The asymmetric responses of aggregate earnings and stock returns to oil shocks and policy uncertainty

ABSTRACT

This paper utilizes a structural VAR model to investigate the asymmetric responses of aggregate earnings and stock market returns to oil price shocks and economic policy uncertainty. It finds that aggregate earnings contain information about oil price fluctuations. The effects of oil shocks on the earnings and returns are amplified by endogenous policy uncertainty responses. Oil shocks and policy uncertainty explain 29.7% and 11.2% of the variation in the aggregate earnings in the long run. The covariance of aggregate earnings and stock market returns is negative and driven by the demand-side oil shock and news coverage/CPI forecast uncertainty.

KEYWORDS:

• Aggregate earnings
• oil prices
• economic policy uncertainty
• stock returns
• structural VAR

Acknowledgments

The author would like to thank the editor Suresh Radhakrishnan and the anonymous referees for helpful comments on the earlier draft of this article. The author would like to acknowledge the collaboration of Jing Wang in the preliminary performance of the research and the earlier draft of the paper.

Disclosure statement

No potential conflict of interest was reported by the author.

Notes

1. Note that the terms covariance responses and responses of the earnings-returns contemporaneous association are used interchangeably for the exposition purpose in the text.

2. The monthly energy review is available at http://www.eia.gov/totalenergy/data/monthly/.

3. The monthly data can be found at http://www-personal.umich.edu/~lkilian/paperlinks.html.

4. The index of policy-related economic uncertainty created by Baker, Bloom, and Davis (2016) is based on the weighted average of the news coverage that discusses both the economy and the fiscal/monetary policy-related uncertainty via the month-by-month searches of Google News, the CPI forecast interquartile range, the number of federal tax code provisions set to expire in the contemporaneous calendar year and future 10 years, and the federal/state/local expenditure forecast interquartile range. The index is available at http://www.policyuncertainty.com/.

5. When I conduct the tests for the model selection, AIC and FPE suggest choosing a model with five lags, whereas the likelihood-ratio tests select a model with more than 24 lags. Based on Sims (1998) and Sims, Stock, and Watson (1990) who argue that even variables that display no inertia do not necessarily show the absence of long lags in regressions on other variables, I therefore use the long lag of 24 as in common with prior literature (for example, Kilian 2009; Kilian and Park 2009). The robustness check shows similar results of impulse response functions and variance decompositions when I perform some variations on the analysis with respect to the lag length, the ordering of the VAR, and using the first difference of real price of oil in the VAR model. Additionally, Kilian and Murphy (2012) argue that when it is not clear a prior whether a variable should be first-differenced, the impulse responses are reasonably precisely estimated using the (log-) levels of variables in the VAR model, given the standard nature of these tests.

6. While the ADF, PP, and KPSS tests on $\mathrm{\Delta }pro{d}_t$, $re{a}_t$, $p{u}_t$, $er{n}_t$, $re{t}_t$ and $co{v}_t$ reject the hypothesis containing a unit root at the 1% significant level, the PP and KPSS tests on $rp{o}_t$ accept the hypothesis containing a unit root at the 5% significant level. Kilian and Murphy (2012) argue that when it is not clear a prior whether the real price of oil should be first-differenced, the impulse responses are reasonably precisely estimated using the (log-) levels of the variables in the VAR model, given the standard nature of these statistics tests.

7. Define $z_t={\left(er{n}_t,re{t}_t\right)}^{\prime }$ in the model, $z_t=A{z}_{t-1}+{\epsilon }_t$, where A is a $2\times 2$ coefficient matrix, ${\epsilon }_t{ }_{}N\left(0,H_t\right)$, $H_t=C+B\prime {\epsilon }_{t-1}{\epsilon }_{t-1}^{\prime }B$, and C and B are $2\times 2$ parameter matrices. The element of covariance of $H_t$ generates the conditional covariance of aggregate earnings and returns.

8. Conducting the Granger causality test by using the VAR model (1), the Wald statistics are 35.28 and 47.76, respectively.