Reassessing the role of precious metals as safe havens–What colour is your haven and why?

Journal of Commodity Markets xxx (xxxx) xxx–xxx

Contents lists available at ScienceDirect

Journal of Commodity Markets journal homepage: www.elsevier.com/locate/jcomm

Reassessing the role of precious metals as safe havens–What colour is your haven and why?☆ ⁎

Sile Li , Brian M. Lucey Trinity Business School, Trinity College Dublin, Dublin 2, Ireland

A R T I C L E I N F O

ABSTRACT

JEL classification: G10 G19

The role of gold as a safe haven asset has been extensively studied in recent years. This article extends previous literature and examines time varying safe haven properties versus equities and bonds of four precious metals (gold, silver, platinum and palladium) across eleven countries. Results suggest that the metals each play safe haven roles; there are times when one metal is not while another may be a safe haven against an asset. The second part of this article attempts to identify robust economic and political determinants of precious metals safe haven properties applying zero-inflated Poisson regression (ZIP) and extreme bound analysis (EBA). Economic Policy Uncertainty is found to be a positive and robust determinant of a precious metal being a safe haven. This holds across countries. Stock volatility, exchange rates, interest rate and credit spreads are also found to be significant, but results are quite mixed for different markets and are fragile of model specification.

Keywords: Safe Haven Precious metals Gold Silver Platinum Palladium Market stress Zero-inflated Poisson Extreme bound analysis.

1. Introduction The precious metals market has attracted many studies in the last half decade. There are a number of studies focusing on the role of gold as a hedge in portfolio diversification, starting from Jaffe (1989) and Chua et al. (1990). Gold as a safe haven has also been examined (see as examples Baur and Lucey (2010); Baur and McDermott (2010); Coudert and Raymond (2010) and Beckmann et al. (2015) with a general finding that it can act as such. The safe haven status of other precious metals is less well studied. It is well established, in for example Hillier et al. (2006) that gold, silver and platinum all have low correlations with stock index returns, particularly during periods of high stock market volatility; therefore, portfolios that contain precious metals perform better than equity portfolios which do not. The term safe haven often refers to the assets suggested to investors to ”park their money” during periods of market stress. A very large number of assets have been suggested as safe havens at various times in various studies (see Lucey and Li, 2015). Baur and Lucey (2010) provides the first operational definition of a safe haven which refers to an asset (e.g. stocks) that is uncorrelated or negatively correlated with another asset or portfolio in times of market stress (e.g. stock market crash or political disturbance). Assets which are hedges may not necessarily be safe havens. This concept is further developed into weak and strong safe havens in Baur and McDermott (2010). This paper evaluates the safe haven nature of the four main precious metals across a variety of countries (in local currency), shows how these vary across time, and provides for the first time evidence of the determinants of a precious metal being a safe haven or otherwise.

☆ The authors would like to thank great feedback received from Robert Czudaj, Jean-Louis Bertrand, Montfort Mlachila, Juhani Raatikainen and all participants at INFINITI 2015 and IFABS 2016 conferences. ⁎ Corresponding author. E-mail addresses: [email protected] (S. Li), [email protected] (B.M. Lucey).

http://dx.doi.org/10.1016/j.jcomm.2017.05.003 Received 27 January 2017; Received in revised form 15 May 2017; Accepted 20 May 2017 2405-8513/ © 2017 Published by Elsevier B.V.

Please cite this article as: Li, S., Journal of Commodity Markets (2017), http://dx.doi.org/10.1016/j.jcomm.2017.05.003

Journal of Commodity Markets xxx (xxxx) xxx–xxx

S. Li, B.M. Lucey

2. Safe haven evidence for precious metals 2.1. Model and data We mainly follow the standard approach to test for a safe haven, as outlined in Baur and Lucey (2010) and Baur and McDermott (2010). In identifying a daily safe haven status for metals, however, we apply a minor modification on the final safe haven dummy as explained below. For example, to test gold safe haven properties during stock market crashes, the following equations are estimated using OLS.

rgold , t = α + βt rstock , t + ϵt

(1)

βt = C0 + C1 D (rstock q5) + C2 D (rstock q2.5 ) + C3 D (rstock q1)

(2)

Eq. (1) models the relation of gold and stock returns. The parameter βt is modeled as a dynamic process by Eq. (2). The dummy variables denoted as D () capture extreme stock market movements and

⎧1 if rstock , t < rstock qx D (rstock qx ) = ⎨ ⎩ 0 if rstock , t ≥ rstock qx

(3)

rstock qx = x% threshold given by the 5%, 2.5% and 1% quantile of the return distribution over the full sample period. Please see Table 2 for the values. The decision rules follow Baur and McDermott (2010) that:

• •

if all coefficients, including intercept C0 in (2), are negative, then gold is a weak safe haven; if all coefficients, including intercept C0, are negative and significant at 10% level, then gold is a strong safe haven.

In some cases when the daily return falls under the 5% quantile but above 2.5% and 1%, hence D (rstock q2.5 ) and D (rstock q1) are zeros, the above rule only applies to C0 and C1. In order to obtain daily safe haven status, we made a few minor modifications.

• • •

first, we run the above regressions over a quarter (i.e. 62 days1). If the results meet the above requirement, we assign a preliminary safe haven dummy value as 1 to each day within this quarter; next, we sum up all these preliminary safe haven dummy results over the whole sample, by which accumulating results for each day from all the regressions containing this day, approximately 62 regressions on average. The result is a count falling within the range of [0, 62]; finally, for those days in which the market return is not in the lower 5% quantile, the safe haven result is modified to 0. Hence the existence of a safe haven is aligned with exact dates of market stress.

We examine the safe haven properties of gold versus equity market movements across a wide variety of countries: United States S & P500 index, United Kingdom FTSE100 index, Germany DAX30 index, France CAC40 Index, Italy FTSE MIB Index, Switzerland Swiss Market Index (SMI), Canada S & P/TSX Composite Index, Japan NIKKEI225 index, China Shanghai Stock Exchange Composite Index (SHCOMP), Indian NIFTY50 Index, South Africa FTSE/JSE Africa Top40 Index and the benchmark 10-year government bond indices of each country respectively.2 Our precious metal data are spot market prices in dollar for US, and in local currency for other countries. Prices in local currency are calculated by using the daily spot exchange rate against US dollar. We thus assume an unhedged investor. We cover the period from January 1994 to July 2016 for most countries. The descriptive statistics of the data are summarized in Table 1. Table 2 and Figs. 1–6. This paper expands the Lucey and Li (2015) analysis of time-varying safe haven status of gold, silver, platinum and palladium against US stock and bond indices to eleven countries. We firstly obtain daily weak and strong safe haven results. Since strong safe haven status is extremely rare, an interesting finding in and of itself, we conduct the explanatory analysis only on weak safe haven results in the second part. For ease of exposition we show the accumulated daily and final safe haven results as a series of graphs in Appendix A (See Figures Appendix A.1–11). In these graphs we highlight on what day we find each precious metal to be a weak or strong safe haven against bonds or equities in each country. In addition, we calculate the average percentage of time when a precious metal is as safe haven for each country (see Table 3 and 4.) We also show the number of daily incidents when precious metals act as weak and strong safe havens across countries (see Figs. 7 and 8) and list the top ten days along with market news on that day (see Table 5). 2.2. Results of safe haven analysis We show in Table 3 and 4 the average weak and strong safe haveness for each metal in each country, which refers to the proportion of time when a metal functions as a weak or strong safe haven over the examined time frame. As Table 3 shows, precious 1 2

The average number of business days of a quarter is 62 for our sample. all data are sourced from Bloomberg

2

Journal of Commodity Markets xxx (xxxx) xxx–xxx

S. Li, B.M. Lucey

Table 1 Summary of data for safe haven analysis. Country

Data

Mean

S.D.

