1
University of Amsterdam
MSc Business Economics, Finance track
Master Thesis
Supervisor: Dr. P.J.P.M. Versijp 7 July 2016
Is gold losing its status as a safe-haven asset?
A research on the effect of the 2015 Chinese stock market crash Arend Griffioen10086692
Abstract
This paper investigates the safe-haven property of gold and how well it stood against the 2015 Chinese stock market crash by using a GARCH model. Daily data is used on the price of gold and the MSCI World Index for the period 2006-2015. The main conclusion is that gold actually served as a (weak) safe haven during the 2015 Chinese stock market crash. The full sample is also divided into two bull and two bear markets. The conclusion from this analysis is that gold acted as a weak safe haven in every subsample. Another finding is that gold can serve as a hedge against the MSCI World Index. Further analysis with time dummies suggest that the case for gold being a safe haven was actually stronger during the 2015 Chinese stock market crash compared to other periods in the sample. Alternative regressions show evidence that gold acts as a strong safe haven in bull market periods and as a weak safe haven in bear market periods.
2
Statement of Originality
3
1. Introduction
Traditionally, gold has been seen as a safe-haven asset. Baur and McDermott (2010) make a distinction between a hedge and a safe haven. They define that a hedge is an asset that is uncorrelated with another asset on average. They furthermore define that an asset is a strong (weak) safe haven when it is negatively correlated (uncorrelated) with another asset in times of falling stock markets (2010, p. 1889). These definitions are adopted in this thesis.
The safe-haven property of gold has been tested and confirmed by several recent papers (e.g. Baur & Lucey, 2010, Reboredo, 2013). However, the 2015 Chinese stock market crash, which also affected other markets like the major European and US stocks, has shown signs that gold is losing its traditional role as a safe-haven asset. Specifically, MacDonald and Shumsky (2015) state in their article that fears about China’s economy led to an 11% decrease in the S&P 500 stock index, when gold just gained 5.9% before falling back. They contrast these figures to the global financial crisis when the price of gold increased by 150% from October 2008 to September 2011 in the wake of the financial crisis (MacDonald & Shumsky, 2015). Therefore, the focus of this thesis will be on the following research question: Has gold lost its safe-haven status during the 2015 Chinese stock market crash?
It should be made clear that this thesis is not a research on the Chinese stock market alone, but actually on the increased worldwide stock market volatility during the Chinese crash and its effect on the price of gold. Therefore, a world stock market index is used. This choice is justified by the fact that this crash spilled over into other markets around the world. A graphical representation of the MSCI World Index in 2015 can be found in Figure 1 below. It can be seen that the world index declined from approximately 1770 to 1550 from June to October 2015, which is a decline of approximately 12.4%.
The contribution of this thesis is that it investigates whether the correlation between gold and the MSCI World Index was more positive during the Chinese stock market crash than during the global financial crisis, as MacDonald and Shumsky (2015) imply in their article.
Additionally, subsamples are used to analyze different bull and bear market periods and German government bond yields (Bunds) are used to test the robustness of the results.
The main findings of this thesis are that gold appears to be a weak safe-haven asset in every period analyzed, for both the 10 and 5percent worst stock returns on the MSCI Index. Another finding is that gold actually was a better safe haven during the 2015 Chinese stock
4
market crash compared to other periods, while no such evidence is found for the period of the global financial crisis. Alternative regressions show that gold actually serves as a strong safe haven in bear market periods and as a weak safe haven in bull market periods.
This paper is organized as follows. Section 2 gives a short introduction on gold as an asset and concludes with an overview of some of the existing literature in which the safe-haven
property of gold is investigated. Section 3 describes the data, hypotheses and research methods used. Section 4 presents the results and additional robustness tests, whereas Section 5 discusses the results and main drawbacks of this thesis.
Figure 1: The MSCI World Index in 2015
This figure shows the MSCI World Index in 2015. It can be seen that global stocks declined approximately 12.4% during the period of the Chinese stock market crash in the months from June to October.
5
2. Literature
This section gives a short introduction on gold as a financial asset and subsequently outlays some of the recent findings on the safe-haven property of gold in the existing literature and links these results to the findings of this thesis.
2.1. Gold
Gold has been a valuable commodity throughout history. It has been used as a medium of exchange and as a store of value (Baur and McDermott, 2010, p. 1887). Many currencies were linked to gold during the eras of the gold standard and the Bretton Woods system in the past two centuries (for further discussion on this topic, see e.g. Giovannini, 1988). Today, gold is mainly used as a store of value and as an investment asset. It attracts considerable attention from
investors, researchers, and the media (Beckmann et al., 2015, p. 16) and is seen as a safe haven in times of economic distress (Baur & Glover, 2012, p. 1).
2.2. Previous findings
There are a number of papers that investigate the safe-haven property of gold in times of economic distress. Baur and Lucey (2010) are the first in the literature to make a distinction between a hedge and a safe haven. They define that a hedge is an asset that is uncorrelated or negatively correlated with another asset or portfolio on average. On the other hand, they define a safe haven to be an asset that is uncorrelated or negatively correlated with another asset or portfolio in times of market declining stock markets (2010, p. 210). This definition is adopted in other recent papers that study the correlation of gold with other assets (e.g. Reboredo, 2013, Hood & Malik, 2013, and Beckmann et al., 2015). Baur and McDermott (2010) extend this definition by making a further distinction between a strong and weak safe-haven asset. They define that a strong (weak) safe haven is an asset that is negatively correlated (uncorrelated) with another asset in times of falling stock markets (2010, p. 1889). As mentioned before, this
definition is also used in this thesis.
Capie et al. (2005) research whether gold has acted as an exchange rate hedge using weekly data for the period 1971-2004 on the gold price and sterling-dollar and yen-dollar exchange rates. They find a negative, typically inelastic, relationship between gold and those exchange rates, with varying strength of this relationship over time (2005, p. 343).
6
using the CAPM for the period 1970-2003. They find that gold has approximately the same mean return as a Treasury Bill and bears no markets risk. Their conclusion is that gold acts as an inflation hedge in their sample (2006, p. 2).
Baur and Lucey (2010) analyze whether gold works as a safe haven in financial markets. Their finding is that gold is a safe haven for stocks and not for bonds. They also find that the safe-haven property only lasts for about 15 trading days after an extreme negative market shock (2010, p. 228). The finding that gold is a safe haven for stocks is in line with the findings of this thesis.
Baur and McDermott (2010) test the hypothesis that gold represents a safe haven against stocks of major emerging and developing countries for the period 1979-2009. Their analysis shows that gold is both a hedge and safe haven for major European stock markets and the US, but not for Australia, Canada, Japan, and large emerging countries such as the BRIC countries. They furthermore looked at specific crisis periods and found that gold was a strong safe haven for most developed markets during the peak of the recent financial crisis (2010, p. 1886). While the
finding for the European and US stock markets are in line with the findings of this thesis, this is not the case for the other countries analyzed. Moreover, their finding that gold was a strong safe haven for most developed markets during the peak of the financial crisis is confirmed by the alternative regressions provided in Table 10 in Appendix B.
Coudert and Raymond-Feingold (2011) look into the role of gold as a safe haven or hedge against stocks by examining stock market indices of France, Germany, the UK, the US, and the G7. They find that gold qualifies as a safe haven for all these stock indices and that gold seems to be able to hedge against stock losses in most cases, although the evidence for the latter is less clear-cut (2011, p. 1613). Their finding that gold qualifies as a safe haven for the stock indices is again in line with the findings of this thesis.
Joy (2011) uses a model of dynamic conditional correlation to test whether gold is a hedge or a safe haven against the US dollar for the period 1986-2008. Although he finds evidence that gold behaves as a hedge against the US dollar, he finds that gold does not act as an effective safe haven from market stress in the entire sample. He does find that gold has increasingly acted as an effective safe haven in more recent years of the sample (2011, pp. 120-131).
Ciner et al. (2013) use quantile regressions to examine the behavior of prices in asset markets during large and unlikely events. They particularly focus on the ability of gold and oil to
7
provide protection during extreme declines in more traditional asset classes. The finding of their study is that gold can be regarded as a safe haven against exchange rates in both the US and the UK. However, they find that gold is not a safe haven for equities (2013, pp. 202-208).
Reboredo (2013a) assesses the role of gold as a safe haven or hedge against the US dollar (USD). He finds evidence of positive and significant dependence between gold and USD
depreciation against different currencies, which implies that gold can act as an effective safe haven. He also finds symmetric tail dependence, indicating that gold can act as an effective safe haven in periods of extreme USD market movements (2013, p. 2675).
