Forecasting Gasoline Demand in Indonesia Using Time Series

Fuel is an essential commodity in both the economy and society. Indonesian fuel demand continues to increase annually, whereas fuel production has decreased. Gasoline accounts for more than 50% of fuel consumption for transportation. A reliable gasoline product demand forecast is required to plan the gasoline supply. The objective of this study is to forecast the demand for total gasoline and its three components, which are gasoline 88, gasoline 90, and gasoline 92. This study compared the Holt–Winters additive model and autoregressive integrated moving average for the time-series data for the 2017-2019 period. Because the Holt–Winters additive model generates more accurate results, it was applied to predict the total demand for gasoline during 2020-2022. The results of the combination of the Holt–Winters model and a neural network to forecast gasoline 92 demand had lower errors than the individual Holt–Winters method. The forecast results show that total gasoline demand is forecasted to increase, but the components indicate a different trend. Gasoline 92 and gasoline 88 decreased, but gasoline 90 increased.


INTRODUCTION
The petroleum industry plays an essential role in the world economy, and disruptions in its supply chain have significant impacts on the economy and society (Lima et al., 2016). Over the years, oil consumption in Indonesia has increased significantly, but oil production had been decreasing. As a result, Indonesia has become an importer of crude oil and refined oil products (Sa'ad, 2009). Among the factors that influence the escalation in petroleum demand, population and economic growth receive significant attention (Zhang et al., 2009). Indonesia's population was 211 million in 2000, which increased to 267 million in 2018. Moreover, the country's economic growth over the years has trended upward. Indonesia's GDP was US$165 billion in 2000, increasing to US$1042 trillion in 2018 (World Bank, 2018). These statistics indicate a tendency for a widening gap between oil demand and domestic supply capabilities that requires attention.
Forecasting gasoline demand, which is part of the energy, plays an essential role in gasoline supply planning (Zhao and Chen, 2014). Accurate forecasting assists decision-makers in understanding the volume of and trends in gasoline demand for supply system planning (Ghalehkhondabi et al., 2017). Demand forecasting errors result imbalances in supply and demand, which negatively affects operating costs, network security, and service quality (De Felice et al., 2013).
Transportation as a backbone of the economy is strongly dependent on petroleum, which is why the sector received significant attention (He et al., 2005;Chai et al., 2016). The transportation sector records the highest petroleum consumption in Indonesia. In 2018, the transportation sector accounted for 45.06% of energy consumption, industry 33.51%, households 14.765% and commercial establishments 4.82%. Fuel accounted for 37.78% of energy consumption, followed by electricity 18.07%, biofuel 13.11%, coal 11.58%, natural gas 11.01%, and other 7.44% (Ministry of Energy and Mineral Resource Republic of Indonesia, 2018).
Issues in the transportation system, especially fuel demand have become concerns of the government (Zhang et al., 2009). Given this importance, we forecast the fuel demand for transportation or gasoline and its components for 2020-2022 using a case study on the Indonesian state oil company. Gasoline is a fuel used in road transportation-primarily vehicles. The components of gasoline to be studied are gasoline 88, gasoline 90 and gasoline 92. They gasoline types have certain RON (Research Octane Number) levels. In this study, gasoline 88, gasoline 90, and gasoline 92 refer to gasoline with RON 88, 90, and 92, respectively. Higher RON numbers indicate higher quality. Gasoline with RON 88 and RON 90 is medium grade, whereas gasoline with RON 92 or higher is premium grade (US Energy Information Administration, 2019). These three types of gasoline are consumed the most by vehicles in Indonesia. Figure 1 provides 2018 fuel demand by type and indicates that these three gasoline components account for 53% of the total fuel provided for transportation, including air and railway transportation. The other types of gasoline, namely gasoline 95, gasoline 98, and gasoline 100, account for only 1% of all fuel used.
Few studies analyzed Indonesia's petroleum demand. Dahl and Kurtubi (2001) analyzed petroleum product demand and price elasticity, Sa'ad (2009) studied total gasoline and diesel demand and price elasticity, and Akhmad and Amir (2018) analyzed the supply and demand of kerosene, solar, and total gasoline. However, no study has forecasted component gasoline demand or has identified an appropriate model for the demand of gasoline products of the Indonesian state oil company. Hence, this study is the first on component gasoline forecasts in Indonesia.
Companies use product demand forecasts for tactical planning such as production planning, and for inventory strategic planning to determine the necessity to build a new plant (Chopra and Meindl, 2016). Therefore, the component demand forecast for gasoline is very important in determining inventory levels for each product, production, and imports. This paper is organized as follows: Section 2 presents an overview of Indonesia's gasoline supply, demand, and policies. Section 3 presents related studies. Section 4 discusses the contribution made by this study, and Section 5 presents a theoretical description of the method for energy forecasting models such as Holt-Winters' additive, autoregressive integrated moving average (ARIMA), linear regression, and neural networks. Section 6 provides the data description and Section 7 presents the analysis. Section 8 presents the forecasting results and a discussion. The findings and subsequent studies are provided in Section 9 as the conclusion.

