. International Journal of Energy Economics and Policy | Vol 8 • Issue 3 • 201814 International Journal of Energy Economics and Policy ISSN: 2146-4553 available at http: www.econjournals.com International Journal of Energy Economics and Policy, 2018, 8(3), 14-21. The Performance of Hybrid ARIMA-GARCH Modeling and Forecasting Oil Price Chaido Dritsaki* Department of Accounting and Finance, School of Management and Economics, Western Macedonia University of Applied Sciences, Greece. *Email: dritsaki@teiwm.gr ABSTRACT Modeling and forecasting oil prices is an important issue for many researchers. One of the methods used in forecasting oil prices is Box-Jenkins methodology through ARIMA models. Although these models provide accurate forecasting over a short time period, they are not able to handle the volatility and nonlinearity presented on data series. For this reason, on this paper we examine a hybrid ARIMA-GARCH model in order to forecast the volatility in the return of oil prices. Moreover, on this paper, the Box-Cox transformation is used for data smoothing for the stabilization of variance and reduction of heteroscedasticity. Parameters’ estimation in the hybrid ARIMA-GARCH model is employed by ML (Maximum Likelihood) method using the steps of Marquardt’s Algorithm (1963) and Broyden-Fletcher-Goldfarb-Shanno algorithm for optimization. The results of the paper showed that the hybridation of ARIMA (33,0,14)-GARCH (1,2) model following normal distribution is the most suitable for forecasting the returns of oil prices. Finally, we use both the dynamic and static procedure for forecasting. The results showed that the static procedure provides with better forecasting than the dynamic. Keywords: ARIMA, GARCH, Oil price forecasting, Hybrid ARIMA-GARCH, Box-Cox transformation JEL Classifications: C33, O13, Q43 1. INTRODUCTION Oil is considered to be one of the most important goods in the world. Its use is ubiquitous in everyday life. Due to its uniqueness, researchers need to develop a better understanding of its price dynamics so that industries that are supplied with or consume vast quantities of oil can take optimal decisions. Oil’s dynamic modeling is a difficult task to accomplish because its price cannot be predicted in various time periods and it depends on many factors. During the past decades, oil price shows volatility. In 1999, during the Asian crisis, Iraq’s decision to increase oil production caused the decrease of oil price in the lowest level. In 2001, the dot-com bubble caused panic, reducing oil price until the beginning of 2002. Following this upheaval, global economy regained its momentum, resulting in an upward trend of oil price. The factors that led to the reduction of oil production were the hostile relationships between the U.S. with the production countries. Price oil reached its highest peak, when the housing bubble burst in the U.S. causing credit crisis. The reduction of oil price that followed until its stabilization after the economic crisis of 2008, is a challenge for researchers investigating for a model who would forecast these situations. Moreover, forecasting returns of oil prices, which is considered a basic economic variable, influences consumers’ decisions, businesses and financial institutions, as well as governments. Timely and reliable predictions of oil prices provide important information on those in the financial markets. This paper tries to develop a hybrid ARIMA-GARCH model in order to investigate and forecast the characteristics of volatility in oil price using daily data from 20 October 1997 until 31 May 2017 for a total of 4980 observations. The rest of the paper is as follows: Section 2 provides a brief literature review. Section 3 presents the analysis of methodology. Section 4 summarizes the data and the descriptive statistics. The empirical results are provided in Section 5 and Section 6 proposes the forecasting results. Finally, the last section offers the concluding remarks. Dritsaki: The Performance of Hybrid ARIMA-GARCH Modeling and Forecasting Oil Price International Journal of Energy Economics and Policy | Vol 8 • Issue 3 • 2018 15 2. LITERATURE REVIEW During last years, there is great interest from researchers for modeling volatility and forecasting oil price. Hansen and Lunde (2005) argued that information related to oil price is necessary for modeling producers’ oil price as well as calculating risk measures. Also, others found out that variations in oil prices have consequences not only in those countries that import but also in those countries that export oil. Rothemberg and Woodford (1996) have studied oil price instability in relation to inflation. Hamilton and Herrera (2004) in their paper argue the relationship between oil price and exports. Yang et al. (2002) refer to the relationship between investment and oil price and Elder and Serletis (2009) examine the relationship between oil price and monetary policy. Furthermore, there are a number of papers relating oil price with inflation, exchange rate, production and employment decrease and competitiveness loss. In all these papers, techniques using autoregressive models are employed (Mirmirani and Li, 2004), cointegration and error correction models (Mohammadi 2009), GARCH models (Sadorsky, 2006, and Agnolucci, 2009), as well as neural networks (Yu, et al. 2008). Finally, it can be said that all papers reach the same conclusion that the returns of oil prices present a unit root, have excessive kurtosis, are negatively skewed and don’t follow Gaussian distribution. Also, the volatility on the returns of oil prices is clustering and persistent, consistent with predictions of GARCH variety models. 3. THEORETICAL BACKGROUND The development and designing of ARIMA models as forecasting tools of financial-economic variables is known as Box-Jenkins Methodology (1976). This methodology tries to find an ARIMA (p, d, q) model which satisfies the stochastic procedure where the sample derived from. Box-Jenkins methodology consists of four repetitive steps: Identification model, parameters’ estimation, diagnostic tests and model’s forecasting ability. 3.1. ARIMA Models The autoregressive integrated moving average model of order p and q, ARIMA (p, d, q) is one of the time series forecasting methods for the nonstationary data series. The ARIMA (p, q, d) can be expressed as: ϕp(L)(1-L) d(yt-μ)=ϑq(L)et (1) or 1 L 1 L y ¼ = 1 L ei i i=1 p d t j j j=1 q t−       −( ) −( ) −        ∑ ∑ϕ ϑ (2) Where,  p i i i=1 p L =1- L( ) ∑ and  q j j j=1 q L =1- L( ) ∑ are polynomials in terms of L of degree p and q. yt is. the time series, and et is the random error at time period t, with μ is the mean of the model. d is the order of the difference operator. φ1, φ 2,…, φp and ϑ1,ϑ2,…, ϑq are the parameters of autoregressive and moving average terms with order p and q respectively. L is the difference operator defined as Δyt=yt-yt-1=(1-L)yt. ARIMA models can be estimated following the Box-Jenkins approach. Given that the stationary procedure is essential for an ARIMA model, during the identification step data are transformed so that the time-series will become stationary. The stationary procedure is a necessary condition in building an ARIMA model. 3.2. Box-Cox Transformation Method Box-Cox (1964) on their paper used a mathematical formula for the data transformation so that these can be more normally distributed and the variance equation corrects normality, linearity and reduces heteroscedasticity. The formula of the Box-Cox transformation is: y y if y if t t t * - ln = ≠ ( ) =      λ λ λ λ 1 0 0 (3) Where, yt yt are actual data in time t. yt * are the transformed data in time t. λ is the minimum value of mean square error of residuals. The transformation of the above equation is valid only for positive values of time series yt>0. If the values of time series contain also negative values, then the transformation will get the following form: y y if y if t t t * ( ) ln λ λ λ λ λ λ λ = +( ) − ≠ +( ) =       2 1 1 2 1 1 1 0 0 (4) Where, λ1 is the transformation parameter, and λ2 is chosen such that yt>−λ2. The main objective in the analysis of data transformation in Box- Cox (1964) technique is to calculate (estimate) λ parameter. For this reason, two approaches are necessary. The first approach is using the maximum likelihood method to estimate data because it facilitates the calculation of likelihood function. Also, maximum likelihood method is easy to obtain an approximate confidence interval for λ. The second approach uses Bayesian method to confirm if the model is fully specified. 