.00] (0.73) [0.00] (1.17) [0.04] (1.67) [0.12] (4.11) [0.34] 0.00 0.00 -0.01 -0.02 -0.03 -0.02 (0.04) [-0.02] (-0.14) [-0.02] (-0.28) [-0.02] (-0.42) [-0.01] (-0.47) [-0.01] (-0.44) [-0.02] 12 16 Regressand 1 2 sample period: 1985Q1-2004Q4 Regression Horizons 4 8 Rt k a b cay t et k et k Rt k et k 0.01 0.01 0.02 -0.04 -0.22 -0.35 (0.68) [-0.01] (0.30) [-0.01] (0.28) [-0.01] (-0.47) [-0.01] (-2.20) [0.08] (-3.65) [0.19] -0.02 -0.04 -0.05 -0.02 -0.01 -0.04 (-2.51) [0.10] (-2.21) [0.10] (-2.00) [0.10] (-2.12) [0.10] (-2.16) [0.10] (-1.92) [0.09] Rt k a b cay t c ct 1 et k et k Rt k et k 0.01 0.01 0.02 -0.04 -0.22 -0.35 (0.66) [-0.01] (0.25) [-0.01] (0.27) [-0.01] (-0.48) [-0.01] (-2.22) [0.08] (-3.65) [0.19] -0.02 -0.03 -0.05 -0.02 -0.01 -0.04 (-2.17) [0.06] (-2.11) [0.07] (-2.05) [0.09] (-2.20) [0.08] (-1.51) [0.00] (-1.92) [0.09] 36 Country Japan Regressand 1 2 sample period: 1985Q1-2004Q4 Regression Horizons 4 8 12 16 Rt k a b cay t et k et k Rt k et k 0.02 0.03 0.07 0.14 0.20 0.22 (1.00) [0.00] (1.16) [0.02] (1.34) [0.06] (2.10) [0.20] (3.02) [0.32] (3.21) [0.30] -0.01 -0.02 -0.05 -0.11 -0.15 -0.16 (-4.90) [0.22] (-4.06) [0.22] (-3.26) [0.22] (-2.52) [0.20] (-2.09) [0.17] (-1.67) [0.12] Rt k a b cay t c ct 1 et k et k Rt k et k Country U.K. 0.01 0.03 0.06 0.13 0.19 0.21 (0.71) [-0.01] (1.01) [0.01] (1.15) [0.04] (1.94) [0.17] (2.88) [0.30] (2.97) [0.28] 0.00 -0.02 -0.04 -0.10 -0.14 -0.14 (-1.44) [0.00] (-2.89) [0.10] (-2.47) [0.12] (-2.16) [0.15] (-1.92) [0.14] (-1.52) [0.10] 12 16 Regressand 1 2 sample period: 1985Q1-2004Q4 Regression Horizons 4 8 Rt k a b cay t et k et k Rt k et k 0.01 0.01 -0.04 -0.20 -0.34 -0.49 (0.57) [-0.01] (0.22) [-0.01] (-0.66) [-0.01] (-2.12) [0.10] (-2.39) [0.22] (-3.16) [0.38] -0.02 -0.03 -0.03 -0.01 0.02 0.05 (-3.18) [0.12] (-2.72) [0.12] (-2.38) [0.12] (-2.36) [0.12] (2.52) [0.11] (2.16) [0.10] Rt k a b cay t c ct 1 et k et k Rt k et k 0.01 -0.01 -0.04 -0.22 -0.31 -0.45 (0.25) [-0.01] (-0.24) [-0.01] (-0.70) [0.00] (-2.35) [0.13] (-2.19) [0.18] (-2.90) [0.33] -0.01 -0.01 -0.03 0.02 -0.01 0.02 (-1.40) [0.02] (-0.85) [0.00] (-2.09) [0.09] (2.10) [0.05] (-0.56) [-0.01] (0.74) [-0.01] 37 Country U.S. Regressand 1 2 sample period: 1965Q1-2004Q4 Regression Horizons 4 8 12 16 Rt k a b cay t et k et k Rt k et k -0.01 -0.01 -0.03 -0.05 -0.07 -0.11 (-1.42) [0.01] (-1.10) [0.01] (-1.32) [0.02] (-1.74) [0.06] (-2.18) [0.12] (-3.72) [0.22] 1.4E-03 -0.01 -0.02 -0.05 -0.07 -0.08 (3.28) [0.12] (-2.76) [0.12] (-2.28) [0.12] (-1.89) [0.12] (-1.72) [0.12] (-1.62) [0.12] Rt k a b cay t c ct 1 et k et k Rt k et k -0.01 -0.02 -0.03 -0.06 -0.08 -0.11 (-1.89) [0.02] (-1.65) [0.02] (-1.80) [0.04] (-1.99) [0.08] (-2.35) [0.14] (-4.00) [0.25] 4.1E-03 -2.1E-03 -0.02 -0.05 -0.07 -0.08 (2.51) [0.04] (-0.48) [0.00] (-1.36) [0.03] (-1.63) [0.09] (-1.60) [0.11] (-1.52) [0.11] 38 Table 5. Relative Explanatory Power of ct 1 Notes: This table presents the results of the long horizon regressions in world’s eight major markets: Australia, Canada, France, Germany, Italy, Japan, U.K., and U.S. K denotes the return horizon in quarters, spanning from 1 to 16. In the first sub-panel of each panel, first we run the regression, Rt k a b cay t c scrt et