In another paper, Reboredo (2013b) also assesses the role of gold as a hedge or safe haven against oil price movements. The main finding is that there is evidence of tail independence between the two markets, indicating that gold can act as an effective safe haven against oil price movements (2013b, p. 130).
Likewise, Hood and Malik (2013) evaluate the role of gold and other precious metals relative to the Volatility Index (VIX) as a hedge and safe haven. They find that gold, unlike other precious metals, serves as a hedge and a weak safe haven for the US stock market (2013, p. 51). The finding that gold serves as a hedge and a weak safe haven for the stock market are yet again in line with the findings of this thesis.
Similarly, Beckmann et al. (2015) perform a broad study on gold as a hedge and/or safe haven that includes data from 18 individual economies and five regional indices. They show that gold generally serves as both a hedge and a safe haven, but that that ability depends on the specific economic environment under observation (2015, p. 23). The alternative regressions in Appendix B show evidence that the safe-haven ability of gold indeed depends on the specific economic environment under observation, namely whether it is a bull or bear market period.
The consensus in the literature seems to be that gold is an effective safe haven in many circumstances. The contribution of this thesis is that it shows whether gold has lost that safe-haven status during the recent 2015 market turmoil, when fears about the Chinese economy led to a global decrease of stocks.
8
3. Empirical analysis
This section describes the data, hypotheses, and empirical method used for this research in detail and concludes with robustness checks.
3.1. Hypotheses
In order to answer the research question, the following hypotheses will be tested:
1. The price of gold is negatively correlated with the MSCI World Index during the 2015 Chinese stock market crash.
2. The price of gold is uncorrelated with the MSCI World Index during the 2015 Chinese stock market crash.
3. The correlation between gold and the MSCI World Index increased during the 2015 Chinese stock market crash compared to other periods.
Using the same definition as Baur and McDermott (2010), if evidence for the first (second) hypothesis is found, then gold is said to be a strong (weak) safe haven asset. The third hypothesis tests whether there is a change in the correlation between gold and the stock market. The
correlations will be estimated using a GARCH model as will be explained in Subsection 3.4. The expectation is that both the first and second hypotheses will be rejected. The reason is that both the price of gold and the world index declined during the 2015 Chinese stock market crash as will be shown Subsection 3.3. For the same reason, it is expected that the third hypothesis will not be rejected. In other words, it is expected that gold did not serve as a safe-haven asset during the 2015 Chinese stock.
3.2. Data
The data used consists of daily returns for a period of 10 years from January 2006 to December 2015. This specific period is chosen to include the global financial crisis that started in 2007 and the main period of interest, the year 2015. The price of gold is obtained from Quandl, the MSCI World Index from Stooq, the Dollar Index from the FED website, and the German government bond (Bund) yields from the Bundesbank website. The Dollar Index is a weighted average of the US dollar against the currencies of a group of 26 major US trading partners1.
Table 1 shows summary statistics of the data. The total number of observations is 2430
1
See https://research.stlouisfed.org/fred2/series/DTWEXB/downloaddata for more information on which currencies are included.
9
spread over 10 years, so each year is assumed to have 243 trading days. The means are calculated by taking the average daily return and compounding it by the number of trading days per year. It can be seen that gold has an average annual return of 10.33%, whereas the world index has an average annual return of 4.53% in this sample. The standard deviations are 1.31% and 11.73%, respectively. Thus, in this period gold has both a higher annual return and a smaller standard deviation than the world index.
Table 1: Summary Statistics
Observations Mean Std. deviation Minimum Maximum
Gold (in US$) 2430 0.10331 0.01314 -0.08527 0.10025
World Index (in US$) 2430 0.04529 0.11732 -0.08103 0.12332
Dollar Index 2430 0.01222 0.00366 -0.22756 0.17483
Bund yield 2430 0.03800 0.43048 -0.38461 0.66667
This table presents the descriptive statistics of the gold and world index returns in US dollars, the return on the dollar index, and the return on the Bund yield. The statistics are based on daily returns where 2430/10 = 243 trading days per annum is assumed.
3.3. Descriptive analysis
This subsection gives an overview of the data to get a first impression of the relationship between the gold price and the world index. Figure 2 below shows the relationship between the gold price and the world index for the whole sample period of 2006-2015. When looking at the world index, it can be seen that there were several bull and bear markets in this period.
The figure shows a positive relationship between gold and the world index for the bull market of 2006-2007, and a negative relationship in the year 2008 in the midst of the financial crisis. This observation is consistent with findings in the existing literature that gold acts as a safe haven in crisis periods. There is a positive relationship again from 2009 on up until mid-2011. In the next bull market from mid-2011 to approximately mid-2015, there is a negative relationship again. Then, in the summer of 2015, the world index seems to decrease at the time the Chinese stock market crashed. At this point, both the gold price and world index seem to be declining.
10 Figure 2: World Index and Gold Price from 2006-2015
This figure shows how the world index and the price of gold in US dollars evolved over a period of ten years from 2006-2015. The world index is labeled on the left y-axis and the gold price on the right y-axis.
Figure 3: World Index and Gold Price in 2015
This figure shows how the world index and the price of gold in US dollars evolved in the year 2015. The world index is labeled on the left y-axis and the gold price on the right y-axis.
11
Figure 3 above zooms in at the year 2015 to take a closer look at what happened to the gold-world index relationship. This figure shows that the gold-world index starts to decline from June to October from a value of approximately 1770 to 1550, a decline of about 12.4%. In the same period, the gold price declines from about $1680 to $1650, or 1.8%. Thus, this graph shows some visual evidence of a positive relationship between the gold price and the world index during the 2015 Chinese stock market crash. This finding challenges the findings in the existing literature that gold acts as a safe haven in times of market stress. Of course, this is not a definitive conclusion as the hypotheses will be tested more formally in the next subsections.
3.4. Research method
This subsection presents the econometric model used to formally test whether gold has served as a safe-haven asset. For this analysis, the generalized autoregressive conditional heteroscedasticity (GARCH) model is used. This approach allows for modeling the time series to be observed. The GARCH model is used to forecast the time-varying volatility of returns of assets observed at high sampling frequencies (Stock & Watson, 2012, p. 703). The approach is similar to that from Baur and McDermott (2010). An additional feature is that this research will include the return on the U.S. dollar index as a control variable. This index measures the strength of the U.S. dollar against a basket of currencies. The reason to include this index is that the price of gold is denominated in U.S. dollars. A change in the strength of the U.S. dollar therefore affects the price of gold and should be included in the regression to account for omitted variable bias. An appreciating (depreciating) dollar makes gold more (less) expensive for foreign investors which decreases (increases) their demand and is thus expected to have a downward (upward) effect on the price of gold.
Figure 4 below depicts the volatility of the gold price in the sample period under investigation. The volatility is obtained by predicting the errors of the gold price return. The figure shows some evidence of clustering volatility. Furthermore, the Lagrange Multiplier test for autoregressive conditional heteroscedasticity in Table 2 shows a Chi-squared statistic significant at the 1% confidence level. This means that there is enough evidence to reject the null hypothesis of no ARCH effects in the data. These results justify the choice of using a GARCH model for this research.
12 Figure 4: Volatility Clustering
The volatility is obtained by predicting the errors of the gold price return. The figure shows some visual evidence of volatility clustering.
Table 2: Lagrange Multiplier test for autoregressive conditional heteroscedasticity
Lags Chi2 Degrees of Freedom Prob > Chi2
1 53.794 1 0.000***
H0: no ARCH effects vs. H1: ARCH(p) disturbance. *** indicates statistical significance at the 1% level. The null hypothesis is rejected at the 1% significance level.
Following up on existing literature, a GARCH(1,1) model is used. Higher-order GARCH(p,q) models are only useful when using a data set with more observations like several decades of daily data (Engle, 2001, pp. 157-168). Specifically, the following regressions will be jointly estimated using Maximum Likelihood:
𝑟𝑔𝑜𝑙𝑑,𝑡 = 𝛽𝑖+ 𝛽𝑖𝑟𝑀𝑆𝐶𝐼,𝑡+ 𝛽𝑖𝑟𝑈𝑆𝐷,𝑡+ 𝜀𝑡 (1)
𝛽𝑖 = 𝛾0,𝑖+ 𝛾1,𝑖𝐷(𝑟𝑀𝑆𝐶𝐼𝑝10) + 𝛾2,𝑖𝐷(𝑟𝑀𝑆𝐶𝐼𝑝5) (2)
𝜎𝑡2 = 𝜋 + 𝜂𝜀
𝑡−12 + 𝜙𝜎𝑡−12 (3)
Equation (1) estimates the relationship between gold and index returns. Equation (2) shows the 10th and 5th percentile dummies along with their interaction terms with the MSCI and Dollar Index return. The dummy variables D(…) refer to the 10th and 5th percentile of the stock index return distribution. These dummies are equal to 1 when the stock market exceeds these thresholds and 0 otherwise. Finally, Equation (3) estimates the time-varying volatility of returns to account
13
for heteroscedasticity.
