Gold price ratios and aggregate stock returns

Tong Fang
School of Economics, Shandong University, Jinan, China, [email protected]
Zhi Su, Libo Yin
School of Finance, Central University of Finance and Economics, Beijing, China
This version: Oct 27, 2021

Abstract: We show that most gold price ratios, which represent the relative valuations of gold,

positively and significantly predict aggregate stock returns. These ratios fail to generate significant

predictive ability after controlling for a series of return predictors described in Welch and Goyal (2008)

or to display significant out-of-sample forecasting performance, except for the gold -oil price ratio (GO).

GO is the most powerful predictor. A one-standard-deviation increase is associated with a 6.60%

increase in the annual excess return for the next month. GO generates the most sizable out-of-sample

𝑅 2 and utility gains for a mean-variance investor. We find that the economic source of GO’s predictive

ability originates from the cash flow channel using stock return decomposition and positive predictive

power on economic conditions. The effect of GO is likely to be reversed only in periods when gold is

valuable relative to oil, but the reversal is found to be insignificant. Return predictability from gold

price ratios provides us with a new perspective for understanding gold price dynamics.

Keywords: gold price ratio, predictive regression, return predictability, cash flow channel

JEL: G12, G17, G53

1. Introduction
Aggregate stock return predictability has drawn much attention among academics and practitioners.

A voluminous literature discusses whether stock returns are predictable (Cochrane, 2008; Ang and

Bekaert, 2007; Campbell and Thompson, 2008). However, Welch and Goyal (2008), in their seminal

paper, indicate that despite significant in-sample predictive ability, traditional predictors fail to predict

future excess returns out-of-sample. Most of the following studies attempt to uncover new predictors

that can significantly predict returns out-of-sample, such as the technical indicators of Neely et al.

(2014), aligned investor sentiment index of Huang et al. (2015), the short interest index of Rapach et al.

(2016) and the asymmetric oil return of Wang et al. (2019). In this paper, we contribute to the previous

1
literature by empirically investigating the predictive ability of a set of gold price ratios in-sample and

out-of-sample.

Historically, gold served as the basis of the monetary system and as the ultimate standard of value

and international settlement (Capie et al., 2005). Over the 20th century, more of gold’s traditional role

in the international monetary system was assumed by the US dollar (Katz and Holmes, 2012). Gold is

now more likely to serve as a financial asset that is priced in US dollars than as the basis for a monetary

system. Gold is also considered a safe haven or hedging assets (Baur and Lucey, 2010; Baur and

McDermott, 2010; Reboredo, 2013; Bekiros et al., 2017). However, Baur et al. (2020) argue that gold

prices are not well understood, and no commonly accepted pricing models exist in the previous literature.

In fact, most studies directly use the US dollar price of gold in their empirical analyses and ignore the

other relative values of gold that may contain different information from the dollar price of gold (Capie

et al., 2005; Baur and Lucey, 2010; Reboredo, 2013; Agyei-Ampomah et al., 2014). In recent works,

Huang and Kilic (2019) construct a gold platinum price ratio (GP) to value gold and indicate that GP

has powerful predictive ability both in-sample and out-of-sample. More generally, Baur et al. (2020)

consider a set of gold price ratios reflecting whether gold is a relative store of value and an inflation

hedge or a measure of the opportunity cost of holding physical gold.1 Baur et al. (2020) find that the

statistical properties are distinctive across gold price ratios, implying that the information of gold price

ratios varies. Since Huang and Kilic (2019) demonstrate that the information of return predictability in

GP is significant, we are interested in a question that combines the return predictability and relative

valuations of gold: do gold price ratios predict aggregate stock returns?

Inspired by Baur et al. (2020), we begin by constructing nine gold price ratios, including the gold-

oil ratio, gold-silver ratio, gold-CPI ratio, gold-corn ratio, gold-copper ratio, gold-Dow Jones Industrial

Average ratio, gold-yield dividend ratio, gold-treasury bond yield ratio and gold-federal fund rate ratio

from 1975M1 to 2020M9.2 We also include the GP ratio studied in Huang and Kilic (2019). The gold

price ratio is calculated as the natural logarithm of the US dollar price of gold to the price of an

individual asset. From Pearson correlation coefficients for gold price ratios and traditional predictors,

we find that five gold price ratios (GS, GCO, GDJ, GTB and GFR), are highly correlated with the

1 Baur et al. (2020) show that a stable long-run ratio is consistent with gold’s store of value and inflation hedge property, and
a stable ratio in periods of uncertainty is consistent with gold’s safe haven property.
2 We refer to these ratios as GO, GS, GCPI, GCO, GCP, GDJ, GYD, GTB and GFR, respectively.

2
traditional predictors considered in Welch and Goyal (2008), while GO and GP are largely unrelated to

traditional predictors. The correlations imply that GO and GP may contain substantially different

information from traditional predictors.

We use univariate and bivariate predictive regressions to investigate the predictive ability of gold

price ratios for stock returns and examine whether information provided by gold price ratios overlaps

that of the traditional predictors studied in Goyal Welch (2008). Statistically, there is an issue of

statistical inferences. Because gold price ratios are highly persistent, the coefficient estimates may be

biased in finite samples, and a spurious regression concern also emerges (Stambaugh, 1999; Ferson et

al., 2003). To address this issue, we use Newey and West (1987) heteroskedasticity and autocorrelation

robust t-statistics and calculate wild-bootstrapped p-values for coefficient estimates following Huang

et al. (2015) and Rapach et al. (2016). We are more interested in the out-of-sample forecasting

performance of gold price ratios using the Campbell and Thompson (2008) out-of-sample 𝑅2 statistic

and the Clark and West (2007) adjusted mean squared forecast errors (MSFE-adj). An asset allocation

exercise is employed to explore whether the predictive ability of gold price ratios is economically

significant.

In-sample tests indicate that seven out of ten gold price ratios significantly and positively predict

excess stock returns. Specifically, the coefficient estimate of GO is the largest among gold price ratios

and traditional predictors. A one-standard-deviation increase in GO is associated with a 6.600% increase

in the annual excess return for the next month. GO also delivers the highest monthly 𝑅2 statistic of

2.225%. The coefficient estimates of GO and GP are always significant after controlling traditional

predictors of Welch and Goyal (2008). The out-of-sample forecast evaluations display that gold price

ratios have higher values of the out-of-sample 𝑅2 statistic than those of traditional predictors on

average, and GO still provides the highest out-of-sample 𝑅2 statistic that is significantly larger than

zero. The out-of-sample 𝑅2 statistics for GO are 1.571% and 3.038%, both significant at the 1% level

according to Clark and West (2007) MSFE-adj statistics, when using two initial estimation samples.

The forecast encompassing tests confirm the superior performance of GO. In addition, we investigate

the economic significance of the predictive ability of gold price ratios through an asset allocation

exercise and find that gold price ratios generate larger utility gains and Sharpe ratios on average than

those of traditional predictors. GO remains the most powerful predictor with the largest utility gain and

Sharpe ratio, confirming that GO’s predictive ability is both statistically and economically significant.
3
 Our results are robust and consistent. First, we employ three alternative estimation procedures to

check the robustness of the predictive ability of gold price ratios, including the weighted least squares

using ex ante variance proposed by Johnson (2019), the reduced-bias estimation of Amihud and Hurvich

(2004), and the instrument variable approach developed by Kostakis et al. (2015). The predictive ability

of GO and GP remains significant when using these estimation approaches. Second, we consider a

principal component analysis that is used to extract information from gold price ratios. The components

with higher loadings on GO have better forecasting performance. Third, GO remains a powerful

predictor after controlling for newly proposed predictors, such as the short interest index of Rapach et

al. (2016), the news-based implied volatility of Manela and Moreira (2017), the aligned investor

sentiment index of Huang et al. (2015) and the asymmetric oil return of Wang et al. (2019).

Why does GO predict future aggregate stock returns? First, we answer this question of interest by

investigating whether the return predictability from GO stems from the cash flow channel, discount rate

channel or both. Using the vector autoregressive (VAR) approach as in Campbell (1991) and Campbell

and Ammer (1993), we decompose aggregate stock returns into expected returns, the discount rate news

component and the cash flow news component. We regress these three components on GO and find that

GO’s predictive ability mainly comes from its predictive power for the cash flow news component, and

GO provides information that is useful for predicting excess returns beyond the information contained

in traditional predictors. Second, we link the predictive ability of GO to economic conditions. We

empirically reveal that GO positively predicts future economic variables. Through a rolling window

estimation and a state-switching predictive regression of Jacobsen et al. (2019), a reversed effect of GO

is found during periods when gold prices increase and oil prices decrease. Only when gold prices

increase and oil price decrease does GO lead to a worse stock market performance. Meanwhile, oil may

partially isolate the hedge or safe have role of gold and reveal the information of gold that is strongly

related to economic fundamentals. Third, we find no evidence of investor learning.

Our paper contributes to previous literature in several ways. First, we contribute to the literature

on return predictability by comprehensively evaluating the return predictive ability of gold price ratios.

We find that GO is a powerful return predictor in-sample and out-of-sample that outperforms traditional

predictors in Welch and Goyal (2008) and other gold price ratios, including GP proposed by Huang and

Kilic (2019). Second, we combine return predictability and relative values of gold and thus extend the

work of Baur et al. (2020), who focus on the statistical properties and influential factors of gold price
4
ratios. Third, we explore the driving forces of GO’s predictive ability. The link between GO and

economic variables provides us with a new perspective for deeply understanding gold price movement,

and also corrects the biased cognition from financial media and investors that GO always indicates

following periods of high risk or economic recession.

The remainder of this paper is organized as follows. Section 2 describes the data. Section 3

comprehensively illustrates the in-sample analysis and out-of-sample forecasting evaluations using gold

price ratios. Section 4 investigates the driving forces of return predictability from GO. Section 5 presents

our robustness checks. Section 6 concludes.

2. Data
2.1 Data description

The monthly US aggregate stock market return is calculated as the excess return, which is the

continuously compounded log return on the S&P 500 index (including dividends) minus the risk-free

rate. The risk-free rate is obtained from Amit Goyal’s website.

Gold price ratios are calculated as the natural logarithm of the ratio of gold to other asset prices:

𝐺𝑃𝑅𝑡 = ln(𝑃𝑡𝐺𝑜𝑙𝑑 ⁄𝑃𝑡𝐴𝑠𝑠𝑒𝑡 ). (1)

Following Baur et al. (2020), we consider the following nine gold price ratios: the gold-oil ratio

(GO), gold-silver ratio (GS), gold-CPI ratio (GCPI), gold-corn ratio (GCO), gold-copper ratio (GC),

gold-Dow Jones Industrial Average ratio (GDJ), gold-dividend yield ratio (GYD), gold-Treasury bond

yield ratio (GTB) and gold-federal fund rates ratio (GFR). We also include the gold-platinum ratio (GP)

studied in Huang and Kilic (2019). Gold and silver prices are collected from the London Bullion Market

Association (LMBA). Crude oil prices are collected from the US Energy Information Administration

(EIA). Dow Jones Industrial Average index and corn and copper prices are obtained from Datastream.

The Dividend yield ratio for S&P 500 is from Amit Goyal’s website. The CPI, US 10-year Treasury

bond yields and federal funds rates are from the FRED database of the Federal Reserve Bank of St.

Louis. All of these asset prices except for the monthly dividend yield ratio are in daily frequency, and

we take their monthly average to compute gold price ratios. The ten gold price ratios are shown in

Figure 1.

[Insert Figure 1 Here]

For interest of return predictability comparisons, we consider the following traditional predictors
5
studied in Welch and Goyal (2008). The Dividend Price Ratio (DP) is the natural logarithm of the ratio

of dividends and prices. The Dividend Yield (DY) is the difference between log dividends and log

lagged prices. The Price Earnings Ratio (PE) is the difference between log prices and log earnings. The

Dividend Payout Ratio (DE) is the difference between log dividends and log earnings. Treasury bill

rates (TBL) are the 3-month Treasury Bill: secondary market rates. Inflation rates (INFL) are the first

differences of the natural logarithms of the CPI. The Long-term Yield (LTY) is the long-term

government bond yield. The Default Spread (DS) is the difference between the yields of Baa and Aaa

corporate bonds. The Term Spread (TS) is the difference in yields between a 10-year constant maturity

US government bond and a 3-month Maturity US Treasury Bill. The Book-to-Market Ratio (BM) is the

ratio of book value to market value for the S&P 500 composite index. The Net Equity Expansion (NTIS)

is the ratio of 12-month moving sums of net issues by NYSE listed stocks divided by the total end-of-

year market capitalization of NYSE stocks. The Stock Variance (SVAR) is calculated as the sum of

squared daily returns on the S&P 500.

The data for DP, DY and PE are collected from Robert Shiller’s website. BM, NTIS and SVAR are

obtained from Amit Goyal’s website. The other predictors are available from the FRED database of the

Federal Reserve Bank of St. Louis.

2.2 Summary statistics

The sample period for BM and NTIS spans from 1975M1 to 2019M12, and the period for the rest

of the predictors and gold price ratios spans from 1975M1 to 2020M9. The summary statistics are

reported in Table 1. The AR(1) coefficients for gold price ratios are larger than 0.980, indicating the

persistence of these price ratios.

[Insert Table 1 Here]

Panel A of Table 2 displays Pearson correlation coefficients for traditional predictors and gold

price ratios. Based on the average absolute correlation coefficient (ACC), GS, GCO, GDJ, GTB and

GFR are highly correlated with traditional predictors.3 The ACCs of GO (0.129) and GP (0.135) show

that GO and GP are largely unrelated to traditional predictors, implying that GO and GP appear to

contain substantially different information from traditional predictors. Panel B reports correlation

3 To clearly interpret the correlations for gold price ratios and traditional predictors, we use an average absolute correlation
coefficient (ACC), which is calculated as the average of the absolute value of correlation coefficients (correlation in magnitude)
between a gold price ratio with all traditional predictors.
6
coefficients for gold price ratios. The correlation coefficient for GO and GP is 0.660, and the

information contained in GO and GP may also be different.

[Insert Table 2 Here]

3. Empirical analysis
3.1 In-sample analysis

We first examine in-sample stock market return predictability. A typical specification regresses

excess stock market returns on an independent lagged predictor (Welch and Goyal, 2008):

𝑟𝑡+1 = 𝛼 + 𝛽𝑋𝑡 + 𝜀𝑡+1 , (2)

where 𝑟𝑡 is the log stock market return in excess of the risk-free rate at month 𝑡 and 𝑋𝑡 is one of the

gold price ratios and return predictors described in Section 2. Gold price ratios and traditional predictors

are demeaned and scaled by their standard deviations. It is noteworthy that gold price ratios and most

of the traditional predictors are highly persistent, as shown in Table 1, suggesting a potentially spurious

regression concern (Ferson et al., 2003). Statistical inferences are also complicated by the well-known

Stambaugh (1999) bias. To make more reliable inferences, we utilize a Newey and West (1987)

heteroskedasticity and autocorrelation-robust t-statistic and calculate a wild-bootstrapped p-value to test

𝐻0 : 𝛽 = 0 against 𝐻𝐴 : 𝛽 > 0 following Rapach et al. (2016).4

[Insert Table 3 Here]

We estimate Equation (2) via OLS, and the estimation results are reported in Table 3.5 Panel B of

Table 3 shows that two of the traditional predictors (INFL and LTY) have significant predictive ability,

and INFL has the largest estimated 𝛽̂ of 0.355. We are more interested in the predictive ability of gold

price ratios, as presented in Panel A of Table 3. First, the estimated 𝛽̂ of gold price ratios are positive,

demonstrating that a higher gold price ratio leads to a higher stock market excess return. Second, seven

of the ten gold price ratios have significant estimated 𝛽̂ : GO, GS, GCO, GCP, GTB, GFR and GP. More

importantly, the estimated 𝛽̂ of GO is the largest among the gold price ratios and is also larger than

those of all traditional predictors. A one-standard-deviation increase in GO is associated with a 6.60%

increase in the annual excess return for the next month. The estimated 𝛽̂ of GP is also economically

4 Based on financial theory and previous empirical evidence, we take the negative (-) of TBL, INFL, LTY, BM, NTIS and
SVAR following Rapach et al. (2016). The wild-bootstrapped p-value estimation procedure can be found in Huang et al. (2015).
5 We also use gold price ratios to predict CRSP value-weighted returns and find similar results.