k , K where Rt k is the cumulative excess market returns over K period: Rt k rt e k , cay t ct a at y yt is the k 1 consumption-wealth ratio, where ct is log consumption, at is the log of household net asset wealth, and yt is log after-tax labor income, a and y are estimated by dynamic least square method, and six lags are used, scrt is the log surplus consumption ratio, which is derived from modified Campbell and Cochrane (1999) model, where habit persistent 0.90 and investors’ risk aversion 2 , a, b and c are slope coefficients, and et k is the scrt and the consumption-wealth ratio cay t are et k separately on ct 1 , where ct 1 ct 1 ct is the regression residual; both the log surplus consumption ratio lagged one quarter period. Then we run et k and Rt k consumption growth rate. The results of the second step are reported in the table. Significant coefficients at the 5% significance level are highlighted in bold face. Newey-West adjusted t-statistics are placed in parenthesis under the estimated coefficients. The adjusted R2 of the regression is placed in the square bracket below the tstatistics. Country Australia Regression Horizons 4 8 Regressand 1 2 sample period: 1979Q1-2004Q2 12 16 Rt k a b cay t c scrt et k et k Rt k et k Country Canada -0.27 (-0.27) [-0.01] -2.24 (-1.30) [0.01] -2.72 (-1.04) [0.00] 0.02 (0.01) [-0.01] -1.03 (-0.42) [-0.01] -3.28 (-1.05) [0.00] 0.17 (0.73) [-0.01] -0.13 (-0.34) [-0.01] -0.99 (-1.65) [0.01] -2.75 (-3.52) [0.04] -2.70 (-4.70) [0.08] -2.05 (-5.24) [0.12] 12 16 Regression Horizons 4 8 Regressand 1 2 sample period: 1985Q1-2004Q4 Rt k a b cay t c scrt et k et k Rt k et k 2.46 (2.95) [0.04] 2.28 (1.51) [0.01] 4.76 (2.29) [0.05] 3.92 (1.71) [0.02] 1.29 (0.59) [-0.01] 2.41 (0.91) [0.00] 0.16 (0.89) [0.00] 0.35 (0.91) [0.00] 0.70 (0.78) [0.00] 0.50 (0.31) [-0.01] 0.06 (0.03) [-0.02] -1.59 (-0.54) [-0.01] 39 Country France Regression Horizons 4 8 Regressand 1 2 sample period: 1985Q1-2004Q4 12 16 Rt k a b cay t c scrt et k et k Rt k et k Country Germany -1.34 (-0.71) [0.00] 0.43 (0.29) [-0.02] 1.65 (1.27) [0.00] -0.03 (-0.02) [-0.02] -3.14 (-2.49) [0.00] -4.68 (-2.48) [0.01] -0.78 (-2.74) [0.09] -1.46 (-2.19) [0.05] -2.60 (-1.88) [0.03] -2.35 (-1.18) [0.00] -1.16 (-0.61) [-0.02] -0.85 (-0.61) [-0.02] 12 16 Regression Horizons 4 8 Regressand 1 2 sample period: 1985Q1-2004Q4 Rt k a b cay t c scrt et k et k Rt k et k Country Italy -2.24 (-3.36) [0.04] -2.70 (-2.51) [0.03] -1.97 (-1.31) [0.00] -1.52 (-0.72) [-0.02] -1.54 (-0.54) [-0.02] -0.28 (-0.08) [-0.02] -0.16 (-0.48) [-0.01] -0.32 (-0.54) [-0.01] -0.47 (-0.61) [-0.01] -0.91 (-0.89) [-0.01] -1.70 (-1.67) [-0.01] -3.16 (-2.52) [0.00] 12 16 Regression Horizons 4 8 Regressand 1 2 sample period: 1985Q1-2004Q4 Rt k a b cay t c scrt et k et k Rt k et k 1.12 (1.76) [0.03] 1.56 (1.70) [0.02] 1.10 (0.68) [-0.01] 0.55 (0.19) [-0.01] 1.45 (0.57) [-0.01] 0.51 (0.20) [-0.02] -0.28 (-1.72) [0.01] -0.48 (-1.71) [0.01] -0.70 (-1.64) [0.01] -0.26 (-1.46) [0.01] -0.29 (-0.51) [-0.01] -0.94 (-1.20) [-0.01] 40 Country Japan Regression Horizons 4 8 Regressand 1 2 sample period: 1985Q1-2004Q4 12 16 Rt k a b cay t c scrt et k et k Rt k et k Country U.K. 1.90 (1.62) [0.02] 1.30 (1.02) [-0.01] 