OVERVIEW OF INDONESIA'S GASOLINE SUPPLY, DEMAND, AND POLICY
The transportation sector is the primary consumer of gasoline. From 2000 to 2018, the average gasoline consumption increased by 5.6% per year, with a maximum increase of 10.83% and a minimum of 1.35%. However, Indonesia's gasoline production does not show a significant increase. The average annual production was 76,905 MB with a maximum of 91,640 MB and a minimum of 67,642 MB. Since 2000, Indonesia's gasoline consumption has exceeded its domestic supply, resulting in an imbalance between supply and demand. As Figure 2 indicates, the difference between supply and demand widened. The shortage caused Indonesian gasoline imports to increase over the years.
Initially, retail fuel sales may only be carried out by national oil companies owned by the government. In 2000, Indonesia liberalized its oil and gas businesses by passing an oil and gas law that allowed foreign companies to sell gasoline in its country.
Gasoline 88 is a product under government supervision, and companies are required to sell it without exceeding the specified sales quota determined by the government. Every price change for all gasoline products must be approved by the government. Specifically for gasoline 88, the government issued a "one price" policy that mandated that this product be sold at the same price throughout Indonesia. The government bears the costs of transporting these products. Although the policy reduces the burden on gasoline consumers in remote areas who initially had to buy gasoline at higher prices than other regions, it certainly had the consequence of increasing the government's financial burden.
Although energy demand prediction is an important issue in all countries, limited gasoline demand forecasting studies exist on Indonesia. Sa'ad (2009) analyzed the demand for petroleum used for transportation with econometric techniques that forecast per capita petroleum consumption. The study concluded that petroleum demand would increase. Akhmad and Amir (2018) used an econometric method with simultaneous equation to predict the supply and consumption of fuel and the factors influencing supply and demand. These authors concluded that the demand, price, and import of fuel oil would increase, and the influencing factors of fuel consumption are fuel price and the previous year's fuel consumption. To the best of our knowledge, no research exists on gasoline component demand forecasting in Indonesia.
Researchers have paid significant attention to studies on the demand for fuel transportation, which commonly used the time series analysis approach (Houri and Baratimalayeri, 2008). A variety of energy demand forecasting methods have been applied over the years. Studies on the forecasting method revealed that no one best method exists for all conditions (Ghalehkhondabi et al., 2017). The most common models in the energy-related demand forecasting area are time series models, regression-based formulations, and artificial neural network (ANN) (Kuster et al., 2017;Wang et al., 2018). Time series forecasting is a forecasting method in which past data for the variable are analyzed to generate a model (Zhang, 2003;Akpinar and Yumusak, 2016). The method is widely used to predict energy needs. A linear regression model also had been used in gasoline consumption forecasting (Sapnken et al., 2018) as well as ANN (Lai et al., 2014).
