Connectedness between oil price shocks and US sector returns: Evidence
from TVP-VAR and wavelet decomposition
María Caridad Sevillano , Francisco Jareno˜, Raquel Lopez ´, Carlos Esparcia
ARTICLE INFO
JEL classification: C22, C51, G01, G15, Q4
Keywords: Crude oil prices, US sector returns, Connectedness, Wavelets, Portfolio
ABSTRACT
This paper examines the dynamic return and volatility connectedness between oil price shocks
(demand, supply, and risk shocks) and US sector returns from October 2001 to January 2022. For
this purpose, we combine the decomposition of the time series in time scales through the
wavelet approach with the application of the TVPVAR model proposed by Antonakakis et al.
(2020). Our results show the high dynamic connectedness between markets and allow the
identification of the role of all sector indices (except Communication Services, Utilities and Real
Estate) and risk shocks as net contributors of shocks to the system, whereas demand and supply
shocks are net receivers of spillovers. We further explore and document from a portfolio
performance perspective the benefits of diversified portfolios comprised of all consider sector
indices that include assets linked to the calculation of oil price shocks according to Ready (2018).
1. Introduction
Crude oil prices have important implications for financial markets due to their ability to affect
corporate earnings. Thus, investigating the links between oil and stock markets and their
transmission mechanisms has attracted a large body of previous literature (Guesmi and Fattoum,
2014; Bouri, 2015; Ferrer et al., 2018; Degiannakis et al., 2018.; Arampatzidis et al., 2021).To date,
there is widespread consensus among researchers about the interdependences between the two
markets and their time-varying nature, especially during episodes of financial turmoil (Jammazi
et al., 2017; Jiang and Yoon, 2020; Heinlein et al., 2021; Hung and Vo, 2021). Despite the extended
evidence on the links between oil prices and equity markets, Kilian (2009) and Kilian and Park
(2009) expand this issue to the context that it is necessary to know the underlying causes of
changes in oil prices to understand its impact in stock markets. In this line, a number of studies
adopt the methodology developed by Kilian (2009) to decompose oil prices into demand and
supply shocks for lowfrequency data or the alternative approach introduced by Ready (2018),
which improves Kilian’s methodology by classifying oil price changes into demand, risk and supply
shocks based on high-frequency data. If crude oil prices rise due to an increase in demand, the
sector in which the price of oil represents an output will increase its profits and, therefore, its
returns. Similarly, if crude oil prices rise due to a reduction in oil production, we would expect
corporate profits (and returns) to rise as well. Conversely, in both, companies that use oil as an
input would see their profits (and returns) fall. Changes in demand and supply for oil may be
motivated by changes in business activity, especially during different phases of the economic
cycle. For example, in times of economic expansion, increased economic activity will lead to an
increase in the demand for crude oil by companies that use crude oil as an input, so that oil prices
will rise. Conversely, in times of crisis, companies will slow down their activities and crude oil
prices will fall due to reduced demand. Changes in crude oil prices due to changes in the supply
side can also be motivated by changes in corporate production, depending on the phase of the
economic cycle. For example, in times of financial turmoil, the decline in oil production may
hamper the extraction process and it may be difficult to resume previous production at the
beginning of the economic recovery. Regarding risk shocks, their effect is largely tied to the
volatility of crude oil prices, suggesting a direct influence solely on the returns of oil producing
companies: a rise in oil price volatility implies increased risk for these companies, potentially
resulting in diminished investor confidence and decreased returns.
This paper investigates the dynamic return and volatility connectedness between oil price shocks,
distinguishing demand, risk and supply shocks (Ready, 2018), and the US equity market at the
sector level by using >20 years of daily data spanning from October 11, 2001, to January 14, 2022.