2.95 (1.47) [0.01] 2.51 (1.34) [0.00] 0.77 (0.42) [-0.01] 2.36 (0.87) [0.00] 0.17 (0.70) [-0.01] 0.32 (0.59) [-0.01] 0.43 (0.31) [-0.01] 1.00 (0.33) [-0.01] 1.42 (0.37) [-0.01] 1.39 (0.40) [-0.01] 12 16 Regression Horizons 4 8 Regressand 1 2 sample period: 1985Q1-2004Q4 Rt k a b cay t c scrt et k et k Rt k et k Country U.S. 0.78 (0.45) [-0.01] 2.37 (0.94) [0.01] 1.15 (0.44) [-0.01] 6.84 (2.19) [0.05] 3.04 (1.73) [0.00] 5.33 (2.28) [0.02] 0.63 (2.65) [0.05] 0.82 (1.74) [0.02] -0.14 (-0.24) [-0.01] -3.76 (-3.70) [0.12] -7.09 (-4.39) [0.14] -10.77 (-4.14) [0.16] 12 16 Regression Horizons 4 8 Regressand 1 2 sample period: 1965Q1-2004Q4 Rt k a b cay t c scrt et k et k Rt k et k 11.10 (3.34) [0.04] 20.10 (3.68) [0.07] 27.01 (2.63) [0.08] 25.01 (3.38) [0.05] 21.14 (1.46) [0.03] 24.96 (2.16) [0.05] -0.32 (-0.56) [0.00] -4.37 (-3.16) [0.09] -11.94 (-2.75) [0.09] -26.01 (-2.44) [0.09] -37.53 (-2.27) [0.09] -46.10 (-2.22) [0.09] 41 Appendix A: Description of Australian Dataset Macroeconomic data usually only exist in lower frequency such as quarterly or annually. In Australia, Australia Bureau of Statistics (ABS) maintains a good record of macroeconomic data. Our consumption, wealth and labor income data are constructed from the time series spreadsheets from AusStats Database 7. These variables are only available in quarterly or annually. For this paper, we use the data dating back to the fourth quarter of 1976 (1976Q4) and until the second quarter of 2004 (2004Q2), which yields 111 observations. The data used here are quarterly, seasonally adjusted, real per capita data, measured in 1989-90 dollar. The consumption data are for non-durables goods and services. Using non-durables goods and services for consumption data has been used by many researchers (e.g. Lettau and Ludvigson, 2001a; Li, Lu and Zhong, 2004) It is calculated as Total Household Final Consumption Expenditure less Clothing and Footwear, Furnishings and Household Equipment, and Purchase of Vehicles ($m, seasonally adjusted in current prices), which is taken from series 5206058.1 of AusStats Time Series Spreadsheets. CPI (Consumer Price Index), from Series 640101b of AusStats Time Series Spreadsheets, is used to deflate all nominal variables in this paper. It is weighted average of all groups index of eight capital cities, with base index 1989-90 = 100. Aggregate wealth, particularly human capital, is not directly observable. In order to use it for forecasting asset returns, we must find a proxy for human capital. To overcome this obstacle, we use after tax labor income to proxy human capital. We use a number of series from AusStats Time Series Spreadsheets to construct after-tax labor income. After-tax labor income is defined as wages and salaries plus transfer payment minus labor income tax. INCOME WAGES TRANSFERS * TAX (20) Wages is quarterly non-farmer wage & salary earners’ average earnings. It is constructed from Non-farmer Wage & Salary Earner’s average weekly earnings and measures of employment. Average weekly earnings (AWE) are from AusStats Series 1364019 (seasonally adjusted, in $). WSE is the total number of non-farm civilian wage and salary earners, available from AusStats