To test the hypotheses presented in Subsection 3.1, the same approach as in the paper of Baur and McDermott (2010) is used. The first and second hypotheses are tested by looking whether the coefficients on the dummies (including 𝛾0) are negative and statistically significantly different from zero during the 2015 Chinese stock market crash. If this is the case, the first
hypothesis is accepted and gold can be seen as a strong safe-haven asset. When these coefficients are not significant (either positive or negative), then the second hypothesis is accepted and gold can be seen as a weak safe haven. To test the third hypothesis, time dummies are used to
investigate whether the correlation between the MSCI Index and gold increased during the 2015 crash. For this part Equation (2) is replaced by Equation (2b):
𝛽𝑖 = 𝛾0,𝑖+ 𝛾1,𝑖𝐶𝑟𝑎𝑠ℎ(𝑟𝑀𝑆𝐶𝐼) + 𝛾2,𝑖𝐶𝑟𝑖𝑠𝑖𝑠(𝑟𝑀𝑆𝐶𝐼) (2b)
where ‘crash’ and ‘crisis’ are time dummies that refer to the 2015 crash and the recent global financial crisis, respectively. The crash dummy equals 1 in the period 12 June 2015 to 29 September 2015 and the crisis dummy equals 1 from 12 October 2007 to 9 March 2009.
As mentioned before, this thesis also improves upon the existing literature by using 10-year German government bond (Bund) yields as a robustness test. For this part, the procedure is the same as the procedure described above, but instead the return on gold is replaced by the return on Bund yields. The choice for these bonds is justified by the fact that financial markets view Bunds as a safe haven (Bernoth & Erdogan, 2012, p. 640). Assuming both Bunds and gold are safe-haven assets, the estimated coefficient on Bunds must be of the opposite sign to that of the coefficient on gold. This is because Bund yields decrease when their demand increases (in times of declining stock markets). For the analysis, the return on these Bund yields is used. Thus, for Bunds to qualify as a strong (weak) safe haven they must have a positive (no) correlation with the MSCI Index, whereas gold must have a negative (no) correlation.
14
4. Results
In this section the estimation results are presented. In the first part the results of the entire sample are discussed. The second part shows the results of the different subsamples. Additional tests with time dummies are shown in the third part. The section concludes with some robustness tests using Bund yield return instead of gold return.
Table 3: Estimation results 01/2006-12/2015
Dependent variable: Gold return Period: 01/2006 – 12/2015 Full sample n = 2430
[1] [2] [3] [4] [5] [6] [7]
Conditional mean equation Regressors: MSCI return 0.1513 -0.0794 -0.0588 -0.0567 -0.0567 -0.0573 -0.0572 (0.000)*** (0.022)** (0.163) (0.175) (0.173) (0.139) (0.140) Dollar return -1.2590 -1.2588 -1.2509 -1.2536 -1.2515 -1.2510 (0.000)*** (0.000)*** (0.000)*** (0.000)*** (0.000)*** (0.000)*** MSCI return*pct10 -0.5430 -0.1278 -0.0588 (0.473) (0.311) (0.358) MSCI return*pct5 0.5114 -0.0530 -0.0774 (0.514) (0.784) (0.272) Dollar return*pct10 -0.0545 (0.888) Dollar return*pct5 0.1928 (0.758) pct10 -0.0076 -0.0015 (0.489) (0.533) pct5 0.0082 0.0007 (0.502) (0.901) Intercept 0.0003 0.0003 0.0002 0.0002 0.0002 0.0002 0.0002 (0.239) (0.136) (0.295) (0.301) (0.336) (0.279) (0.269) Conditional variance equation
ARCH L1. 0.0670 0.0743 0.0746 0.0742 0.0747 0.0744 0.0743 (0.001)*** (0.001)*** (0.001)*** (0.001)*** (0.001)*** (0.001)*** (0.001)*** GARCH L1. 0.9208 0.9132 0.9128 0.9133 0.9128 0.9130 0.9132 (0.000)*** (0.000)*** (0.000)*** (0.000)*** (0.000)*** (0.000)*** (0.000)*** Intercept 0.0000 0.0000 0.0000 0.0000 0.0000 0.0000 0.0000 (0.049)** (0.037)** (0.036)** (0.037)** (0.037)** (0.037)** (0.040)*** Summary statistics Wald Chi2 22.79 181.39 198.16 187.54 187.55 193.02 191.21 (0.000)*** (0.000)*** (0.000)*** (0.000)*** (0.000)*** (0.000)*** (0.000)*** Log pseudo-likelihood 7368.89 7474.36 7475.92 7475.38 7475.08 7475.49 7475.48
15
4.1. Full sample analysis
Table 3 above shows the estimation results of the entire sample. In the upper part the conditional mean equation is presented and the middle part shows the conditional variance equation.
Summary statistics can be found in the bottom part.
4.1.1. Conditional mean equation
The regression in column (1) of Table 3 shows the base specification. Column (2) shows the regression with the control variable Dollar Index return and columns (3) to (7) show the regressions of the gold return on the MSCI percentiles and their interaction terms. To avoid issues of multicollinearity, gold is regressed separately on each percentile with interaction term, (see Table 4 in Appendix A for the correlation matrix).
The output in column (1) gives the regression results of the gold return on the MSCI Index return. The coefficient on the MSCI return is positive (0.1513) and statistically significant at the 1% level. This positive coefficient means that gold and the MSCI Index co-move and contradicts findings in the existing literature that gold can serve as a hedge for stocks.
Column (2) includes the Dollar Index return as a control variable. The coefficient on the Dollar Index return is -1.259 and is statistically significant at the 1% level. This indicates that it is a necessary control variable. This coefficient stays roughly the same in all subsequent
regressions. Adding this variable also changes the sign on the MSCI return coefficient, which is now -0.0794 with a p-value of 2.2%. Thus, the results are now similar to findings in the existing literature that gold can serve as a hedge for stocks.
In the regression in column (3), all dummy variables and their interaction terms are added. As mentioned before, there is a high correlation (>0.7) between these variables. They are
therefore separated in the following regressions. The Dollar Index interaction term is removed in the subsequent regressions for the same reason (again, refer to Appendix A for the correlations).
The regression in column (4) displays the output for the 10th percentile dummy and its interaction term with the MSCI Index return. The coefficient on the MSCI is still negative (-0.0567) but insignificant. Nevertheless, it still indicates that gold can serve as a hedge for the world index, because there is no positive correlation. The coefficient on the interaction term is also negative (-0.1278) and insignificant. Thus, according to the definition used in this thesis, there is evidence that gold acted as a weak safe haven in this period. Removing the 10th percentile dummy in column (5) does not bring any notable changes to the coefficients or their p-values.
16
In columns (5) and (6) the analysis of the previous two columns is repeated, but now for the dummy of the 5th percentile. The regression in column (5) displays the output for the 5th percentile dummy and its interaction term with the MSCI Index return. The coefficients on the MSCI and the Dollar return stay roughly the same again. Removing the 5th percentile dummy from column (6) changes the p-value of the interaction term from 0.784 to 0.272. The coefficient on the interaction term of the 5th percentile in column (6) is negative (-0.0774) and insignificant, indicating that gold also acted as a weak safe haven for the 5 percent worst stock returns of the world index. Thus, for the entire sample, there is evidence that gold acted as a weak safe haven for both the 10 and 5 percent worst stock returns of the MSCI Index.
4.1.2. Conditional variance equation
The middle part of Table 3 shows the ARCH/GARCH specifications. Stock and Watson (2012) explain that a sum close to one indicates that changes in the conditional variance are persistent (2012, p. 704). Summing the ARCH and GARCH coefficients of each regression in Table 3 gives a number almost 0.99. Thus, the volatility of the gold return seems to be persistent.
4.2. Subsample analysis
In this subsection, the sample will be divided into four different subsamples. This is less than in the descriptive analysis of Subsection 3.3 to minimize the number of subsamples. Specifically, the following bull and bear markets are assumed: 01/2006 – 10/2007 bull market, 10/2007 – 03/2009 bear market, 03/2009 – 04/2014 bull market, and finally, 04/2014 – 12/2015 bear market. A summary of the regression results can be found in Table 5 below. The complete regression results can be found in Tables 6 to 9 in Appendix B.