ARIMA methods are widely used in time series data analysis (Suganthi and Samuel, 2012). However, Holt-Winters, an extension of exponential smoothing, is also broadly used to predict energy consumption and remains a reliable approach (Kays et al., 2018). Taylor (2003) used the Holt-Winters method to forecast electricity demand. Jónsson et al. (2014) predicted a real-time electricity market using the Holt-Winters method. Jiang et al. (2020) applied the enhanced Holt-Winters exponential smoothing to predict electricity consumption in China and concluded that this method generated accurate results with few sample data points. In contrast, Ediger and Akar (2007) used ARIMA and seasonal ARIMA (SARIMA) to forecast energy demand in Turkey. Another energy consumption forecasting method was conducted by Ozturk and Ozturk (2018) using the ARIMA model. Many authors used a combination of ARIMA and other models to forecast energy demand, such as ARIMA and ANFIS (Barak and Sadegh, 2016), ARIMA and ANN (Babazadeh, 2017), ARIMA and MGM (Wang et al., 2018), and ARIMA and NMGM . They concluded that a combination model provides better results than a single model.
In forecasting, using a shorter period of data provides a more accurate result than a longer period of data (As'ad, 2012). Although ARIMA is widely used for demand forecasting, it does not always generate more accurate results than simpler methods such as linear regression, a logistic model (Melikoglu, 2013), or a quadratic regression (Li et al., 2010). Chai et al. (2016) revealed that the exponential smoothing and the ARIMA prediction results were very close but that the exponential smoothing result was more accurate than the ARIMA result. Akpinar and Yumusak (2016) concluded that ARIMA error rates decrease as the computation complexity of the method increases.
Many authors compared the Holt-Winters method and ARIMA. Li et al. (2010) compared various time series methods to forecast petrol demand in Australia and concluded that the quadratic and linear regression methods outperformed other methods, including Holt-Winters and ARIMA, and that ARIMA provided a better result than Holt-Winters. Taylor (2003) applied the Holt-Winters method to forecast electricity demand and concluded that the method outperforms ARIMA (Hussain et al., 2016). Oliveira and Oliveira (2018) conducted a study of electricity consumption in several developed and developing countries by comparing ARIMA with exponential smoothing. ARIMA showed better results for developing countries cases, such as Brazil and Mexico, whereas exponential smoothing performed well for Canada, France, and Italy.
The neural network or ANN is another forecasting method that is based on the machine learning approach and can accommodate nonlinearity and linearity models. This methodology predicts the causal effect of variables (Chattopadhyay et al., 2019). Many forecasting studies used the neural network (NN) method because it generates accurate prediction results (Ryu et al., 2017). Although the NN has many advantages relative to the multiple regression method, it has limitations. Some of these limitations are that the model parameters cannot be identified, meaning that the functional relationship between variables is unrevealed (Detienne et al., 2003). Moreover, to obtain the minimum error, trial and error must be exercised many times (Ayyoub and Riaz, 2017).
Although accurately forecasting demand is not possible, several studies have been successful. Therefore, researchers always make their best attempts to minimize forecast errors (Hussain et al., 2016). Table 1 presents a summary of the energy forecasting studies using time series and their accuracy measurement results.