Kurt.

Skew.

Min.

Max.

Obs.

US

AUUSD AGUSD PTUSD PDUSD SP500 US10YR

726.32 12.20 936.78 408.32 1213.70 267.98

466.59 9.03 485.58 230.45 407.73 105.20

−0.89 0.93 −1.10 −0.98 −0.11 −1.19

0.74 1.28 0.40 0.56 0.27 0.29

252.55 4.06 335.00 114.75 438.92 111.93

1897.46 48.44 2250.50 1110.50 2163.75 477.14

5564 5564 5564 5564 5564 5564

UK

AUGBP AGGBP PTGBP PDGBP GBP FTSE100 UK10YR

449.28 7.50 569.28 254.35 1.64 5294.94 347.30

300.43 5.68 288.97 152.25 0.15 1076.94 141.78

−0.87 0.84 −1.16 −0.93 0.34 −0.85 −1.02

0.81 1.31 0.43 0.63 0.95 −0.50 0.36

157.19 2.51 196.63 68.12 1.29 2876.60 131.83

1178.93 29.13 1186.96 763.02 2.11 7103.98 655.21

5628 5628 5628 5628 5628 5628 5628

Germany

AUEUR AGEUR PTEUR PDEUR EUR G10YR DAX

580.71 9.68 750.82 344.25 1.22 292.31 5693.76

345.01 6.49 326.58 206.99 0.16 113.69 2356.44

−0.87 0.59 −1.24 0.35 −0.31 −0.93 −0.33

0.81 1.22 0.11 0.89 −0.39 0.41 0.43

237.70 3.35 280.62 88.99 0.83 123.16 1911.70

1378.11 32.68 1474.29 1201.32 1.60 535.37 12374.73

5611 5611 5611 5611 5611 5611 5611

France

AUEUR AGEUR PTEUR PDEUR EUR F10YR CAC

582.39 9.71 752.41 344.81 1.22 313.35 3893.41

345.24 6.50 326.30 207.11 0.16 125.41 1134.37

−0.89 0.57 −1.24 0.33 −0.31 −0.78 −0.40

0.80 1.21 0.10 0.88 −0.40 0.45 0.09

237.70 3.35 280.62 88.99 0.83 124.48 1721.14

1378.11 32.68 1474.29 1201.32 1.60 594.65 6922.33

5620 5620 5620 5620 5620 5620 5620

Italy

AUEUR AGEUR PTEUR PDEUR EUR IT10YR FTSEMIB

639.87 10.84 840.03 391.65 1.21 27702.84 482.31

352.17 6.57 288.56 197.99 0.17 9044.54 156.82

−1.23 0.12 −0.97 0.37 −0.59 −0.93 −0.18

0.52 1.01 −0.09 0.81 −0.32 0.38 0.76

237.70 3.80 280.62 123.09 0.83 12362.51 260.83

1378.11 32.68 1474.29 1201.32 1.60 50109.00 863.70

4628 4628 4628 4628 4628 4628 4628

Switzerland

AUCHF AGCHF PTCHF PDCHF CHF SMI SWISS10YR

793.52 13.23 1058.46 479.85 1.23 6386.90 192.04

373.90 7.41 415.20 270.83 0.24 1786.84 54.34

−0.99 0.53 −0.74 3.22 −0.78 −0.46 −1.00

0.65 1.14 0.27 1.48 0.35 −0.50 0.21

380.20 5.28 447.03 143.40 0.72 2450.30 98.10

1668.72 42.33 2363.01 1833.44 1.82 9531.46 301.42

5594 5594 5594 5594 5594 5594 5594

Canada

AUCAD AGCAD PTCAD PDCAD CAD SPTSX CAN10YR

845.27 14.00 1098.52 499.05 1.26 313.80 9732.81

450.07 8.42 428.29 273.41 0.19 128.68 3320.62

−1.10 0.49 −1.17 0.50 −1.31 −1.14 −1.30

0.71 1.15 0.10 0.90 −0.02 0.25 −0.08

375.55 5.80 470.51 153.26 0.92 113.17 3959.85

1879.43 46.05 2215.39 1670.53 1.61 571.33 15657.63

5575 5575 5575 5575 5575 5575 5575

Japan

AUYEN AGYEN PTYEN PDYEN YEN NIKKEI225 JP10YR

73609.82 1222.10 96644.31 42898.78 107.33 14288.74 201.77

40902.94 745.53 44347.49 23443.52 14.37 3933.61 40.76

−1.32 −0.49 −0.83 −0.22 −0.36 −1.21 −0.74

0.54 0.78 0.30 0.71 −0.33 0.08 −0.17

26980.84 411.56 36028.74 12778.74 75.82 7054.98 117.61

157087.65 3950.11 234097.01 130128.39 147.26 22666.80 280.59

5455 5455 5455 5455 5455 5455 5455

China

AUCNY AGCNY PTCNY PDCNY CNY

6555.29 112.79 8788.33 3082.26 7.11

2386.83 53.32 2058.71 1242.48 0.80

−0.97 0.75 0.14 −1.25 −1.49

0.18 1.04 0.54 0.24 0.33

2671.47 36.42 5011.72 1131.96 6.04

12145.45 2907 314.96 2907 15996.55 2907 5632.73 2907 8.28 2907 (continued on next page)

3

Journal of Commodity Markets xxx (xxxx) xxx–xxx

S. Li, B.M. Lucey

Table 1 (continued) Country

Data

Mean

S.D.

Kurt.

Skew.

Min.

Max.

Obs.

CN10YR SHCOMP

2456.67 99.79

997.72 6.42

1.00 −0.46

0.98 −0.31

1011.50 84.66

6092.06 113.16

2907 2907

India

AUINR AGINR PTINR PDINR INR NIFTY50 IND10YR

49657.93 807.30 59395.02 25842.97 51.38 4634.63 188.20

27601.62 475.94 20415.34 14219.78 8.46 2465.06 77.67

−1.61 −0.82 −1.17 −1.37 −0.98 −1.20 0.51

0.05 0.44 −0.22 0.32 0.68 0.02 −0.83

11917.57 194.78 20031.65 7079.35 39.27 854.20 7.27

97924.95 2152.57 105825.32 54983.18 68.83 8996.25 328.28

4120 4120 4120 4120 4120 4120 4120

South Africa

AUZAR AGZAR PTZAR PDZAR ZAR JSE40 SA10YR

7996.554 133.9439 9950.66 4417.256 8.576227 24445.77 158.7283

4763.54 81.09 3801.90 2450.75 2.03 12239.71 44.02

−1.37 −1.24 −1.29 −0.85 1.58 −0.94 −1.08

0.27 0.32 −0.16 0.57 1.27 0.27 0.51

1699.15 30.96 2507.51 1043.13 5.65 5942.88 98.12

19780.78 327.51 17582.03 9930.14 16.87 49081.01 248.23

3583 3583 3583 3583 3583 3583 3583

Fig. 1. Stock Market Return. Note that data for China and South Africa are not continuous. There is missing data of government 10 year bond indices for both countries, which are from October 7, 2010 to December 1, 2010 for China, and from August 29, 2003 to January 3, 2005 for South Africa.