Series 1364010. ω is the number of weeks in one quarter, calculated as (1/7)*(365/4). (21) WAGES * AWE *WSE Transfers are constructed as Total Secondary Income Receivable less Social Contributions for Workers Compensation, available from AusStats Series 5206036. γ is the proportion of labor income in the total household income, which is calculated as WAGES/(Total Primary Income). Tax is calculated as the sum of Income Tax Payable and Other Current Tax on income, wealth etc. All above series are available in AusStats Series 5206036. The labor income we constructed above is aggregate labor income. We deflate this series with Australian Population (available from AusStats Series 1364010) and CPI to get per capita real after-tax labor income. Quarterly net household wealth data for the period from 1976Q4 to 1999Q3 are taken from Tan & Voss (2003)8, which includes financial wealth and non-financial wealth. Using annual net wealth of household balance sheet from AusStats Series 5204050, we extend household wealth data from 1999Q4 to 2004Q2 by interpolation. We employed augmented Dickey-Fuller (1979) and Philips-Perron (1988) Unit Root test in the series of household consumption, labor income and net household wealth. We find there is only one unit root in each of these three series. Next we move to test whether there is any cointegration relationship among these three variables. To do this, we conduct two kinds of cointegration test: one is Phillips-Ouliaris (1990) residual-based cointegration test, which is to discern whether there is at least one cointegration vector among these three variables; the other test is a more popular one: Johansen’s (1988, 1991) L-Max test and trace statistics test, which will tell us the number of cointegration vectors of the long-term relationship. Both tests suggest there is only one cointegration vector among the three variables. The results of Phillips-Ouliaris cointegration test and Johansen cointegration test are given in Panel A and Panel B of Appendix B, respectively. Cointegration tests suggest that there is a shared trend in consumption, labor income and net asset wealth. In order to examine the predictability of stock market returns using cay t , we should estimate the parameters of the cointegration relation. Due to endogenously determined nature of ct, at and yt series, we use the single equation procedure suggested in Stock and Watson (1993) and use the dynamic least squares (DLS) estimates with Newey and West (1987) to correct for any residual serial correlation, k k i 1 i 1 ct a at y yt bai at i byi yt i t . Ordinary Least Square (OLS) estimate of (22) produces the super-consistent estimate of (22) a and y of cointegration parameters (Lettau and Ludvigson 2001). Adding leads and lags of the first difference of net asset wealth and labor income will capture the effects of regressor endogeneity in the linear regression of consumption 7 8 These series can be retrieved at http://www.abs.gov.au/ausstats/abs@.nsf/w2.3.1?OpenView. Real per capita net household wealth data can be downloaded at http://web.uvic.ca/~gvoss/. 