Table 5: Summary subsample regressions
Dependent variable: Gold return
Bull market Bear market Bull market Bear market
01/2006 – 10/2007 10/2007 – 03/2009 03/2009 – 04/2014 04/2014 – 12/2015
n = 437 n = 339 n = 1231 n = 423
Regressors Coeff. est. p-value Coeff. est. p-value Coeff. est. p-value Coeff. est. p-value
MSCI return 0.0128 (0.894) -0.1793 (0.163) -0.0890 (0.145) -0.0601 (0.428)
Dollar return -1.5559 (0.000)*** -1.9335 (0.000)*** -1.3741 (0.000)*** -0.7433 (0.000)***
MSCI return*pct10 0.0226 (0.905) -0.0825 (0.619) -0.0466 (0.623) -0.0952 (0.365)
MSCI return*pct5 -0.2652 (0.157) -0.1141 (0.436) -0.0221 (0.831) -0.1415 (0.149) This table summarizes the most important regression coefficients for the subsamples. The full regressions can be seen in Tables 6 to 9 in Appendix B. The coefficients are statistically significant at the *10%, **5%, or ***1% level. The coefficients on the MSCI and Dollar Index return in each regression of every period are approximately the same, so the coefficients of the fourth regression of each period in Appendix B are displayed here.
17
The first period in Table 5 is the bull market period of 01/2006 – 10/2007. In this period the coefficient on the MSCI Index return has a positive sign and is highly insignificant. This
insignificance indicates that there was no clear relationship between the MSCI Index and the gold price in this period. This finding is again in line with findings in the existing literature that gold can serve as a hedge for stocks. The coefficient on the Dollar Index return is negative (-1.5559) and statistically significant at the 1% level. The coefficient on the interaction term of the MSCI return with the 10th percentile dummy has a positive value (0.0226), but is also insignificant. Thus, this indicates that for both the 10% and 5% worst stock returns, no significant relationship was found between the gold price and the MSCI Index. This leads to the conclusion that gold served as a weak safe haven for both the 10th and 5th percentile worst stock returns in this period.
The second period in Table 5 summarizes the coefficients of the regressions for the bear market period 10/2007 – 03/2009. This is the period in which the biggest stock market declines occurred during the global financial crisis. The coefficient on the MSCI Index return is now negative (-0.1793) and significant at the 1% level. Just like in the previous period, this finding is in line with the existing literature that gold can serve as a hedge for stocks, but now the case is stronger with a negative and significant coefficient. The coefficient on the Dollar Index return is again negative and statistically significant at the 1% level. In this bear market period, both the interaction term for the 10th and 5th percentile have negative coefficients (-0.0825 and -0.1141), though both values are insignificant. Thus, the conclusion for this period is the same as the previous one, namely that gold served as a weak safe-haven asset for both percentiles.
The third period is a relatively long period (03/2009 – 04/2014) where several short bull and bear markets are actually combined to total one long bull market. In this combined period, the coefficients on the interaction terms are again negative and insignificant. The last period in Table 5 is from 04/2014 to 12/2015 and includes the period of the Chinese stock market crash. The coefficients on the interaction terms again do not change signs and are still insignificant. This leads to the following conclusion. In every subsample analyzed, gold acted as a weak safe haven for both the 10and 5 percent worst stock returns on the world index. Furthermore, the hedge property of gold seems to be present in every period analyzed. An alternative analysis is presented in Table 10 at the end of Appendix B, where the percentile dummies are redefined for each subsample.
18
4.3 Time dummy tests
To test the hypothesis whether the correlation between the gold price and the MSCI World Index increased during the 2015 Chinese stock market crash compared to other periods, the time dummy ‘crash’ is used. This dummy equals 1 during the period 12 June 2015 to 29 September 2015, and 0 otherwise. This starting date is chosen as 12 June was the day the Chinese stock market started to decline (Riley & Yan, 2015). From Figure 1 in the introduction, it can be seen that the MSCI Index declined approximately 12.4% in this whole period. Additionally, the time dummy ‘crisis’ is added to the regression. This dummy equals 1 during the period of stock market declines in the global financial crisis, which is assumed to be from 12 October 2007 to 9 March 2009. The output of the time dummy regressions can be found in Table 11 below.
Column (1) shows the regression including all variables. The interaction term of the ‘crash’ dummy and the MSCI Index return is negative (-0.1477) and significant at the 10% level. This means that there is some evidence that the correlation between the MSCI Index and the gold price was different compared to other periods. In fact, this indicates that the safe-haven property of gold was stronger during the Chinese stock market crash compared to other periods. The interaction of the ‘crisis’ dummy with the MSCI Index return is also negative (-0.1573), but not significant. The interaction term of the crash dummy with the Dollar Index return is positive and insignificant. On the other hand, the interaction term for the crisis dummy with the Dollar Index return is negative and significant at the 10% level. Therefore, only the latter is kept in the subsequent regression. The insignificant crash and crisis dummy without interaction terms are also removed.
Column (2) shows the regression without the dummies and crisis interaction term with the Dollar Index. There are no notable changes in the estimated coefficients or their p-values.
Column (3) also excludes the insignificant crash dummy interaction term with the Dollar Index. The only notable change now is that the p-value of the coefficient on the interaction term of the crash dummy with the MSCI Index changes from 9% to 5.4%.
Columns (4) and (5) display the regressions where the crash dummy with its interaction term on the MSCI and the crisis dummy with its interaction terms on the Dollar and MSCI Index are separated. Still, the conclusions do not change as the coefficients and their p-values stay roughly equal as in the previous regressions. Thus, this analysis shows evidence that the safe-haven property of gold was more present during the 2015 Chinese stock market crash than during
19
the global financial crisis and compared to other periods. This means that the hypothesis the correlation between gold and the MSCI Index return increased during the 2015 crash compared to other periods cannot be rejected. In fact, there is evidence that the reverse is true.
Table 11: Estimation results with time dummies
Dependent variable: Gold return Period: 01/2006 – 12/2015 n = 2430
[1] [2] [3] [4] [5]
Conditional mean equation Regressors: MSCI return -0.0518 -0.0516 -0.0436 -0.0670 -0.0606 (0.167) (0.169) (0.237) (0.072)* (0.077)* Dollar return -1.1966 -1.1963 -1.1564 -1.2388 -1.1836 (0.000)*** (0.000)*** (0.000)*** (0.000)*** (0.000)*** MSCI return*crash -0.1477 -0.1365 -0.1517 -0.1302 (0.061)* (0.090)* (0.054)* (0.098)* Dollar return*crash 0.4577 0.4149 (0.233) (0.259) crash -0.0008 (0.357) MSCI return*crisis -0.1573 -0.1650 -0.1725 -0.1550 (0.216) (0.175) (0.157) (0.200) Dollar return*crisis -0.8991 -0.9238 -0.9633 -0.9346 (0.093)* (0.071)* (0.060)* (0.067)* crisis 0.0006 (0.695) Intercept 0.0003 0.0003 0.0003 0.0003 0.0003 (0.198) (0.213) (0.198) (0.161) (0.161)
Conditional variance equation
ARCH L1. 0.0714 0.0715 0.0719 0.0745 0.0722 (0.001)*** (0.001)*** (0.001)*** (0.001)*** (0.001)*** GARCH L1. 0.9159 0.9157 0.9154 0.9129 0.9152 (0.000)*** (0.000)*** (0.000)*** (0.000)*** (0.000)*** Intercept 0.0000 0.0000 0.0000 0.0000 0.0000 (0.049)** (0.049)** (0.047)** (0.038)** (0.045)** Summary statistics Wald Chi2 198.85 192.18 190.74 186.67 184.74 (0.000)*** (0.000)*** (0.000)*** (0.000)*** (0.000)*** Log pseudo-likelihood 7484.43 7480.88 7479.98 7475.41 7478.59
The p-values are given in parentheses. The coefficients are statistically significant at the *10%, **5%, or ***1% level. The variable ‘crash’ is a time dummy which equals 1 during the period 12 June 2015 to 29 September 2015, and 0 otherwise. The variable ‘crisis’ is a time dummy which equals 1 during the period 12 October 2007 to 9 March 2009, and 0 otherwise.