metals could provide weak safe haven hedge about 33% of the time against stocks and 31% of the time against bonds, when the asset return fall is in the lower 5% region. On average, gold is not the most frequent safe haven across eleven countries. In general, precious metals perform better as weak safe haven in the US, Germany, Italy, Switzerland and South Africa. In the US, silver is the best safe haven against S & P500 and bond market falls (50% and 28% of the time respectively), followed by palladium (35% against stock and 19% against bond), gold (31% against stock and 26% against bond) and platinum (26% against stock and 24% against bond). However gold is the best safe haven in UK, against both stock and bond market events, 34% and 36% of the time respectively, followed by the others. In Europe, precious metals as a safe haven against stock market events is slightly more common in Germany and Italy than Switzerland and France; while against bond market 5% drawdowns, it is more common for France and Italy than Germany and Switzerland. Gold is the most common safe haven against stock market in Germany and Italy, but not in France and Switzerland. Gold is more usual than the other three metals as safe haven against Canadian bond market declines (42% of the time), while silver is more common against Canadian stock market falls (32% of the time). Gold is the most common against Nikkei225 index falls (36% of the time) but the least common against Japanese bond index declines (30% of the time). Results for China are at the average level across countries. Interestingly gold is the least common safe haven for both stock and bond market declines in China. Gold is a safe haven against India stock market declines 30% of the time, less frequent than palladium and platinum, however it is more likely than the others to hedge bond falls. Precious metals are much more a safe haven against South African stock market falls (43% of the time), than bond market (only 25% of the time). Gold is not as frequently a haven as are

4

Journal of Commodity Markets xxx (xxxx) xxx–xxx

S. Li, B.M. Lucey

Fig. 2. Bond Market Return. Note that data for China and South Africa are not continuous. There is missing data of government 10 year bond indices for both countries, which are from October 7, 2010 to December 1, 2010 for China, and from August 29, 2003 to January 3, 2005 for South Africa.

Fig. 3. Gold Return in Local Currency. Note that data for China and South Africa are not continuous. There is missing data of government 10 year bond indices for both countries, which are from October 7, 2010 to December 1, 2010 for China, and from August 29, 2003 to January 3, 2005 for South Africa.

platinum and palladium in South Africa. Strong safe haven results are very sparse across all countries as shown in Table 4, on average less than 1% of the time. In sum, precious metals provided protection at different times across countries. No precious metal provides safe haven status consistently, over time or across asset classes. Strong safe haven status is rare. Given these findings, we now turn to explanation for why safe haven status does or does not occour. 3. Economic determinants of precious metals safe haveness 3.1. Market distress The safe haven dummies are calculated, by definition, via examining time-varying relationships between precious metals and equity or bond market when these markets show major drawdowns. However a market fall could be triggered by a variety of different 5

Journal of Commodity Markets xxx (xxxx) xxx–xxx

S. Li, B.M. Lucey

Fig. 4. Silver Return in Local Currency. Note that data for China and South Africa are not continuous. There is missing data of government 10 year bond indices for both countries, which are from October 7, 2010 to December 1, 2010 for China, and from August 29, 2003 to January 3, 2005 for South Africa.

Fig. 5. Platinum Return in Local Currency. Note that data for China and South Africa are not continuous. There is missing data of government 10 year bond indices for both countries, which are from October 7, 2010 to December 1, 2010 for China, and from August 29, 2003 to January 3, 2005 for South Africa.

incidents such as political turmoil or a major corporate bankruptcy, and may be associated with characteristics such as liquidity flights or increasing risk aversion. It seems reasonable then to examine the extent to which stress indicators may be operating around these falls. Generally there are three types of market stress indicators: financial market stress, political stress and consumer sentiment. Financial crises are no longer seen as binary - crisis/no crisis - events. Financial Stress indices (FSI) have bee created to address that point, yielding a continuous variable with a spectrum of values where extreme values are called a crisis. One of the earliest indices is Illing and Liu (2006)'s FSI for Canada which combines stress measures of the banking sector, foreign exchange market, debt market and equity market. Their methodology was later adapted and further developed by Federal Reserve Bank of Kansas City (Hakkio and Keeton, 2009), International Monetary Fund (Cardarelli et al., 2011), by Balakrishnan et al. (2011) for emerging economies,by Corbet and Twomey (2014) for the UK and by Bjørnskov and Dreher (2008) for Germany. The impact of political tension on the gold price is examined back in 1983 by Ariovich (1983). They find that there is a positive relationship between political tension and the gold price. Here proxy political tension by the Economic Policy Uncertainty (EPU) 6

Journal of Commodity Markets xxx (xxxx) xxx–xxx

S. Li, B.M. Lucey

Fig. 6. Palladium Return in Local Currency. Note that data for China and South Africa are not continuous. There is missing data of government 10 year bond indices for both countries, which are from October 7, 2010 to December 1, 2010 for China, and from August 29, 2003 to January 3, 2005 for South Africa. Table 2 Values of 1%, 2.5% and 5% quantile in market returns.

US Bond US Stock UK Bond UK Stock Germany Bond Germany Stock French Bond French Stock Italian Bond Italian Stock Switzerland Bond Switzerland Stock Canadian Bond Canadian Stock Japanese Bond Japanese Stock Chinese Bond Chinese Stock Indian Bond Indian Stock South Africa Bond South Africa Stock

1% Quantile

2.5% Quantile

5% Quantile

−1.1% −3.3% −0.9% −3.2% −0.9% −4.4% −0.9% −4.1% −1.1% −4.8% −0.6% −3.6% −0.9% −3.2% −0.7% −4.1% −1.1% −5.5% −1.2% −4.2% −1.2% −3.8%

−0.9% −2.4% −0.7% −2.5% −0.6% −3.2% −0.6% −3.0% −0.8% −3.5% −0.5% −2.5% −0.7% −2.3% −0.5% −3.1% −0.7% −4.0% −0.8% −3.3% −0.8% −3.0%

−0.7% −1.8% −0.6% −1.9% −0.5% −2.4% −0.5% −2.3% −0.6% −2.6% −0.4% −1.9% −0.5% −1.7% −0.4% −2.4% −0.4% −2.9% −0.5% −2.3% −0.6% −2.2%

Index published by Baker et al. (2013). This index is based on newspaper coverage frequency as a proxy for movements in policyrelated economic uncertainty. Some studies define safe haven properties as when investors risk aversion suddenly and sharply increased which fundamentally change the transmission channel of risk between assets such as Vayanos (2004) and Baur and McDermott (2012). Baur and Glover (2012) argues that the investor behavior has the potential to undermine and possibly destroy the safe haven property of gold via increasing holding of it when investors face uncertainties. Measurements of consumer sentiment include the monthly Michigan Index of Consumer Sentiment (ICS) for US, the European Economic Sentiment Indicator published by the European Commission for UK and Germany, and the Zew Indicator of Economic Sentiment for Japan. In addition to market stress indicators, we also examine if key economic variables are associated with a high probability of a precious metal functioning as safe haven. A comprehensive discussion of the economic issues involved in gold and hypothesised to influence gold prices is contained in Connor et al. (2015), and we thus merely note the series here. Precious metals are believed to be leading indicators of inflationary expectations by many including Frankel (1986) and Taylor

7

Journal of Commodity Markets xxx (xxxx) xxx–xxx

S. Li, B.M. Lucey

Table 3 Precious metals average weak safe haveness across countries.