42 on asset wealth and labor income. From DLS estimation of (22), we can obtain the estimated trend deviation cayt ct a at y yt , where ‘hat’ means estimated parameter. Using quarterly data from 1976Q4 to 2004Q2, we obtain the estimated coefficients of the trend deviation (Corresponding t-statistics are given below in the parenthesis) of constant, net asset wealth and labor income: ct 1.476 0.405at 0.239 yt (3.06) (11.85) Thus (23) (2.29) cayt ct 0.405at 0.239 yt . For Australian stock market index, there are two commonly used indices: AllOrd Index and ASX/S&P 200 Index. AllOrd Index is from the Centre for Research in Finance (CRIF) database of the Australian Graduate School of Management (AGSM), which is dating back to December 1979 with base index of 500 in that month. ASX/S&P 200 index is collected from the bulletin statistics of Reserve Bank of Australia (RBA) with base index of 500 at December 1979. They are monthly data, so we convert them into quarterly data using the average value of the three months in each quarter. AllOrd Index is the value weighted price index for the companies traded in Australian Stock Exchange (ASX) which should provide a better proxy for nonhuman components of household asset wealth than ASX/S&P 200 index. We try stock market returns using both AllOrd Index and ASX/S&P 200 index, and the empirical results are not sensitive to the type of index used. We deflate the stock market index by CPI to get the real stock market return. Stock market returns are log returns: rt = log(AllOrdt) – log(AllOrdt-1) , (24) where AllOrd is the series of All Ord Index. For risk-free rate, we use the short-term 2-year government bond as a proxy. Quarterly and monthly 2year government short-term bond yields are also available in IFS until June 2004. Let us denote rmt as the log real return of the stock market index, and rft as the log real return on the risk-free rate, then log real excess return on the stock market is rte rmt rft . Appendix B: Cointegration Tests of Consumption, Household Income and Household Wealth in Eight Countries Notes: Panel A examines the presence of unit root from the cointegrating regression of consumption on after-tax labor income and net household wealth using Philips-Ouliaris (1990) residual-based cointegration test for Australia. Augmented Dickey-Fuller test is applied to the residuals from the regression of consumption (ct) on wealth (at) and household income (yt). The optimal lag is chosen by the AIC criterion. The critical values are assuming trending series. Significant t-test statistics at the 5% significance level are highlighted in bold face. Panel B reports Johansen L-Max and Trace Statistics for cointegration tests among consumption, income and asset wealth. A constant is included in the cointegration space. The column labeled “L-max90” and “L-max95” denote the 90% and 95% confidence level of L-max statistics, respectively; the “Trace90” and “Trace95” gives the 90% and 95% confidence level of trace statistics, respectively. “r” is the number of the cointegration relation. Optimal lag is chosen by AIC criterion. Significant test statistics at 95% (90%) confidence level are highlighted in bold (italics) face. The critical values of the Johansen cointegration tests are obtained