The current research contributes to the literature in several ways. First, to the best of our
knowledge, the present study is the first that assesses the dynamic interactions between
decomposed oil shocks and US sector indices by combining the disaggregation of time scales
through the wavelet approach and the analysis of connectedness through the Time-Varying
Parameter Vector Autoregressive (TVP-VAR) model. Thus, at the first stage, we use the Maximum
Overlap Discrete Wavelet Transform (MODWT) to identify changes in the connectedness
between oil price shocks and US sector stocks across time scales and to understand the
transmission mechanism at different frequencies (Reboredo and Rivera-Castro, 2014;
Maghyereh et al., 2019; Jiang and Yoon, 2020; Khraief et al., 2021). In the second step, we apply
the TVPVAR method proposed by Antonakakis et al. (2020), which extends and improves the
approach of Diebold and Yilmaz (2009, 2012, 2014) to study the dynamic return and volatility
spillovers between oil shocks and stock indices. In this sense, the TVP-VAR model explores the
dynamic transmissions between oil and the US stock market in a more flexible and robust way,
as it does not include rolling-window analysis. As a result, the TVP-VAR model is more accurate
in capturing changes in parameters without losing valuable observations. Combining wavelet and
TVP-VAR approaches allows one not only to determine whether there are interrelations between
the selected variables but also to detect the changes over time, exposing the variations in the
short and long term, as well as detecting the role of each type of oil shock. Another main
contribution is that we explore the diversification benefits of including assets linked to the
calculation of crude oil shocks in an equity-based portfolio composed of all sector indices by
considering return decomposition across time scales based on the wavelet method. Moreover,
for the portfolio performance analysis, we use different measures that consider not only total
risk, but also downside and upside risk. Finally, our sample covers different crisis periods, namely,
the Global Financial Crisis (GFC) of 2008 and the crisis caused by the COVID-19 pandemic, as well
as episodes of market specific shocks such as the collapse of oil prices in July 2014 through early
2016.
Our study offers several significant results. First, we find a considerable level of connectedness
between oil price shocks and US sector returns in terms of return and volatility. Second, the
interdependences vary over time and are stronger for the long-term horizon for the return
analysis, whereas we find the opposite in terms of volatility. With respect to the impact of the
GFC and COVID-19 crisis, our results show that they not only trigger a higher connectedness in
the system but can also change the role of receivers and transmitters of shocks of each variable.
Third, Industrials and Consumer Discretionary adopt a dominant position as transmitters of
shocks to the system in terms of both return and volatility regardless of the time horizon. The
role of demand and supply shocks (risk shocks) as net receivers (transmitters) of spillovers from
(to) the system is also consistent when the connectedness analysis is conducted in terms of both
return and volatility. Fourth, this work confirms that the magnitude of the connectedness
between each type of oil shock and every sector index differs across sectors. Thus, interactions
between demand shocks and sector stock returns are more pronounced for firms that rely on oil
as an input. For instance, we find the highest net pairwise return connectedness is reported for
the pair demand Shock-Energy. Supply shocks also exhibit the highest net pairwise return
connectedness with Energy, closely followed by sectors of companies linked to consumer
spending (i.e., Consumer Discretionary and Consumer Staples). In turn, risk shocks are more
closely connected to energy-intensive sectors, such as Industrials and Utilities. Overall, our
findings have several implications for portfolio management. On the one hand, the time-varying
nature of connectedness implies that investment strategies should be adaptive, requiring
continuous monitoring and adjustment to evolving market conditions. Moreover, we prove that
connectedness varies across time scales, thus affecting the design of trading strategies. On the
other hand, the heterogeneity in the connectedness results for the different oil shocks across
sectors suggests hedging opportunities for sectors that are not closely connected to oil prices.
Finally, our portfolio analysis of already diversified sector equity portfolios when assets linked to
oil shocks are included reveals the outperformance of such type of diversification by asset class,
both in terms of classical performance and downside and upside risk measures.
The rest of the paper is organised as follows. Section 2 provides a review of the relevant literature
on the interaction between oil and stock markets. In Section 3 we present the methodological
approach that we use in this study. Section 4 describes the dataset. Section 5 reports and
discusses the most remarkable empirical results. Finally, Section 6 offers some concluding
remarks.
3. Methodological approach
Following a strand of recent financial literature (Adekoya et al., 2022; Chen et al., 2022; Ren and
Lucey, 2022; Umar et al., 2022; Antonakakis et al., 2023), this paper applies the TVP-VAR model
to explore the mechanism of risk transmission between oil and the US sector stock market.