20
4.4. Robustness tests
In this subsection some robustness tests with Bund yield returns are presented. As explained before, the regression equations are similar to the equations in the previous subsections with Bund yield returns replacing the gold returns (the interaction terms with the Dollar Index return are omitted here, because they were insignificant for the gold case). Assuming Bunds are safe havens, the demand for Bunds increases in times of declining stock markets. When demand for Bunds increases, their yields decrease. As this research uses the return on those Bund yields, the relationship between the MSCI Index and Bund yield returns should be the opposite of the relationship between the MSCI Index and gold price returns, assuming both Bunds and gold are safe-haven assets. A summary of the full sample together with the subsamples can be found in Table 12 below. Please refer to Tables 13 to 17 in Appendix C for the complete regressions.
The first sample in Table 12 is the full sample of 2006-2015. The coefficient on the MSCI Index return is positive (0.3887) and statistically significant at the 1% level. This positive
coefficient is indeed the opposite of the negative sign on when gold return is used as the
dependent variable. Thus, there is enough evidence to conclude that Bunds can act as a (strong) hedge for the MSCI Index in the full sample. The coefficient on the Dollar Index return is also positive, but insignificant. Thus, while the dollar had a negative relationship with gold, it has no significant correlation with Bunds. The coefficients on the 10th and 5th percentile interaction terms are both positive but insignificant. This means there is enough evidence to conclude that Bunds acted as a weak safe haven for the MSCI Index in the full sample.
Table 12: Summary Bund yield regressions full- and subsamples
Dependent variable: Bund return
Full sample Bull market Bear market Bull market Bear market
2006 – 2015 01/2006 – 10/2007 10/2007 – 03/2009 03/2009 – 04/2014 04/2014 – 12/2015
n = 2430 n = 437 n = 339 n = 1231 n = 423
Regressors Coeff. est. p-value Coeff. est. p-value Coeff. est. p-value Coeff. est. p-value Coeff. est. p-value
MSCI return 0.3887 (0.000)*** 0.3639 (0.000)*** 0.2487 (0.001)*** 0.5686 (0.000)*** 0.8362 (0.023)**
Dollar return 0.1934 (0.197) 0.6302 (0.003)*** -0.2528 (0.324) 0.2925 (0.285) 1.4512 (0.134)
MSCI return*pct10 0.0021 (0.981) 0.0459 (0.729) -0.1634 (0.128) 0.1921 (0.216) 0.6496 (0.417)
MSCI return*pct5 0.0245 (0.796) -0.0138 (0.926) -0.0785 (0.447) 0.1736 (0.298) 0.9299 (0.228) This table summarizes the regression coefficients for the Bund yield. The full regressions can be seen in Table 13 to 17 in Appendix C. The coefficients are statistically significant at the *10%, **5%, or ***1% level. The coefficients on the MSCI and Dollar Index return in each regression of every period are approximately the same, so the coefficients of the fourth regression of each period in Appendix C are displayed here.
21
The second sample is the bull market period of 01/2006 – 10/2007. Like in the full sample, the coefficient on the MSCI Index return is again positive and significant at the 1% level. The coefficient on the Dollar Index return is also positive and now also significant at the 1% level. The coefficient on the 10th percentile interaction term is positive and insignificant; the coefficient on the 5th percentile interaction term is negative and insignificant. Because both percentiles have insignificant coefficients, it can be concluded that Bunds acted as a weak safe-haven asset for both the 10 and 5 percent worst stock returns in this bull market period. This result is the same as in the regression with the gold return.
The third sample is the bear market period of 10/2007 – 03/2009. The coefficient on the MSCI return changes from 0.3639 to 0.2487 and is still significant at the 1% level. The
coefficient on the Dollar Index return is now negative but insignificant. The coefficients on the 10th and 5th percentile interaction terms are now negative, but still insignificant. This means that also for this period, there is enough evidence to conclude that Bunds acted as a weak safe-haven asset for both percentiles. The conclusion for this period is again the same as in the regression with gold.
The fourth sample is the bull market period of 03/2009 – 04/2014. Again, the coefficient on the MSCI Index return is positive and significant at the 1% level. The coefficient on the Dollar Index is positive but insignificant. The coefficients on the 10th and 5th percentile interaction terms are positive and insignificant. This means that Bunds acted as a weak safe haven for both the 10th and 5th percentile and is again similar to the conclusion for the regression with gold as the dependent variable for this period.
The last sample is the bear market period of 04/2014 – 12/2015. The coefficient on the MSCI Index return is still positive, but now only significant at the 5% level. The coefficient on the Dollar Index return is again positive and insignificant. The coefficients on the 10th and 5th percentile interaction terms are again also positive and insignificant. Thus, for this bear market period, there is enough evidence to conclude that Bunds acted as a weak safe-haven asset for both the 10th and 5th percentile. The result for this period is yet again similar to the conclusion for the regression with gold as the dependent variable.
In summary, both gold and Bunds serve as a hedge in every sample analyzed. Moreover, in every sample, both assets consistently appear to be weak safe-haven assets for both the 10 and 5 percent worst stock returns of the MSCI Index. The finding that gold serves as a hedge for
22
stocks is in line with the findings of Baur and McDermott (2010), Coudert and Raymond-Feingold (2011), and Hood and Malik (2013). The finding that gold acts as a safe haven for stocks is in line with the findings of Baur and Lucey (2010), Coudert and Raymond-Feingold (2011), and Hood and Malik (2013).
4.4.1. Time dummies
To complete the analysis, this subsection uses the same time dummies as in Subsection 4.3, but now with the return on Bund yields as the dependent variable. The output can be found in Table 18 below.
The first column shows the regression which includes all variables. In the second and third regressions, the crash and crisis dummy, and the interaction term of the crash dummy with the Dollar Index return are removed. Doing so does not change the remaining coefficients or their p-values notably. The third column shows a positive but insignificant coefficient on the
interaction term of the crash dummy with the MSCI Index. This means that there is not enough evidence to conclude that the safe-haven property of Bunds was stronger during the period of the 2015 Chinese stock market crash. This contrasts the finding when using gold return as the dependent variable, where evidence was found that gold acted as a stronger safe haven in that period. The coefficient on the interaction term of the crisis dummy with the MSCI Index is surprisingly negative and significant at the 1% level. This indicates that Bunds were actually a worse safe-haven asset during the global financial crisis compared to other periods. The fourth and fifth columns show the regressions where the Bund is separately regressed on the crash and crisis dummy interaction terms. While the coefficients change slightly in magnitude, these regressions do not alter the conclusion.
Thus, this analysis shows that Bunds were not significantly a better safe haven during the 2015 Chinese stock market crash compared to other periods and even a worse safe haven during the global financial crisis. This contrasts the finding that gold acted as a better safe-haven asset during the 2015 crash compared to other periods and compared to the global financial crisis.
23
Table 18: Estimation results with time dummies and Bund return as dependent variable
Dependent variable: Bund return Period: 01/2006 – 12/2015 n = 2430
[1] [2] [3] [4] [5]
Conditional mean equation Regressors: MSCI return 0.5426 0.5433 0.5440 0.3855 0.5501 (0.000)*** (0.000)*** (0.000)*** (0.000)*** (0.000)*** Dollar return 0.4128 0.4117 0.4154 0.1869 0.4262 (0.011)** (0.011)** (0.011)** (0.205) (0.009)*** MSCI return*crash 1.0741 1.0698 1.0634 1.2228 (0.263) (0.265) (0.238) (0.167) Dollar return*crash 1.0548 1.0662 (0.700) (0.678) crash 0.0002 (0.982) MSCI return*crisis -0.3539 -0.3489 -0.3496 -0.3558 (0.000)*** (0.000)*** (0.000)*** (0.000)*** Dollar return*crisis -0.4635 -0.4544 -0.4579 -0.4695 (0.155) (0.167) (0.163) (0.153) crisis -0.0005 (0.592) Intercept -0.0001 -0.0002 -0.0002 0.0000 -0.0002 (0.756) (0.557) (0.559) (0.922) (0.547)
Conditional variance equation
ARCH L1. 0.1029 0.1026 0.1028 0.1026 0.1022 (0.000)*** (0.000)*** (0.000)*** (0.000)*** (0.000)*** GARCH L1. 0.9047 0.9049 0.9047 0.9051 0.9052 (0.000)*** (0.000)*** (0.000)*** (0.000)*** (0.000)*** Intercept 0.0000 0.0000 0.0000 0.0000 0.0000 (0.106) (0.106) (0.105) (0.104) (0.104) Summary statistics Wald Chi2 134.68 134.44 134.67 101.62 133.93 (0.000)*** (0.000)*** (0.000)*** (0.000)*** (0.000)*** Log pseudo-likelihood 5945.71 5945.50 5945.42 5933.41 5944.11
The p-values are given in parentheses. The coefficients are statistically significant at the *10%, **5%, or ***1% level. The variable ‘crash’ is a time dummy which equals 1 during the period 12 June 2015 to 29 September 2015, and 0 otherwise. The variable ‘crisis’ is a time dummy which equals 1 during the period 12 October 2007 to 9 March 2009, and 0 otherwise.