CONTRIBUTION OF THIS STUDY
This study fills the gap in the research on component gasoline demand forecasting in Indonesia. Moreover, most petroleum or gasoline demand forecasting in a time series analysis used previous demand or consumption as a time series variable. The contribution of this study is that forecasting gasoline is performed using a ratio variable. The following four methods are studied: Holt-Winters' additive, ARIMA, linear regression, and NN. The NN method is applied to model the correlation between the ratio of gasoline 92 to total demand and the price difference between gasoline 92 and gasoline 90. Demand forecasts for gasoline 90 and gasoline 88 are carried out using the variable for the demand ratio of gasoline 88 to gasoline 90.

METHODS
Gasoline demand forecasting was carried out using a top-down approach through which aggregated gasoline was first predicted. The components of gasoline demand were then calculated using their predicted demand ratio. This approach was applied because aggregate demand forecasting was more reliable than the summation of the individual component predictions (Ediger and Akar, 2007).
In this study, we used the Holt-Winters additive, ARIMA, linear regression, and NN methods. Holt-Winters additive and ARIMA are the time series methods that are widely used in demand forecasting because they can capture trends and seasonality. These two popular tools are used by researchers and practitioners to forecast studies (Xu et al., 2018). NN can predict nonlinear correlation among variables (Sharma and Chopra, 2013).
The Holt-Winters additive and ARIMA model are compared, and the best one is selected to apply the appropriate model to forecast gasoline demand. We analyzed the linear regression and NN to select the most accurate model to be applied to model the correlation between the price difference and the demand ratio. A combination of the Holt-Winters' additive prediction result and the forecasting result based on the correlation between the price difference and the demand ratio model is used to predict the component gasoline demand.
This study used IBM SPSS version 26 software to apply the Holt-Winters additive, ARIMA, linear regression, and NN model.

Holt-Winters Method
Holt-Winters is an exponential smoothing development that includes trends and seasonal data. Exponential smoothing assigns a different weight to each observation. Previous period data are given higher weights than older period data (Xu et al., 2018).
The Holt-Winters method uses three equations for level, trend, and seasonal. The seasonal component can also be treated additively in the formulation. The formulation includes α, β, and γ as smoothing parameters. The Holt-Winters additive method is as presented in equations (1), (2), (3), and (4) (Chase, 2013): where L t = the level of the series s = length of seasonality (e.g., number of the month in a year) b t = trend S t = seasonal component F = forecast for m periods ahead Y t = actual demand in period t α = constant between 0 and 1 β = constant between 0 and 1 γ = constant between 0 and 1.

Linear Regression
The linear regression is commonly used method in fuel demand studies (Li et al., 2010). This study compared the linear regression and NN methods to examine the correlation between the price difference and the demand for gasoline 92.
The regression analysis method estimates the parameter of the relationship between two or more variables. Typically, the modeler seeks to discover the cause and effect of one variable on another (Chase, 2013), such as the effect of the price difference between two products on sales.
The linear relationship between variables Y and X is given in equation (5): where Y = dependent variable c = intercept b = slope of a line X = independent variable

Autoregressive Integrated Moving Average (ARIMA)
ARIMA models were developed by Box and Jenkins (Gujarati and Porter, 2014). The requisite for applying ARIMA is that the data must be stationary. Stationary means that the data have constant mean and constant variance, and it can be determined using a graph. If no trend exists, then the time series is stationary. The way to remove the nonstationary data is by differencing, which is done by applying the differences among the observations (Li et al., 2010).
ARIMA is a combination of (1) auto regressive (AR), (2) integrated average (IA), and (3) moving average (MA). The IA is used to make the series stationary. In ARIMA (p, d, q), p expresses the number of autoregressive terms, q is the number of lagged forecast errors and d is the number of nonseasonal differences. The autocorrelation function (ACF) and the partial autocorrelation function (PACF) should be analyzed to determine the order of p and q (Brown and Rozeff, 1979;Fan and Yao, 2003;Gottman, 1981;Hussain et al., 2016).
Three steps are involved in applying ARIMA: (1) model identification, (2) parameter estimation, and (3) model diagnostics and forecasting (Ediger and Akar, 2007;Asuamah and Ohene, 2015;Barak and Sadegh, 2016). If seasonality is contained in the ARIMA model, SARIMA is used. SARIMA is represented as ARIMA (p, d, q) (P, D, Q), where P is the number of the seasonal autoregressive (SAR), D is the number of the seasonal differences and Q is the number of the seasonal moving average (SMA) (Debnath and Mourshed, 2018;Oliveira and Oliveira, 2018).
Bayesian Information Criterion (BIC) is a criterion for ARIMA models selection. The model to be selected is the one with the lower BIC value (Clement, 2014). The formula used to compute BIC is: where k = the number of free parameters to be estimated n = the number of observations σ e 2 = error variance Under the normality assumption, the following formula may be more tractable (Clement, 2014): The general form of ARIMA forecasting is: (Rachev et al., 2007).