Specifically, the TVP-VAR connectedness estimation process is conducted in four steps and
applies to both returns and volatilities, but omitting the first two steps for the former. First, we
implement an AR (1) model to fit the time-dependence in the mean return equation:
where Rt denotes the assets’ return, εt are the residuals of current period, σt is the returns’
volatility and ηt is the standardized innovation of the process.
Second, we use the residuals from the AR model (innovations) and estimate some univariate
standard GARCH (1,1) specifications to model the returns’ dynamic volatility, where the
univariate parameters are estimated under the assumption that the standardized innovations
are normally distributed (Bollerslev, 1986):
where ω is the constant of the model, α is the parameter of the ARCH component model, β is the
parameter of the GARCH component model, εt− 1 is the model’s residual at me t − 1 and σ2 t−
1 is the variance of the previous period. The GARCH parameters are estimated using the Quasi-
Maximum Likelihood Estimation (QMLE) fitting method.
Third, the returns or GARCH volatilities series are decomposed at different frequencies (from high
to low) through the wavelet MODWT approach (see, e.g., Berger, 2015; Dewandaru et al., 2016;
Yang et al., 2018; Yang, 2019, among others). We perform the series decomposition via a set of
convolutions using the MODWT wavelet and the chosen decomposition level (in our case, we use
up to seven levels of decomposition). The decomposed signals based on multiresolution
decomposition in MODWT can be calculated as follows: (For further details regarding the
MODWT, please see Appendix A.)
where the functions SJ (t) and DJ (t) indicate the smooth and detail signals and constitute a
decomposition of a signal into its orthogonal components in terms of specific frequency.
Specifically, following Lim (2020) and Esparcia et al. (2022), among others, we decompose our
return and volatility series into seven different time scales, namely, d1 (2–4 days), d2 (4–8 days),
d3 (8–16 days), d4 (16–32 days), d5 (32–64 days), d6 (64–128 days) and d7 (128–256 days). The
short-term horizon is defined as {D1 = (d1 + d2 + d3)} and represents high-frequency movements
from 2 to 16 days. In turn, the long-term horizon is defined as {D7 = (d4 + d5 + d6 + d7)} and
represents fluctuations occurring from 16 to 256 days or low-frequency movements. The
MODWT produces detail coefficients and approximation coefficients at each decomposition
level. The detail coefficients represent the finer or high-frequency changes in the series, while
the approximation coefficients represent the smoother or low-frequency changes. The detail and
approximation coefficients obtained can be analysed to identify patterns, trends or changes at
different time scales.
Fourth, TVP-VAR connectedness (Antonakakis et al., 2020) is calculated for both the
undecomposed return or volatility series and their high-frequency D1 and low-frequency D7
wavelet decomposition. One lag and 10-step ahead forecasts are considered for the TVP-VAR
computation. The TVP-VAR model for each of the undecomposed series and decompositions is
defined as follows:
where yt represents the m× 1 vector of endogenous variables and zt− 1 represents the mp× 1
vector. At is an m × mp-dimensional matrix, and Ait is an m× m matrix, whereas vec (At) is the
vectorisation of At, which is an m2p× 1-dimensional vector. ϵt and et are m × 1 and m2p × 1-
dimensional vectors, respectively. Ωt− 1 represents all informa on available un l t – 1. Finally,
the time-varying variance–covariance matrices ∑t and Et are m × m and m2p × m2p matrices,
respectively.
The connectedness theory, as introduced by Diebold and Yilmaz (2014), relies on the use of
generalized forecast error variance decomposition (GFEVD) (Koop and Korobilis, 2014) and
generalised impulse response functions (GIRF). To implement this approach, it is essential to
transform the TVP-VAR model into its vector moving average (VMA) representation. This
transformation is expressed by the following equation:
This index allows us to estimate how much a shock in one variable directly affects other variables.