24
5. Discussion
This paper has investigated whether gold has lost its safe-haven status during the 2015 Chinese stock market crash. In their article, MacDonald and Shumsky (2015) imply that this was indeed the case. The findings of this thesis opposes this and even provides evidence that gold acted as a stronger safe-haven asset during the 2015 crash compared to the recent global financial crisis and other periods. The implication for investors and risk managers is that they can still rely on the traditional role of gold as a safe-haven asset in times of market stress.
The subsamples analyzed in this thesis all show the same result. Namely that gold can act as a weak safe-haven asset for the MSCI World Index. This finding is in line with the findings of Baur and Lucey (2010), Coudert and Raymond-Feingold (2011), and Hood and Malik (2013). Another finding of this thesis is that gold can act as an effective hedge for stocks. This finding is in line with the papers of Baur and McDermott (2010), Coudert and Raymond-Feingold (2011), and Hood and Malik (2013).
That fact that the coefficients on the percentile dummy interaction terms with the MSCI Index are insignificant means there is no evidence for a nonlinear relationship between gold and the MSCI Index. In other words, the analysis with the percentile dummies shows no evidence that the safe-haven effect of gold changes in times of declining stock markets.
Alternative regressions shown in Table 10 in Appendix B provide evidence that gold actually served as a strong safe-haven asset in bear market periods and as a weak safe haven in bull market periods, which is in line with the findings of Baur and McDermott (2010) and Beckmann et al. (2015). This conclusion should be taken with caution though, because in this analysis the percentile dummies are redefined for each period.
A shortcoming of this thesis is that it only uses a world index in the analysis. Baur and McDermott (2010), for example, make a distinction between different countries and find that gold is both a hedge for major European stock markets and the US, but not for Australia, Canada, Japan, and large emerging countries such as the BRIC countries (2010, p. 1886). Thus, although this thesis finds that gold is a safe haven for the MSCI Index in every sample analyzed, this might not be true on a country-by-country basis. It would be interesting to examine this in future
25
6. References
Baur, Dirk G., and Kristoffer J. Glover, 2012, The Destruction of a Safe Haven Asset?,
University of Technology, Sydney, SSRN: http://ssrn.com/abstract=2142283, 1-19. Baur, Dirk G., and Thomas K.J. McDermott, 2010, Is gold a safe haven? International evidence,
Journal of Banking & Finance 34, 1886-1898.
Baur, Dirk G., and Brian M. Lucey, 2010, Is Gold a Hedge or a Safe Haven? An Analysis of Stocks, Bonds and Gold, The Financial Review 45, 217-229.
Beckmann, Joscha, Theo Berger, and Robert Czudaj, 2015, Does gold act as a hedge or a safe haven for stocks? A smooth transition approach, Economic Modelling 48, 16-24. Bernoth, Kerstin, and Burcu Erdogan, 2012, Sovereign bond yield spreads: A time-varying
coefficient approach, Journal of International Money and Finance 31, 639-656.
Capie, Forrest, Terence C. Mills, and Geoffrey Wood, 2005, Gold as a hedge against the dollar,
Journal of International Financial Markets, Institutions & Money 15, 343-352.
Ciner, Cetin, Constantin Gurdgiev, and Brian M. Lucey, 2013, Hedges and safe havens: An examination of stocks, bonds, gold, oil and exchange rates, International Review of
Financial Analysis 29, 202-211.
Coudert, Virginie, and Hélène Raymond-Feingold, 2011, Gold and financial assets: Are there any safe havens in bear markets?, Economic Bulletin 31, 1613-1622.
Engle, Robert F., 2001, GARCH 101: The Use of ARCH/GARCH Model in Applied Econometrics, Journal of Economic Perspectives 15, 157-168.
Giovannini, Alberto, 1988, How Do Fixed-Exchange-Rates Regimes Work: The Evidence From The Gold Standard, Bretton Woods and The EMS, NBER Working Paper 2766, 1-53. Hood, Matthew, and Farooq Malik, 2013, Is gold the best hedge and a safe haven under changing
market volatility?, Review of Financial Economics 22, 47-52.
Joy, Mark, 2011, Gold and the US dollar: Hedge or haven?, Finance Research Letters 8, 120-131.
MacDonald, Alistair, and Tatyana Shumsky, 2015, Gold’s Role as Safe-Haven Investment Wanes, The Wall Street Journal. Retrieved from:
http://www.wsj.com/articles/golds-role-as-safe-haven-investment-wanes-1445250762. McCown, James R., and John R. Zimmerman, 2006, Is gold a zero-beta asset? Analysis of the
26
http://ssrn.com/abstract=920496.
Reboredo, Juan C., 2013a, Is gold a safe haven or a hedge for the US dollar? Implications for risk management, Journal of Banking & Finance 37, 2665-2676.
Reboredo, Juan C., 2013b, Is gold a hedge or a safe haven against oil price movements?,
Resources Policy 38 (2), 130-137.
Riley, Charles, and Sophia Yan, 2015, China’s stock market crash…in 2 minutes, CNN Money. Retrieved from: http://money.cnn.com/2015/07/09/investing/china-crash-in-two-minutes/. Stock, James H., and Mark M. Watson, 2012, Introduction to Econometrics, (Pearson, Harlow,
27
Appendix A
Table 4: Correlation matrix regression variables
goldret msciret dollarret msciret*pct10 msciret*pct5 dollarret*pct10 dollarret*pct5 pct10 pct5 goldret 1.0000 msciret 0.0745 1.0000 dollarret -0.5697 -0.5877 1.0000 msciret*pct10 0.0215 0.7025 -0.4063 1.0000 msciret*pct5 0.0080 0.6296 -0.3586 0.9117 1.0000 dollarret*pct10 -0.0882 -0.5192 0.5334 -0.7386 -0.6604 1.0000 dollarret*pct5 -0.0301 -0.5056 0.4410 -0.7326 -0.8034 0.8289 1.0000 pct10 -0.0216 -0.6223 0.3609 -0.8614 -0.6236 0.6414 0.4964 1.0000 pct5 0.0059 -0.5751 0.3238 -0.8203 -0.9041 0.5883 0.7198 0.6897 1.0000 This table shows the correlation matrix of the regression variables. A correlation greater than |0.7| indicates multicollinearity (Stock & Watson, 2012)
28
Appendix B
Table 6: Estimation results 01/2006 – 10/2007
Dependent variable: Gold return Period: 01/2006 – 10/2007 Bull market n = 437
[1] [2] [3] [4] [5] [6]
Conditional mean equation Regressors: MSCI return 0.2231 0.0191 0.0144 0.0128 0.0543 0.0581 (0.000)*** (0.812) (0.880) (0.894) (0.547) (0.518) Dollar return -1.5526 -1.5572 -1.5559 -1.5410 -1.5260 (0.000)*** (0.000)*** (0.000)*** (0.000)*** (0.000)*** MSCI return*pct10 -0.3132 0.0226 (0.614) (0.905) MSCI return*pct5 4.0537 -0.2652 (0.176) (0.157) pct10 -0.0057 (0.619) pct5 0.0935 (0.159) Intercept 0.0007 0.0004 0.0004 0.0005 0.0003 0.0003 (0.188) (0.392) (0.384) (0.394) (0.542) (0.526)
Conditional variance equation
