A detailed look at crude oil price volatility prediction using macroeconomic variables

• Published date: 01 November 2020
• Date: 01 November 2020
• DOI: http://doi.org/10.1002/for.2679
• Author: Nima Nonejad

Received: 16 April 2019 Revised: 4 November 2019 Accepted: 22 February 2020

DOI: 10.1002/for.2679

RESEARCH ARTICLE

A detailed look at crude oil price volatility prediction using

macroeconomic variables

Nima Nonejad

Danske Bank, Høje Taastrup,Denmark

CREATES, Aarhus, Denmark

Correspondence

Nima Nonejad, Danske Bank,

Dalbergstrøget 17, Høje Taastrup, 2630

Taastrup,Denmark.

Email: nimanonejad@gmail.com

Abstract

We investigate whether crude oil price volatility is predictable by conditioning

on macroeconomic variables. We consider a large number of predictors, take

into account the possibility that relative predictive performance varies over the

out-of-sample period, and shed light on the economic drivers of crude oil price

volatility.Results using monthly data from 1983:M1 to 2018:M12 document that

variables related to crude oil production, economic uncertainty and variables

that either describe the current stance or provide information about the future

state of the economy forecast crude oil price volatility at the population level 1

month ahead. On the other hand, evidence of finite-sample predictability is very

weak. A detailed examination of our out-of-sample results using the fluctuation

test suggests that this is because relative predictive performance changes dras-

tically over the out-of-sample period. The predictive power associated with the

more successful macroeconomic variables concentratesaround the Great Reces-

sion until 2015. They also generate the strongest signal of a decrease in the price

of crude oil towards the end of 2008.

KEYWORDS

Crude oil price volatility, forecast evaluation, macroeconomic variables, realized volatility

1INTRODUCTION

Our aim with this study is to quantify the predictive impact

of macroeconomic variables on crude oil price volatility

by way of a forecasting evaluation. Specifically, we inves-

tigate whether inclusion of macroeconomic variables in

vector autoregressions of monthly crude oil price realized

volatility improves out-of-sample forecasts relative to the

benchmark. We do so conditional on recent studies, such

as Conrad et al. (2014), Degiannakis and Filis (2017), Pan

et al. (2017), Ma et al. (2018), Meng and Liu (2019), and

Zang et al. (2019).1For example, Conradet al. (2014) inves-

1Interestingly, compared to the large number of studies that forecast

crude oil price volatility at the daily frequency using parametric condi-

tional volatility models, such as generalized autoregressive conditional

heteroskedasticity (GARCH) or stochastic volatility (see, among many

tigated the macroeconomic determinants of long-term

volatility and correlation between equity and crude oil

price returns from an in-sample perspective.2The authors

found that changes in long-term crude oil price volatility

could be explained by various measures of US macroe-

conomic activity. Pan et al. (2017) went beyond Conrad

et al. and evaluated whether incorporating macroeco-

nomic information further improved the accuracy of crude

oil price volatility forecasts. While the Markov-switching

others, Chan & Grant, 2016; Kang et al., 2009;Wanget al., 2016; Wei et al.,

2010), the topic of whether incorporating information from macroeco-

nomic variables can help forecast crude oil price volatility has received

less attention in the literature.

2As correctly pointed out by a reviewer,there are a number of studies that

focus on in-sample analysis (see, e.g., Antonakakis et al., 2017; Bu, 2014;

Gong & Lin, 2018).

Journal of Forecasting. 2020;39:1119–1141. wileyonlinelibrary.com/journal/for © 2020 John Wiley & Sons, Ltd. 1119

1120 NONEJAD

GARCH-MIDAS model considered by the authors out-

performed the single regime specification, it was more

difficult to find evidence that employing macroeconomic

variables, such as the first difference of log-global crude oil

production or Kilian's real global activity index (proxying

crude oil demand) led to out-of-sample forecast accuracy

gains relative to the benchmark. The remaining cited stud-

ies evaluate the predictive power afforded by combining

individual forecasts, for example, using Lasso regressions

and DMA as in Ma et al. (2018) or simple combinations as

in Meng and Liu (2019).