There are additional connectedness measures that can be calculated based on the total
connectedness index. First, the total directional connectedness to others (TO) shows the
transmission of a shock from variable i to all other variables j:
Second, the total directional connectedness from others (FROM) illustrates the transmission of a
shock on variable i, i.e., which is received from all variables j:
Finally, the net total directional connectedness (NET) is obtained by subtracting Eq. (9) from Eq.
(8), showing the effect that variable i has on the system. Thus, if Cgi,t > 0, variable i impacts the
system more than it could be influenced by that; if Cgi,t < 0,variable i is driven by the system;
and, if Cgi,t = 0, variable i has no influence, nor is it influenced by the system.
4. Data
To achieve the purpose of this paper, we use two different datasets. On the one hand, the first
dataset consists of daily trading closing prices of 11 sector indices of the US stock market:
Consumer Discretionary, Consumer Staples, Energy, Financials, Health Care, Industrials,
Information Technology, Materials, Real Estate, Communication Services, and Utilities. We
collect equity market data at the aggregate and sector level from Thomson Reuters DataStream.
On the other hand, Ready (2018) introduces a novel methodology to decompose oil price changes
into demand, supply and risk shocks. Demand shocks are identified as the portion of
contemporaneous returns on a global stock index of oil-producing firms orthogonal to VIX
innovations, which represent changes in the market discount rate. Supply shocks are computed
from oil price changes orthogonal to both demand and risk shocks. Risk shocks are derived from
an ARMA (1,1) model on the VIX index, with residuals serving as innovations (Umar et al., 2021a).
These orthogonal shocks together account for all variations in oil prices. Following Ready (2018),
we decompose oil price shocks based on three variables: (i) the World Integrated Oil and Gas
Producer Index (WIOGPI), (ii) one-month crude oil futures returns on the second nearest maturity
for the NYMEX and (iii) the CBOE volatility index (VIX). We use information in daily asset prices to
better capture the dynamic interactions between stock returns and decomposed oil price shocks
(Umar et al., 2021a; Zhou and Geng, 2021; Al-Fayoumi et al., 2023; Elsayed et al., 2023).2 The
WIOGPI is used as a proxy for an index that collects oil-producing firms and is obtained from
Thomson Reuters DataStream. NYMEX crude-light sweet oil futures contracts are included as a
measure that captures crude oil changes and are retrieved from [Link]. Finally, we use
the VIX index as a proxy for changes in risk with data collected from the CBOE website. Our
sample period ranges from October 11, 2001, to January 14, 2022, based on the longest data
availability of all series.
Table 1 provides the descriptive statistics and results of unit-root tests for the variables under
study. To account for heteroscedasticity patterns, we obtain time series of daily index continuous
returns as Rt = ln (Pt/Pt-1), where Pt and Pt-1 refer to closing prices of the US sector indices at
business days t and t − 1, respectively. Over the full sample period, Consumer Discretionary and
Information Technology saw the highest average daily returns (0.0004), while the highest
volatility (standard deviation) was registered by Real Estate and Financials (standard deviation of
0.019).3 Concerning oil price shocks, we observe that demand shocks have a negative mean (−
0.0004), whereas the mean of supply shocks is positive (0.0006), with values close to zero though.
Risk shocks stand out as the most dynamically fluctuating type of shocks, boasting a standard
deviation of 71.778. Furthermore, they exhibit the highest mean value in absolute terms (−
0.0112). Nonetheless, throughout the sample, discernible instances emerge where both demand
and supply shocks exhibit significantly greater (negative) magnitudes. In these episodes, the
negative mean of risk shocks undergoes a transformation, shifting towards positive values of
even greater magnitude.4 All variables included in this research show negative skewness (except
demand and risk shocks) and excess kurtosis. In addition, the Jarque-Bera test statistics also show
a non-normal distribution. Finally, the standard unit root (augmented Dickey-Fuller (ADF) (Dickey
and Fuller, 1981) and Phillips-Perron (PP)) (Phillips and Perron, 1988) and stationarity
(Kwiatkowski-Phillips-Schmidt-Shin (KPSS) (Kwiatkowski et al., 1992)) test results confirm that all
log-returns of US sector indices and demand, risk and supply shock series are stationary.