3.3. GARCH Models GARCH models are used mainly for modeling financial time series that present time-varying volatility clustering. The general Dritsaki: The Performance of Hybrid ARIMA-GARCH Modeling and Forecasting Oil Price International Journal of Energy Economics and Policy | Vol 8 • Issue 3 • 201816 GARCH (r, s) model for the conditional heteroskedasticity according to Bollerslev (1986) has the following form: yt=μt+zt (5) Where, μt is conditional mean of yt. z is the shock at time t. zt=σtet (6) Where, et→iid N(0,1). σ α α β σt i t i i r i t i i s z 2 0 2 1 2 1 = + +− = − = ∑ ∑ (7) Where, σt 2 is the conditional variance of yt. α0 is a constant term. r is the order of the ARCH terms. s is the order of the GARCH terms. αi and βi are the coefficients of the ARCH and GARCH parameters, respectively. with constrains: α α β α β 0 11 0 0 1 2 0 1 2 > ≥ = ≥ = + == ∑ i i i i i s i r for i r for i s , , ,..., , , ,..., ∑∑ <          1 3.4. Hybrid ARIMA-GARCH Model In order to recommend a hybrid ARIMA-GARCH model, two stages should be applied. In the first stage, we use the best ARIMA model that fits on stationary and linear time series data while the residuals of the linear model will contain the non-linear part of the data. In the second stage, we use the GARCH model in order to contain non-linear residuals patterns. This hybrid model, which combines ARIMA and GARCH model containing non linear residuals patterns, is applied to analyze and forecast the returns of oil prices. 3.5. Estimation of Hybrid ARIMA-GARCH Model The hybrid ARIMA-GARCH model is a non linear time series model which combines a linear ARIMA model with the conditional variance of a GARCH model. The estimation procedure of ARIMA and GARCH models are based on maximum likelihood method. Parameters’ estimation in logarithmic likelihood function is done through nonlinear Marquardt’s algorithm (Marquardt, 1963). The logarithmic likelihood function has the following equation: ln [( ), ln[ ( ( )), ] - ln[ ( )]L y D zt t t t T θ θ υ σ θ=      = ∑ 1 2 2 1 (8) Where θ is the vector of the parameters that have to be estimated for the conditional mean, conditional variance and density function, zt denoting their density function, D(zt(θ),υ), is the log-likelihood function of [yt(θ)], for a sample of T observation. The maximum likelihood estimator θ̂ for the true parameter vector is found by maximizing (8) (Dritsaki, 2017). 3.6. Diagnostic Checking of Hybrid ARIMA-GARCH Model The diagnostic tests of hybrid ARIMA-GARCH models are based on residuals. Residuals’ normality test is employed with Jarque and Bera (1980) test. Ljung and Box (1978) (Q-statistics) statistic for all time lags of autocorrelation is used for the serial correlation test. Also, for the conditional heteroscedasticity test we use the squared residuals of autocorrelation function. 3.7. Forecast Evaluation On hybrid ARIMA-GARCH models we use both the static and dynamic forecast. The dynamic forecast, also known as n-step ahead forecast, uses the actual lagged value of Y variable in order to compute the first forecasted value. The static forecast (one-step ahead forecast) of Yt+1 based on an hybrid ARIMA-GARCH model is defined as: ( ) p q t t 1 t t -1 0 i t 1-i j t 1- j i 1 j 1 Ŷ (1) Y Y , Y ,... Y+ + + = = = Ε = φ + φ + θ ε∑ ∑ (9) Where the εs follow the stated GARCH model. To evaluate the forecast efficiency, we use two statistical measures, mean squared error (MSE) and mean absolute error (MAE). MSE it computes the squared difference between every forecasted value and every realised value of the quantity being estimated, and finds the mean of them afterwards. MSE has the following formula: ( ) n 2 i i i=1 1 ˆMSE= Y -Y n ∑ (10) Where, Yi is the vector of observed values of the variable being