Effect of outliers on volatility forecasting and Value at Risk estimation in crude oil markets

• Date: 01 September 2016
• DOI: http://doi.org/10.1111/opec.12079
• Published date: 01 September 2016
• Author: Selvamuthu Dharmaraja,Himanshu Sharma

Effect of outliers on volatility forecasting

and Value at Risk estimation in crude oil

markets

Himanshu Sharma* and Selvamuthu Dharmaraja**

*Graduate Student, Department of Mathematics, Indian Institute of Technology Delhi, New Delhi 110016,

India.

**Professor, Department of Mathematics, Indian Institute of Technology Delhi, New Delhi 110016, India.

Email: dharmar@maths.iitd.ac.in

Abstract

Crude oil markets are one of the most volatile commodity markets. The effect of shocks on

volatility is of concern to policy makers and market participants. A better understanding of how

shocks affect volatility over time would be helpful to participants of financial markets. This article

investigates effect of outliers in three markets—Brent, West Texas Intermediates (WTI) and

Organization of Petroleum Exporting Countries (OPEC). We compare forecasting and Value at

Risk (VaR) estimation accuracy of GARCH family models with and without outlier adjustment

and conclude that conditional volatility models work better on outlier adjusted data with respect to

out-of-sample performance but not necessarily with respect to VaR estimation. Furthermore, we

conclude that VaR violations in crude oil markets are independent.

1. Introduction

Crude oil is the most actively traded commodity in world, both in spot as well as futures

market and its price is a significant factor in modelling macroeconomic variables

including inflation, GDP, stock market returns, interest rates and exchange rates for

importing as well as exporting countries. Apart from macroeconomic importance, oil

price volatility is a crucial factor for portfolio selection, managing Value at Risk (VaR)

and option pricing for investors. In this context accurate modelling and forecasting of

crude oil price and volatility are foundations for determining risk in managing oil

supply and derivatives apart from being a matter of concern to energy researchers and

policy makers. In this regard time-series models have been used intensively to model

volatility of crude oil markets. However, for modelling volatility in financial market

simple Autoregressive Moving Average (ARMA) models fall short due to their

homoscedastic nature. Volatility in financial market follows time-varying clustering

property, i.e. periods of swings followed by periods of calm. In such scenarios,

©2016 Organization of the Petroleum Exporting Countries. Published by John Wiley & Sons Ltd, 9600 Garsington

Road, Oxford OX4 2DQ, UK and 350 Main Street, Malden, MA 02148, USA.

276

conditional heteroscedastic models come into picture. The first of these models, the

Autoregressive Conditional Heteroscedasticity, or ARCH model was introduced by

Engle (1982) and it was followed by Generalized Autoregressive Conditional

Heteroscedasticity, or GARCH model by Bollerslev (1986). These models consider

evolution of volatility at time tas a function of information available before t.

Furthermore, there is an intuition of asymmetric leverage effect in financial market due

to investor’s choice implying that volatility is higher in times of negative returns as

compared with times of positive returns. There are variants of GARCH model (described

later) which helps in capturing this asymmetric leverage effect.

Historicallyfinancial markets have been hugelyaffected by outliers or jumps such as the

Operation Desert Storm with a 34% return for WTI market. Financial institutions and

economists can help from a better understanding of how these shocks affect volatility over

time. In general, thesejumps relate to sudden change in sentiment of buyersand sellers due

to some political announcement, criminal attack, war or policy change in the markets.

However, in modelling volatility of crude oil, or any commodity in general, these outliers

should be identifiedand data be adjusted appropriately.For investors in crudeoil market one

of the crucialtasks is to measure VaR to access theirportfolio. VaR is a quantitativemeasure

of the value that is at risk for long (or short) positions with a given level of confidence.

However,shocks in oil market can bias volatilityforecast and VaR prediction, thus affecting

oil market investors. Our goal is to study the effect of shocks on forecasting accuracy of

conditional volatility models and VaR estimation in crude oil market.

2. Literature review

Engle and Patton (2001) outlined some stylized facts about volatility that should be

captured in a model. These include persistence, mean reverting nature, leverage effect,

non-normality and effect of exogenous variables. They used raw data of Dow Jones

Industrial Index to illustrate these facts and the ability of GARCH family models to

capture these features. Except influence of exogenous variables, each model framework

that is applied in this article is tested for these trends, unlike most of the previous articles

on crude oil markets where all trends are not tested simultaneously. Apart from these

features of volatility, studies suggest that financial data is affected by contaminated

observations, called outliers, which reflect extraordinary, infrequently occurring events or

shocks that have important effect on macroeconomic time series. These types of event

include changes in policy regimes,natural disasters and periods of wars. Such events have

undesirable effects on estimates of the parameters of volatility models as discussed in

Charles (2008). To the best of our knowledge, only Charles and Darne (2014) takes into

account presence of outliers using GARCH-based outlier detection scheme in analysing

crude oil market volatility and no work has been done in crude oil market which analyses

©2016 Organization of the Petroleum Exporting Countries OPEC Energy Review September 2016

Effect of outliers on crude oil volatility & VaR 277