7
large, confirming the results of Huang and Kilic (2019). Moreover, GO delivers the highest monthly

𝑅 2 statistic (2.225%), which represents an economically meaningful degree of return predictability

(Campbell and Thompson, 2008).

̂ statistics, which are asymptotically point-
Table 3 also reports the Elliott and Müller (2006) 𝑞𝐿𝐿

optimal for testing the instability or breaks in coefficients of regression models. It is necessary to test

whether regression coefficient 𝛽 is stable across our sample period. The null hypothesis of the test is

that coefficient 𝛽 remains stable over time. As presented in Table 3, we do not find evidence of

structural instability in the regression coefficient of gold price ratios.

In summary, the univariate regression results indicate that gold price ratios have higher predictive

power than traditional predictors, and GO displays the highest degree of predictive ability among gold

price ratios and traditional predictors.

The predictive information in gold price ratios is significant. However, if the information overlaps

with that from traditional return predictors, it is not necessary to consider gold price ratios as new return

predictors. Therefore, we consider a bivariate regression to determine if gold price ratios (especially the

seven price ratios) still have significant predictive power for excess returns when the traditional

predictors are included in the predictive regression. The bivariate regression is shown in Equation (3):

𝑟𝑡+1 = 𝛼 + 𝛽𝐺𝑃𝑅𝑡 + 𝜃𝑋𝑡 + 𝜀𝑡+1 . (3)

where 𝑋𝑡 is one of the traditional return predictors considered in this paper. We utilize a Newey and

West (1987) heteroskedasticity and autocorrelation-robust t-statistic and calculate a wild-bootstrapped

p-value to test 𝐻0 : 𝛽 = 0 (𝜃 = 0) against 𝐻𝐴 : 𝛽 > 0 (𝜃 > 0).

Table 4 reports the estimated 𝛽̂ of gold price ratios (the estimated 𝜃̂ for traditional predictors are

shown in Appendix Table A1). We find that the estimated 𝛽̂ of only two of the seven gold price ratios

(GO and GP) remain sizable and are still highly significant after controlling traditional predictors.

Including traditional predictors does not influence the predictive ability of GO and GP. However, the

estimated 𝛽̂ of the remaining five ratios (GS, GCO, GCP, GTB and GFR) are not always statistically

significant, because these gold price ratios are highly correlated with some of the traditional predictors.

Collectively, GO and GP contain information that is quite different from that contained in numerous

popular predictors from the previous literature.

[Insert Table 4 Here]

3.2 Out-of-sample forecasting evaluations

8
3.2.1 Out-of-sample 𝑹𝟐

Kandel and Stambaugh (1996) argue that if in-sample predictability can be sustained out-of-sample,

it will be of substantial economic significance. Consequently, an interesting question to address in this

paper is whether the in-sample predictability of future returns from gold price ratios can be sustained

out-of-sample. We perform 1-month-ahead out-of-sample forecasting by generating excess return

forecasts using a sequence of expanding windows as follows:

𝑟̂𝑡+1 = 𝛼̂ + 𝛽̂ 𝑋𝑡 , (4)

where 𝛼̂ and 𝛽̂ are the OLS estimates of 𝛼 and 𝛽 in Equation (2) based on data from the beginning

of the sample through month 𝑡 . Specifically, we use two initial estimation periods to assess the

robustness of out-of-sample forecasting evaluations. The first initial estimation period spans from

1975M1 to 1984M12 (120 months), and the second period spans from 1975M1 to 1994M12 (240

months).6

To evaluate out-of-sample forecasting performance, we employ the Campbell and Thompson

2
(2008) 𝑅𝑂𝑂𝑆 statistic, as shown in Equation (5):

2 ∑𝑇 𝑖 (𝑟 −𝑟̂ 𝑖 )2
𝑅𝑂𝑂𝑆 = 1 − ∑𝑖=𝑡+1
𝑇 2. (5)
𝑖=𝑡+1 𝑖 −𝑟̅𝑖 )
(𝑟

2
A positive 𝑅𝑂𝑂𝑆 indicates a better out-of-sample forecast than the historical mean model

(Benchmark), and a negative statistic shows the opposite. To test whether predictive regression has a

lower MSFE than the historical mean model, we use the Clark and West (2007) adjusted mean squared
2 2
forecast errors (MSFE-adj). The null hypothesis is 𝐻0 : 𝑅𝑂𝑂𝑆 ≤ 0 and the alternative is 𝐻𝐴 : 𝑅𝑂𝑂𝑆 > 0.

The MSFE-adj statistic is defined as:

𝑓̂𝑡+1 = (𝑟𝑡+1 − 𝑟̅𝑡+1 )2 − [(𝑟𝑡+1 − 𝑟̂𝑡+1 )2 − (𝑟̅𝑡+1 − 𝑟̂𝑡+1 )2 ]. (6)

We regress 𝑓̂𝑡+1 on a constant and obtain the associated Newey and West (1987) t-statistic. The

MSFE-adj statistic has an approximately standard normal distribution when comparing forecasts from

nested models (Clark and West, 2007).

[Insert Table 5 Here]

The out-of-sample forecasting results are reported in Table 5. For out-of-sample forecasting using
2
the first initial sample, the 𝑅𝑂𝑂𝑆 statistics of GO, GS, GCO, GCP and GP are positive. However, only

6 Huang et al. (2015) use an initial estimation period of 234 months. Huang and Kilic (2019) consider two initial estimation
periods of 120 and 180 months. Wang et al. (2019) use an initial period of 240 months. Considering the selection of initial
estimation period lengths used in the previous literature, we use initial estimation periods of 120 and 240 months.
9
 2
the 𝑅𝑂𝑂𝑆 statistics of GO (1.571%) and GP (1.113%) are statistically significant. As indicated by the
2 2
sizable 𝑅𝑂𝑂𝑆 , GO and GP deliver a much lower MSFE than the historical mean model. The 𝑅𝑂𝑂𝑆

statistics of traditional return predictors are all negative and insignificant, confirming the results of

Welch and Goyal (2008). In this way, gold price ratios exhibit better out-of-sample forecasting

performance than traditional predictors. For the initial estimation sample from 1975M1 to 1994M12,

the out-of-sample results are similar. GO and GP continue to significantly outperform the historical

mean model.
2
Most importantly, GO has the largest 𝑅𝑂𝑂𝑆 statistic among the predictors. When we estimate
2
using the second initial estimation sample, the 𝑅𝑂𝑂𝑆 statistic is 3.038%, significant at the 1% level.

Huang et al. (2019) find that GP performs better than traditional predictors considered in Welch and

Goyal (2008). We further reveal that GO is even more powerful than GP and other predictors in terms

of return predictability.

3.2.2 Forecast encompassing test

To directly compare the information content of the predictive regression forecast based on GO to

that of the predictive regression forecasts based on the remaining predictors, we apply forecast

encompassing tests of Harvey et al. (1998). We form an optimal combination forecast as a convex

combination of a predictive regression forecast based on one of the other predictors and a forecast based

on GO:
𝐶 𝑖 𝐺𝑂
𝑟̂𝑡+1 = (1 − 𝜆)𝑟̂𝑡+1 + 𝜆𝑟̂𝑡+1 , (7)
𝐶 𝐺𝑂 𝑖
where 𝑟̂𝑡+1 is a forecast combination and 0 ≤ 𝜆 ≤ 1 . 𝑟̂𝑡+1 and 𝑟̂𝑡+1 denote the predictive

regression forecast based on GO and one of the other predictors, respectively. We estimate 𝜆̂ using the

approach of Harvey et al. (1998). If 𝜆̂ = 0, the optimal combination forecast excludes the forecast

based on GO, which means that GO fails to contain useful information on return predictability beyond

information on other gold price ratios and traditional predictors. If 𝜆̂ > 0, the optimal combination

forecast includes the forecast based on GO, and GO contains information different from that of other

gold price ratios and traditional predictors.

As displayed in Table 5, the estimated 𝜆̂ (ENCT-𝜆̂) are all sizable and significant. When the initial

estimation sample covers 120 months, all of the estimated 𝜆̂ are larger than 0.70. Most of these values

are larger than 0.80 except for predictive regression forecasts based on GP (0.608) and INFL (0.738).

In particular, the majority of the estimated 𝜆̂ are equal to one when the initial sample period covers
10
240 months, meaning that the optimal forecast only incorporates information of GO. Therefore, the

predictive regression forecasts based on GO encompass the forecasts based on any one of the other gold

price ratios and traditional predictors. We provide strong evidence that GO significantly outperforms

the other gold price ratios and traditional predictors with respect to out-of-sample forecasting.

3.3 Out-of-sample forecasting over business cycles

The predictive ability of a return predictor is also influenced by business cycles (Cochrane, 2007;

Neely et al., 2014). We continue to investigate the out-of-sample forecasting performance of gold price

ratios during economic expansions and recessions. The dates for the expansions and recessions are

determined by NBER business cycles.

2
We use the following 𝑅𝑂𝑂𝑆 to compare the forecast performance of the predictors over business

cycles:

2 ∑𝑇 𝐼𝑖𝑏𝑐 (𝑟𝑖 −𝑟̂ 𝑖 )2

𝑅𝑏𝑐,𝑂𝑂𝑆 = 1 − ∑𝑖=𝑡+1
𝑇 𝑏𝑐 2
, 𝑏𝑐 = 𝐸𝑋𝑃, 𝑅𝐸𝐶, (8)
𝑖=𝑡+1 𝐼𝑖 (𝑟𝑖 −𝑟̅𝑖 )

where 𝐼𝑖𝐸𝑋𝑃 (𝐼𝑖𝑅𝐸𝐶 ) is an indicator variable that equals 1 when month 𝑖 is an economic expansion

(recession) period and zero otherwise (Henkel et al., 2011; Neely et al., 2014). EXP and REC denote

economic expansion and recession, respectively. We also consider two initial estimation samples.

[Insert Table 6 Here]

The out-of-sample forecasting results over NBER business cycles are presented in Table 6. The
2 2
𝑅𝑅𝐸𝐶,𝑂𝑂𝑆 and 𝑅𝐸𝑋𝑃,𝑂𝑂𝑆 statistics of GO and GP are positive, showing that predictive regressions

based on GO and GP outperform the historical mean model during economic recessions and expansions.

Most of the traditional predictors fail to outperform the benchmark during recessions, expansions or
2 2
both. For example, the 𝑅𝑅𝐸𝐶,𝑂𝑂𝑆 and 𝑅𝐸𝑋𝑃,𝑂𝑂𝑆 statistics of PE, DS, TS and SVAR are negative when
2
estimated using the first initial sample period. Moreover, we find that the 𝑅𝑅𝐸𝐶,𝑂𝑂𝑆 of GO is larger
2 2 2
than the 𝑅𝐸𝑋𝑃,𝑂𝑂𝑆 . For the second initial sample period, the 𝑅𝑅𝐸𝐶,𝑂𝑂𝑆 and 𝑅𝐸𝑋𝑃,𝑂𝑂𝑆 statistics are

2.693% and 5.081%, respectively. GO may not be the most powerful predictor among gold price ratios

and traditional predictors during expansions, but it substantially outperforms the other gold price ratios

and traditional predictors during recessions. The performance of GO displays a countercyclical behavior

of return forecasting (Wang et al., 2019). As pointed by the NBER business cycles, the most recent

economic peak occurred in February 2020, and the economic downturn caused by the COVID-19

pandemic is identified as a recession. If this economic recession lasts for a longer period of time, the

11
predictive ability of GO will be enhanced.7

3.4 Forecasting with longer horizons

The AR(1) coefficients shown in Table 1 suggest the high persistence of gold price ratios, and they

may exhibit long-term predictive ability for excess stock market returns. In this subsection, we examine

the predictive ability of gold price ratios with a longer horizon. The predictive regression is described

as follows:
1
∑ℎ𝑖=1 𝑟𝑡+𝑖 = 𝛼 + 𝛽𝑋𝑡 + 𝜀𝑡+ℎ , (9)
ℎ

where h=3, 6, and 12 months are forecast horizons. Forecasting results with long horizons are reported

in Table 7. Similar to the results given in Table 3, the estimated 𝛽̂ of GO, GS, GTB, GFR and GP are

positive and significant for all forecast horizons. GO provides the highest 𝑅2 statistics for h=3 and 6

months (5.095% and 7.882%, respectively) and the largest coefficient estimates. For h=12, the

estimated 𝛽̂ of GP is 0.504, which is larger than that of GO (0.477). The 𝑅2 statistic for GP also

becomes the largest among gold price ratios.

[Insert Table 7 Here]

The out-of-sample forecasting results are similar to the in-sample results (the out-of-sample
2
forecasting results with longer horizons are provided in Appendix Table A2). Specifically, the 𝑅𝑂𝑂𝑆

statistics of GO for h=3, 6, and 12 are 5.439%, 8.832% and 14.028%, respectively. Meanwhile, the
2
𝑅𝑂𝑂𝑆 statistics of GP are 4.172%, 8.515% and 15.928%, respectively. GO and GP are the two gold

price ratios that generate the best out-of-sample results, and GP outperforms GO for longer horizons.

Therefore, we empirically show that gold price ratios (GO and GP in particular) have powerful

significant predictive ability both in-sample and out-of-sample.

3.5 Asset allocation implications

To explore the economic value of the predictive ability of gold price ratios from an asset allocation

perspective, we consider an investor with mean-variance preferences who monthly allocates between

equities and risk-free bills using a predictive regression forecast of excess stock market returns

(Campbell and Thompson, 2008; Rapach et al., 2010; Neely et al., 2014). At the end of month 𝑡, the

7 We also consider several additional tests using subsample analysis, including a pre-crisis subsample (1975M1-2008M8), a
post-crisis subsample (2008M10-2020M9), and a pre-COVID subsample (1975M1-2019M12). GO performs better for a
subsample that covers more economic recessions. If a subsample contains more expansion periods, the predictive ability of
GO will be heavily compromised. For example, the estimate coefficient of GO is not highly significant for the post-crisis
subsample, because a long period of economic expansion covers the majority of this subsample. These subsample analyses
confirm that GO performs better during economic recessions.
12
investor optimally allocates the following share of the portfolio to equities during month 𝑡 + 1:
1 𝑟̂𝑡+1
𝑤𝑡 = ( 2 ), (10)
̂𝑡+1
𝛾 𝜎

2
where 𝑟̂𝑡+1 is a predictive regression simple excess return forecast, 𝜎̂𝑡+1 is a forecast of the excess

return variance, and 𝛾 is the investor’s coefficient of relative risk aversion.8 Following Campbell and

Thompson (2008), Huang et al. (2015), and Rapach et al. (2016), among others, we calculate the

variance forecast using a ten-year rolling window of past returns. The share of the portfolio to equities

𝑤𝑡 is restricted to lie between 0 and 1.5, which imposes realistic portfolio constraints and produces

better-behaved portfolio weights (Neely et al., 2014). We use 𝑟𝑓,𝑡+1 to represent the risk-free rates, and

the portfolio return is written as:

𝑟𝑝,𝑡+1 = 𝑤𝑡 𝑟𝑡+1 + 𝑟𝑓,𝑡+1 . (11)

The investor realizes an average utility or certainty equivalent return (CER) as:

𝐶𝐸𝑅 = 𝑟̅𝑝 − 0.5𝛾𝜎𝑝2 , (12)

where 𝑟̅𝑝 and 𝜎𝑝2 are the mean and variance of the portfolio returns for the out-of-sample forecast

period, respectively. The CER is interpreted as the risk-free rate of return that an investor is willing to

accept instead of adopting the given risky portfolio. After calculating the CER for the historical mean

model forecast, we obtain the CER gain by the difference between the CER for an investor who uses

the predictive regression forecast to guide asset allocation and the CER for an investor who uses the

historical mean model. We also calculate the monthly Sharpe ratio, which is the mean portfolio return

in excess of the risk-free rate divided by the standard deviation of the excess portfolio return.