predicted. iŶ is the vector of n predictions. MAE it computes the mean of all the absolute, instead of squared, forecast errors. The formula is the following: n i i i 1 1 ˆMAE Y - Y n = = ∑ (11) 4. DATA AND DESCRIPTIVE STATISTICS The data used in our paper come from energy information administration. Data are daily covering the period 20 October 1997 Dritsaki: The Performance of Hybrid ARIMA-GARCH Modeling and Forecasting Oil Price International Journal of Energy Economics and Policy | Vol 8 • Issue 3 • 2018 17 until 31 May 2017 including 4980 observations. Using the value of λ = 0.3374 which was calculated from XLSTAT of Excel and Equation (3) from Box-Cox, we transformed oil price time-series. In the following Figure 1, the closing values as well as transformed values in oil prices are presented. From Figure 1 we can see that transformed data are less volatile from the initial ones. Then, we examine normality in oil prices before and after the transformation. From Figure 2, we can see that the transformed data have a better adjustment in relation to normal distribution. In the following Table 1, the descriptive statistics of Brent index are presented before and after Box-Cox transformation. The Table 1 show that standard deviation of transformed series has reduced from 33.78242 to 2.392691. Also, we can see that in the transformed data there is no normality even though it has been reduced. The daily return of oil price index after the data transformation is calculated as follows: r =ln y y ×100= ln y -ln y ×100t t t-1 t t-1       (12) Where, yt is the daily closing price of oil price index at day t. yt-1 is the daily closing price of oil price index at previous day. rt are the daily returns of oil price index. In the following Figure 3 we present the closing prices, returns and volatility of oil price index after the Box-Cox transformation. The volatility of oil prices is estimated from the daily squared returns (Sadorsky, 2006). From Figure 3 we note that daily closing prices of oil follow a random walk whereas returns from oil prices seem to be stationary. The confirmation in stationarity of the returns of oil price index is done with Dickey-Fuller (1979; 1981) and Phillips-Perron (1998) unit root tests. The results of Table 2 confirm that returns of oil prices are stationary in their levels. Consequently, for an ARIMA (p, d, q) model the value for d = 0. On Table 3, the descriptive statistics of the returns on Brent are presented. The results on Table 3 show that the mean of daily return on oil price is quite small in relation to its standard deviation. Also, the return in oil prices appear small positive asymmetry and leptokurtosis with fat tails and Jarque and Bera (1980) statistic proves that the return of oil prices don’t follow normal distribution. On Table 4, the autocorrelation diagram on the return of oil price is presented. Figure 1: Daily closing prices of oil before and after the transformation Figure 2: Normality test in daily prices of oil before and after transformation Table 1: Descriptive Statistics of Brent index before and after Box-Cox transformation Brent Box-Cox brent Mean 58.86256 Mean 8.281836 Median 53.15500 Median 8.362000 Maximum 143.9500 Maximum 12.88700 Minimum 9.100000 Minimum 3.280000 Standard deviation 33.78242 Standard deviation 2.392691 Skewness 0.436870 Skewness −0.080189 Kurtosis 1.951764 Kurtosis 1.919715 Jarque-Bera 386.4104 Jarque-Bera 247.4930 Probability 0.000000 Probability 0.000000 Sum 293135.6 Sum 41243.54 Sum Sq. Dev. 5682294.0 Sum Sq. Dev. 28504.63 Observations 4980 Observations 4980 Table 2: Unit root tests of the returns of Brent Brent Augmented Dickey-Fuller Phillips-Perron C C, T C C, T −69,3773*(0) −69,3816*(0) −69,3794*[9] −69,3809*[8] *,**,***show significant at 1%, 5% and 10% levels respectively. The numbers within parentheses followed by ADF statistics represent the lag length of the dependent variable used to obtain white noise residuals. The lag lengths for ADF equation were selected using Schwarz Information Criterion. MacKinnon (1996) critical value for rejection of hypothesis of unit root