A change in the time-varying correlation between oil prices and the stock market

ABSTRACT

We investigate the time-varying correlation between oil prices and stock prices. Estimation results from a multivariate DCC-GARCH model reveal that the correlation has changed since the financial crisis. Historically, the correlation has been close to zero or slightly negative. However, the correlation changed to positive during the Great Recession and continued to be positive through the first half of 2017. We investigate the role quantitative easing played in this change in correlation using a threshold model.

KEYWORDS:

• Multivariate GARCH
• oil prices
• stock returns
• nonlinear models

JEL CLASSIFICATION:

• C32
• G10

Disclosure statement

No potential conflict of interest was reported by the authors.

Notes

¹ The article, ‘Oil, stocks at tightest correlation in 26 years’, calculates recent correlations as high as 0.97.

² https://www.brookings.edu/blog/ben-bernanke/2016/02/19/the-relationship-between-stocks-and-oil-prices/.

³ A full literature review looking at the relationship between oil prices and the stock market is beyond the scope of this article.

⁴ For days when stock market/oil market daily data are not available, we linearly interpolate between the nearest available days. Oil data were retrieved from the Federal Reserve Economic Database and S&P 500 data were retrieved from Bloomberg.

⁵ For a further explanation of the methodology, see Jones and Olson (2013) and Antonakakis, Chatziantoniou and Filis (2013).

⁶ Coefficients from the model are not shown due to space constraints. The residuals from the model were tested for serial correlation and there is no remaining serial correlation in the residuals or squared residuals.

⁷ Panel B of Figure 3 re-estimates the time-varying correlation using the procedure outlined at the beginning of Section 3 with weekly data. The size of the Federal Reserve’s balance sheet is only available staring at the end of 2002 in weekly frequency. The correlations estimated using a weekly frequency are very close to the correlations using a daily frequency.

⁸ The Bayesian Information Criteria suggests 0 lags in the $\mathrm{\Delta }sdDC{C}_t$ series, while the AIC suggests 1 lag. The results are unchanged using either 0 or 1 lag.

⁹ We eliminate the highest and lowest 10% of the ordered values of the threshold.

¹⁰ Panel B of Table 2 contains the Ljung–Box Q-statistics of the residuals from our threshold model. Autocorrelation in the residuals in both of the regimes is insignificant.