ARCH L1. 0.0598 0.0911 0.0937 0.0914 0.0839 0.0914 (0.076)* (0.097)* (0.116) (0.102) (0.165) (0.113) GARCH L1. 0.9233 0.8757 0.8724 0.8753 0.8867 0.8758 (0.000)*** (0.000)*** (0.000)*** (0.000)*** (0.000)*** (0.000)*** Intercept 0.0000 0.0000 0.0000 0.0000 0.0000 0.0000 (0.380) (0.286) (0.293) (0.290) (0.371) (0.299) Summary statistics Wald Chi2 7.29 26.08 26.07 26.26 27.55 25.58 (0.007)*** (0.000)*** (0.000)*** (0.000)*** (0.000)*** (0.000)*** Log pseudo-likelihood 1332.50 1347.98 1348.12 1347.99 1349.93 1348.74
29 Table 7: Estimation results 10/2007 – 03/2009
Dependent variable: Gold return Period: 10/2007 – 03/2009 Bear market n = 339
[1] [2] [3] [4] [5] [6]
Conditional mean equation Regressors: MSCI return 0.0052 -0.2187 -0.1814 -0.1793 -0.1662 -0.1711 (0.948) (0.008)*** (0.161) (0.163) (0.143) (0.128) Dollar return -1.9167 -1.9389 -1.9335 -1.9017 -1.9366 (0.000)*** (0.000)*** (0.000)*** (0.000)*** (0.000)*** MSCI return*pct10 -0.1074 -0.0825 (0.575) (0.619) MSCI return*pct5 0.0116 -0.1141 (0.954) (0.436) pct10 -0.0009 (0.792) pct5 0.0052 (0.416) Intercept 0.0011 0.0010 0.0007 0.0007 0.0006 0.0007 (0.324) (0.310) (0.519) (0.544) (0.567) (0.482)
Conditional variance equation
ARCH L1. 0.0377 0.0478 0.0491 0.0492 0.0479 0.0490 (0.012)** (0.005)*** (0.003)*** (0.003)*** (0.005)*** (0.004)*** GARCH L1. 0.9362 0.9214 0.9186 0.9185 0.9186 0.9182 (0.000)*** (0.000)*** (0.000)*** (0.000)*** (0.000)*** (0.000)*** Intercept 0.0000 0.0000 0.0000 0.0000 0.0000 0.0000 (0.136) (0.135) (0.145) (0.145) (0.139) (0.137) Summary statistics Wald Chi2 0.00 36.90 37.60 37.17 42.60 37.68 (0.9481) (0.000)*** (0.000)*** (0.000)*** (0.000)*** (0.000)*** Log pseudo-likelihood 843.06 864.59 864.81 864.79 865.28 865.02
30 Table 8: Estimation results 03/2009 – 07/2014
Dependent variable: Gold return Period: 03/2009 – 07/2014 Bull market n = 1231
[1] [2] [3] [4] [5] [6]
Conditional mean equation Regressors: MSCI return 0.2103 -0.1063 -0.0891 -0.0890 -0.1012 -0.1003 (0.000)*** (0.051)* (0.143) (0.145) (0.072)* (0.082)* Dollar return -1.3768 -1.3745 -1.3741 -1.3773 -1.3753 (0.000)*** (0.000)*** (0.000)*** (0.000)*** (0.000)*** MSCI return*pct10 -0.0356 -0.0466 (0.887) (0.623) MSCI return*pct5 0.0467 -0.0221 (0.921) (0.831) pct10 0.0002 (0.961) pct5 0.0017 (0.881) Intercept 0.0003 0.0003 0.0002 0.0002 0.0003 0.0003 (0.340) (0.327) (0.547) (0.539) (0.408) (0.397)
Conditional variance equation
ARCH L1. 0.0746 0.0876 0.0881 0.0880 0.0882 0.0878 (0.028)*** (0.019)** (0.018)** (0.019)** (0.015)** (0.019)** GARCH L1. 0.8839 0.8746 0.8739 0.8739 0.8739 0.8743 (0.000)*** (0.000)*** (0.000)*** (0.000)*** (0.000)*** (0.000)*** Intercept 0.0000 0.0000 0.0000 0.0000 0.0000 0.0000 (0.119) (0.046)** (0.046)** (0.045)** (0.043)** (0.044)** Summary statistics Wald Chi2 25.75 107.56 110.77 110.50 111.63 111.06 (0.000)*** (0.000)*** (0.000)*** (0.000)*** (0.000)*** (0.000)*** Log pseudo-likelihood 3786.97 3841.03 3841.28 3841.28 3841.13 3841.08
31 Table 9: Estimation results 07/2014 – 12/2015
Dependent variable: Gold return Period: 07/2014 – 12/2015 Bear market n = 423
[1] [2] [3] [4] [5] [6]
Conditional mean equation Regressors: MSCI return -0.0182 -0.0949 -0.0596 -0.0601 -0.0585 -0.0601 (0.755) (0.104) (0.433) (0.428) (0.401) (0.388) Dollar return -0.7551 -0.7293 -0.7433 -0.7291 -0.7386 (0.000)*** (0.000)*** (0.000)*** (0.000)*** (0.000)*** MSCI return*pct10 -0.3284 -0.0952 (0.120) (0.365) MSCI return*pct5 -0.3635 -0.1415 (0.262) (0.149) pct10 -0.0050 (0.274) pct5 -0.0062 (0.481) Intercept -0.0004 -0.0001 -0.0002 -0.0002 -0.0002 -0.0002 (0.278) (0.719) (0.611) (0.559) (0.591) (0.591)
Conditional variance equation
ARCH L1. 0.0607 0.0776 0.0773 0.0793 0.0773 0.0770 (0.300) (0.286) (0.316) (0.274) (0.310) (0.311) GARCH L1. -0.4327 -0.3340 -0.3261 -0.3500 -0.3314 -0.3342 (0.011)** (0.112) (0.134) (0.112) (0.124) (0.123) Intercept 0.0001 0.0001 0.0001 0.0001 0.0001 0.0001 (0.000)*** (0.000)*** (0.000)*** (0.000)*** (0.000)*** (0.000)*** Summary statistics Wald Chi2 0.10 22.48 30.10 24.03 30.01 27.97 (0.7546) (0.000)*** (0.000)*** (0.000)*** (0.000)*** (0.000)*** Log pseudo-likelihood 1433.51 1447.70 1448.73 1448.10 1448.62 1448.45
32 Table 10: Alternative subsample regressions
Dependent variable: Gold return
Bull market Bear market Bull market Bear market
01/2006 – 10/2007 10/2007 – 03/2009 03/2009 – 04/2014 04/2014 – 12/2015
n = 437 n = 339 n = 1231 n = 423
Regressors
Coeff.
est. p-value Coeff. est. p-value Coeff. est. p-value Coeff. est. p-value
MSCI return 0.0191 (0.812) -0.2187 (0.008)*** -0.1063 (0.051)* -0.0949 (0.104)
Dollar return -1.5827 (0.000)*** -1.7667 (0.000)*** -1.2289 (0.000)*** -0.7071 (0.000)***
MSCI return*pct10 -0.0120 (0.932) -0.2625 (0.029)** -0.1128 (0.178) -0.1423 (0.064)*
MSCI return*pct5 0.0333 (0.834) -0.2034 (0.084)* -0.0584 (0.544) -0.1691 (0.025)** This table presents a summary of an alternative to the regressions for the subsamples in Table 5. In these regressions, the
percentile dummies are redefined for each period. The coefficients are statistically significant at the *10%, **5%, or ***1% level.
The first period of Table 10 above is a bull market period. The insignificant coefficients on the interaction terms indicate that gold acted as a weak safe haven for both the 10th and 5th percentile. The second period is a bear market period, in which both the interaction term for the 10th and 5th percentile are now significant at the 5% and 10%, respectively. This indicates that gold acted as a strong safe-haven asset during the global financial crisis. The third period again shows negative values for both the 10th and 5th percentile interaction terms, but now they are not significant. Thus, there is evidence that gold acted as a weak safe haven in this bull market period. The final period is a bear market with negative and significant coefficients, with p-values of 6.4% for the 10th percentile and 2.5% for the 5th percentile. This leads to the conclusion that gold also acted as a strong safe-haven asset in this bear market period.
Summarizing, this alternative specification, where the percentile dummies are redefined for each period, shows a different picture of the safe-haven property of gold. Specifically, gold appears to act as a weak safe-haven asset in bull market periods and a strong safe-haven asset in bear market periods. Baur and McDermott (2010) similarly find that gold was a strong safe haven for most developed markets during the peak of the recent financial crisis (2010, p. 1886). This finding is especially appealing to investors and risk managers, because it implies that gold acts as a stronger safe-haven when it is needed most. The findings are also in line with the findings of Beckmann et al. (2015), namely that gold generally serves as both a hedge and a safe haven, but that that ability depends on the specific economic environment under observation (2015, p. 23). One should remain cautious though in comparing these periods with each other, because the dummies are redefined for each period.