applied. The numbers within brackets followed by PP statistics represent the bandwidth selected based on Newey-West (1994) method using Bartlett Kernel. C=Constant, T=Trend Dritsaki: The Performance of Hybrid ARIMA-GARCH Modeling and Forecasting Oil Price International Journal of Energy Economics and Policy | Vol 8 • Issue 3 • 201818 The above results indicate that autocorrelation coefficients on lags 14,15,18,33 and 35 and partial autocorrelation on lags 14,15,18,28 and 33 on the return of oil price is larger than both standard errors ( . )± = ± = ± 2 2 4980 0 02934 n . Furthermore, according to Ljung-Box statistic on Table 4, the existence of ARCH or GARCH cannot be rejected after the 13th lag. 5. EMPIRICAL RESULTS Following, we define the form of ARIMA (p, d, q) model given the results of autocorrelation and partial autocorrelation diagram on Table 4. The p and q parameters of ARIMA model are determined from the coefficients of partial autocorrelation and autocorrelation respectively, comparing them with critical value ± 2 n =± 2 4980 =±0.02934 . From the values of coefficients of partial autocorrelation and autocorrelation on Table 4 we see that the value of p will be p = 14 or p = 15 or p = 18 or p = 28 or p = 33 and for q will be q = 14, or q = 15 or q = 18, or q = 33 or q = 35. Using the above values, we choose the best ARIMA (p, 0, q) model from the smaller values of Schwarz criterion. Table 5 provides the values of p and q. The results of Table 5 show that ARIMA (33,0,14) model is the most suitable for the returns of oil index. In the following Table 6 we get the estimations of this model. Given the ARCH effects on the returns of oil price index, we proceed with the estimations of hybrid ARIMA-GARCH models to examine the volatilities that exist in the related returns of oil price. Moreover, from Figure 3 the returns in oil prices show cluster in volatility. To catch this cluster we should use ARIMA as well as GARCH models. Thus, in the levels this time-series on returns of oil prices we have to find out the appropriate hybrid ARIMA- GARCH model. Estimation parameters’ is held with Maximum Likelihood method using the steps of Marquardt’s algorithm (1963) and also Broyden-Fletcher-Goldfarb-Shanno algorithm optimization. Estimation parameters’ as well as diagnostic tests Figure 3: Closing prices, returns and volatility of brent index Table 3: Descriptive statistics of the returns of brent Indice Brent Mean 0.0093 Median 0.0104 Maximum 9.3977 Minimum −10.2945 Standard deviaton 1.1389 Skewness 0.0288 Kurtosis 9.0624 Jarque and Bera 7625.40 Probability 0.000000 Q (24) 54.619* Observations 4979 *indicate statistical significance at 1% Dritsaki: The Performance of Hybrid ARIMA-GARCH Modeling and Forecasting Oil Price International Journal of Energy Economics and Policy | Vol 8 • Issue 3 • 2018 19 of normality, autocorrelation and conditional heteroscedasticity are presented on Table 7. From the Table 7 we can see that the hybrid ARIMA (33,0,14)-GARCH (1,1) with normal distribution is the most appropriate. All coefficients are statistical significant and there seems no problem in diagnostics tests (except normality). Persistence gets the value 0.99996 indicating high persistence in volatility in the oil price index. Table 5: Comparison of ARIMA models within the range of exploration using Schwarz criteria Returns of Brent ARIMA (14,0,14) 3.098784 ARIMA (14,0,15) 3.098888 ARIMA (14,0,18) 3.099006 ARIMA (14,0,33) 3.097087 ARIMA (14,0,35) 3.099275 ARIMA (15,0,14) 3.098691 ARIMA (15,0,15) 3.101659 ARIMA (15,0,18) 3.100680 ARIMA (15,0,33) 3.098897 ARIMA (15,0,35) 3.100866 ARIMA (18,0,14) 3.098830 ARIMA (18,0,15) 3.100695 ARIMA (18,0,18) 3.101842 ARIMA (18,0,33) 3.099034 ARIMA (18,0,35) 3.100882 ARIMA (28,0,14) 3.099227 ARIMA (28,0,15) 3.101020 ARIMA (28,0,18) 3.101108 ARIMA (28,0,33) 3.099381 ARIMA (28,0,35) 3.101431 ARIMA (33,0,14) 3.096988 ARIMA (33,0,15) 3.098987 ARIMA (33,0,18) 3.099104 ARIMA (33,0,33) 3.099902 ARIMA (33,0,35) 3.099255 Table 6: Estimations of the ARIMA (33,0,14) model of the returns of Brent Variables ARIMA (33,0,14) AR (33) −0.054488 (0.000) MA (14) 0.058173 (0.000) SIGMASQ 1.289110 (0.000) Log likelihood −7697.181 Jarque and Bera 7443.852 (0.000) Q2 (5) 414.48 (0.000) X2 (10) 329.4274 (0.000) X2 (20) 387.4811 (0.000) X2 (30) 407.5273 (0.000) AR and MA denote the autoregressive and moving average terms respectively. SIGMASQ is the coefficient of variance error. Q2(5)is the Q-Statistic of correlogram of squared residuals at fifth lag. X2 is the value of Chi-square of ARCH test and (10), (20), (30) are the corresponding lags. P values in parentheses denote probability. Table 4: Correlogram on the return of oil prices Autocorrelation Partial correlation S. no AC PAC Q-Stat Prob 1 0.017 0.017 1.3596 0.244 2 0.006 0.006 1.5696 0.456 3 0.005 0.005 1.6877 0.640 4 −0.009 −0.009 2.0574 0.725 5 0.000 0.000 2.0576 0.841 6 −0.029 −0.029 6.3057 0.390 7 0.024 0.025 9.1804 0.240 8 0.008 0.007 9.5011 0.302 9 0.013 0.013 10.340 0.324 10 −0.005 −0.006 10.458 0.401 11 −0.023 −0.023 13.065 0.289 12 −0.011 −0.011 13.654 0.323 13 0.018 0.020 15.226 0.293 14 0.054 0.053 29.679 0.008 15 0.036 0.034 36.041 0.002 16 0.012 0.009 36.787 0.002 17 −0.011 −0.014 37.393 0.003 18 −0.035 −0.034 43.354 0.001 19 −0.013 −0.009 44.149 0.001 20 −0.021 −0.017 46.445 0.001 21 −0.024 −0.024 49.405 0.000 22 −0.013 −0.016 50.268 0.001 23 0.008 0.005 50.574 0.001 24 0.028 0.027 54.619 0.000 25 0.014 0.017 55.558 0.000 26 0.013 0.015 56.405 0.001 27 −0.002 −0.003 56.417 0.001 28 −0.027 −0.032 60.169 0.000 29 −0.008 −0.012 60.483 0.001 30 0.010 0.010 61.023 0.001 31 −0.002 −0.001 61.035 0.001 32 −0.013 −0.011 61.949 0.001 33 −0.053 −0.052 76.224 0.000 34 −0.004 0.000 76.292 0.000 35 −0.031 −0.024 81.064 0.000 36 0.003 0.010 81.100 0.000 Dritsaki: The Performance of Hybrid ARIMA-GARCH Modeling and Forecasting Oil Price International Journal of Energy Economics and Policy | Vol 8 • Issue 3 • 201820 6. FORECASTING For the forecasting of hybrid ARIMA (33,0,14)-GARCH (1,1) model we use both the dynamic and static procedure. The dynamic procedure computes forecasting for periods after the first sample period, using the former fitted values from the lags of dependent variable and ARMA terms. This procedure is called n-step ahead forecasts. The static procedure uses actual values and not forecasted values of the dependent variable. This procedure is called one step-ahead forecast. In the following Figure 4, we present the criteria for the evaluation of forecasting the returns of oil price using the dynamic and static forecast respectively. The Figure 4 indicates that the static procedure gives better results rather than the dynamic because both mean squared error and MAE are lower in the static process. 7. DISCUSSION AND CONCLUSION This paper aims to create a hybrid model combining ARIMA model with GARCH models of high volatility in order to analyze and forecast the return of oil price. According to various papers, the returns of oil prices have unit root, excessive kurtosis and negative skewness so they don’t follow the Gaussian distribution. Instead, the transformation of Box-Cox was used to smooth the data. This resulted in the stabilization of variance and the decrease in heteroscedasticity. The empirical results of the paper showed that the hybrid ARIMA (33,0,14)-GARCH (1,1) provides the optimal results and improves estimation and forecasting in relation to previous methods. In conclusion, the combination of robust and flexible linear ARIMA models and the power of non linear GARCH models in handling volatility and the risk return of oil price, made hybrid models to be the most suitable for analysis and forecasting of time series. Figure 4: Dynamic and static forecast of the ARIMA (33,0,14) - GARCH model of the returns of brent Table 7: Estimates of the ARIMA (33,0,14) - GARCH models of the returns of Brent Distribution ARCH (1) GARCH (1,1) Normal t-student GED Normal t-student GED Mean equation AR (33) −0.048140* −0.031854* −0.02396** −0.03227** −0.02503** −0.021140 MA (14) 0.055167* 0.059645* 0.051390* 0.043054* 0.046523* 0.045317* Variance equation α0 1.070661* 1.051116* 1.015478* 0.002279* 0.001956* 0.002073** α1 0.176842* 0.232728* 