US UK GERMANY FRANCE ITALY SWISS CANADA JAPAN CHINA INDIA SOUTHAFRICA Avg.%

AU_STOCK

AG_STOCK

PT_STOCK

PD_STOCK

Avg.%

AU_BOND

AG_BOND

PT_BOND

PD_BOND

Avg.%

31% 34% 40% 31% 40% 32% 29% 36% 26% 30% 40% 34%

50% 25% 37% 38% 39% 38% 32% 33% 36% 33% 34% 36%

26% 28% 26% 29% 35% 35% 31% 24% 34% 23% 51% 31%

35% 26% 40% 28% 31% 32% 23% 25% 30% 44% 47% 33%

35% 28% 36% 31% 36% 34% 29% 30% 32% 32% 43% 33%

26% 36% 23% 30% 25% 33% 42% 30% 29% 29% 20% 29%

28% 26% 29% 30% 43% 36% 39% 39% 38% 25% 23% 32%

24% 21% 25% 47% 38% 32% 36% 33% 43% 27% 30% 32%

19% 26% 32% 35% 36% 22% 32% 33% 30% 24% 26% 29%

24% 27% 27% 35% 35% 31% 37% 34% 35% 26% 25% 31%

Notes: Weak safe haveness is calculated as a percentage of time when precious metals are weak safe haven at extreme market events (i.e. when market return falls under 5% quantile of return distribution). For instance, gold has been a safe haven against S & P500 for 86 days among those about 278 days when the stock market crashed from January 4 to July 15, 2016, hence 86/278 = 31%. Bold indicates that the number is above average safe haveness across eleven countries. Table 4 Precious metals average strong safe haveness across countries.

US UK GERMANY FRANCE ITALY SWISS CANADA JAPAN CHINA INDIA SOUTHAFRICA Avg.%

AU_STOCK

AG_STOCK

PT_STOCK

PD_STOCK

Avg.%

AU_BOND

AG_BOND

PT_BOND

PD_BOND

Avg.%

1.1% 0.0% 0.0% 1.1% 0.0% 1.1% 0.0% 0.4% 1.4% 1.6% 0.0% 0.6%

1.4% 1.4% 0.0% 1.1% 0.0% 1.1% 0.0% 0.4% 2.8% 1.6% 2.8% 1.1%

0.7% 0.0% 0.0% 0.4% 0.0% 0.0% 0.7% 2.6% 2.1% 1.6% 1.1% 0.8%

0.0% 0.4% 0.0% 0.4% 0.0% 1.1% 1.4% 0.4% 0.0% 1.1% 1.7% 0.6%

0.8% 0.4% 0.0% 0.7% 0.0% 0.8% 0.5% 0.9% 1.5% 1.5% 1.4% 0.8%

0.0% 2.1% 0.0% 2.8% 0.0% 0.4% 0.0% 0.0% 0.0% 0.0% 0.6% 0.5%

0.4% 2.8% 0.0% 0.0% 0.0% 0.7% 0.4% 1.5% 0.0% 0.0% 0.6% 0.6%

0.0% 1.8% 0.0% 1.8% 0.0% 0.7% 1.8% 0.7% 0.7% 0.0% 1.1% 0.8%

0.0% 0.0% 0.0% 0.0% 0.0% 0.0% 0.0% 0.0% 0.0% 0.0% 0.0% 0.0%

0.1% 1.7% 0.0% 1.2% 0.0% 0.4% 0.5% 0.5% 0.2% 0.0% 0.6% 0.5%

Notes: Strong safe haveness is calculated as a percentage of times when precious metals are a strong safe haven at extreme market events (i.e. when market return falls under 5% quantile of return distribution). For instance, gold has been a safe haven against S & P500 for only 3 days among those about 278 days when the stock market crashed from January 4 to July 15, 2016, hence 3/278 = 1.1%. Bold indicates that the number is above average safe haveness across eleven countries.

Fig. 7. Total Number of Incidents When Precious Metals as Weak Safe Haven Across Countries. Note: the aggregated number of weak safe haven incidents on a particular day is calculated as the sum of weak safe haven dummy indicating whether a precious metal acted as a weak safe haven on this day. For instance, on December 3, 2015, there are 23 incidents, including gold, platinum and palladium as safe haven against US bond index (3 incidents), gold, silver, platinum and palladium as safe haven against FTSE100 (4 incidents), and so on. The full list can be found in Table 5.

8

Journal of Commodity Markets xxx (xxxx) xxx–xxx

S. Li, B.M. Lucey

Fig. 8. Total Number of Incidents When Precious Metals as Strong Safe Haven Across Countries. Note: the aggregated number of strong safe haven incidents on a particular day is calculated as sum of strong safe haven dummy indicating whether a precious metal acted as a strong safe haven on this day. For instance, on August 11, 2013, there are 7 incidents, including gold, silver, platinum and palladium as strong safe haven against UK bond market crash, and gold, silver, and palladium as strong safe haven against French bond market fall. Table 5 Top ten dates with highest safe haven incidents across countries. 03 December 2015

15 January 2016

05 January 2015

14 May 1999

02 April 2004

Mario Draghi, ECB failed to deliver the major stimulus packages as expected.

Friday market slide was triggered by a crash in crude oil prices and China's stock market tumbling into a bear market. US Stock(Ag,Pt,Pd) UK Stock(Au,Ag,Pt,Pd) Germany Stock(Au,Ag,Pd) France Stock(Au,Ag,Pt) Italy Stock(Au,Ag,Pt,Ad) Swiss Stock(Au,Ag,Pt,Pd) Canada Stock(Ag,Pt,Pd)

Falling oil prices and discouraging economic data e.g. lower than expected exports on US manufacturing. US Stock(Au,Ag,Pt,Pd) Germany Stock(Ag,Pt,Pd) France Stock(Ag,Pt,Pd) Italy Stock(Ag,Pt,Pd) Italy Bond(Au,Ag,Pt,Pd) Canada Stock(Ag,Pt) S.Africa Stock(Au,Ag,Pt,Pd)

Increasing speculation that the Federal Reserve is close to raising its target rate since job growth beat expectation. Selloff in the bond market. US Bond(Au) UK Bond(Au,Ag,Pt,Pd) Germany Bond(Au,Ag,Pt,Pd) France Bond(Ag,Ag,Pt,Pd) Italy Bond(Au,Ag,Pt,Pd) Swiss Bond(Pt,Pd) Canada Bond(Au,Pt,Pd)

04 September 2001 U.S. stocks early gains were evaporate as tech sector worries overshadow manufacturing data. US Bond(Au,Ag,Pt,Pd) UK Bond(Au,Ag,Pt,Pd)

14 March 2007 The housing slump was spreading to the financial industry in US.

Unexpectedly big jump in US April CPI, which leads to speculation that official interest rate could soon increase. US Stock(Ag,Pt,Pd) UK Stock(Au,Ag,Pt,Pd) UK Bond(Au) Germany Bond(Ag,Pd) France Bond(Ag,Pd) Italy Bond(Au,Ag,Pd) Canada Stock(Au,Pt,Pd) Canada Bond(Au,Ag,Pt,Pd) 04 September 2008 Dow loses 345 points as mixed retail sales reports, weak job market news and an oil price slide. US Stock(Au,Ag,Pt) UK Stock(Au,Pt) Germany Stock(Au,Ag,Pt, Pd) France Stock(Ag,Pt,Pd) Italy Stock(Au,Ag,Pt,Pd) Swiss Stock(Ag,Pt) Japan Bond(Au,Ag,Pt)