[Insert Table 8 Here]

The CER gains and Sharpe ratios are reported in Table 8. When we estimate using an initial sample

of 120 months, eight of the ten gold price ratios, and three of the traditional predictors have positive

CER gains: DE (0.449%), TBL (0.208%) and NTIS (0.358%). For the initial estimation period of 240

months, all of the gold price ratios have positive CER gains ranging from 1.068% to 7.022%, and three

of the traditional predictors generate negative CER gains. The average CER gain of gold price ratios is

0.311% (2.989%), which is substantially higher than the average CER gain of traditional predictors -

0.926% (1.076%), for the initial estimation sample of 120 (240) months. With respect to the Sharpe

ratio, gold price ratios also provide a higher average Sharpe ratio than traditional predictors. More

8 We use risk-aversion coefficient 𝛾=3 in this subsection. The results are similar when we take 𝛾=5.
13
interestingly, GO has the largest CER gain and Sharpe ratio among gold price ratios and traditional

predictors. For example, the CER gain and Sharpe ratio of GO reach 7.022% and 0.209, respectively,

when estimated with the second initial sample. Hence, GO performs the best among all predictors from

an asset allocation perspective. In addition, GP generates a CER gain of 5.404%, and thus, we provide

the economic significance of the predictive ability of GP in Huang and Kilic (2019).

The results listed in Table 8 demonstrate that the information provided by gold price ratios (more

specifically, GO) has substantial economic value for an investor of risk-aversion. Combined with the

results in Section 3.2, GO predicts excess returns out-of-sample with both statistical and economic

significance. The predictive information from GO is much more valuable than that from the other gold

price ratios and traditional predictors.

4. Economic explanations
We investigate the role of gold price ratios in forecasting excess stock returns and find that GO

and GP perform better than the other ratios. More importantly, GO is the most powerful return predictor

both in-sample and out-of-sample. Since Huang and Kilic (2019) have discussed the sources of return

predictability from GP, a question of interest is as follows: why does GO predict future stock returns?

In this section, we mainly investigate the economic underpinnings of the predictive ability of GO.

4.1 Stock return decomposition

The valuation models demonstrate that asset prices should equal expected discounted cash flows,

which indicates that asset prices are determined by both future expected cash flows and discount rates

(Cochrane, 2011). In this way, the predictive ability of GO to forecast aggregate returns may come from

either the cash flow channel or the discount rate channel or both. To investigate whether the economic

driving force of GO originates from one or both channels, we employ the framework of Campbell (1991)

and Campbell and Ammer (1993). We first decompose stock returns into cash flow and discount rate

news and then investigate the relationship between GO and news components.

As in Campbell and Shiller (1988), the log-linear approximation of log return 𝑟𝑡+1 is given by:

𝑅𝑡+1 ≈ 𝑘 + 𝜌𝑝𝑡+1 + (1 − 𝜌)𝑑𝑡+1 + 𝑝𝑡 , (13)

where 𝑟𝑡+1 = ln[(𝑃𝑡+1 + 𝐷𝑡+1 )⁄𝑃𝑡 ]. 𝑃𝑡 and 𝐷𝑡 represent the stock price and dividend, respectively,

and 𝑝𝑡 and 𝑑𝑡 denote the corresponding log price and log dividend. ̅̅̅̅̅̅̅
𝑑 − 𝑝 is the mean of 𝑑𝑡 − 𝑝𝑡 .
̅̅̅̅̅̅̅
The parameter 𝜌 = 1⁄[1 + 𝑒𝑥𝑝(𝑑 − 𝑝)] and 𝑘 = −𝑙𝑜𝑔(𝜌) − (1 − 𝜌)𝑙𝑜𝑔[(1⁄𝜌) − 1] . We rewrite
14
Equation (13) as:

𝑝𝑡 ≈ 𝑘 + 𝜌𝑝𝑡+1 + (1 − 𝜌)𝑑𝑡+1 − 𝑟𝑡+1 . (14)

Solving Equation (14) forward and imposing the no-bubble transversality condition

lim 𝜌 𝑗 𝑝𝑡+𝑗 = 0, the canonical stock price decomposition of Campbell and Shiller (1988) is expressed
𝑗→∞

as:
∞ ∞ 𝑘
𝑝𝑡 = ∑𝑗=0 𝜌 𝑗 (1 − 𝜌)𝑑𝑡+1+𝑗 − ∑𝑗=0 𝜌 𝑗 𝑟𝑡+1+𝑗 + . (15)
1−𝜌

Letting 𝐸𝑡 denote the expectation operator conditional on information through month 𝑡 ,

Equations (13) and (15) show the following decomposition for the log return innovation:

∞ 𝑗 𝑗 ∞
𝑟𝑡+1 − 𝐸𝑡 𝑟𝑡+1 = (𝐸
⏟ 𝑡+1 − 𝐸𝑡 ) ∑𝑗=0 𝜌 Δ𝑑𝑡+1+𝑗 − (𝐸
⏟ 𝑡+1 − 𝐸𝑡 ) ∑𝑗=0 𝜌 𝑟𝑡+1+𝑗 . (16)
𝐶𝐹
𝐶𝑎𝑠ℎ 𝑓𝑙𝑜𝑤 𝑛𝑒𝑤𝑠, 𝜂𝑡+1 𝐷𝑅
𝐷𝑖𝑠𝑐𝑜𝑢𝑛𝑡 𝑟𝑎𝑡𝑒 𝑛𝑒𝑤𝑠, 𝜂𝑡+1

Based on Equation (16), the stock return innovation can be decomposed into cash flow news and

discount rate news components:

𝑟 𝐶𝐹 𝐷𝑅
𝜂𝑡+1 = 𝑟𝑡+1 − 𝐸𝑡 𝑟𝑡+1 = 𝜂𝑡+1 − 𝜂𝑡+1 . (17)

Campbell and Shiller (1991) and Campbell and Ammer (1983) use a VAR(1) model to extract the

cash flow news and discount rate news of stock return innovations:

𝑌𝑡+1 = 𝐴𝑌𝑡 + 𝑢𝑡+1 , (18)

where 𝑌𝑡 = (𝑟𝑡 , 𝑑𝑡 − 𝑝𝑡 , 𝑥𝑡′ )′, 𝑥𝑡 is an n-vector of predictors representing a proxy for the market

information set, 𝐴 is a (𝑛 + 2) × (𝑛 + 2) matrix of VAR coefficients, and 𝑢𝑡 is an (𝑛 + 2)-vector

of zero-mean innovations. We define 𝑒1 = [1,0, … ,0] as an (𝑛 + 2)-vector, then the stock return

innovation and discount rate news component can be described as:

𝑟
𝜂𝑡+1 = 𝑒1′ 𝑢𝑡+1 , (19)
𝐷𝑅
𝜂𝑡+1 = 𝑒1′ 𝜌𝐴(𝐼 − 𝜌𝐴)−1 𝑢𝑡+1 . (20)

According to Equation (17), the cash flow news component is given by:
𝐶𝐹 𝑟 𝐷𝑅
𝜂𝑡+1 = 𝜂𝑡+1 + 𝜂𝑡+1 . (21)

In terms of Equation (18), the expected return at 𝑡 + 1 conditional on information through 𝑡 is

shown as follows:

𝐸𝑡 𝑟𝑡+1 = 𝑒1′ 𝐴𝑦𝑡 . (22)

𝐶𝐹 𝐷𝑅
Using Equation (17), the log return is decomposed as 𝑟𝑡+1 = 𝐸𝑡 𝑟𝑡+1 + 𝜂𝑡+1 − 𝜂𝑡+1 . We can

estimate the VAR(1) model via OLS and obtain 𝐸̂𝑡 𝑟𝑡+1 𝜂̂ 𝑡+1
𝑟 𝐶𝐹
, 𝜂̂ 𝑡+1 𝐷𝑅
and 𝜂̂ 𝑡+1 .
15
 We regress individual estimated components on GO as follows:
𝑍
𝑍𝑡+1 = 𝛼 + 𝛽𝑍 𝐺𝑂𝑡 + 𝜀𝑡+1 , 𝑍𝑡+1 = 𝐸̂𝑡 𝑟𝑡+1 , ̂𝜂𝑡+1
𝐶𝐹 𝐷𝑅
, 𝜂̂ 𝑡+1 . (23)

From comparisons between the estimated 𝛽̂𝑍 given in Equation (23), we can determine the extent

to which GO’s ability to predict aggregate stock returns relates to its ability to predict the individual

components. Moreover, the properties of OLS imply that the relationship between 𝛽̂ and 𝛽̂𝑍 is 𝛽̂ =

𝛽̂𝐸̂ +𝛽̂𝐶𝐹
̂ − 𝛽̂𝐷𝑅
̂ , where 𝛽̂ is the estimated coefficient for Equation (2) when GO is the predictor. We
9

take the set of traditional predictors in Section 2 as a proxy for the market information set.

[Insert Table 9 Here]

Table 9 reports the estimation results. We find that all of the estimates 𝛽̂𝐸̂ are significant. However,

they are limited in magnitude, implying that they contribute little to the size of 𝛽̂ . Nearly all of the 𝛽̂𝐷𝑅
̂

are insignificant, and the estimates in magnitude also suggest a relatively small contribution of 𝛽̂ . In

contrast, the estimated 𝛽̂𝐶𝐹

̂ are more sizable and nearly all significant. For example, the estimated 𝛽̂𝐶𝐹
̂

is 1.241 with a Newey and West (1987) t-statistic of 2.458, significant at the 5% level, when the VAR

model includes stock returns and DP. The estimated coefficients 𝛽̂𝐶𝐹
̂ indicate that the return

predictability of GO is derived from changes in expectations of cash flows. Our results for the 𝛽̂𝐶𝐹
̂

estimates remain almost the same when we include different variables proxied as different market

information sets in the VAR model. Overall, GO predicts future excess returns though the cash flow

channel, and this conclusion is robust.

The previous literature usually argues that discount rate news contributes a larger share of stock

return innovation (Cochrane, 2008). Wang et al. (2019) show that AOR predicts future excess returns

through the discount rate channel. Møller and Rangvid (2020) find that the industrial production growth

rate in the fourth quarter of a year significantly predicts global asset returns, and the predictive ability

also results from the discount rate channel. However, recent studies suggest that the return predictability

of a predictor can come from cash flow news. The predictive ability of aligned investor sentiment

studied in Huang et al. (2015), SII of Rapach et al. (2016) and cash conversion cycle of Lin and Lin

(2021) specifically originate from the cash flow channel. We also empirically uncover a powerful

predictor that forecasts aggregate returns through the cash flow channel.

9 Huang et al. (2015) suggest that it is better to use low-frequency data to avoid spurious predictability. In Robert Shiller’s
dataset, some of monthly dividends are interpolated from quarterly dividends. Therefore, we directly use quarterly data instead
of monthly data in estimating Equation (18) and (23). The OLS estimated 𝛽̂ is 1.407 with a t-statistic of 2.601 for quarterly
data. We also consider yearly data and find the return predictability comes from the cash flow channel.
16
4.2 Link to economic conditions

The relationship between GO and economic conditions may be helpful to explain the predictive

ability of GO (Fama and French, 1989; Fama, 1990; Chen et al., 2019; Wang et al., 2019; Møller and

Rangvid, 2020). We continue to examine the relationship between GO and economic indicators. We use

the following predictive regression:

𝑌𝑡+1 = 𝛼 + 𝛽𝐺𝑂𝑡 + 𝜃𝑌𝑡 + 𝜀𝑡+1 , (24)

where 𝑌𝑡 denotes one of the economic variables. We consider the following economic variables: the

smoothed recession probability (SRP), which provides reliable signals of the start of a new recession

(Chauvet and Piger, 2008); the default spread (DS); the Chicago Fed National Activity Index (CFNAI),

which gauges overall US economic activity and related inflationary pressure; the Kansas City Financial

Stress Index (FSI), a monthly measure of stress in the US financial system; and macro (MU) and

financial uncertainty (FU) proposed by Ludvigson et al. (2015).

[Insert Table 10 Here]

The estimation results of Equation (24) are shown in Table 10. We take negative values of all these

economic variables except for CFNAI based on the expected signs of 𝛽̂ . We find that GO negatively

predicts the default spread, financial stress, and macro and financial uncertainty. GO also positively

predicts the future CFNAI but the coefficient estimate is insignificant at any level. The partial-𝑅2

statistics indicate that GO retains good marginal predictive ability in the presence of the lagged

economic variable. Therefore, a higher GO indicates better economic conditions. Good economic

conditions usually suggest higher dividend growth (see Appendix Table A3), which is regarded as a

cash flow proxy (Campbell and Shiller, 1988; Fama and French, 2000; Lettau and Ludvigson, 2005;

Huang et al., 2015), the economic variables link GO to dividend growth rates, confirming the findings

shown in Section 4.1 that GO predicts future returns through the cash flow channel.

4.3 Rolling window regressions

Our results seem to contradict a long-held view among financial media and investors that increases

in GO signal future financial downturns or economic recessions. Why does this contradiction occur?

To address this question, we use a rolling window estimation for Equation (24) and examine how the

estimated 𝛽̂ changes across the full sample periods. We apply a rolling window length of 120 months.

[Insert Figure 2 Here]

The rolling window 𝛽̂ estimates are plotted in Figure 2. The solid lines denote the 𝛽̂ estimates,
17
and the dashed lines denote the 95% confidence intervals.10 Figure 2 clearly displays the signs of 𝛽̂

estimates depending on the subsample period. For example, most of the estimates for the predictive

regression with SRP as the dependent variable are negative and significant at the 5% level. However,

the signs of the 𝛽̂ estimates are positive during the 2008 financial crisis, and then an increase in GO

leads to stock market declines. When the US economy starts undergoing economic expansion in 2009,

the signs of the 𝛽̂ estimates become negative. The time-varying 𝛽̂ estimates the contradiction found

between our results and the cognitions of financial media and investors. For most of our sample, GO is

positively related to future economic conditions, and the relationship between economic indicators and

GO reverses only in periods involving extraordinarily high risks or significant economic recessions (not

all recessions or periods of high risk).

The results of rolling regressions are similar to those of the S&P 500 composite index. The 𝛽̂

estimates for Equation (2) are significantly positive for most rolling windows and insignificantly

negative for the financial crisis and COVID-19 pandemic periods. The signs of 𝛽̂ estimates do not

always change when a recession starts. As displayed in Figure 3, GO decreased during the 2nd oil crisis,

the 1981 economic recession and the Gulf War I, while the S&P 500 index decreased following GO.

During the dot-com bubble (at least before 2001M9), financial crisis and COVID-19 pandemic, the

relationship between GO and the S&P 500 index was negative. The negative relation was much more

significant during the 2008 financial crisis. The COVID-19 pandemic signaled a new recession and GO

reached its highest historical level since 1975M1, leading to worse economic conditions and stock

market declines.