33
Appendix C
Table 13: Bund yield estimation results 01/2006 – 12/2015
Dependent variable: Bund return Period: 01/2006 – 12/2015 Full sample n = 2430
[1] [2] [3] [4] [5] [6]
Conditional mean equation Regressors: MSCI return 0.3591 0.3895 0.3846 0.3887 0.3786 0.3825 (0.000)*** (0.000)*** (0.000)*** (0.000)*** (0.000)*** (0.000)*** Dollar return 0.1935 0.1885 0.1934 0.1658 0.1914 (0.189) (0.197) (0.187) (0.246) (0.189) MSCI return*pct10 -0.1642 0.0021 (0.252) (0.981) MSCI return*pct5 -0.4064 0.0245 (0.029)** (0.796) pct10 -0.0040 (0.160) pct5 -0.0131 (0.019)** Intercept -0.0001 0.0000 0.0000 0.0000 0.0001 0.0000 (0.850) (0.912) (0.910) (0.928) (0.827) (0.978)
Conditional variance equation
ARCH L1. 0.1014 0.1020 0.1016 0.1020 0.1024 0.1022 (0.000)*** (0.000)*** (0.000)*** (0.000)*** (0.000)*** (0.000)*** GARCH L1. 0.9060 0.9056 0.9058 0.9056 0.9048 0.9054 (0.011)** (0.000)*** (0.000)*** (0.000)*** (0.000)*** (0.000)*** Intercept 0.0000 0.0000 0.0000 0.0000 0.0000 0.0000 (0.100)* (0.103) (0.101) (0.103) (0.094)* (0.103) Summary statistics Wald Chi2 99.03 100.06 144.90 105.00 117.69 105.55 (0.000)*** (0.000)*** (0.000)*** (0.000)*** (0.000)*** (0.000)*** Log pseudo-likelihood 5930.47 5931.67 5933.06 5931.67 5937.49 5931.73
34 Table 14: Bund yield estimation results 01/2006 – 10/2007
Dependent variable: Bund return Period: 01/2006 – 10/2007 Bull market n = 437
[1] [2] [3] [4] [5] [6]
Conditional mean equation Regressors: MSCI return 0.2829 0.3774 0.3640 0.3639 0.3778 0.3794 (0.000)*** (0.000)*** (0.000)*** (0.000)*** (0.000)*** (0.000)*** Dollar return 0.6323 0.6310 0.6302 0.6259 0.6337 (0.003)*** (0.003)*** (0.003)*** (0.004)*** (0.003)*** MSCI return*pct10 0.0068 0.0459 (0.986) (0.729) MSCI return*pct5 0.6866 -0.0138 (0.538) (0.926) pct10 -0.0007 (0.922) pct5 0.0150 (0.561) Intercept 0.0005 0.0006 0.0006 0.0006 0.0001 0.0006 (0.223) (0.146) (0.138) (0.138) (0.159) (0.156)
Conditional variance equation
ARCH L1. 0.0120 0.0287 0.0318 0.0319 0.0258 0.0284 (0.741) (0.581) (0.561) (0.561) (0.623) (0.589) GARCH L1. -0.7348 -0.2668 -0.2620 -0.2619 -0.2827 -0.2675 (0.127) (0.429) (0.426) (0.426) (0.423) (0.428) Intercept 0.0001 0.0001 0.0001 0.0001 0.0001 0.0001 (0.000)*** (0.001)*** (0.000)*** (0.001)*** (0.001)*** (0.001)*** Summary statistics Wald Chi2 25.42 38.43 40.40 38.72 55.41 39.12 (0.000)*** (0.000)*** (0.000)*** (0.000)*** (0.000)*** (0.000)*** Log pseudo-likelihood 1468.03 1472.46 1472.52 1472.52 1472.55 1472.46
35 Table 15: Bund yield estimation results 10/2007 – 03/2009
Dependent variable: Bund return Period: 10/2007 – 03/2009 Bear market n = 339
[1] [2] [3] [4] [5] [6]
Conditional mean equation Regressors: MSCI return 0.1965 0.1684 0.2554 0.2487 0.2004 0.2022 (0.000)*** (0.008)*** (0.001)*** (0.001)*** (0.004)*** (0.004)*** Dollar return -0.2291 -0.2339 -0.2528 -0.2561 -0.2426 (0.361) (0.375) (0.324) (0.328) (0.344) MSCI return*pct10 -0.0865 -0.1634 (0.544) (0.128) MSCI return*pct5 -0.1354 -0.0785 (0.409) (0.447) pct10 0.0031 (0.349) pct5 -0.0025 (0.631) Intercept -0.0003 -0.0004 -0.0012 -0.0010 -0.0005 -0.0006 (0.651) (0.623) (0.155) (0.230) (0.518) (0.460)
Conditional variance equation
ARCH L1. 0.0273 0.0255 0.0256 0.0259 0.0258 0.0254 (0.001)*** (0.001)*** (0.001)*** (0.001)*** (0.001)*** (0.001)*** GARCH L1. 0.9807 0.9832 0.9837 0.9832 0.9830 0.9836 (0.000)*** (0.000)*** (0.000)*** (0.000)*** (0.000)*** (0.000)*** Intercept 0.0000 0.0000 0.0000 0.0000 0.0000 0.0000 (0.686) (0.610) (0.573) (0.579) (0.595) (0.578) Summary statistics Wald Chi2 15.27 18.03 21.37 20.77 18.89 18.77 (0.000)*** (0.000)*** (0.000)*** (0.000)*** (0.001)*** (0.000)*** Log pseudo-likelihood 940.92 941.47 943.32 942.88 941.94 941.84
36 Table 16: Bund yield estimation results 03/2009 – 04/2014
Dependent variable: Bund return Period: 03/2009 – 04/2014 Bull market n = 1231
[1] [2] [3] [4] [5] [6]
Conditional mean equation Regressors: MSCI return 0.5798 0.6542 0.5674 0.5686 0.5881 0.5910 (0.000)*** (0.000)*** (0.000)*** (0.000)*** (0.000)*** (0.000)*** Dollar return 0.3229 0.2610 0.2925 0.2836 0.2879 (0.252) (0.375) (0.285) (0.290) (0.292) MSCI return*pct10 -0.2020 0.1921 (0.595) (0.216) MSCI return*pct5 -0.7778 0.1736 (0.172) (0.298) pct10 -0.0087 (0.209) pct5 -0.0240 (0.075)* Intercept -0.0007 -0.0007 -0.0002 -0.0003 -0.0004 -0.0005 (0.172) (0.166) (0.698) (0.588) (0.446) (0.383)
Conditional variance equation
ARCH L1. 0.0702 0.0704 0.0706 0.0150 0.0718 0.0719 (0.000)*** (0.000)*** (0.000)*** (0.000)*** (0.000)*** (0.000)*** GARCH L1. 0.9239 0.9237 0.9239 0.9230 0.9228 0.9226 (0.000)*** (0.000)*** (0.000)*** (0.000)*** (0.000)*** (0.000)*** Intercept 0.0000 0.0000 0.0000 0.0000 0.0000 0.0000 (0.088)* (0.087)* (0.088)* (0.087)* (0.086)* (0.086)* Summary statistics Wald Chi2 77.23 77.52 95.38 78.49 87.95 79.97 (0.000)*** (0.000)*** (0.000)*** (0.000)*** (0.000)*** (0.000)*** Log pseudo-likelihood 2982.14 2983.16 2985.75 2984.51 2987.43 2984.23
37 Table 17: Bund yield estimation results 04/2014 – 12/2015
Dependent variable: Bund return Period: 04/2014 – 12/2015 Bear market n = 423
[1] [2] [3] [4] [5] [6]
Conditional mean equation Regressors: MSCI return 0.9871 1.1020 0.9237 0.8362 0.8690 0.7955 (0.050)** (0.011)** (0.014)** (0.023)** (0.006)*** (0.014)** Dollar return 1.5365 1.3848 1.4512 1.3535 1.3225 (0.139) (0.164) (0.134) (0.168) (0.166) MSCI return*pct10 -3.1823 0.6496 (0.015)** (0.417) MSCI return*pct5 -7.0603 0.9299 (0.001)*** (0.228) pct10 -0.0510 (0.062)* pct5 -0.1740 (0.002)*** Intercept -0.0052 -0.0057 -0.0053 -0.0053 -0.0055 -0.0055 (0.009)*** (0.002)*** (0.001)*** (0.002)*** (0.001)*** (0.001)*** Conditional variance equation
ARCH L1. 0.4584 0.4534 0.4838 0.5150 0.5302 0.5104 (0.116) (0.085)* (0.069)* (0.037)** (0.016)** (0.017)** GARCH L1. 0.6569 0.6604 0.6413 0.6208 0.6101 0.6264 (0.000)*** (0.000)*** (0.000)*** (0.000)*** (0.000)*** (0.000)*** Intercept 0.0001 0.0001 0.0001 0.0001 0.0001 0.0001 (0.406) (0.409) (0.369) (0.260) (0.223) (0.233) Summary statistics Wald Chi2 3.83 6.79 11.38 14.39 17.62 14.04 (0.050)** (0.034)** (0.023)** (0.002)*** (0.002)*** (0.003)*** Log pseudo-likelihood 600.55 602.50 603.71 603.13 607.28 603.94