Italy Bond(Au,Ag,Pt,Pd)

US Bond(Au,Pt,Pd) UK Stock(Au,Ag,Pt,Pd) Germany Stock(Au,Pd) France Stock(Ag,Pt,Pd) France Bond(Pd) Italy Bond(Au,Ag,Pt,Pd) Swiss Bond(Au,Pt,Pd) Canada Bond(Au,Ag,Pt) 08 November 2013 Stocks ended sharply lower on Thursday, after weak earnings from Whole Foods and Qualcomm. US Bond(Au,Ag,Pt,Pd) UK Bond(Au,Ag,Pt,Pd) Germany Bond(Au,Ag,Pt,Pd)

Germany Bond(Au,Ag,Pt,Pd)

UK Stock(Au,Ag,Pt,Pd) Germany Stock(Au,Ag,Pt, Pd) France Stock(Au,Ag,Pt,Pd)

France Bond(Au,Ag,Pt,Pd) Canada Bond(Ag,Pt,Pd) S.Africa Bond(Ag,Pt,Pd)

France Bond(Au,Ag,Pt,Pd) Italy Bond(Au,Ag,Pt,Pd) Canada Bond(Ag)

Swiss Stock(Au,Ag,Pt,Pd) Japan Stock(Au,Ag) S.Africa(Au,Ag)

09 December 2014 Slump in oil prices and discouraging China's trade numbers, contraction of Japan's economy and so on. France Stock(Ag,Pt,Pd) Italy Stock(Ag,Pt,Ad)

China Stock(Au,Ag,Pt,Pd) China Bond(Au,Ag,Pt) S.Africa Stock(Au,Ag,Pt,Pd)

(1998). Many other papers have shown the (complex) relationship between gold and inflation (see Bampinas and Panagiotidis, 2015a for example). Frankel (1986) argues that it is because gold and other minerals prices move faster than the sluggish prices of manufactured goods and services in the short run. We include the nominal interest rate, and the value of domestic currency which are measured by weighted value against major currencies for US dollar, and for other currencies its exchange rate with US dollar. We also consider industrial production due to the industrial usage of precious metals, credit spreads which are associated with business cycle, Goldman Sachs Commodity Index (GSCI) and oil (See Ciner et al., 2013 and Bampinas and Panagiotidis, 2015b). The above six economic variables normally have some empirical relationship with precious metals price or volatility movements (see studies of Batten et al., 2010 and Baur, 2013), it would be interesting to examine if these relationships also influence on precious metals safe haveness during market distress periods. Some of the explanatory variables have a shorter history (See Table M.8 for data summary for eleven countries and Appendix M. for data source details). In order to examine all explanatory variables, the safe haven results are matched against it. The time period examined across countries are as follows: US (January 1997 - March 2016), UK (February 2000 - May 2016), Germany (February

9

Journal of Commodity Markets xxx (xxxx) xxx–xxx

S. Li, B.M. Lucey

2000 - May 2016), France (February 2000 - February 2016), Italy (April 2010 - March 2016), Switzerland (July 2005 - May 2016), Canada (October 2009 - April 2016), Japan (October 2000 - May 2016), China (February 2011 - December 2015), India (November 2011 - March 2016), South Africa (May 2007 - December 2015). All the data are dis-aggregated into daily.3 We also conduct Augmented Dickey-Fuller test for stationary, and results are shown in Table 1. We convert, via differences, data to stationary series. 3.2. Model and results Zero-inflated Poisson (ZIP) regression The Poisson distribution is used to model variation in count data (i.e. data that can equal 0,1,2,…). The usual transformation g used is the logarithmic, so that g (u ) = exp (u ) transforms a continuous linear predictor βΔ (Xi ) to a positive yli . In this case, we model the modified daily safe haven count result for market i at time t (yi, t ) as a function of a number of market stress and economic indicators. Results would show whether an increase in one economic indicator has a positive or negative impact on the safe haven count result, given all the other predictors remain at the same level. In addition, since the modified daily safe haven result is conditional to whether market return is in the lower 5% quantile range, we apply the Zero-Inflated Poisson (ZIP) regression to model this independent process of excess zeros. These excess zeros are those days where we do not have a market drawdown in the lower 5% quantile. The ZIP regression thus include two parts, a Poisson regression on count data and an inflation model which predict excess zeros. The model is as follows:

⎧= 0 if Ri , q5% = 0 yi, t = ⎨ ⎩ ∼ Poisson (exp (α + β1 Δ (X1, i, t ) + β2 Δ (X2, i, t ) + …)) if Ri , q5% < 0

(4)

in which Δ (Xn, i ), n = 1, 2, 3, … are first differences of explanatory variables of market stress or economic indicators for market i. Here we use Ri , q5% as the predictor in the inflation model, which equals to:

⎧ ri, t if ri, t < ri q5% Ri , q5% = ⎨ ⎩ 0 if ri, t ≥ ri q5%

(5)

The model is ran in R via maximum likelihood with the package ‘pscl’. Table 6 provides results for the USA while others are shown in the online appendix. Similar patterns can be seen across countries for a number of explanatory variables. For example, an increase of Economic Policy Uncertainty is found to be positively related to the likelihood of gold being a safe haven against stock market falls in France, Japan and India, and against bond market extreme events in US, UK, Germany, France, Italy and India, albeit the coefficients are very small. The largest positive coefficient is 1.01 for Japan. It is also found to be significant for other precious metals such as silver against stock market crashes in US, UK, France, Italy, Canada, Japan and China and so on. It suggests that precious metals are more likely to become safe haven during market distress triggered by policy instability across countries. The Economic Policy Uncertainty Index is not published for Switzerland or South Africa. Table 7 A Financial stress index is only available for US and European countries. Results are quite mixed. Take gold as an example, Δfsi is a negative significant predictor of gold being a safe haven against US stock market, but is a positive and significant predictor against US bond market. It is, however, negative and significant for gold against both stock and bond markets in UK. In Europe, it also shows different results across countries. When Δfsi increases, gold is more likely to be a safe haven against stock markets in Germany and Italy but negative in France. It has some quite large coefficients, especially as a positive determinant for gold against Italian stocks and Germany bonds, which is probably due to the small number of safe haven incidents. Similarly, stock volatility as measured by VIX in US, VFTS in UK and so on also show mixed results across countries. Δstockvolatility has a negative relationship with gold as safe haven against stock markets in US, Germany, Italy, Switzerland and South Africa. However, it is a positive predictor for gold as safe haven in UK, France, Canada, Japan and India. Furthermore, it is a positive predictor for gold as safe haven against bond markets in US, Germany, France, Italy, Canada, Japan and South Africa. As to credit spreads, another indicator of economic uncertainty, its change is a negative predictor for gold, platinum and palladium as safe haven against US stocks, however it is positive for silver. The change of consumer confidence is a significant predictor for gold as a safe haven against both stock and bond markets in US and China. However it has opposite results - it is positive in US but negative in China. Therefore given other variables remain the same level, a further decline of consumer sentiment is likely to have a negative impact on gold as safe haven in US. In China, however, gold is more likely to be a safe haven when consumer confidence is further impaired.4 Change of CPI, or inflation, is mainly a positive predictor for precious metals as a safe haven against stock market extreme events, suggesting that precious metals are more likely to be a safe haven against stock market declines under higher inflationary environment. However results are mixed for the bond markets. The currency value as measured by the trade weighted value of U.S. dollar or the exchange rate with U.S. dollar for other currencies shows an interesting divergence for stock and bond markets. It is a mainly positive determinant for precious metals as safe haven against stocks across countries, except for Germany and Canada. For bond markets, it is mostly positive for gold, 3 For example, to convert monthly Economic Policy Uncertainty Index into daily, we apply the same value to each day of the month. No other frequency conversion technique is used here. 4 The change of benchmark interest rate is rare relatively to other explanatory variables, possibly one of the reasons why it caused so many linear singularity problem in the Poisson regressions, which means that it almost perfectly separates or predicts the safe haven count results in some cases, such as Canada, China, India and South Africa. In other cases, despite not causing singularly problem, it has extreme large coefficients such as in US, UK, Switzerland, Japan and China. It is mostly positive when it is significant.