Neural Network
The NN is a prediction method that resembles the work of the human brain in processing information. NN form a specific structure consisting of several process units called neurons. This structure helps neurons solve problems by communicating with one another. Neurons are the fundamental operational unit of a NN. Each neuron performs the following tasks: receive signals from other neurons, signals are multiplied by a certain weight, the results of the multiplication of neurons with each weight are added up, the sum is transferred by the transfer function, and the number of transformations is sent to other neurons.
The neuron typology consists of three layers. The first layer is the input layer, the last layer is the output layer and a hidden layer is between the input and output layers. The input layer contains predictive variables. A network of neurons must have at least one independent variable as a factor. The data provided are called the values of the input variable.
The hidden layer consists of nodes that function as "black boxesˮ of NN. The value of each node is the result of the activation function, which is the sum of the input weights and biases. The output layer is the target variable. A minimum of one dependent variable exists as a target variable with nominal, ordinal or scale categories. (Ayyoub and Riaz, 2017). The neuron is a real function of input vector (x j ,…, x k ). The output is obtained as 1 ) ( where x j represents the jth input to the kth neuron, w kj is the weight of the neuron k and its jth input, y k is the neuron output, and b k is a bias constant (Haykin, 2005). The activation function φ is usually sigmoid (Sharma and Chopra, 2013).
A graphical presentation of a neuron is provided in Figure 3.
The sum square error (Szoplik, 2015), which measures the network error is calculated using the following formula: where d i = real value z i = value calculated by NN

Forecast Evaluation
The root mean square error (RMSE), mean absolute percentage error (MAPE) and mean absolute error (MAE) are commonly used to measure the accuracy for goodness of fit. The formula to compute RMSE, MAPE, and MAE are as follows (Hussain et al., 2016): where Y t = actual value in time period t F t = forecasted value in time period t N = total number of observations where Y t = actual value in time period t F t = forecasted value in time period t N = total number of observations All the parameter descriptions are the same as for MAPE. MAE and MAPE are similar, MAE is the absolute error and MAPE is the error in percentage.
The scale of judgement based on MAPE criteria developed by Lewis (Melikoglu, 2013) given in Table 2.   The R square value is used to measure the goodness of fit of a model. R square is the squared correlation between the forecast variable Y and the estimated value Ŷ . The formula is: (Chase, 2013).
R square values are between 0 and 1, and an R square of 1 indicates a perfect fit. When using time series data, an R square higher than 0.75 indicates a fairly good model fit (Chase, 2013).
P-values describe the exact significance level associated with an explanatory variable. If the P-value is 0.05 or less at a 95% confidence level, the explanatory variable is significant in predicting variable Y (Chase, 2013). The mean of total gasoline demand was 17,477 MB with a standard deviation of 1059, a minimum of 14,842 MB, and a maximum of 19,487 MB. The mean of gasoline 92 demand was 424 MB, the standard deviation was 768, the minimum was 1180 MB, and the maximum was 3671 MB. The mean ratio of gasoline 92 demand to total gasoline is 13.86%. Figure 5 provides a plot of the ratio of gasoline 88 and gasoline 90 demand and indicates a downward trend in this ratio. At the beginning of 2017, gasoline 88 demand was about 1 times that of gasoline 90 demand-approximately the same. However, at the end of 2019, gasoline 88 demand was approximately 0.5 times that of gasoline 90 demand. This ratio is modeled to forecast the demand for these two products.