[Insert Figure 3 Here]

4.4 State-switching forecasting performance

The negative relationship between GO and economic activities or aggregate stock returns is more

likely to be found in recent economic recessions or crises, during which GO experiences a sharp increase,

as shown in Section 4.3. To clearly investigate when the impact of GO changes, we employ a state-

switching regression proposed by Jacobsen et al. (2019):

𝑟𝑡+1 = 𝛼 + 𝛽𝐺𝑂𝑡 ∗ 𝑆𝑡𝑎𝑡𝑒1 + 𝜃𝐺𝑂𝑡 ∗ 𝑆𝑡𝑎𝑡𝑒2 + 𝜀𝑡+1 , (25)

where 𝑆𝑡𝑎𝑡𝑒 is a state variable determined by NBER business cycles, GO dynamics scenarios (See

10 We cannot precisely determine the signs of each estimated coefficient, so we estimate Equation (24) via OLS and calculate
the Newey and West (1987) t-statistics to test 𝐻0 : 𝛽 = 0 against 𝐻𝐴 : 𝛽 ≠ 0.
18
more details in Appendix Table A4) and thresholds of GO and GO’s changes ΔGO. For example, when

we use NBER expansion and recession as state variables, 𝑆𝑡𝑎𝑡𝑒1 = 1 and 𝑆𝑡𝑎𝑡𝑒2 = 0 if the US

economy is in an expansion, and 𝑆𝑡𝑎𝑡𝑒1 = 0 and 𝑆𝑡𝑎𝑡𝑒2 = 1 if the US economy is in a recession.

[Insert Table 11 Here]

The estimation results for state-switching regressions are reported in Table 11. Unfortunately, it is

quite difficult to uncover the negative relationship between GO and stock returns from the state-

switching regressions. For example, GO positively and significantly predicts future stock market returns

during economic expansions and recessions as shown in Panel A. GO also has a positive predictive

ability when GO is beyond its 90%, 95% and 99% quantiles. Interestingly, we only find a negative but

insignificant relationship between GO and stock returns for Scenarios 5 and 6 as listed in Appendix

Table A4, when gold prices increase and oil prices decline. The estimated coefficient is also negative

for combined scenario 5&6. Such sample periods perfectly correspond to the financial crisis and

COVID-19 pandemic.

Moreover, the estimated 𝛽̂ for Scenarios 1&2 is 0.641, positive and significant at the 1% level.

The estimated coefficient for Scenario 1&2&3&4 is 0.684, significant at the 1% level. Scenarios 1, 2,

3 and 4 indicate that gold and oil prices simultaneously change. A strand of literature finds a hedge role

of crude oil (Arouri et al., 2011; Junttila et al., 2018; Battern et al., 2021). Therefore, oil may partially

isolate the hedge or safe have role of gold and enhance the information of gold that is strongly related

to economic fundamentals (Huang and Kilic, 2019). Only when gold prices increase and oil prices

decrease does GO signal a bad stock market performance.

We believe that the reversed effect (Scenarios 5 and 6) may only occur for a short period and fails

to cover sufficient samples for estimation. The negative relationship may be concealed by the positive

relationship of the full sample. For investment implications, investors may consider adjusting their asset

allocations to avoid risk when they observe an extremely higher GO (or an increase in gold prices and

a decrease in oil prices simultaneously).

4.5 Learning

McLean and Pontiff (2016) argue that the predictive ability of a variable becomes weaker over

after it is considered in academic publications. Since many studies investigate the relationship between

stock returns, gold prices and oil prices, we are interested in the question whether there is evidence that

investors gradually learn of the predictive ability of GO over time. We use the predictive regression
19
described in Bakshi et al. (2013):

𝑟𝑡+1 = 𝛼 + (𝛽 + 𝛽 𝑇𝑟𝑒𝑛𝑑 𝑇𝑟𝑒𝑛𝑑)𝐺𝑂𝑡 + 𝜀𝑡+1 , (26)

where Trend is a variable computed from an observation number series. Because GO positively predicts

future returns, there should be evidence of learning if the 𝛽̂ 𝑇𝑟𝑒𝑛𝑑 estimate is significant and negative.

We find that the estimated 𝛽̂ 𝑇𝑟𝑒𝑛𝑑 is 0.005 with a Newey and West (1987) t-statistic of 0.126, which

is insignificant at any level. The estimation results suggest no evidence of learning.

5. Robustness checks
5.1 Alternative estimation procedures

We consider three alternative estimation procedures for Equation (2) to examine whether the

results are robust to other estimation methods. (1) The first procedure is the weighted least squares

method using ex ante variance (WLS-EV) proposed by Johnson (2019). Heteroscedasticity in return

predictability regressions is usually addressed using White (1980) heteroscedasticity-consistent

standard errors. However, Westerlund and Narayan (2014) document that using generalized least

squares (GLS) results in a more efficient estimator that is less noisy and has more power for finite

samples. Following Johnson (2019), we first use AR(1) to estimate the conditional variance of the next
̂ 𝑡2 and then estimate 𝛽̂ via weighted least squares with 𝜎̂𝑡 used
month’s unexpected returns 𝜎̂𝑡2 = 𝑅𝑉

as the weights. (2) The second approach is a reduced-bias estimation method (RBM) proposed by

Amihud and Hurvich (2004). Stambaugh (1999) indicates that the coefficient estimate of the predictive

regression can be biased in small samples when the predictor is persistent and correlated with excess

returns. We address the small-sample bias by using a reduced-bias estimator from Amihud and Hurvich

(2004). The reduced-bias estimator is obtained by estimating an augmented regression and adding a

proxy for the errors in AR(1) for predictors. (3) The third approach is the Wald test (IVX-Wald)

proposed by Kostakis et al. (2015), which is a powerful Wald test that is robust to the predictor’s degree

of persistence (unit root, local-to-unit root, near stationary, or stationary). The Wald test is necessary

because gold price ratios are highly persistent.

[Insert Table 12 Here]

Table 12 reports the estimation results obtained from WLS-EV, ARM and IVX. The estimated

coefficients of GO, GS, GFR and GP are still positive and significant. The predictive ability of these

gold price ratios is a statistical artifact of their high persistence. The three alternative estimation
20
procedures confirm the robustness of our results.

5.2 Controlling for new return predictors

As discussed in Section 3, we examine whether the predictive ability of gold price ratios is still

significant after controlling for traditional predictors using Equation (3). In fact, our results are robust

to inclusions of return predictors that are recently proposed. The new return predictors are shown as

follows: the short interest index (SII) in Rapach et al. (2016), news-based implied volatility (NVIX) in

Manela and Moreira (2017), aligned investor sentiment (AIS) in Huang et al. (2015), and the

asymmetric oil price return (AOR) in Wang et al. (2019). The sample periods for the SII, NVIX, AIS

and AOR are 1975M1-2014M12, 1975M1-2016M3, 1975M1-2018M12 and 1975M1-2019M12,

respectively. Previous literature shows that these new predictors negatively forecast future excess

aggregate returns. Therefore, we take their negative values to test 𝐻0 : 𝜃 = 0 against 𝐻𝐴 : 𝜃 > 0.

[Insert Table 13 Here]

As shown in Table 13, the estimated coefficients of GO and GP are significant after controlling for

these new predictors. Some gold price ratios that significantly predict future returns, such as GS, GYD

and GFR, do not have significant predictive ability. In sum, directly controlling for new predictors from

the recent literature does not affect the predictive ability of GO and GP.

5.3 Predictive ability controlling for other gold price ratios

We indicate that GO and GP contain the information that is distinctive from the information of

traditional predictors in Section 3.1 using a bivariate regression as shown in Equation (3), and also

reveal that GO is arguably the most powerful predictor among all the gold price ratios using out-of-

sample forecasting evaluations. However, we do not show whether the predictive ability of GO remain

significant after controlling for other gold price ratios. Therefore, we use a bivariate regression to test

whether GO outperforms other gold price ratios. The estimation results are shown in Appendix Table

A5. We find that the 𝛽̂ estimates for GO remain significant after controlling for other gold price ratios.

Although the Pearson correlation coefficient is 0.660 for GO and GP, the 𝛽̂ estimate is still significant

at the 10% level when using GP as a control variable. The results in Appendix Table A5 confirm that

GO is the most powerful predictor among gold price ratios.

5.4 Principal components of gold price ratios

In this subsection, we employ a principal component analysis to extract information from these ten

gold price ratios and examine the predictive ability of the components. Because the first three
21
components explain more than 85% of the variation in the gold price ratios, we mainly consider the

predictive ability of the first three components (PC1/2/3). The loadings of the three components are

plotted in Appendix Figure A1. Table A6 reports the in-sample and out-of-sample forecasting results.

The first principal component significantly predicts excess returns with an estimated 𝛽̂ of 0.172% at
2
the 5% significance level and generates a 𝑅𝑂𝑂𝑆 statistic of 0.772% for the first initial estimation period.

The first component has stronger predictive ability than traditional predictors. However, the second and

third components fail to significantly outperform the historical mean model in out-of-sample forecasting.

These results can be explained by the loadings of the three components. The first component has higher

loadings on GO, GS and GP. These three ratios show the most stable predictive ability for returns and

thus improve the predictive power of PC1. GO already presents much greater predictive ability than

other gold price ratios, and incorporating information on other gold price ratios in a component actually

destroys the predictive ability of GO. In this way, including GO in predictive regressions is already

sufficient to generate good predictive performance.11

5.4 Is GO driven by oil prices?

Some readers may conjecture that the predictive ability of GO for economic conditions is mainly

derived from incorporation with oil prices because oil prices are found to be negatively associated with

economic activities by a large body of literature (Darby, 1982; Hamilton, 1983; Mork, 1989; Davis and

Haltiwanger, 2001; Killian, 2009; Oladosu, 2009; Killian and Park, 2009). If this is the case, we can

directly predict stock returns using indicators that are closely related to oil prices and considering gold

priced in oil may not be necessary. Such a concern can be alleviated by the following reasons. First, the

correlation coefficient for GO and the natural logarithm of gold prices is 0.388, whereas the coefficient

for GO and the natural logarithm of oil prices is -0.187. The information of GO is more related to gold

prices in US dollars. Second, GO is a powerful predictor and contains information that is substantially

different from the AOR of Wang et al. (2019). Third, some recent works indicate that the predictive

ability of oil prices for economic activities is generally weaker than it seems to be (Hamilton, 2011;

Killian and Vigfusson, 2013; Ravazzolo and Rothman, 2013). Finally, we show how gold and oil prices

11 We also use the partial least squares (PLS) described in Huang et al. (2015) to extract predictive information contained in
GPR. However, the variable constructed by PLS does not perform better than GO, GP or the first principal component. Huang
et al. (2015) employ PLS to extract an aligned investor sentiment from several sentiment indicators considered in Baker and
Wurgler (2006). However, the information from gold price ratios used in this paper may be different, indicating that we cannot
determine what the extracted information represents. Based on its predictive ability, the extracted indicator may not contain
information on return predictability.
22
drive the movement of GO in Appendix Table A4. If the absolute change in the gold price (oil price) is

greater than that of the oil price (gold price), we assume that it is the gold price (oil price) that mainly

drives the movement of GO. As displayed in Appendix Table A4, there are only 59 days on which GO

is mainly driven by oil price changes. Gold and oil prices jointly determine the movement of GO.

6. Conclusion
We empirically investigate the predictive ability of ten gold price ratios for US excess stock returns.

Gold price ratios are constructed as the natural logarithm of gold to other asset prices. We find that gold

price ratios positively predict future stock returns and have higher predictive ability than traditional

predictors studied in Welch and Goyal (2008) on average. Among these ratios, the gold-oil ratio (GO)

is the most powerful return predictor, and the information contained in GO does not overlap with that

contained in traditional predictors and other gold price ratios. A one-standard-deviation increase in GO

is associated with a 6.60% increase in the annual excess return for the next month in-sample. GO also

significantly outperforms the historical mean model out-of-sample and generates substantial economic

gains for a mean-variance investor. Therefore, the predictive ability of GO is both statistically and

economically significant.

Through stock return decomposition using the VAR framework of Campbell (1991) and Campbell

and Ammer (1993), we demonstrate that the predictive ability of GO mainly originates from GO’s

anticipation of aggregate cash flow news. The predictive power of GO for economic variables also

confirms the channel of cash flow. Using rolling window and state-switching regressions, we find that

the relationship between economic indicators and lagged GO is reversed when GO is extremely high

during periods when gold prices increase and oil prices decrease (for example, the financial crisis and

COVID-19 pandemic). However, such relationship is found to be insignificant. When gold and oil

prices change simultaneously, oil may isolate the hedge or safe haven role of gold and reveal the

information of gold that is related to economic fundamentals.

Our paper has meaningful implications for investors, policymakers and academics. Investors may

make decisions or adjust their asset allocations based on the movements of GO. It may not be necessary

to pay more attention to increases in GO unless GO is driven by increases in gold prices and decreases

in oil prices. Policymakers may consider GO a reliable leading indicator of future economic

environment and prepare for policy interventions when there is a decrease in GO. Academics may
23
reconsider the impact of gold price ratios, instead of the dollar price of gold, on economies and financial

markets.

References
Amihud, Y., Hurvich, C.M., 2004. Predictive regression: A reduced-bias estimation method. Journal of

Financial and Quantitative Analysis 39: 813-841.

Ang, A., Bekaert, G., 2007. Stock return predictability: is it there? Review of Financial Studies 20(3):

651-707.

Arouri, M.E.H., Jouini, J., Nguyen, D.K., 2011. Volatility spillovers between oil prices and stock sector

returns: Implications for portfolio management. Journal of International Money and Finance 30:

1387-1405.

Baker, M., Wurgler, J., 2006. Investor sentiment and the cross-section of stock returns. Journal of

Finance 61: 1645-1680.

Bakshi, G., Panayotov, G., Skoulakis, G., 2013. In search of explanation for the predictive ability of the

Baltic dry index for global stock returns, commodity returns, and global economic activity.

University of Maryland Working Paper.

Battern, J.A., Kinateder, H., Szilagyi, P.G., Wagner, N.F., 2021. Hedging stocks with oil. Energy

Economics 93: 104422.

Baur, D.G., Lucey, B.M., 2010. Is gold a hedge or a safe haven? An analysis of stocks, bonds and gold.

The Financial Review 45: 217-229.

Baur, D.G., McDermott, T.K., 2010. Is gold a safe haven? International evidence. Journal of Banking

and Finance 34: 1886-1898.

Baur, D.G., Beckmann, J., Czudaj, R.J., 2020. The relative valuation of Gold. Macroeconomic

Dynamics 24(6): 1346-1391.

Bekiros, S., Boubaker, S., Nguyen, D.K., Uddin, G.S., 2017. Black swan events and safe havens: the

role of gold in globally integrated emerging markets. Journal of International Money and Finance

73: 371-334.

Capie, F., Mills, T.C., Wood, G., 2005. Gold as a hedge against the dollar. International Journal of

Financial Markets, Institutions and Money 15: 343-352.

Campbell, J.Y., Shiller, R.J., 1988. The dividend-price ratio and expectations of future dividends and
24
 discount factors. Review of Financial Studies 1(3): 195-228.

Campbell, J.Y., Thompson, S.B., 2008. Predicting excess stock returns out of sample: can anything beat

the historical average? Review of Financial Studies 21(4): 1509-1531.

Chauvet, M., Piger, J., 2008. A comparison of the real-time performance of business cycle dating

methods. Journal of Business and Economic Statistics 26: 42-49.

Chen, J., Jiang, F., Xue, S., Yao, J., 2019. The world predictive power of U.S. equity market skewness

risk. Journal of International Money and Finance 96: 210-227.

Clark, T.E., West, K.D., 2007. Approximately normal tests for equal predictive accuracy in nested

models. Journal of Econometrics 138: 291-311.

Cochrane, J.H., 2008. The dog that did not bark: A defense of return predictability. Review of Financial

Studies 21(4): 1533-1575.

Cochrane, J.H., 2011. Presidential address: Discount rates. Journal of Finance 66(4): 1047-1108.

Driesprong, G., Jacobsen, B., Maat, B., 2008. Striking oil: Another puzzle? Journal of Financial

Economics 89: 307-327.

Fama, E.F., 1990. Stock returns, expected returns, and real activity. Journal of Finance 45(4): 1089-

1108.