10

Journal of Commodity Markets xxx (xxxx) xxx–xxx

S. Li, B.M. Lucey

Table 6 Summary of zero-inflated Poisson regression results - US. Stock

Count Model (Intercept)

Δvix

Δfsi

Δsentiment Δepu Δr

Δcpi

ΔUSD Δind production

Δcreditspreads Δgsci

Δoil

Zero-inflated Model (Intercept)

Ri, q5%

Bond

Gold Coef. (p value)

Silver Coef. (p value)

Platinum Coef. (p value)

Palladium Coef. (p value)

Gold Coef. (p value)

Silver Coef. (p value)

Platinum Coef. (p value)

Palladium Coef. (p value)

2.88* (0) −0.17* (0) −0.85* (0) 0.04* (0) 0.00 (0.117) 0.35 (0.869) 0.36 (0.021) 0.34* (0.001) 0.04 (0.524) −0.71* (0) 0.15* (0) −0.65* (0)

2.79* (0) −0.10* (0) −0.06 (0.596) 0.05* (0) 0.00 (0.205) 0.78 (0.517) 0.29* (0.005) 0.17 (0.029) 0.26* (0) 0.02 (0.727) 0.04* (0) −0.12 (0.013)

3.03* (0) −0.11* (0) −0.40* (0.01) 0.03* (0.001) 0.00* (0) −1.88* (0) 0.27 (0.06) 0.31* (0.001) 0.09 (0.051) −0.95* (0) 0.08* (0) −0.34* (0)

2.24* (0) −0.06 (0.04) −0.20 (0.209) 0.02 (0.123) 0.00* (0) −2.24 (0.037) −0.25 (0.134) 1.09* (0) 0.49* (0) −1.57* (0) 0.02* (0.141) −0.13 (0.103)

2.52* (0) 0.06 (0.025) 0.58* (0.001) 0.02* (0.01) 0.00 (0.114) 32.27 (0.774) −0.11 (0.462) 0.29* (0.004) 0.95* (0) −1.85* (0) −0.02 (0.368) 0.22* (0.004)

2.56* (0) 0.12* (0) 0.86* (0) −0.01 (0.442) 0.00* (0) 6.29* (0.003) −0.22 (0.127) 0.47* (0) 0.40* (0) −1.64* (0) −0.02 (0.462) 0.28* (0.004)

2.67* (0) 0.07* (0.002) 0.66* (0) 0.00 (0.803) 0.00 (0.36) −0.97 (0.079) −0.13 (0.255) 0.51* (0) 0.18 (0.021) 0.17 (0.382) −0.05* (0.004) 0.26* (0.003)

2.11* (0) 0.05 (0.227) −0.37 (0.176) −0.08* (0) 0.01* (0) 5.45* (0.008) −1.72* (0) 0.29 (0.035) 0.67* (0) −1.66* (0) 0.06 (0.02) −0.13 (0.282)

5.49* (0) 170.84* (0)

5.04* (0) 178.64* (0)

5.35* (0) 140.81* (0)

5.34* (0) 173.82* (0)

5.88* (0) 450.62* (0)

5.84* (0) 457.88* (0)

5.62* (0) 409.87* (0)

5.97* (0) 434.57* (0)

Notes: This table shows Poisson regression coefficients for each of the variables along with standard p-values. The second part corresponds to the inflation model, which includes logit coefficients for predicting excess zeros along with the standard p-values. * denotes a significant level of 1%. N/A denotes a variable dropped due to linear singularity problem.

however, for silver, platinum and palladium, it is positive only for US, and in some cases for UK, Germany, and South Africa, it has been mainly a negative predictor. Although we might expect the general commodity price environment to have more of an impact on the industrial metals - silver, platinum, and palladium - it is significant for gold as well. An increase of the oil price is a negative predictor for gold as safe haven against US stocks but positive for bond markets. While it has opposite results for UK and Germany that it is positive as predictor for gold as safe haven against stock market and negative for bond market. For the Asian countries, the oil price is found to be a positive determinant for gold, platinum and palladium as a safe haven against China stock market, for platinum and palladium against the Indian stock market, and for gold as a safe haven against China and Indian bond markets declines. Extreme bound analysis (EBA) The results of the ZIP analysis provide little consistent evidence on cross country determinants of safe haven status. Considering the relatively large number of potential determinants and potential various model specification, we attempt to isolate robust determinants. To do this we apply extreme bound analysis (EBA) in R with the package ”ExtremeBounds”. EBA identifies the variables most robustly associated with the outcome variable and tests whether a variable retains sign and significance across all possible combinations of variables. For instance, VIX is found to be significant with gold as a safe haven against the S & P500 by logistic regression, EBA allows us to examine how robust this significant positive relationship is when the model changes. EBA has been applied to deal with model uncertainty in many economics and finance studies. Some early studies include Sala-i Martin (1997) on economic growth, Hafner-Burton (2005) on the repression to globalization, Hegre (2006) on civil wars, and more recently, Chakrabarti and Zeaiter (2014) on the determinants of sovereign default, and Raputsoane (2014) on the relationship between financial stress indicator variables and monetary policy interest rate in South Africa. Given a large number of possible variables, Extreme Bounds analysis allows us to see the extent to which each variable impacts across a wide variety of possible specifications. Set X as the vector of possible explanatory variables which can impact on the 11

Coef. (G.CDF)

12

free

focus

free

focus

focus

focus

focus

focus

focus

focus

free

Δvix

Δfsi

Δepu

Δsentiment

Δcreditspreads

Δr

Δcpi

ΔUSD

Δind production

Δgsci

Δoil

0.20* (100%) 0.18* (100%) −0.04 (68%) 0.23* (98%) 0.01* (92%) 0.03 (70%) 0.06 (61%) −0.08 (82%) −0.09 (81%) −0.04* (91%) 0.03* (100%) −0.07 (76%)

focus

focus

free

focus

focus

focus

focus

focus

free

focus

free

free

0.39* (100%) 0.37* (100%) −0.02 (56%) 0.19* (92%) 0.01 (79%) 0.07 (81%) 0.09 (59%) 0.02 (56%) 0.06 (66%) −0.09* (94%) 0.02* (96%) −0.01 (45%)

Notes: * denotes a robust determinant when the G.CDF is above 90%. coefficients for EPU is multiplied by 100 for illustration purpose.

free

(Intercept)

focus

focus

focus

focus

focus

free

focus

focus

free

free

free

free

type

Coef. (G.CDF) 0.24* (100%) 0.25* (100%) 0.08 (75%) 0.63* (99%) 0.00 (59%) 0.00 (51%) −1.74 (84%) 0.02 (55%) −0.02 (58%) −0.06* (91%) 0.01 (84%) −0.04 (61%) focus

focus

focus

focus

focus

free

focus

focus

free

free

free

free

type

Coef. (G.CDF) 0.16* (100%) 0.15* (100%) 0.02 (60%) 0.14* (93%) −0.01* (93%) −0.04 (84%) −0.24 (82%) −0.07 (77%) 0.06 (80%) −0.02 (71%) 0.01 (73%) −0.04 (73%) focus

focus

focus

free

focus

focus

focus

free

focus

focus

focus

free

type

type

type

Coef. (G.CDF)

Gold

Palladium

Silver

Gold

Platinum

Bond

Stock

Table 7 Summary of sophisticated extreme bound analysis-US.