ANALYSIS
This section consists of five stages. The first stage compares two forecasting methods-the Holt-Winters additive and ARIMA. The second stage is to construct a model of gasoline 92 demand on the basis of the relationship between gasoline 92 and gasoline 90 price differences with the ratio of gasoline 92 demand to total gasoline.
The third, fourth, and fifth stages forecast the total gasoline, gasoline 92, gasoline 90, and gasoline 88 demand, respectively.

Stage 2
In this stage, the correlation between the price difference and the demand ratio is calculated. The price difference is between gasoline 90 and gasoline 92, whereas the demand ratio is the ratio of gasoline 92 demand to total gasoline demand. The price of gasoline 90 and gasoline 92 increased periodically, and the price difference over time between the two products increases. Figure 6 indicates that if the price difference is IDR 1000, then the gasoline consumption is approximately 18% of total consumption. In contrast, if the price difference is IDR 2200, then the gasoline 92 consumption is 11%-14% of total consumption. Therefore, the larger the price difference between gasoline 90 and gasoline 92 the smaller the ratio of gasoline 92 demand to total gasoline demand. In other words, the larger the price difference between the two products, the more consumers will switch to products at lower prices.
The correlation between the price difference of the two products and the ratio is examined using two models. The first model is a linear regression, and the second model is a NN. The model with the lower MAPE will be selected.
The process of modeling the relationship between the price differences and the ratio of gasoline consumption 92 to total gasoline is carried out as follows: 1. Data are processed to determine the linear regression equation.
The independent variable is price difference and the dependent variable is the ratio of gasoline 92 demand to total gasoline demand. 2. The dependent variable value is calculated using a linear regression formula with parameters generated by the software. 3. The data are processed to find a model for the relationship between price differences with the ratio of gasoline 92 consumption to total gasoline using the NN method. 4. The data processing for NN modeling is carried out by iteration to obtain the smallest error generated by the software. In each iteration the MAPE is calculated and compared one to another. The input layer variable is the price difference, and the output layer variable is the consumption ratio as the dependent variable. After several trials, the sigmoid activation function is chosen because it produced a lower error than the other activation function 5. The MAPE of the linear regression model and NN model is calculated and the results are compared. Table 5 provides a summary of the linear regression model, and Table 6 provides the result of the linear regression model.
Then, the formula of the linear regression model for the correlation of the differences in prices and consumption is as follows: where Y = percentage of gasoline 92 consumption to total gasoline consumption X = price difference between gasoline 92 and gasoline 90.    After several iterations of the NN, the model with the lowest error is obtained. The summary is as provided in Table 7, and the correlation between the price difference and the ratio of gasoline 92 demand to total gasoline demand is provided in Table 8.
The results of the linear regression model and the ANN model are compared. Table 9 shows that the predicted and the actual demand ratio of gasoline 92 using the ANN model is more accurate than the linear regression. The MAPE of the ANN model is 4.89% whereas the linear regression model is 8.48%. Therefore, the ANN model will be applied to forecast gasoline 92 demand using its ratio to total gasoline demand.

Stage 3
In this stage, total gasoline demand is forecasted using Holt-Winters additive as a suitable method. The result is shown in Figure 7.
As Figure 7 indicates, the trend in the demand for total gasoline, which comprises gasoline 88, gasoline 90, and gasoline 92, increased.

Stage 4
At this stage, gasoline 92 is forecasted on the basis of the combined weight calculation of 20% of the calculation using the ANN model in Table 8 plus 80% of the calculation using the Holt-Winters additive. The results of these analyses are presented in Figure 8.  Gasoline 92 demand is forecasted using the combined model. The forecast is done using certain assumptions of the price difference between gasoline 92 and gasoline 90, as listed in Table 10.