Fama, E.F., French, K., 1989. Business conditions and expected returns on stocks and bonds. Journal

of Financial Economics 25: 23-49.

Fama, E.F., French, K., 2000. Forecasting profitability and earnings. Journal of Business 73: 161-175.

Hamilton, J.D., 1983. Oil and the macroeconomy since World War II. Journal of Political Economy 91:

228-248.

Henkel, S.J., Martin, J., Nardari, F., 2011. Time-varying short-horizon predictability. Journal of

Financial Economics 99: 560-580.

Huang, D., Kilic, M., 2019. Gold, platinum, and expected stock returns. Journal of Financial Economics

132: 50-75.

Huang, D., Jiang, F., Tu, J., Zhou, G., 2015. Investor sentiment aligned: A powerful predictor of stock

returns. Review of Financial Studies 28(3): 791-837.

Johnson, T.L., 2019. A fresh look at return predictability using a more efficient estimator. Review of

Asset Pricing Studies 9(1): 1-46.

Junttila, J., Pesonen, J., Raatikainen, J., 2018. Commodity market based hedging against stock market
25
 risk in times of financial crisis: The case of crude oil and gold. Journal of International Financial

Markets, Institutions and Money 56: 255-280.

Kandel, S., Stambaugh, R., 1996. On the predictability of stock returns: An asset allocation perspective.

Journal of Finance 51: 385-424.

Katz, J., Holmes, F., 2012. The Goldwatcher: Demstifying Gold Investing (John Wiley & Sons, New

York).

Killian, L., 2009. Not all oil price shocks are alike: disentangling demand and supply shocks in the

crude oil market. American Economic Review 99(3): 1053-1069.

Killian, L., Park, C., 2009. The impact of oil prices shocks on the US stock market. International

Economic Review 50: 1267-1287.

Killian, L., Vigfusson, R.J., 2013. Do oil prices help forecast U.S. real GDP? The role of nonlinearities

and asymmetries. Journal of Business and Economic Statistics 31(1): 78-93.

Manela, A., Moreira, A., 2017. News implied volatility and disaster concerns. Journal of Financial

Economics 123(1): 137-162.

McCracken, M.W., 2007. Asymptotics for out of sample tests of Granger causality. Journal of

Econometrics 140: 719-752.

Møller, S.V., Rangvid, J., 2018. Global economic growth and expected returns around the World: The

end-of-the-year effect. Management Science 64(2): 573-591.

Mork, K.A., 1989. Oil and the macroeconomy when prices go up and down: An extension of Hamilton’s

results. Journal of Political Economy 97(3): 740-744.

Neely, C.J., Rapach, D.E., Tu, J., Zhou, G., 2014. Forecasting the equity risk premium: the role of

technical indicators. Management Science 60(7): 1772-1791.

Newey, W., West, K., 1987. A simple, positive semi-definite, heteroskedasticity and autocorrelation

consistent covariance matrix. Econometrica 59(3): 817-858.

Rapach, D.E., Ringgenberg, M.C., Zhou, G., 2016. Short interest and aggregate stock returns. Journal

of Financial Economics 121(1): 46-65.

Rapach, D.E., Strauss, J.K., Zhou, G., 2010. Out-of-sample equity premium prediction: combination

forecasts and links to the real economy. Review of Financial Studies 23: 821-862.

Ravazzolo, F., Rothman, P., 2013. Oil and U.S. GDP: A real-time out-of-sample examination. Journal

of Money, Credit and Banking 45: 449-463.

26
Reboredo, J.C., 2013. Is gold a safe haven or a hedge for the US dollar? Implications for risk

management. Journal of Banking and Finance 37: 2665-2676.

Stambaugh, R.F., 1999. Predictive regressions. Journal of Financial Economics 54: 375-421.

Wang, Y., Pan, Z., Liu, L., Wu, C., 2019. Oil price increases and the predictability of equity premium.

Journal of Banking and Finance 102: 43-58.

Welch, I., Goyal, A., 2008. A comprehensive look at the empirical performance of equity premium

prediction. Review of Financial Studies 21(4): 1455-1508.

Westerlund, J., Narayan, P., 2014. Testing for predictability in conditionally heteroskedastic stock

returns. Journal of Financial Econometrics 13: 342-375.

White, H., 1980. A heteroskedasticity-consistent covariance matrix estimator and a direct test for

heteroskedasticity. Econometrica 48: 817-838.

27
Table 1
Descriptive statistics
Mean. Std. Min. Max. AR(1) ADF-p Sample
Panel A: Gold price ratios
GO 2.738 0.383 1.893 4.620 0.977 0.020 1975M01-2020M09
GS 4.064 0.318 2.843 4.716 0.985 0.072 1975M01-2020M09
GCPI 1.205 0.436 0.388 2.160 0.998 0.848 1975M01-2020M09
GCO 5.116 0.497 3.677 6.404 0.989 0.223 1975M01-2020M09
GCP 5.963 0.315 5.047 6.599 0.981 0.033 1975M01-2020M09
GDJ -2.287 0.823 -3.764 -0.244 0.996 0.702 1975M01-2020M09
GYD 4.911 0.646 3.505 6.202 0.999 0.904 1975M01-2020M09
GTB 4.543 1.210 2.651 8.017 1.004 0.991 1975M01-2020M09
GFR 5.299 2.100 3.032 10.443 1.000 0.939 1975M01-2020M09
GP -0.152 0.320 -0.850 0.796 0.998 0.663 1975M01-2020M09
Panel B: Traditional predictors
DP -3.638 0.444 -4.491 -2.710 0.995 0.629 1975M01-2020M09
DY -3.660 0.432 -4.509 -2.773 0.994 0.667 1975M01-2020M09
PE 2.946 0.467 1.893 3.789 0.996 0.659 1975M01-2020M09
DE -0.795 0.333 -1.244 1.380 0.986 0.000 1975M01-2020M09
TBL 0.044 0.035 0.000 0.163 0.993 0.347 1975M01-2020M09
INFL 0.294 0.363 -1.934 1.509 0.615 0.085 1975M01-2020M09
LTY 6.181 3.251 0.620 15.320 0.998 0.870 1975M01-2020M09
DS 1.097 0.453 0.550 3.380 0.959 0.000 1975M01-2020M09
TS 1.647 1.224 -2.650 4.150 0.954 0.000 1975M01-2020M09
BM 0.453 0.278 0.121 1.207 0.993 0.546 1975M01-2019M12
NTIS 0.005 0.020 -0.058 0.046 0.981 0.009 1975M01-2019M12
SVAR 0.254 0.570 0.015 7.474 0.378 0.000 1975M01-2020M09
Panel C: Excess stock market returns (%)
S&P500 0.344 3.667 -22.860 11.339 0.240 0.000 1975M01-2020M09
Notes: This table reports descriptive statistics for gold price ratios, traditional predictors in Welch and
Goyal (2008) and stock market returns. AR(1) denotes the autoregression coefficient estimate. ADF-p
is the p-value for ADF unit-root tests. The sample period for BM and NTIS spans from 1975M01 to
2019M12, and the sample period of the rest predictors is from 1975M01 to 2020M09.

28
Table 2
Predictor correlations
GO GS GCPI GCO GCP GDJ GYD GTB GFR GP
Panel A: Correlations for traditional predictors and gold price ratios
DP 0.002 -0.472 0.245 -0.397 0.173 0.933 -0.182 -0.522 -0.453 0.315
DY 0.013 -0.469 0.249 -0.394 0.173 0.933 -0.178 -0.519 -0.449 0.323
PE 0.138 0.586 -0.238 0.429 -0.171 -0.933 0.199 0.530 0.423 -0.192
DE 0.222 0.263 -0.025 0.033 0.275 0.143 -0.001 -0.047 0.021 0.185
TBL -0.193 -0.476 -0.131 -0.499 0.081 0.645 -0.452 -0.750 -0.799 -0.016
INFL -0.182 -0.445 0.054 -0.289 -0.084 0.377 -0.188 -0.326 -0.334 -0.109
LTY -0.227 -0.469 -0.167 -0.567 0.116 0.716 -0.504 -0.823 -0.785 -0.010
DS -0.066 -0.262 0.300 -0.061 0.386 0.538 0.099 -0.107 -0.046 0.193
TS -0.067 0.158 -0.064 -0.060 0.063 0.019 -0.022 0.008 0.256 0.021
BM -0.237 -0.715 0.180 -0.495 0.075 0.852 -0.274 -0.565 -0.477 0.146
NTIS -0.103 -0.194 -0.368 -0.448 -0.117 0.067 -0.490 -0.450 -0.313 -0.089
SVAR 0.097 0.114 0.047 0.117 0.134 -0.045 0.092 0.124 0.105 0.024
ACC 0.129 0.385 0.172 0.316 0.154 0.517 0.223 0.398 0.372 0.135
Panel B: Correlation for gold price ratios
GO 1.000
GS 0.451 1.000
GCPI 0.389 -0.063 1.000
GCO 0.453 0.551 0.613 1.000
GCP 0.506 0.196 0.446 0.438 1.000
GDJ 0.072 -0.493 0.472 -0.208 0.322 1.000
GYD 0.406 0.317 0.868 0.862 0.413 0.042 1.000
GTB 0.391 0.442 0.633 0.844 0.248 -0.354 0.893 1.000
GFR 0.300 0.354 0.605 0.729 0.219 -0.277 0.827 0.929 1.000
GP 0.660 0.174 0.548 0.280 0.526 0.337 0.430 0.308 0.245 1.000
Notes: This table reports the Pearson correlation coefficients for gold price ratios and traditional
predictors. Panel A reports correlation coefficients of gold price ratios and traditional predictors. Panel
B reports correlation coefficients of gold price ratios. ACC is the average absolute correlation coefficient
(ACC), which is calculated as the average of absolute value of correlation coefficients. ACC is used to
measure the correlation in magnitude.

29
Table 3
Univariate US stock market return predictability
𝛽(%) t-stat 𝑅2 (%) ̂
𝑞𝐿𝐿
Panel A: Gold price ratios
GO 0.550*** 2.923 2.225 -3.603
GS 0.290** 1.805 0.630 -1.515
GCPI 0.188 1.081 0.263 -4.293
GCO 0.231* 1.369 0.396 -4.598
GCP 0.232* 1.327 0.401 -3.071
GDJ 0.067 0.363 0.033 -5.097
GYD 0.224 1.330 0.373 -3.020
GTB 0.273* 1.591 0.549 -2.105
GFR 0.312** 1.960 0.720 -2.584
GP 0.468*** 2.679 1.614 -2.714
Panel B: Traditional predictors
DP 0.050 0.284 0.019 -4.740
DY 0.124 0.698 0.115 -5.110
PE 0.002 0.014 0.000 -3.947
DE 0.055 0.227 0.023 -1.324
TBL(-) 0.246 1.393 0.450 -2.332
INFL(-) 0.355** 1.796 0.941 -4.914
LTY(-) 0.288* 1.579 0.619 -1.912
DS 0.065 0.269 0.032 -3.539
TS 0.150 0.842 0.169 -6.009
BM(-) 0.141 0.767 0.143 -2.616
NTIS(-) 0.066 0.296 0.033 -8.583**
SVAR(-) 0.540 1.463 2.184 -5.820
Panel C: Kitchen sink model: gold price ratios and traditional predictors
Kitchen sink - - 14.808 -
Notes: This table reports the estimation results for the in-sample predictive regression:
𝑟𝑡+1 = 𝛼 + 𝛽𝑋𝑡 + 𝜀𝑡+1 ,
where 𝑟𝑡+1 is the log stock market return in excess of the risk-free rate at month 𝑡, 𝑋𝑡 is one of gold
price ratios and return predictors, and (-) indicates that we take the negative value of the predictor. Each
predictor is standardized. We use the Newey and West (1987) heteroskedasticity and autocorrelation
robust t-statistics and compute a wild-bootstrapped p-value for testing 𝐻0 : 𝛽 = 0 against 𝐻𝐴 : 𝛽 > 0.
***, ** and * indicate significance at the 1%, 5% and 10% level, respectively.

30
Table 4
Bivariate US stock market return predictability
𝛽(%) DP DY PE DE TBL(-) INFL(-)
GO 0.549*** 0.548*** 0.559*** 0.565*** 0.522*** 0.501***
GS 0.404** 0.446*** 0.440** 0.296** 0.224 0.165
GCPI 0.187 0.167 0.201 0.190 0.159 0.209
GCO 0.298* 0.331** 0.281* 0.230* 0.144 0.140
GCP 0.230 0.217 0.240 0.235 0.255** 0.204
GDJ 0.152 -0.380 0.537 0.060 0.384 0.233
GYD 0.241* 0.254* 0.232* 0.224 0.142 0.163
GTB 0.411** 0.462** 0.377** 0.276* 0.200 0.175
GFR 0.420** 0.459*** 0.378** 0.311** 0.318 0.217*
GP 0.504*** 0.479*** 0.488*** 0.473*** 0.465*** 0.434***
𝛽(%) LTY(-) DS TS BM(-) NTIS(-) SVAR(-)
GO 0.511*** 0.556*** 0.561*** 0.554*** 0.561*** 0.607***
GS 0.199 0.330** 0.273* 0.363* 0.279** 0.356***
GCPI 0.145 0.185 0.198 0.235* 0.190 0.213
GCO 0.100 0.236* 0.240* 0.212 0.255 0.298**
GCP 0.271* 0.243 0.223 0.250* 0.224 0.309**
GDJ 0.561** 0.045 0.064 0.664** 0.060 0.042
GYD 0.106 0.220 0.227 0.208 0.260* 0.275*
GTB 0.107 0.283* 0.271* 0.272* 0.306** 0.345**
GFR 0.221 0.315** 0.292** 0.260* 0.277* 0.372**
GP 0.467*** 0.473*** 0.464*** 0.553*** 0.485*** 0.480***
Notes: This table reports the coefficient estimates for the bivariate predictive regression:
𝑟𝑡+1 = 𝛼 + 𝛽𝐺𝑃𝑅𝑡 + 𝜃𝑋𝑡 + 𝜀𝑡+1 .
̂
The 𝛽 estimates and significance levels are shown in this table, and the 𝜃̂ estimates are found in
Appendix. (-) indicates that we take the negative value of the predictor. Each predictor is standardized.
We use the Newey and West (1987) heteroskedasticity and autocorrelation robust t-statistics and
compute a wild-bootstrapped p-value for testing 𝐻0 : 𝛽 = 0 (𝜃 = 0) against 𝐻𝐴 : 𝛽 > 0 (𝜃 > 0). ***,
** and * indicate significance at the 1%, 5% and 10% level, respectively.

