VALIDACIÓN DE INSTRUMENTO

The automatic lag length tests are carried out as a starting point again. As in the US, there is evidence of a lag structure with 14 lags and thus we allow for 14 lags in the

25 _{Theory suggests that markets are forward looking and that the stock market rises on the back of} expectations of continuing economic expansion. We therefore assume that current economic growth rates are a good proxy for expected economic growth.

cointegration analysis (see table 15 on page 183). Two cointegration vectors are found via the trace statistics and by the maximum eigenvalue statistic at the 5% level (see table 16 on page 184). Thus we conclude that we might have two cointegration vectors in the Japanese data set. As a first step we have considered only one vector and taken the estimated long- term relationships for one cointegration vector into account. In contrast to the US data set, only CPI appears to be significant in the Japanese case. This would mean that all other variables drop out of the relationship and stock prices only depend on the CPI (see table 17 on page 184). We believe this cannot represent a proper macroeconomic model for stock prices. For this reason, one vector has not been an option for Japan and two vectors have been considered. In the case of two vectors, we allow for a financial vector normalised on the stock market and an output vector normalised on industrial production. Further, we test the probability of restricting individual variables in the two different vectors. We find a high p-value of 0.71 that there exist those two vectors. One vector is normalised on the stock market and shows a positive effect of industrial production and a negative impact of money supply on stock prices. The second vector is normalised on industrial production with a negative effect of CPI and the discount rate on industrial production (see table 18 on page 185). Furthermore, both vectors show some error correction behaviour with the major drivers being M1, CPI and the Nikkei. However, only M1 is statistically significant in both error correction vectors.

To evaluate the relationship further, we consider the plot of the two cointegration vectors against time. Graph 16 on page 186 visualises that the stock market vector appears to trend down until the early 1990s, whereas the industrial production vector (see graph 17 on page 187) seems to shift up until the early 1990s. Thus there is only “one” mean reversion effect in both vectors over the whole period and this indicates a structural shift rather than good error correction behaviour in the cointegrated system. We would expect more mean-

reversion effects in the cointegration vector that indicate recessions, shocks and changes in risk aversion. For this reason, we think the calculated long-term relationship via cointegration analysis might not be very strong in Japan. One reason may be a structural break. Therefore, we will now split the data set. It is generally known that Japan was in a high economic growth period during the 1980s and in a low or no growth period thereafter. Thus it is likely to have an economic structural shift at the end of the 1980s or early 1990s. As our theoretical stock price model assumes that corporate earnings are mainly correlated with economic output and we have used industrial production as the main fundamental driver of stock prices, we first test for a structural break in industrial production between 1985 and 1995. The structural break test, explained in chapter 3.6.4, finds good evidence of a break for industrial production in March 1993 (see table 19 on page 189). Furthermore, the downshifting stock market vector hits a low during the early 1990s whereas the up-shifting output vector reaches a high after 1992. During the early 1990s, Japan experienced a down-shift in economic growth and CPI. We therefore expect a structural break to be most likely during the early part of the 1990s. This gives us good reason to test for a structural break in March 1993 in the Japanese cointegration system. The bootstrapped Chow break point test, as explained in chapter 3.6.4, is applied to the Japanese dataset for March 1993. We set the number of bootstrapped replications to 1000. The p-value for no structural break in March 1993 is found to be zero (see table 20 on page 189) and we conclude that there is reasonable evidence that a structural break occurred in March 1993.

Hence, the next step is to split the data set into two parts, before and after March 1993. First of all, we carry out a unit root test in order to verify the order of integration in the sub periods. In both periods we find good evidence that all series are I(1) (see table 21 and 22 on pages 189 and 190). Furthermore, the optimal lag length is determined via lag length

criteria tests (see table 23 and 24 on pages 190 and 191). We then start with the earlier period between January 1965 and March 1993 with the cointegration analysis. Two vectors are found via the maximum eigenvalue statistics and five26 cointegration vector are found by the trace statistic at the 5% level (see table 25 on page 192). Thus we conclude that we might have two vectors or possibly even more cointegration vectors in the early part of the Japanese data set. As a first step we consider only one vector and take the estimated long- term relationships for one cointegration vector into account. We only find the CPI variable being significant in the long-term relationship whereas M1 is the most insignificant variable (see table 26 on page 192). Given that money supply M1 has been the most insignificant variable in the long term relationship, we test for M1 being equal to zero and the test shows a p-value of 0.65 of that being the case (see table 27 on page 193). Hence we drop money supply and re-estimate the relationship. Now we find two vectors for the trace and for the maximum eigenvalue statistic (see table 28 on page 193). First, we only look at one vector and find that all variables are statistically significant in the long-term relationship27 (see table 29 on page 194). Industrial production gives a large positive coefficient, thus supporting the view that the Nikkei225 rose on the back of economic expansion until March 1993. In contrast, the 10-year bond yield shows a very large negative coefficient und thus supports the assumption of rising real interest rates having a negative impact on real stock prices during that period. Also the price level CPI has a negative impact on real stock prices until 1993 and indicates that rising consumer prices were bad news for real stock prices. Except for CPI, the size of the coefficients is generally larger than the estimates we have found in the US data set between 1965 and 2004. This indicates that until 1993, the Japanese stock market was indeed more cyclical and more

26 _{This is not in line with economic theory and we make a rational decision to consider at most two}

cointegration relationships.

27 _{A scond vector could be given by the money equation. However, we could not support the money equation}

as second vector. As we do not have a theoretical motivation for a second vector, we only allow for one vector.

sensitive to macroeconomic variables. Also the error correction term has a negative impact on the Nikkei as well as consumer prices and the cointegration vector shows mean reversion behaviour (see graph 25 on page 200).

Secondly, we now look at the period between March 1993 and June 2004. Multiple cointegration vectors are found via the maximum eigenvalue statistics and by the trace statistic at the 5% level (see table 30 on page 194). Thus we conclude that we might have more cointegration vectors in the later part of the Japanese data set. As a first step we consider only one vector and take the estimated long-term relationships for one cointegration vector into account (see table 31 on page 195). We find all variables are statistically significant in the long-term relationship. All variables show a positive coefficient except money supply, which indicates a negative impact. Contrary to the early period in Japan, consumer prices show now a large positive coefficient. Industrial production still shows a positive coefficient and the size of the coefficient is comparable to the US or Japan during the 1965M1 to 1993M3 period. Contrary to our expectations, the discount rate also shows a large positive coefficient. Finally, money supply yields a negative impact on the Japanese stock market between March 1993 and June 2004. The error-correction term shows a negative relationship with the Nikkei, the discount rate and money supply M1. A look at the cointegration vector reveals strong mean reversion behaviour for the later period in Japan (see graph 26 on page 200). Thus, the split into the period before and after March 1993 has resulted in much more reasonable results than for the whole period. Furthermore, the early part of the data set seems to yield long-term coefficients that are comparable to the US data set, whereas the later period in Japan seems to be characterised by a change in the impact of inflation, the discount rate and money supply. In section 3.8 we will discuss the findings and try to find reasons and arguments for the differences and commonalties that we have found. However, before we discuss the

findings in detail, we will gather more information by applying variance decomposition to the data.