Stage 5
In this stage, gasoline 90 and gasoline 88 are forecasted. The steps to generate a forecast are as follows: 1. Forecast the ratio of gasoline 90 and gasoline 88 demand using the Holt-Winters additive method. 2. Compute the demand of gasoline 90 plus gasoline 88, and then calculate the gasoline 90 and gasoline 88 demand forecasts using the previous ratio.
Step 1 As Figure 9 indicates, the demand for gasoline 88 at the beginning of 2017 was 1.06 times that of gasoline 90.
During the observation period, the ratio of the demand for the two products decreases. At the end of 2019, gasoline 88 demand becomes 0.53 times that of gasoline 90 and continues to decline. The prediction is that, at the end of 2022, gasoline 88 demand is predicted to be 0.14 times that of gasoline 90.
At this stage the demand for gasoline 88 and gasoline 90 demand is calculated using the predicted demand ratio.

RESULT AND DISCUSSION
The Holt-Winters additive and the ARIMA models were analyzed to determine a suitable forecasting model. The results indicated that the Holt-Winters additive model is more accurate than the ARIMA model because it has lower MAPE and RMSE. We identified the appropriate correlation model between the price difference and the ratio of gasoline 92 to total demand and compared the linear regression model and the NN model. The results showed that the NN model has better accuracy than the linear model. These results further implied that the Holt-Winters additive is applied to forecast total demand and the ratio of gasoline 88 to gasoline 90. When forecasting gasoline 92 demand, the Holt-Winters additive model and the ratio of gasoline 92 to total gasoline demand are combined. Referring to Table 2, forecasting models of this study produce high accuracy that shows that the MAPE is <10%. The MAPE for total gasoline demand is 1.472, and is 3.45 for gasoline 92, and the ratio of gasoline 88 to gasoline 90 is 5.591. The results of the gasoline forecast and the components are presented in Figure 10 and are summarized in Table 11.
From 2017 to 2022, yearly total gasoline demand is predicted to increase by 13%, whereas gasoline 90 is predicted to increase by 88%. In contrast, gasoline 92 and gasoline 88 are predicted to decrease by 50% and 44%, respectively.   Demand for gasoline 88 has been declining over the years and is predicted to continue to decline. During 2016, gasoline 90 was launched as a substitute for gasoline 88. The expectation is that consumers will buy gasoline 90 instead of gasoline 88. However, because of a significant price difference between gasoline 90 and gasoline 92, many consumers who initially used gasoline 92 changed to gasoline 90.

CONCLUSION
As the population increased and oil production was depleted, Indonesia's domestic oil demand exceeded its production level. Therefore, Indonesia has been importing crude oil and refined oil product. Accurate supply planning predictions are required to balance supply and demand. This study is designed to facilitate the planning of Indonesian fuel oil to reduce the risk of domestic oil supply shortages.
This study concluded that the greater the price difference of two products with a closed quality, the lower the demand for products at a higher price. This study also revealed that forecasting by combining two models provides higher accuracy than by using one model. Moreover, 3 years of data generated a more accurate forecast than 5 years of data.
The time-series forecasting method generally uses the ARIMA model. This study applies the Holt-Winters additive method because it produces higher accuracy than ARIMA. A simple timeseries method was found to generate more accurate forecasting results than a sophisticated method such as ARIMA. Empirical studies supported this finding (e.g., Fildes and Makridakis, 1995;Fildes et al., 1998;Li et al., 2010;Hussain et al., 2016), which concluded that the performance of simple forecasting methods is almost the same as that of sophisticated statistics. The reason is that simple methods can extrapolate the patterns of a time series better than sophisticated methods (Li et al., 2010). The NN model can accommodate the correlation between two variables that are not entirely linear and produces a model with higher accuracy than a linear regression. The conclusion reached is that the NN model is more appropriate than the regression model for nonlinear relationships Because demand forecasting is part of logistic planning, further research is suggested to analyze the supply level. Forecasting results are not always accurate, and derivations from what was planned always exist in the oil supply chain. Therefore, supply analysis requires a simulation model to determine the impact of any deviation in the supply chain factors.

ACKNOWLEDGMENTS
This study was funded by Universitas Indonesia. The authors would like to thank PT. Pertamina (Persero)-Indonesian state oil company for information and valuable support. Also, the authors would like to thank Enago (www.enago.com) for the English language review.