31
Table 5
Out-of-sample forecasting results
Initial sample = 120 months Initial sample = 240 months
2
𝑅𝑂𝑂𝑆 (%) MSFE-adj ENCT-𝜆̂ 2
𝑅𝑂𝑂𝑆 (%) MSFE-adj ENCT-𝜆̂
Panel A: Gold price ratios
GO 1.571 2.230** - 3.038 2.610*** -
GS 0.174 1.155 0.890** 0.544 1.228 1.000***
GCPI -0.130 0.290 0.860*** -0.153 0.293 1.000***
GCO 0.049 0.495 0.921*** 0.342 0.999 1.000***
GCP 0.264 0.914 0.869** 0.384 1.060 1.000***
GDJ -0.476 -0.056 0.854*** -0.448 0.144 1.000***
GYD -0.087 0.223 0.887*** -0.069 0.286 1.000***
GTB -0.405 0.461 0.884*** -0.325 0.514 1.000***
GFR -1.405 0.416 1.000*** -0.629 0.318 1.000***
GP 1.113 2.266** 0.608* 1.131 2.102** 1.000**
Panel B: Traditional predictors
DP -1.168 -0.961 0.950*** -0.657 -0.184 1.000***
DY -1.639 -1.057 0.908*** -0.858 -0.830 1.000***
PE -1.276 -1.296 0.963*** -0.469 -0.173 1.000***
DE -1.689 -0.207 1.000*** -1.618 -0.431 1.000***
TBL -0.246 0.538 0.879*** -0.132 0.396 1.000***
INFL -0.224 1.457* 0.738*** -1.220 0.458 1.000***
LTY -0.271 1.059 0.813*** -0.307 0.743 1.000***
DS -1.657 -1.334 0.952*** -1.731 -2.792 1.000***
TS -1.386 -0.849 0.972*** -0.897 -1.003 1.000***
BM -0.506 0.124 0.980*** -0.066 0.466 1.000***
NTIS -1.912 0.741 0.853*** -2.129 -1.934 1.000***
SVAR -9.643 -0.757 1.000*** -2.268 1.039 0.725***
Notes: This table reports the out-of-sample forecasting evaluations of gold price ratios and traditional
2
predictors. 𝑅𝑂𝑂𝑆 is the out-of-sample 𝑅2 as shown in Equation (5), MSFE-adj is the Clark and West
(2007) adjusted mean squared forecast errors, and ENCT-𝜆̂ is the estimated coefficient for Equation (7)
in forecast encompassing test. We consider two samples for initial estimations, the first sample is from
1975M1 to 1985M12 (120 months), and the second sample spans from 1975M1 to 1994M12 (240
months). ***, ** and * indicate significance at the 1%, 5% and 10% level, respectively.

32
Table 6
Out-of-sample forecasting over business cycles
Initial sample = 120 months Initial sample = 240 months
2 2 2 2
𝑅𝑂𝑂𝑆,𝐸𝑋 (%) 𝑅𝑂𝑂𝑆,𝑅𝐸 (%) 𝑅𝑂𝑂𝑆,𝐸𝑋 (%) 𝑅𝑂𝑂𝑆,𝑅𝐸 (%)
Panel A: Gold price ratios
GO 1.942 3.357 2.693 5.081
GS 2.512 -1.418 2.246 -0.217
GCPI 1.690 -0.815 0.697 0.281
GCO 1.924 -0.735 1.502 0.336
GCP 1.929 -0.162 1.159 0.913
GDJ 1.064 -0.666 0.017 0.524
GYD 1.942 -1.136 1.132 -0.124
GTB 1.967 -2.039 1.395 -1.099
GFR 0.218 -1.716 0.723 -0.886
GP 2.398 1.326 1.697 1.938
Panel B: Traditional predictors
DP -0.222 -0.312 -0.490 0.732
DY -1.389 0.433 -1.221 1.270
PE -0.177 -0.682 0.021 0.471
DE 0.531 -3.028 -0.148 -2.022
TBL 1.974 -1.620 1.537 -0.838
INFL 4.001 -5.075 2.535 -4.805
LTY 2.454 -2.521 1.861 -1.702
DS -0.655 -0.887 -1.304 -0.686
TS -0.486 -0.446 -0.501 0.179
BM 1.371 -1.280 1.119 -0.099
NTIS 0.556 -3.675 -0.271 -3.065
SVAR -8.065 -6.440 1.542 -5.912
2
Notes: This table reports the out-of-sample forecasting results over NBER business cycles. 𝑅𝑂𝑂𝑆,𝐸𝑋
2
and 𝑅𝑂𝑂𝑆,𝑅𝐸 denote the out-of-sample 𝑅2 during economic expansions and recessions, respectively,
as shown in Equation (8). We consider two samples for initial estimations, the first sample is from
1975M1 to 1985M12 (120 months), and the second sample spans from 1975M1 to 1994M12 (240
months).

33
Table 7
In-sample forecasting with longer horizons
h=3 h=6 h=12
2 2
𝛽(%) t-stat 𝑅 (%) 𝛽(%) t-stat 𝑅 (%) 𝛽(%) t-stat 𝑅2 (%)
GO 0.551*** 3.180 5.095 0.519** 2.690 7.882 0.477** 2.737 12.003
GS 0.330** 2.791 1.901 0.339*** 3.161 3.774 0.317** 2.999 6.102
GCPI 0.181 1.176 0.566 0.156 1.031 0.791 0.142 0.835 1.220
GCO 0.267* 1.938 1.212 0.275** 2.454 2.386 0.203 1.849 2.352
GCP 0.272* 1.635 1.276 0.279 1.552 2.524 0.252 1.380 3.807
GDJ 0.070 0.424 0.087 0.063 0.368 0.132 0.058 0.324 0.214
GYD 0.237 1.622 0.968 0.230 1.690 1.706 0.233 1.701 3.209
GTB 0.275* 1.887 1.266 0.245* 1.857 1.849 0.232* 1.816 2.985
GFR 0.307** 2.199 1.615 0.291** 2.364 2.684 0.298** 2.809 5.263
GP 0.504*** 3.102 4.271 0.506** 2.838 7.822 0.504** 2.704 13.875
Notes: This table reports the estimation results for the in-sample predictive regression with longer
horizons:
1
∑ℎ𝑖=1 𝑟𝑡+𝑖 = 𝛼 + 𝛽𝑋𝑡 + 𝜀𝑡+ℎ ,
ℎ

where ℎ = 3,6,12 months is the forecast horizon. Each gold price ratio is standardized. We use the
Newey and West (1987) heteroskedasticity and autocorrelation robust t-statistics and compute a wild-
bootstrapped p-value for testing 𝐻0 : 𝛽 = 0 against 𝐻𝐴 : 𝛽 > 0. ***, ** and * indicate significance at
the 1%, 5% and 10% levels, respectively.

34
Table 8
Asset allocation implications
Initial sample = 120 months Initial sample = 240 months
CER gain (%) SR CER gain (%) SR
Panel: Gold price ratios
GO 2.511 0.156 7.022 0.209
GS -0.220 0.099 2.258 0.107
GCPI 1.623 0.139 1.400 0.095
GCO 0.273 0.109 1.068 0.086
GCP 0.195 0.107 1.636 0.095
GDJ -3.132 0.039 2.061 0.103
GYD 1.466 0.135 1.549 0.096
GTB -0.312 0.100 3.248 0.127
GFR 0.131 0.108 4.247 0.149
GP 0.570 0.115 5.404 0.174
Average 0.311 0.111 2.989 0.124
Panel B: Traditional predictors
DP -1.748 0.062 1.432 0.091
DY -2.016 0.053 1.712 0.095
PE -3.063 0.034 0.119 0.061
DE 0.449 0.113 0.873 0.074
TBL 0.208 0.107 2.562 0.112
INFL -0.531 0.093 0.280 0.071
LTY -0.611 0.097 3.164 0.125
DS -0.270 0.095 -1.122 0.029
TS -0.449 0.093 -0.987 0.026
BM -1.970 0.061 0.526 0.069
NTIS 0.358 0.111 -0.172 0.041
SVAR -1.463 0.074 4.522 0.163
Average -0.926 0.083 1.076 0.080
Notes: This table reports the annualized equivalent return (CER) gain, the monthly Sharpe ratio (SR)
for a mean-variance investor who optimally allocates across equities and the risk-free asset using the
out-of-sample forecasts based on gold price ratios and traditional predictors. We consider two samples
for initial estimations, the first sample is from 1975M1 to 1985M12 (120 months), and the second
sample spans from 1975M1 to 1994M12 (240 months).

35
Table 9
Forecasting market return components using GO
Variables 𝛽̂𝐸̂ t-stat 𝛽̂𝐶𝐹
̂ t-stat 𝛽̂𝐷𝑅
̂ t-stat
r, DP 0.103* 1.734 1.241** 2.458 -0.064 -1.572
r, DP, DY 0.153** 2.447 1.210** 2.388 -0.044 -0.958
r, DP, PE 0.443*** 6.534 1.009** 2.075 0.045 0.213
r, DP, DE 0.194*** 2.924 1.161** 2.399 -0.052 -0.686
r, DP, TBL 0.365*** 3.745 0.439 0.829 -0.603* -1.794
r, DP, INFL 0.220*** 3.374 1.135** 2.240 -0.053 -1.022
r, DP, LTY 0.475*** 4.371 0.348 0.653 -0.585 -1.518
r, DP, DS 0.109* 1.839 1.280** 2.499 -0.019 -0.385
r, DP, TS 0.104* 1.699 1.163** 2.312 -0.140* -2.111
r, DP, BM 0.100* 1.693 1.246** 2.468 -0.061 -1.514
r, DP, NTIS 0.519*** 6.842 0.794* 1.865 -0.095 -0.421
r, DP, SVAR 0.137** 2.254 1.216** 2.411 -0.055 -1.327
Notes: This table reports the coefficient estimates for the predictive regression:
𝑍
𝑍𝑡+1 = 𝛼 + 𝛽𝑍 𝐺𝑂𝑡 + 𝜀𝑡+1 ,
where 𝑍𝑡+1 is one of the three estimated components of the S&P500 log return, including the expected
return component 𝐸̂𝑡 𝑟𝑡+1 , cash flow news ̂𝜂𝑡+1
𝐶𝐹 𝐷𝑅
, and discount rate news 𝜂̂ 𝑡+1 . These components are
estimated using a VAR(1) approach of Campbell (1991) and Campbell and Ammer (1993). The first
column shows the variables that are included in a VAR model. We calculate the Newey and West (1987)
heteroskedasticity and autocorrelation robust t-statistics. ***, ** and * denote significance at the 1%,
5% and 10% levels, respectively.

36
Table 10
Forecasting economic variables using GO
𝛽 t-stat 𝜃 t-stat Partial-𝑅2 (%)
SRP(-) 1.350 1.362 0.895*** 20.719 1.635
DS(-) 0.010* 1.501 0.958*** 40.801 0.664
CFNAI 0.030 0.277 0.239 1.119 0.086
FSI(-) 0.049*** 2.376 0.785*** 24.626 1.092
MU(-, h=1) 0.004** 1.930 0.972*** 48.203 3.233
MU(-, h=3) 0.003*** 2.435 0.981*** 63.506 3.494
MU(-, h=12) 0.001*** 2.297 0.989*** 83.696 2.141
FU (-, h=1) 0.004* 1.517 0.973*** 54.172 1.107
FU (-, h=3) 0.003* 1.533 0.977*** 63.044 1.082
FU (-, h=12) 0.001* 1.463 0.986*** 91.239 0.943
Notes: This table reports the estimation results for the predictive regression:
𝑌𝑡+1 = 𝛼 + 𝛽𝐺𝑂𝑡 + 𝜃𝑌𝑡 + 𝜀𝑡+1 ,
where 𝑌𝑡+1 denotes one of the economic variables described in Section 4.2. (-) indicates that we take
the negative value of the economic variable. We use the Newey and West (1987) heteroskedasticity and
autocorrelation robust t-statistics and compute a wild-bootstrapped p-value for testing 𝐻0 : 𝛽 =
0 (𝜃 = 0) against 𝐻𝐴 : 𝛽 > 0 (𝜃 > 0). ***, ** and * indicate significance at the 1%, 5% and 10%
level, respectively. h is the forecast horizons for the macro and financial uncertainty of Ludvigson et al.
(2015).

37
Table 11
State-switching predictive regression results
𝛽 t-stat 𝜃 t-stat 𝑅2 (%)
Panel A: Business cycles
0.388** 2.415 1.056** 2.058 2.473
Panel B: GO dynamics scenarios
Scenario1 0.574** 2.391 0.539** 2.255 1.870
Scenario2 1.478* 1.746 0.526*** 2.974 2.020
Scenario3 0.591 0.854 0.542*** 3.639 1.870
Scenario4 1.500** 2.393 0.514*** 2.848 2.118
Scenario5 -0.046 -0.121 0.696*** 3.353 2.496
Scenario6 -0.373 -0.583 0.575*** 3.156 2.042
Scenario7 0.703** 2.293 0.521** 2.586 1.900
Scenario8 2.008** 2.604 0.517*** 2.874 2.216
Scenario1&2 0.641*** 2.873 0.503** 2.132 1.899
Scenario3&4 0.758 1.217 0.499*** 3.126 1.946
Scenario5&6 -0.089 -0.258 0.734*** 3.446 2.715
Scenario7&8 0.860*** 2.829 0.482** 2.342 2.024
Scenario1234 0.684** 2.427 0.394* 1.775 2.022
Panel C: Threshold of GO
90% 0.362 1.376 0.651*** 2.612 1.988
95% 0.349 1.061 0.615*** 2.823 1.959
99% 0.990*** 4.516 0.499** 2.446 2.032
Notes: This table reports the results of state-switching predictive regressions as in Jacobsen et al. (2019):
𝑟𝑡+1 = 𝛼 + 𝛽𝐺𝑂𝑡 ∗ 𝑆𝑡𝑎𝑡𝑒1 + 𝜃𝐺𝑂𝑡 ∗ 𝑆𝑡𝑎𝑡𝑒2 + 𝜀𝑡+1 ,
where 𝑟𝑡+1 is the log stock market return in excess of the risk-free rate at month 𝑡, and 𝑆𝑡𝑎𝑡𝑒 denotes
one of the state variables. Panel A reports the predictive ability of GO during NBER expansions and
recessions, and 𝑆𝑡𝑎𝑡𝑒1 equals 1 when there is an expansion and 𝑆𝑡𝑎𝑡𝑒2 equals 1 when there is a
recession. Panel B reports results for eight scenarios with respect to the gold price, oil price and GO
movements in Appendix Table A4, and similarly, 𝑆𝑡𝑎𝑡𝑒1 equals 1 during the periods of a scenario and
𝑆𝑡𝑎𝑡𝑒2 equals 1 when the period is not in the same scenario. We also consider combined scenarios. For
example, Scenario12 indicates that we consider Scenarios 1 and 2 at the same time. Panel C reports
results for the 90%, 95% and 99% quantiles of GO. For example, “90%” in the first row of Panel C
indicates that 𝑆𝑡𝑎𝑡𝑒1 equals 1 when GO is larger than its 90% quantile, and 𝑆𝑡𝑎𝑡𝑒2 equals 1 when
GO is lower than its 90% quantile. We use the Newey and West (1987) heteroskedasticity and
autocorrelation robust t-statistics and compute a wild-bootstrapped p-value for testing 𝐻0 : 𝛽 =
0 (𝜃 = 0) against 𝐻𝐴 : 𝛽 ≠ 0 (𝜃 ≠ 0). ***, ** and * indicate significance at the 1%, 5% and 10%
level, respectively.

38
Table 12
Forecasting returns using alternative estimation approaches
WLS-EV RBM IVX
𝛽(%) t-stat 𝛽(%) t-stat 𝛽(%) IVX Wald-stat
GO 0.438*** 2.417 0.535*** 3.071 0.548*** 12.324
GS 0.288** 1.940 0.270** 1.821 0.292* 2.982
GCPI 0.089 0.537 0.189 1.085 0.169 1.157
GCO 0.178 1.123 0.263* 1.604 0.242 1.919
GCP 0.132 0.825 0.206 1.209 0.211 1.797
GDJ -0.031 -0.172 -0.166 -0.842 0.109 0.326
GYD 0.179 1.138 0.228* 1.358 0.204 0.212
GTB 0.225* 1.352 0.246* 1.407 0.296 3.250
GFR 0.255** 1.693 0.311** 1.897 0.327** 3.932
GP 0.346** 1.994 0.388*** 2.413 0.493*** 8.969
Notes: This table reports the estimation results based on three alternative estimation approaches for the
predictive regression:
𝑟𝑡+1 = 𝛼 + 𝛽𝑋𝑡 + 𝜀𝑡+1 .
The three alternative estimation approaches are the weighted least squares using ex ante variance (WLS-
EV) proposed by Johnson (2019), the reduced-bias estimation method of Amihud and Hurvich (2004),
and the Wald test proposed by Kostakis et al. (2015).