0.204613* 0.055819* 0.041448* 0.047791* β1 --- --- --- 0.944145* 0.958353* 0.951375* T-Dist. DOF/GED parameter 4.144466* 1.101958* --- 6.977499* 1.389278* Diagnostic tests Persistence --- --- --- 0.99996 0.999801 0.999166 Log L −7558.006 −7231.040 −7239.278 −6982.652 −6879.77 −6889.303 Q2 (20) 399.13* 353.11* 368.90* 21.371 28.79*** 24.18 ARCH (10) 164.15* 148.70* 153.64* 11.391 20.69** 15.37 Jarque and Bera 8807.26* 9342.78* 9181.98* 736.55* 913.20* 816.71* The persistence is calculated as (α1+β1) for the GARCH model. Log L is the value of the logarithmic likelihood. Q2 (20) is the Q-Statistic of correlogram of squared residuals at twenty lag. ARCH (10) represents the F-statistic of ARCH test at 10th lag. *,**,***indicate statistical significance at 1%, 5% and 10% respectively Dritsaki: The Performance of Hybrid ARIMA-GARCH Modeling and Forecasting Oil Price International Journal of Energy Economics and Policy | Vol 8 • Issue 3 • 2018 21 REFERENCES Agnolucci, P. (2009), Volatility in crude oil futures: Α comparison of the predictive ability of GARCH and implied volatility models. Energy Economics, 31(2), 316-321. Bollerslev, T. (1986), Generalized autoregressive conditional heteroskedasticity. Journal of Econometrics, 31(3), 307-327. Box, G.E.P., Cox, D.R. (1964), An analysis of transformations. Journal of the Royal Statistical Society, Series B, 26(2), 211-252. Box, G.E.P., Jenkins, G.M. (1976), Time Series Analysis. Forecasting and Control. San Francisco: Holden-Day. Dickey, D.A., Fuller, W.A. (1979), Distributions of the estimators for autoregressive time series with a unit root. Journal of American Statistical Association, 74(366), 427-431. Dickey, DA., Fuller, W.A. (1981), Likelihood ratio statistics for autoregressive time series with a unit root. Econometrica, 49(4), 1057-1072. Dritsaki, C. (2017), An empirical evaluation in GARCH volatility modeling: Evidence from the Stockholm stock exchange. Journal of Mathematical Finance, 7, 366-390. Elder, J., Serletis, A. (2009), Oil price uncertainty. Energy Economics, 31(6), 852-856. Hamilton, J.D., Herrera, A.M. (2004), Oil shocks and aggregate macroeconomic behavior: The role of monetary policy: A comment. Journal of Money, Credit and Banking, 36(2), 265-286. Hansen, P.R., Lunde, A.S. (2005), A forecast comparison of volatility models: Does anything beat a GARCH(1, 1)? Journal of Applied Econometrics, 20(7), 873-889. Jarque, C., Bera, A. (1980), Efficient tests for normality, homoscedasticity and serial independence of regression residuals. Economics Letters, 6, 255-259. Ljung, G.M., Box, G.E.P. (1978), On a measure of a lack of fit in time series models. Biometrika, 65(2), 297-303. MacKinnon J.G. (1996), Numerical distribution functions for unit root and cointegration tests. Journal of Applied Econometrics, 11(6), 601-618. Marquardt, D.W. (1963), An algorithm for least squares estimation of nonlinear parameters. Journal of the Society for Industrial and Applied Mathematics, 11, 431-441. Mirmirani, S., Li, H.C. (2004), A comparison of VAR and neural networks with genetic algorithm in forecasting price of oil. Advances in Econometrics, 19, 203-223. Mohammadi, H. (2009), Electricity prices and fuel costs: Long-run relations and shortrun dynamics. Energy Economics, 31(3), 503-509. Newey, W.K., West, K.D. (1994), Automatic lag selection in covariance matrix estimation. Review of Economic Studies, 61(4), 631-654. Phillips, P.C., Perron, P. (1998), Testing for a unit root in time series regression. Biometrika, 75(2), 335-346. Rothemberg, J.J., Woodford, M. (1996), Imperfect competition and the effects of energy price increases on economic activity. Journal of Money, Credit and Banking, 28(4), 549-577. Sadorsky, P. (2006), Modeling and forecasting petroleum futures volatility. Energy Economics, 28(4), 467-488. Yang, C.W., Hwang, M.J., Huang, B.N. (2002), An analysis of factors affecting price volatility of the US oil market. Energy Economics, 24(2), 107-119. Yu, L., Wang, S., Lai, K.K. (2008), Forecasting crude oil price with an EMD-based neural network ensemble learning paradigm. Energy Economics, 30(5), 2623-2635.