0.17* (100%) −0.01 (74%) 0.15* (98%) −0.01 (51%) 0.02* (94%) 0.04 (83%) 0.05 (69%) −0.08 (85%) 0.18* (95%) 0.07* (98%) 0.01 (87%) 0.01 (72%)

Coef. (G.CDF)

focus

free

focus

focus

free

focus

focus

focus

free

focus

focus

free

type

Silver

0.17* (100%) −0.01 (84%) 0.16* (98%) −0.20 (83%) 0.00 (77%) 0.03 (81%) 0.07 (74%) −0.18* (99%) 0.14* (91%) 0.02 (78%) 0.01* (93%) 0.00 (51%)

Coef. (G.CDF)

focus

focus

focus

free

focus

focus

focus

focus

focus

free

free

free

type

Platinum

0.19* (100%) −0.05* (100%) 0.14* (95%) 0.05 (62%) 0.00 (68%) 0.02 (68%) −0.97 (82%) 0.07 (77%) 0.27* (99%) −0.04 (86%) 0.00 (66%) 0.02 (72%)

Coef. (G.CDF)

focus

focus

focus

free

free

focus

focus

focus

free

focus

focus

free

type

Palladium

0.11* (100%) −0.01* (93%) 0.02 (67%) 0.10 (85%) −0.01* (92%) −0.05* (99%) −0.07 (69%) −0.13* (100%) 0.10* (90%) 0.03* (91%) 0.00* (91%) 0.01 (75%)

Coef. (G.CDF)

S. Li, B.M. Lucey

Journal of Commodity Markets xxx (xxxx) xxx–xxx

Journal of Commodity Markets xxx (xxxx) xxx–xxx

S. Li, B.M. Lucey

dependent variable y; we can then draw D variables from X and estimate the regression, repeating for all possible combinations of D from X . If there are theoretical or a priori empirical reasons to include a set of free variables F in all regressions, these can be specified. In essence, evaluating the coefficients of the elements of D we are interested in seeing whether these retain statistical significance across a wide range of the estimated models. Consider the regression

y = α + βj v + γj F + δj D + ϵ

(6)

where v is the variable of interest, an element of D , the subset of the set of doubtful variables X.k variables are drawn from D in each pass. Typically in the literature k=3 We collect the estimated regression coefficient βlj and the corresponding standard errors σlj for each pass of the regression for later evaluation. Leamer (1985) investigates the upper and lower extreme bounds, defined as

βlj ± τσl

(7)

where τ is the critical value for the 95% confidence level. A severe drawback of Leamer's (1985) approach is the binary nature of the possible outcomes: robust or fragile. A focus variable is identified as robust if all extreme bounds of the vast multitude of estimated models have the same sign; which tends to be rather unlikely when running thousands of regressions. Independent of the field of research, papers that applied a Leamer (1985) testing procedure tend to rush to the conclusion that most variables under study are fragile (Levine and Renelt, 1992; Bartley and Cohen, 1998). Aware of this shortcoming, Sala-i Martin (1997) focuses on the entire distribution of the regression coefficients, hence providing the researcher with a level of confidence for the robustness of each variable. Examination here is not done on the bounds but instead on the cumulative distribution function - the percentage of coefficients that lies either side of zero. The author proposes two variants, one assuming that the coefficients are normally distributed, the other one, less restrictive, allows a more general distribution of the coefficients. Formally, Sala-i Martin (1997) augment Leamer's testing procedure by first computing the weighted mean of both the regression coefficients βlj and the variance σlj2 where the weighting wj is specified by the researcher. The aggregated CDF(0) for the given regression coefficient is given as: M

Φ (0) =

∑ wj ϕj (0|βlj , σlj2)

(8)

j =1

where Φ is the weighted average of all individual CDF(0)'s and ϕj (0|βlj , σlj2 ) is the individual CDF(0) for each estimated regression model. ? propose a multistage approach to Extreme Bounds analysis, akin to the General-to-specific econometric methodology of Hendry (1980). To some extent this addresses the argument advanced in Hendry and Krolzig (2004) - a simplistic model, where all possible determinants are included is run first. Any robust variables are then fixed, and a second stage model uses only those fragile variables. We firstly run a naive extreme bounds analysis which considers all the combinations of potential determinants. Since the dependent variable is a safe haven count, the objective regression is further specified as a Generalized Linear Model. Although the results from naive extreme bounds analysis might be biased since it may not take into account the likely multicollinearity among the included variables as well as some variables measuring similar concepts according to Hlavac (2004), they still provide valuable insights and are also used in the next step to set a more sophisticated extreme bound analysis.5 In sophisticated EBA, those variables determined to be robust determinants in the previous naive analysis are set as free variables which are to be included in all regression models. We consider results from Sala-i-Martins version and define a determinant as robust when its value of CDF (0) the fraction of the variables cumulative distribution - is more than 85%. For example, VIX and EPU are robust variables for gold as safe haven against S & P500, which are then set as a free variable in the sophisticated version. The rest of variables are set as focus variables, which enter as possible variables of interest. In addition we also specify variables which measure similar concepts as exclusive parameters, and thus not to be included in any regression at the same time. We specify one set of mutually exclusive variables: the financial stress indicators which include VIX and FSI for US, FTSE and FSI for UK, VDAX and FSI for Germany and so on. The k argument remains at its default value of 0:3, which means that rather than estimating all possible combinations of the doubtful variables in nave EBA, only up to three doubtful variables are added to the focus variable in each specification - eventually however all focus variables are included in some specification. To further eliminate the influence of coefficient estimates from model specifications that suffer from high multicollinearity, a maximum acceptable variance inflation factor is set at 7. In addition, heteroscedasticity-robust standard error are used via the sandwich package and more weight is given to estimation results from regression models which provide a better fit. Finally the model remains set as general which does not assume any particular distribution of coefficient estimates across different model specifications. Since EBA doesn't model excess zeros separately, we expect results are somewhat different from zero-inflated Poisson regressions. Nevertheless, any common results from zero-inflated Poisson regression and Extreme Bound Analysis are more likely to indicate a robust determinant. In general, results from EBA show a more clear divergence between potential safe haven determinants across countries. Consistent with ZIP results, Δepu is found to be a robust positive determinant for all four metals as safe haven against stock market crashes in US. Coefficient are tend to be very small though. For bond market, however, an increase of Δepu is negatively associated with gold and silver as safe haven. Similarly, Δepu is a positive predictor for gold as safe haven in most other countries including Germany, France, Italy, China and India. Δfsi is always negatively associated with silver as safe haven against stocks in US, UK and European countries, however results are mixed for other metals. Results of stock volatility show a much 5

Although we do control for multicollinearity via selection only of those variables in each pass where the data pass a variance inflation test.