from Osterwald-Lenum (1992, Table 1). Australia Panel A. Phillips-Ouliaris Test for Cointegration Relationship Critical Value Augmented Dickey-Fuller T-statistics 5% Critical Value 10% Critical Value -1.94 -1.62 -2.88 H 0: r 0 1 2 L-max Statistics 34.58 9.77 6.45 L-max90 19.77 13.75 7.52 Panel B. Johansen Cointegration Test L-max95 Trace Statistics 22.00 50.81 15.67 16.23 9.24 6.45 Trace90 32.00 17.85 7.52 43 Trace95 34.91 19.96 9.24 Canada Panel A. Phillips-Ouliaris Test for Cointegration Relationship Critical Value Augmented Dickey-Fuller T-statistics 5% Critical Value 10% Critical Value -0.63 -1.94 -1.62 H 0: r 0 1 2 L-max Statistics 27.02 14.05 2.12 L-max90 19.77 13.75 7.52 Panel B. Johansen Cointegration Test L-max95 Trace Statistics 22.00 43.19 15.67 16.17 9.24 2.12 Trace90 32.00 17.85 7.52 Trace95 34.91 19.96 9.24 France Panel A. Phillips-Ouliaris Test for Cointegration Relationship Critical Value Augmented Dickey-Fuller T-statistics 5% Critical Value 10% Critical Value -1.16 -1.94 -1.62 H 0: r 0 1 2 L-max Statistics 21.19 8.24 3.49 L-max90 19.77 13.75 7.52 Panel B. Johansen Cointegration Test L-max95 Trace Statistics 22.00 32.93 15.67 11.73 9.24 3.49 Trace90 32.00 17.85 7.52 Trace95 34.91 19.96 9.24 Germany Panel A. Phillips-Ouliaris Test for Cointegration Relationship Critical Value Augmented Dickey-Fuller T-statistics 5% Critical Value 10% Critical Value -1.27 -1.94 -1.62 H 0: r 0 1 2 L-max Statistics 19.51 11.51 2.44 Panel B. Johansen Cointegration Test L-max90 L-max95 Trace Statistics 19.77 22.00 33.46 13.75 15.67 13.95 7.52 9.24 2.44 Trace90 32.00 17.85 7.52 Trace95 34.91 19.96 9.24 Italy Panel A. Phillips-Ouliaris Test for Cointegration Relationship Critical Value Augmented Dickey-Fuller T-statistics 5% Critical Value 10% Critical Value -1.94 -1.62 -2.04 H 0: r 0 1 2 L-max Statistics 26.98 9.03 3.17 L-max90 19.77 13.75 7.52 Panel B. Johansen Cointegration Test L-max95 Trace Statistics 22.00 39.19 15.67 12.21 9.24 3.17 Trace90 32.00 17.85 7.52 44 Trace95 34.91 19.96 9.24 Japan Panel A. Phillips-Ouliaris Test for Cointegration Relationship Critical Value Augmented Dickey-Fuller T-statistics 5% Critical Value 10% Critical Value -1.94 -1.62 -1.83 H 0: r 0 1 2 L-max Statistics 35.20 24.83 6.22 L-max90 19.77 13.75 7.52 Panel B. Johansen Cointegration Test L-max95 Trace Statistics 22.00 66.24 15.67 31.05 9.24 6.22 Trace90 32.00 17.85 7.52 Trace95 34.91 19.96 9.24 U.K. Panel A. Phillips-Ouliaris Test for Cointegration Relationship Critical Value Augmented Dickey-Fuller T-statistics 5% Critical Value 10% Critical Value -1.16 -1.94 -1.62 H 0: r 0 1 2 L-max Statistics 30.83 19.64 3.51 L-max90 19.77 13.75 7.52 Panel B. Johansen Cointegration Test L-max95 Trace Statistics 22.00 53.97 15.67 23.15 9.24 3.51 Trace90 32.00 17.85 7.52 Trace95 34.91 19.96 9.24 U.S. Panel A. Phillips-Ouliaris Test for Cointegration Relationship Critical Value Augmented Dickey-Fuller T-statistics 5% Critical Value 10% Critical Value -1.94 -1.62 -2.04 H 0: r 0 1 2 L-max Statistics 50.91 10.31 7.09 L-max90 19.77 13.75 7.52 Panel B. Johansen Cointegration Test L-max95 Trace Statistics 22.00 68.31 15.67 17.40 9.24 7.09 Trace90 32.00 17.85 7.52 45 Trace95 34.91 19.96 9.24