39
Table 13
Bivariate US stock market return predictability using new predictors
SII(-) NVIX(-) AIS(-) AOR(-)
𝛽(%) t-stat 𝛽(%) t-stat 𝛽(%) t-stat 𝛽(%) t-stat
GO 0.481*** 2.443 0.599*** 2.966 0.356** 2.032 0.512*** 2.860
GS 0.441*** 2.729 0.338*** 2.177 0.107 0.649 0.284** 1.867
GCPI 0.038 0.181 0.226 1.275 0.133 0.821 0.178 1.068
GCO 0.222* 1.280 0.276* 1.609 0.085 0.539 0.231* 1.526
GCP 0.085 0.440 0.287* 1.503 0.528*** 2.770 0.217* 1.305
GDJ -0.011 -0.054 0.053 0.291 0.273* 1.456 0.048 0.252
GYD 0.145 0.772 0.292** 1.754 0.105 0.658 0.224* 1.433
GTB 0.198 1.101 0.371** 2.146 0.024 0.147 0.262** 1.717
GFR 0.197 1.029 0.463*** 2.700 0.083 0.513 0.268** 1.722
GP 0.371** 1.864 0.483*** 2.705 0.414*** 2.691 0.450*** 2.875
𝜃(%) t-stat 𝜃(%) t-stat 𝜃(%) t-stat 𝜃(%) t-stat
GO 0.326* 1.613 0.307 1.213 0.540*** 3.163 1.610 0.402
GS 0.537*** 2.586 0.271 1.021 0.601*** 3.256 2.691 0.662
GCPI 0.382* 1.638 0.247 0.911 0.622*** 3.551 2.031 0.504
GCO 0.413** 2.020 0.260 0.972 0.613*** 3.405 2.706 0.676
GCP 0.367** 1.670 0.270 0.971 0.834*** 4.243 1.987 0.497
GDJ 0.396** 1.856 0.208 0.771 0.700*** 3.752 2.023 0.485
GYD 0.377** 1.733 0.282 1.043 0.613*** 3.434 2.397 0.601
GTB 0.380** 1.774 0.324 1.200 0.618*** 3.261 2.648 0.666
GFR 0.352* 1.546 0.391* 1.363 0.603*** 3.257 2.496 0.632
GP 0.207 0.857 0.242 0.939 0.624*** 3.606 0.711 0.171
Notes: This table reports the estimation results for the bivariate predictive regression:
𝑟𝑡+1 = 𝛼 + 𝛽𝐺𝑃𝑅𝑡 + 𝜃𝑋𝑡 + 𝜀𝑡+1 .
where 𝑋𝑡 indicates one of the newly proposed return predictors, including the short interest index (SII)
of Rapach et al. (2016), the news-based implied volatility (NVIX) of Manela and Moreira (2017), the
aligned investor sentiment of Huang et al. (2015), and the asymmetric oil return (AOR) of Wang et al.
(2019). (-) indicate that we take the negative value of the predictor. The sample period for SII, NVIX,
AIS and AOR are 1975M1-2014M12, 1975M1-2016M3, 1975M1-2018M12 and 1975M1-2019M12,
respectively. We use the Newey and West (1987) heteroskedasticity and autocorrelation robust t-
statistics and compute a wild-bootstrapped p-value for testing 𝐻0 : 𝛽 = 0 (𝜃 = 0) against 𝐻𝐴 : 𝛽 >
0 (𝜃 > 0). ***, ** and * indicate significance at the 1%, 5% and 10% level, respectively.

40
 4
5
6
7
8
0
1
2
3
4
5

3
4
5
6
7
8
0
1
2
3

0
1
2
3
4
5
6
7
8
9
10
11
12
1975/1/1 1975/1/1 1975/1/1 1975/1/1
1975/1/1
1976/6/1 1976/6/1 1976/6/1 1976/6/1 1976/6/1
1977/11/1 1977/11/1 1977/11/1 1977/11/1 1977/11/1
1979/4/1 1979/4/1 1979/4/1 1979/4/1 1979/4/1
1980/9/1 1980/9/1 1980/9/1 1980/9/1 1980/9/1

Figure 1
1982/2/1 1982/2/1 1982/2/1 1982/2/1 1982/2/1
1983/7/1 1983/7/1 1983/7/1 1983/7/1 1983/7/1
1984/12/1 1984/12/1 1984/12/1 1984/12/1 1984/12/1
1986/5/1 1986/5/1 1986/5/1 1986/5/1 1986/5/1
1987/10/1 1987/10/1 1987/10/1 1987/10/1 1987/10/1
1989/3/1 1989/3/1 1989/3/1 1989/3/1 1989/3/1
1990/8/1 1990/8/1 1990/8/1 1990/8/1 1990/8/1
1992/1/1 1992/1/1 1992/1/1 1992/1/1 1992/1/1
1993/6/1 1993/6/1 1993/6/1 1993/6/1 1993/6/1
1994/11/1 1994/11/1 1994/11/1 1994/11/1 1994/11/1
1996/4/1 1996/4/1 1996/4/1 1996/4/1 1996/4/1
1997/9/1 1997/9/1 1997/9/1 1997/9/1 1997/9/1
1999/2/1 1999/2/1 1999/2/1 1999/2/1 1999/2/1
2000/7/1 2000/7/1 2000/7/1 2000/7/1 2000/7/1
Gold-oil ratio

2001/12/1 2001/12/1 2001/12/1 Gold-CPI ratio 2001/12/1

Gold-copper ratio
2001/12/1
2003/5/1 2003/5/1 2003/5/1 2003/5/1 2003/5/1

Gold-Dividend yield ratio

2004/10/1 2004/10/1 2004/10/1 2004/10/1 2004/10/1

Gold-Federal fund rates ratio

2006/3/1 2006/3/1 2006/3/1 2006/3/1 2006/3/1
2007/8/1 2007/8/1 2007/8/1 2007/8/1 2007/8/1
2009/1/1 2009/1/1 2009/1/1 2009/1/1 2009/1/1
2010/6/1 2010/6/1 2010/6/1 2010/6/1 2010/6/1
2011/11/1 2011/11/1 2011/11/1 2011/11/1 2011/11/1
2013/4/1 2013/4/1 2013/4/1 2013/4/1 2013/4/1
2014/9/1 2014/9/1 2014/9/1 2014/9/1 2014/9/1
2016/2/1 2016/2/1 2016/2/1 2016/2/1 2016/2/1
2017/7/1 2017/7/1 2017/7/1 2017/7/1 2017/7/1
2018/12/1 2018/12/1 2018/12/1 2018/12/1 2018/12/1
2020/5/1 2020/5/1 2020/5/1 2020/5/1 2020/5/1

0.198
0.398
0.598
0.798
0.998
0.198
0.398
0.598
0.798
0.998

0.198
0.398
0.598
0.798
0.998
0.198
0.398
0.598
0.798
0.998

0.198
0.398
0.598
0.798
0.998
-0.002
-0.002

-0.002
-0.002

-0.002
0
1
2
3
4
5
6
7
8
2
3
4
5

0
1
2
3
4
5
6
7
8
-4
-3
-2
-1
0

41
-1.0
-0.5
0.0
0.5
1.0
1975/1/1 1975/1/1 1975/1/1 1975/1/1
1975/1/1 1976/6/1 1976/6/1 1976/6/1 1976/6/1
1976/7/1 1977/11/1 1977/11/1 1977/11/1 1977/11/1
1978/1/1 1979/4/1 1979/4/1 1979/4/1 1979/4/1
1979/7/1 1980/9/1 1980/9/1 1980/9/1 1980/9/1
1981/1/1 1982/2/1 1982/2/1 1982/2/1 1982/2/1
1982/7/1 1983/7/1 1983/7/1 1983/7/1 1983/7/1
1984/1/1 1984/12/1 1984/12/1 1984/12/1 1984/12/1
1985/7/1 1986/5/1 1986/5/1 1986/5/1 1986/5/1
1987/1/1 1987/10/1 1987/10/1 1987/10/1 1987/10/1
1988/7/1 1989/3/1 1989/3/1 1989/3/1 1989/3/1
1990/1/1 1990/8/1 1990/8/1 1990/8/1 1990/8/1
1991/7/1 1992/1/1 1992/1/1 1992/1/1 1992/1/1
1993/1/1 1993/6/1 1993/6/1 1993/6/1 1993/6/1
1994/7/1 1994/11/1 1994/11/1 1994/11/1 1994/11/1
1996/1/1 1996/4/1 1996/4/1 1996/4/1 1996/4/1
1997/9/1 1997/9/1 1997/9/1 1997/9/1
1997/7/1
1999/2/1 1999/2/1 1999/2/1 1999/2/1
1999/1/1
2000/7/1 2000/7/1 2000/7/1 2000/7/1
2000/7/1
Gold-corn ratio
Gold-silver ratio

2001/12/1 2001/12/1 2001/12/1 2001/12/1

2002/1/1

Gold-Platinum ratio
2003/5/1 2003/5/1 2003/5/1 2003/5/1
2003/7/1
2004/10/1 2004/10/1 2004/10/1 2004/10/1
2005/1/1

Gold-Treasury bond yield ratio

2006/3/1 2006/3/1 2006/3/1 2006/3/1
2006/7/1

This figure plots the gold price ratios. The shades indicate NBER economic recessions.
2007/8/1 2007/8/1 2007/8/1 2007/8/1
2008/1/1
Gold-Dow Jones Industrial Average ratio

2009/1/1 2009/1/1 2009/1/1 2009/1/1

2009/7/1
2010/6/1 2010/6/1 2010/6/1 2010/6/1
2011/1/1
2011/11/1 2011/11/1 2011/11/1 2011/11/1
2012/7/1 2013/4/1 2013/4/1 2013/4/1
2013/4/1
2014/1/1 2014/9/1 2014/9/1 2014/9/1 2014/9/1
2015/7/1 2016/2/1 2016/2/1 2016/2/1 2016/2/1
2017/1/1 2017/7/1 2017/7/1 2017/7/1 2017/7/1
2018/7/1 2018/12/1 2018/12/1 2018/12/1 2018/12/1
2020/1/1 2020/5/1 2020/5/1 2020/5/1 2020/5/1
0.198
0.398
0.598
0.798
0.998
0.198
0.398
0.598
0.798
0.998

0.198
0.398
0.598
0.798
0.998
0.198
0.398
0.598
0.798
0.998

0.198
0.398
0.598
0.798
0.998
-0.002
-0.002

-0.002
-0.002

-0.002
 -1.6
-1.2
-0.8
-0.4
0.0
0.4
0.8
1.2
-1.2
-0.8
-0.4
0.0
0.4
0.8
1.2
1.6
2.0
2.4
-1.2
-0.8
-0.4
0.0
0.4
0.8
1.2
1.6
2.0
-1.2
-0.8
-0.4
0.0
0.4
0.8
-1.2
-0.8
-0.4
0.0
0.4
0.8
1985/2/1 1985/2/1 1985/2/1 1985/2/1 1985/2/1
1986/4/1 1986/4/1 1986/4/1 1986/4/1 1986/4/1
1987/6/1 1987/6/1 1987/6/1 1987/6/1 1987/6/1
1988/8/1 1988/8/1 1988/8/1 1988/8/1 1988/8/1

Figure 2
1989/10/1 1989/10/1 1989/10/1 1989/10/1 1989/10/1
1990/12/1 1990/12/1 1990/12/1 1990/12/1 1990/12/1
1992/2/1 1992/2/1 1992/2/1 1992/2/1 1992/2/1
1993/4/1 1993/4/1 1993/4/1 1993/4/1 1993/4/1
1994/6/1 1994/6/1 1994/6/1 1994/6/1 1994/6/1
1995/8/1 1995/8/1 1995/8/1 1995/8/1 1995/8/1
1996/10/1 1996/10/1 1996/10/1 1996/10/1 1996/10/1
1997/12/1 1997/12/1 1997/12/1 1997/12/1 1997/12/1
1999/2/1 1999/2/1 1999/2/1 1999/2/1 1999/2/1
2000/4/1 2000/4/1 2000/4/1 2000/4/1 2000/4/1
2001/6/1 2001/6/1 2001/6/1 2001/6/1 2001/6/1
2002/8/1 2002/8/1 2002/8/1 2002/8/1 2002/8/1
2003/10/1 2003/10/1 2003/10/1 2003/10/1 2003/10/1
2004/12/1 2004/12/1 2004/12/1 2004/12/1 2004/12/1
Estimates for SRP

2006/2/1 2006/2/1 2006/2/1 2006/2/1 Estimates for CFNAI 2006/2/1

Estimates for FU (h=3)

Estimates for MU (h=1)

Estimates for MU (h=12)

2007/4/1 2007/4/1 2007/4/1 2007/4/1 2007/4/1
2008/6/1 2008/6/1 2008/6/1 2008/6/1 2008/6/1
2009/8/1 2009/8/1 2009/8/1 2009/8/1 2009/8/1
2010/10/1 2010/10/1 2010/10/1 2010/10/1 2010/10/1
2011/12/1 2011/12/1 2011/12/1 2011/12/1 2011/12/1
2013/2/1 2013/2/1 2013/2/1 2013/2/1 2013/2/1
2014/4/1 2014/4/1 2014/4/1 2014/4/1 2014/4/1
2015/6/1 2015/6/1 2015/6/1 2015/6/1 2015/6/1
2016/8/1 2016/8/1 2016/8/1 2016/8/1 2016/8/1
2017/10/1 2017/10/1 2017/10/1 2017/10/1 2017/10/1
2018/12/1 2018/12/1 2018/12/1 2018/12/1 2018/12/1
2020/2/1 2020/2/1 2020/2/1 2020/2/1 2020/2/1

-1.6
-1.2
-0.8
-0.4
0.0
0.4
0.8
1.2
-1.2
-0.8
-0.4
0.0
0.4
0.8
1.2
-1.2
-0.8
-0.4
0.0
0.4
0.8
1.2
1.6
2.0
-1.2
-0.8
-0.4
0.0
0.4
0.8
1.2
-1.2
-0.8
-0.4
0.0
0.4
0.8
1.2
1.6

42
1985/2/1 1985/2/1 1985/2/1 1985/2/1 1985/2/1
1986/4/1 1986/4/1 1986/4/1 1986/4/1 1986/4/1
1987/6/1 1987/6/1 1987/6/1 1987/6/1 1987/6/1
1988/8/1 1988/8/1 1988/8/1 1988/8/1 1988/8/1
1989/10/1 1989/10/1 1989/10/1 1989/10/1 1989/10/1
1990/12/1 1990/12/1 1990/12/1 1990/12/1 1990/12/1

confidence intervals. We take a rolling window of 120 months.