13

Journal of Commodity Markets xxx (xxxx) xxx–xxx

S. Li, B.M. Lucey

clear picture from EBA, that it is dominantly a robust positive determinant for precious metals as safe haven against stock market, with two exceptions of Canada and India. However results for bond markets show an interesting divergence that stock volatility has a negative impact on precious metals as safe haven in US, UK, Germany, India and South Africa, however it has positive impact in Italy, Canada and Japan. An increase of Δcredit spreads is most of the time positively associated with gold, silver and platinum as safe haven against S & P500, however it is not robust as G.CDF is less than 90%. In Europe, it is a positive determinant for gold as safe haven against stock market crashes in Germany and Italy, however it is negative in France and Switzerland. In Asia when gold is a safe haven against stocks, Δcredit spreads is more likely increasing in China and India, however decreasing in Japan and South Africa. Δinterest rate is more often to be a robust positive predictor for precious metals as safe haven against stock markets in US, UK and European countries. However it is mainly a negative predictor for Asian countries. 4. Conclusion This paper extends precious metals safe haven analysis in two aspects. First, we identify during which periods precious metals are safe haven against stock and bond market decline across developed and emerging countries. We see some patterns that precious metals are safe haven across countries clustered during some periods. However, for many times, precious metals provided protection at different times across countries. Second, we attempt to characterize those periods that under which political, economic and financial conditions, gold, silver, platinum and palladium are more likely to become safe haven assets and hedge the risks of market declines. Political risk is found to a positive and robust determinant across countries when precious metals are safe haven against stock and bond markets tail events. Stock volatility, exchange rate, interest rate and credit spreads tend to have large coefficient values but results are quite mixed for different markets and not robust. Appendix A. Supplementary data Supplementary data associated with this article can be found in the online version at http://dx.doi.org/10.1016/j.jcomm.2017. 05.003. References Ariovich, G., 1983. The impact of political tension on the price of gold. J. Stud. Econ. Econ. 16 (1), 17–37. Baker, S.R., Bloom, N., Davis, S.J., 2013. Measuring economic policy uncertainty. Work. Pap., 1–79. Balakrishnan, R., Danninger, S., Elekdag, S., Tytell, I., 2011. The transmission of financial stress from advanced to emerging economies. Emerg. Mark. Financ. Trade 47 (0), 40–68. Bampinas, G., Panagiotidis, T., 2015a. Are gold and silver a hedge against inflation? A two century perspective. Int. Rev. Financ. Anal. 41, 267–276. Bampinas, G., Panagiotidis, T., 2015b. On the relationship between oil and gold before and after financial crisis linear, nonlinear and time-varying causality testing. Stud. Nonlinear Dyn. Econ. 19 (5), 657–668. Bartley, W.A., Cohen, M.A., 1998. The effect of concealed weapons laws: an extreme bound analysis. Econ. Inq. 36, 258–265. Batten, J.a., Ciner, C., Lucey, B.M., 2010. The macroeconomic determinants of volatility in precious metals markets. Resour. Policy 35 (2), 65–71. Baur, D.G. 2013. Gold - Fundamental Drivers and Asset Allocation. Baur, D.G. and K.J. Glover 2012. The Destruction of a Safe Haven Asset? University of Technology Working Paper Series. Baur, D.G., Lucey, B.M., 2010. Is gold a hedge or a safe haven? An analysis of stocks, bonds and gold. CFA Dig. 40 (3), 2–3. Baur, D.G. and T. McDermott 2012. Safe Haven Assets and Investor Behaviour under Uncertainty. Baur, D.G., McDermott, T.K., 2010. Is gold a safe haven? International evidence. J. Bank. Financ. 34 (8), 1886–1898. Beckmann, J., Berger, T., Czudaj, R., 2015. Does gold act as a hedge or a safe haven for stocks? A smooth transition approach. Econ. Model. 48, 16–24. Bjørnskov, C., Dreher, A., Fischer, J.a.V., 2008. Cross-country determinants of life satisfaction: exploring different determinants across groups in society. Social. Choice Welf. 30 (1), 119–173. Cardarelli, R., Elekdag, S., Lall, S., 2011. Financial stress and economic contractions. J. Financ. Stab. 7 (2), 78–97. Chakrabarti, A., Zeaiter, H., 2014. The determinants of sovereign default: a sensitivity analysis. Int. Rev. Econ. Financ. 33, 300–318. Chua, J.H., Sick, G., Woodward, R.S., 1990. Diversifying with Gold Stocks. CFA Inst. 46 (4), 76–79. Ciner, C., Gurdgiev, C., Lucey, B.M., 2013. Hedges and safe havens: an examination of stocks, bonds, gold, oil and exchange rates. Int. Rev. Financ. Anal. 29, 202–211. Corbet, S., Twomey, C., 2014. An index of financial market stress for the United Kingdom. Econ. Bus. Lett. 3 (2), 127–133. Connor, F.A., Lucey, B.M., Batten, J.A., Baur, J.A., 2015. The financial economics of golda survey. Int. Rev. Financ. Anal. 41, 186–205. Coudert, V., Raymond H., 2010. Gold and financial assets: are there any safe havens in bear markets?. Frankel, J. 1986. Expectations and Commodity Price Dynamics the Overshooting Model.Pdf. Hafner-Burton, E.M., 2005. Right or robust? The sensitive nature of repression to globalization. J. Peace Res. 42 (6), 679–698. Hakkio, C., Keeton, W., 2009. Financial Stress: What is it, How Can it be Measured, and Why Does It Matter?. Hegre, H., 2006. Sensitivity analysis of empirical results on civil war onset. J. Confl. Resolut. 50 (4), 508–535. Hendry, D.F., 1980. Econometrics-alchemy or science? Economica 47 (188), 387–406. Hendry, D.F., Krolzig, H.-M., 2004. We ran one regeression. Oxf. Bull. Econ. Stat. 66 (5), 799–810. Hillier, D., Draper, P., Faff, R., 2006. Do precious metals shine? An investment perspective. Financ. Anal. J. 62 (2), 98–106. Hlavac, M. 2004. ExtremeBounds: Extreme Bounds Analysis in R. Illing, M., Liu, Y., 2006. Measuring financial stress in a developed country: an application to Canada. J. Financ. Stab. 2 (3), 243–265. Jaffe, J.F., 1989. Gold and gold stocks as investments for institutional portfolios. Financ. Anal. J. 45 (2), 53–59. Leamer, E.E., 1985. Sensitivity analyses would help. Am. Econ. Rev. 75 (3), 308–313. Levine, R., Renelt, D., 1992. A sensitivity analysis of cross-country growth regressions. Am. Econ. Rev. 82 (4), 942–963. Lucey, B.M., Li, S., 2015. What precious metals act as safe havens, and when? Some US evidence. Appl. Econ. Lett. 22 (1), 35–45. Raputsoane, L. 2014. Financial Stress Indicator Variables and Monetary Policy in South Africa Financial Stress Indicator Variables an Monetary Policy in South Africa. Sala-i Martin, X.X., 1997. I just ran two million regressions.pdf. Am. Econ. Rev. 87 (2), 178–183. Taylor, N.J., 1998. Precious metals and inflation. Appl. Financ. Econ. 8 (2), 201–210. Vayanos, D. 2004. Flight to Quality, Flight to Liquidity, and the Pricing of Risk.

14