Despite the increasing number of studies on this topic, it

is rather difficult to draw general conclusions. The studies

employ different predictors, forecasting approaches, sam-

ple periods and predominately focus on population-level

predictability compared to finite-sample predictability.

The first interpretation focuses on properties of the

data-generating process (DGP) and evaluates whether data

are generated, such that the macroeconomic variable of

interest Granger-causes crude oil price volatility. The lat-

ter explores whether the specification of interest improves

the accuracy of out-of-sample point forecasts relative to

the benchmark. It is well known that the predictive model

of interest can fail to produce more accurate volatility

point forecasts than the benchmark while at the same time

perform well in terms of population-level predictability.3

Another important aspect is that a thorough analysis

of “local-level” predictability is largely neglected in the

current literature. Typically, researchers select specifica-

tions that forecast best, on average, over the out-of-sample

period. Henceforth, we refer to this as “global” predic-

tive evaluation. Evidently, this approach can lead to mis-

leading conclusions. For example, the specification with

the higher mean square error (MSE) computed over the

out-of-sample period can actually produce more accurate

forecasts when considering more recent data. Likewise,

the better performance of one model relative to another

can be entirely due to a specific episode. This is pointed

out in an interesting study by Giacomini and Rossi (2010).

Based on the assumption that relative predictive perfor-

mance of two models can vary substantially over the

out-of-sample period, the authors provide a local forecast

evaluation framework, which allows researchers to eval-

uate the relative predictive power of a predictor at each

out-of-sample observation. In a similar fashion to Rossi

and Sekhposyan (2010), we rely mainly on the fluctuation

test proposed by Giacomini and Rossi.

3This is due to the bias–variance tradeoff: The conditional bias reduction

afforded by including the predictor might not offset increased forecast

variance related to estimating the larger (more complex) model; see Paye

(2012) for an intuitive illustration.

The main contribution of this study is to address the

mentioned weaknesses and provide detailed answers to

important questions, such as: Which macroeconomic vari-

ables help forecast crude oil price volatility? Why is

the evidence of finite-sample predictability different from

population-level predictability? To what degree does rel-

ative performance change throughout the out-of-sample?

Can the evidence of predictability be translated into eco-

nomic meaningful gains? In a nutshell, the unique feature

of this study relies on its comprehensive approach, both in

terms of scope as well as in terms of the applied economet-

ric methodology.

The literature identifies several channels that can

drive time variation in crude oil price volatility; see

Drachal (2016) for a very good overview.4They include

demand/supply, the affinity between equity and crude

oil markets, US dollar exchange rate, investors' uncer-

tainty about fundamentals, and expectations concerning

the future state of the economy. These channels moti-

vate the set of forecasting variables considered in this

study. They include a measure of global crude oil produc-

tion, the default yield spread, variables related to current

and expected economic conditions, equity return realized

volatility, and measures of inflation.

Although we predominately rely on an out-of-sample

econometric approach, in-sample results also appear for

reference and comparison. This is of great importance,

since existence of in-sample predictability does not nec-

essarily imply success in terms of out-of-sample perfor-

mance. This focus follows recent findings in the literature

on equity return volatility prediction, where evidence of

out-of-sample predictability is less evident than in-sample

predictability (see, e.g., Christiansen et al., 2012; Paye,

2012).5The Diebold and Mariano (1995) and Clark and

West (2007) tests are the workhorses with regard to eval-

uating the evidence of finite-sample and population-level

predictability, respectively. The latter adds an adjustment

term to the out-of-sample difference in MSE that accounts

for parameter estimation noise. Likewise, the null hypoth-

esis involves the population difference in MSE between

the benchmark and the predictive model; that is, at the

population values of the parameters, the competing fore-

casts are equally accurate. By contrast, the null hypothesis

4Drachal (2016) evaluated the predictive impact of macroeconomic vari-

ables on West Texas Intermediate (WTI) crude oil price returns. Intu-

itively, we can argue that certain macroeconomic variables that predict

crude oil price returns also might help forecast crude oil price volatility.

5Paye (2012) found little evidence that employing macroeconomic vari-

ables in predictive regressions of equity return volatility succeeded

out-of-sample. Christiansen et al. (2012) found that combining fore-

casts via Bayesian model averagingimproved out-of-sample performance

relative to the benchmark, but not by much.