1992/2/1 1992/2/1 1992/2/1 1992/2/1 1992/2/1
1993/4/1 1993/4/1 1993/4/1 1993/4/1 1993/4/1
1994/6/1 1994/6/1 1994/6/1 1994/6/1 1994/6/1
1995/8/1 1995/8/1 1995/8/1 1995/8/1 1995/8/1
1996/10/1 1996/10/1 1996/10/1 1996/10/1 1996/10/1
1997/12/1 1997/12/1 1997/12/1 1997/12/1 1997/12/1
1999/2/1 1999/2/1 1999/2/1 1999/2/1 1999/2/1
2000/4/1 2000/4/1 2000/4/1 2000/4/1 2000/4/1
2001/6/1 2001/6/1 2001/6/1 2001/6/1 2001/6/1
2002/8/1 2002/8/1 2002/8/1 2002/8/1 2002/8/1
2003/10/1 2003/10/1 2003/10/1 2003/10/1 2003/10/1
2004/12/1 2004/12/1 2004/12/1 2004/12/1 2004/12/1
Estimates for DS

Estimates for FSI

2006/2/1 2006/2/1 2006/2/1 2006/2/1 2006/2/1

Estimates for FU (h=1)

Estimates for MU (h=3)

Estimates for FU (h=12)

2007/4/1 2007/4/1 2007/4/1 2007/4/1 2007/4/1
2008/6/1 2008/6/1 2008/6/1 2008/6/1 2008/6/1
2009/8/1 2009/8/1 2009/8/1 2009/8/1 2009/8/1
2010/10/1 2010/10/1 2010/10/1 2010/10/1 2010/10/1
2011/12/1 2011/12/1 2011/12/1 2011/12/1 2011/12/1
2013/2/1 2013/2/1 2013/2/1 2013/2/1 2013/2/1
2014/4/1 2014/4/1 2014/4/1 2014/4/1 2014/4/1
2015/6/1 2015/6/1 2015/6/1 2015/6/1 2015/6/1
2016/8/1 2016/8/1 2016/8/1 2016/8/1 2016/8/1
2017/10/1 2017/10/1 2017/10/1 2017/10/1 2017/10/1
2018/12/1 2018/12/1 2018/12/1 2018/12/1 2018/12/1
2020/2/1 2020/2/1 2020/2/1 2020/2/1 2020/2/1

Equation (24). The solid lines are the coefficient estimates, and the dashed lines indicate the 95%
This figure reports the rolling window estimates for GO in the predictive regression as shown in
 Recession: The 2nd oil crisis Recession: The 1981 economic crisis Recession: The Gulf War I
3.0 125 2.7 150 3.2 390
GO SP500 GO SP500 GO SP500

2.9 120 2.6 140 3.0 370

115 130 350

2.8 2.5 2.8
110 120 330
2.7 2.4 2.6
105 110 310
2.6 2.3 2.4
100 100 290

2.5 95 2.2 90 2.2 270

2.4 90 2.1 80 2.0 250

1980/1/1 1980/2/1 1980/3/1 1980/4/1 1980/5/1 1980/6/1 1980/7/1 1981/7/1 1981/10/1 1982/1/1 1982/4/1 1982/7/1 1982/10/1 1990/7/1 1990/9/1 1990/11/1 1991/1/1 1991/3/1

Recession: The dot.com bubble Recession: The financial crisis Recession: The COVID-19 pandemic
2.8 1300 4.0 1500 4.8
GO SP500 GO SP500 GO SP500
1250 1400 4.6 3500
2.7 3.5
1200 1300 4.4
2.6 3300
1150
3.0 1200 4.2
1100
2.5 1100 4.0 3100
1050 2.5
2.4 1000 3.8
1000
2900
2.0 900 3.6
2.3 950
900 800 3.4
1.5 2700
2.2
850 700 3.2

2.1 800 1.0 600 3.0 2500

2001/3/1 2001/5/1 2001/7/1 2001/9/1 2001/11/1 2007/12/1 2008/5/1 2008/10/1 2009/3/1 2020/1/1 2020/3/1 2020/5/1 2020/7/1 2020/9/1

Figure 3
This figure plots the relationships between GO and S&P500 index during six US economic recession
periods. GO is on the left vertical axis, and S&P500 is on the right.

43
Appendices
Table A1
Bivariate US stock market return predictability
𝜃(%) DP DY PE DE TBL(-) INFL(-)
GO 0.046 0.114 -0.072 -0.068 0.148 0.265
GS 0.241* 0.333* -0.255 -0.023 0.139 0.281
GCPI 0.004 0.082 0.051 0.060 0.225 0.366**
GCO 0.168 0.253 -0.117 0.048 0.174 0.314*
GCP 0.010 0.086 0.044 -0.009 0.267 0.338*
GDJ -0.092 0.479 0.504 0.046 0.494** 0.442**
GYD 0.093 0.169 -0.043 0.056 0.182 0.324*
GTB 0.263 0.362* -0.196 0.069 0.096 0.298
GFR 0.239 0.329* -0.156 0.049 -0.008 0.283
GP -0.111 -0.033 0.099 -0.031 0.241 0.308*
𝜃(%) LTY(-) DS TS BM(-) NTIS(-) SVAR(-)
GO 0.176 0.101 0.185 0.049 0.016 0.598*
GS 0.195 0.151 0.107 -0.114 0.014 0.581*
GCPI 0.265* 0.009 0.163 0.193 -0.002 0.550
GCO 0.232 0.079 0.164 0.044 -0.046 0.574
GCP 0.321** -0.029 0.136 0.170 0.042 0.581
GDJ 0.691*** 0.041 0.149 0.712** 0.071 0.538
GYD 0.235 0.043 0.155 0.091 -0.060 0.565
GTB 0.201 0.095 0.147 0.000 -0.068 0.582*
GFR 0.115 0.079 0.074 0.025 -0.018 0.579*
GP 0.288* -0.027 0.138 0.262* 0.029 0.551
Notes: This table reports the coefficient estimates for the bivariate predictive regression:
𝑟𝑡+1 = 𝛼 + 𝛽𝐺𝑃𝑅𝑡 + 𝜃𝑋𝑡 + 𝜀𝑡+1 .
This table reports the 𝜃̂ estimates and significance levels. (-) indicates that we take the negative value
of the predictor. Each predictor is standardized. We use the Newey and West (1987) heteroskedasticity
and autocorrelation robust t-statistics and compute a wild-bootstrapped p-value for testing 𝐻0 : 𝛽 =
0 (𝜃 = 0) against 𝐻𝐴 : 𝛽 > 0 (𝜃 > 0). ***, ** and * indicate significance at the 1%, 5% and 10%
level, respectively.

44
Table A2
Out-of-sample forecasting with longer horizons
Length = 120 months Length = 240 months
2 2
𝑅𝑂𝑂𝑆 (%) MSFE-adj 𝑅𝑂𝑂𝑆 (%) MSFE-adj
GO 5.349 3.752*** 7.944 4.224***
GS 1.550 2.565*** 2.017 2.693***
GCPI 0.256 1.255 0.267 1.157
GCO 1.010 1.915** 1.656 2.793***
GCP 1.304 2.497*** 1.837 3.069***
GDJ -0.443 0.602 -0.354 0.805
h=3 GYD 0.632 1.666** 0.868 1.705**
GTB 0.458 1.636* 0.792 1.776**
GFR -0.327 1.443* 0.645 1.602*
GP 4.172 3.958*** 4.754 3.742***
PC1 2.935 3.779*** 3.944 3.952***
PC2 -0.364 0.410 -0.170 0.702
PC3 1.289 2.407*** 2.098 2.040**
GO 8.832 5.167*** 11.812 5.546***
GS 3.998 4.429*** 4.096 4.797***
GCPI 0.512 1.820** 0.347 1.539*
GCO 2.457 3.821*** 3.004 4.965***
GCP 2.737 4.148*** 3.676 5.286***
GDJ -0.427 1.293* -0.276 1.465*
h=6 GYD 1.449 2.934*** 1.795 2.916***
GTB 1.071 2.434*** 1.397 2.463***
GFR 1.268 2.883*** 1.984 3.030***
GP 8.515 5.620*** 9.487 5.355***
PC1 5.566 6.118*** 6.899 6.303***
PC2 -0.300 1.021 -0.276 1.027
PC3 3.372 2.957*** 3.864 2.328**
GO 14.028 7.412*** 17.249 7.718***
GS 6.152 6.050*** 6.177 6.508***
GCPI 0.777 2.424*** -0.411 1.564*
GCO 2.515 5.072*** 3.039 6.127***
GCP 4.344 6.567*** 5.432 7.623***
GDJ -0.353 2.138** -0.486 2.102**
h=12 GYD 3.334 4.710*** 4.010 4.747***
GTB 2.272 3.085*** 3.226 3.310***
GFR 5.048 6.243*** 6.242 6.410***
GP 15.928 7.216*** 16.399 6.738***
PC1 9.491 8.522*** 11.270 8.773***
PC2 -0.039 1.536* -0.518 1.182
PC3 4.824 4.150*** 4.383 2.938***
Notes: This table reports the out-of-sample forecasting results with longer horizons. The predictive
regression is shown as follow:
1
∑ℎ𝑖=1 𝑟𝑡+𝑖 = 𝛼 + 𝛽𝐺𝑃𝑅𝑡 + 𝜀𝑡+ℎ ,
ℎ

where ℎ = 3,6,12 months is the forecast horizon. Each gold price ratio is standardized. The out-of-
sample 𝑅2 statistics as shown in Equation (5), and MSFE-adj is the Clark and West (2007) adjusted
mean squared forecast errors. We consider two samples for initial estimations, the first sample is from
1975M1 to 1985M12 (120 months), and the second sample spans from 1975M1 to 1994M12 (240
months). ***, ** and * indicate significance at the 1%, 5% and 10% level, respectively.

45
Table A3
Forecasting dividend growth using economic variables
𝛽 t-stat 𝜃 t-stat Partial-𝑅2 (%)
SRP(-) 0.076** 1.922 0.790*** 22.600 1.577
DS(-) 0.111** 2.661 0.780*** 22.237 3.311
CFNAI 0.055* 2.056 0.806*** 23.455 0.812
FSI(-) 0.055 1.335 0.794*** 21.600 0.822
MU(-, h=1) 0.106** 2.547 0.775*** 21.658 2.902
MU(-, h=3) 0.105** 2.587 0.774*** 21.484 2.802
MU(-, h=12) 0.097*** 2.809 0.777*** 21.392 2.406
FU (-, h=1) 0.108*** 2.689 0.771*** 21.093 2.908
FU (-, h=3) 0.110*** 2.798 0.769*** 20.931 2.988
FU (-, h=12) 0.115*** 3.097 0.764*** 20.466 3.212
Notes: This table reports the estimation results for the predictive regression:
𝐷𝐺𝑡+1 = 𝛼 + 𝛽𝑌𝑡 + 𝜃𝐷𝐺𝑡 + 𝜀𝑡+1 ,
where 𝐷𝐺𝑡+1 denotes the dividend growth of S&P500, and 𝑌𝑡 denotes one of the economic variables.
(-) indicates that we take the negative value of the economic variable. We use the Newey and West
(1987) heteroskedasticity and autocorrelation robust t-statistics and compute a wild-bootstrapped p-
value for testing 𝐻0 : 𝛽 = 0 (𝜃 = 0) against 𝐻𝐴 : 𝛽 > 0 (𝜃 > 0). ***, ** and * indicate significance
at the 1%, 5% and 10% level, respectively. h is the forecast horizons for the macro and financial
uncertainty of Ludvigson et al. (2015).

46
Table A4
Gold and oil price dynamics
Scenario Δ𝑃𝑡𝐺𝑜𝑙𝑑 Δ𝑃𝑡𝑂𝑖𝑙 Absolute change Δ𝐺𝑂𝑡 Sample days

1 + + |Δ𝑃𝑡𝐺𝑜𝑙𝑑 | > |Δ𝑃𝑡𝑂𝑖𝑙 | + 142

2 + + |Δ𝑃𝑡𝐺𝑜𝑙𝑑 | < |Δ𝑃𝑡𝑂𝑖𝑙 | - 16

3 - - |Δ𝑃𝑡𝐺𝑜𝑙𝑑 | > |Δ𝑃𝑡𝑂𝑖𝑙 | - 107

4 - - |Δ𝑃𝑡𝐺𝑜𝑙𝑑 | < |Δ𝑃𝑡𝑂𝑖𝑙 | + 13

5 + - |Δ𝑃𝑡𝐺𝑜𝑙𝑑 | > |Δ𝑃𝑡𝑂𝑖𝑙 | + 76

6 + - |Δ𝑃𝑡𝐺𝑜𝑙𝑑 | < |Δ𝑃𝑡𝑂𝑖𝑙 | + 17

7 - + |Δ𝑃𝑡𝐺𝑜𝑙𝑑 | > |Δ𝑃𝑡𝑂𝑖𝑙 | - 106

8 - + |Δ𝑃𝑡𝐺𝑜𝑙𝑑 | < |Δ𝑃𝑡𝑂𝑖𝑙 | - 13

Notes: This table reports the gold price, oil price and GO dynamics. Δ𝑃𝑡𝐺𝑜𝑙𝑑 , Δ𝑃𝑡𝑂𝑖𝑙 and Δ𝐺𝑂
indicate the change of the gold price, oil price and GO from t to t-1, respectively. “+” and “-” denote
the positive change (increase) and negative change (decrease), respectively.

47
Table A5
Bivariate US stock market return predictability using gold price ratios
GPR 𝛽(%) t-stat 𝜃(%) t-stat
GO - - - -
GS 0.524** 2.293 0.056 0.276
GCPI(-) 0.559*** 2.722 0.025 0.135
GCO(-) 0.557*** 2.753 0.017 0.095
GCP(-) 0.579*** 2.813 0.057 0.308
GDJ 0.578*** 2.903 0.026 0.142
GYD 0.547*** 0.254 0.006 0.032
GTB 0.521*** 2.466 0.074 0.390
GFR 0.501*** 2.490 0.166 0.979
GP 0.426* 1.640 0.189 0.791
Notes: This table reports the coefficient estimates for the bivariate predictive regression:
𝑟𝑡+1 = 𝛼 + 𝛽𝐺𝑂𝑡 + 𝜃𝐺𝑃𝑅𝑡 + 𝜀𝑡+1 .
The coefficient estimates and significance levels are shown in this table. (-) indicates that we take the
negative value of the predictor. Each predictor is standardized. We use the Newey and West (1987)
heteroskedasticity and autocorrelation robust t-statistics and compute a wild-bootstrapped p-value for
testing 𝐻0 : 𝛽 = 0 (𝜃 = 0) against 𝐻𝐴 : 𝛽 > 0 (𝜃 > 0). ***, ** and * indicate significance at the 1%,
5% and 10% level, respectively.

48
Table A6
Forecasting returns using principal components
Panel A: In-sample analysis
𝛽(%) t-stat 𝑅2 (%) ̂
𝑞𝐿𝐿
PC1 0.172** 2.343 1.105 -2.394
PC2 0.072 0.580 0.084 -5.806
PC3 0.274** 1.825 0.749 -3.791
Initial Sample = 120 months Initial sample = 240 months
Panel B: Out-of-sample forecasting
2 2
𝑅𝑂𝑂𝑆 (%) MSFE-adj 𝑅𝑂𝑂𝑆 (%) MSFE-adj
PC1 0.772 1.810** 1.083 1.898**
PC2 -0.465 -0.488 -0.289 0.011
PC3 -0.099 1.544* 0.337 1.261
Panel C: Out-of-sample forecasting over business cycles
2 2 2 2
𝑅𝑂𝑂𝑆,𝐸𝑋 (%) 𝑅𝑂𝑂𝑆,𝑅𝐸 (%) 𝑅𝑂𝑂𝑆,𝐸𝑋 (%) 𝑅𝑂𝑂𝑆,𝑅𝐸 (%)
PC1 3.585 -1.657 3.561 -0.766
PC2 0.977 -0.486 0.043 0.868
PC3 -0.272 2.670 -1.211 4.094
Notes: This table reports the predictive ability of principal components of gold price ratios. We consider
three principal components of gold price ratios. Panel A, B and C show the results for in-sample, out-
of-sample and out-of-sample over business cycles, respectively. In Panel A, we use the Newey and West
(1987) heteroskedasticity and autocorrelation robust t-statistics and compute a wild-bootstrapped p-
value for testing 𝐻0 : 𝛽 = 0 against 𝐻𝐴 : 𝛽 > 0. ***, ** and * indicate significance at the 1%, 5% and
10% levels, respectively.

49
 The first principal component The second principal component The third principal component
0.80 0.80 0.80

0.60 0.60 0.60

0.40 0.40 0.40

0.20 0.20 0.20

0.00 0.00 0.00

GO GS GCPI GCO GCP GDJ GYD GTB GFR GP GO GS GCPI GCO GCP GDJ GYD GTB GFR GP GO GS GCPI GCO GCP GDJ GYD GTB GFR GP

-0.20 -0.20 -0.20

-0.40 -0.40 -0.40

Figure A1
This figure reports the loadings on the three principal components of gold price ratios.

50