A Novel Methodology for Forecasting Petrochemical Product Prices in East China Market by Applying ARIMAX Time Series and Machine Learning Models

Abstract

Forecasting petrochemical product prices is essential for economic decision making in the petrochemical industry. However, it is a challenging task to achieve accurate forecasts, given the price volatility in East China market, and the fact that the petrochemical product prices can be affected by various factors relevant in the industry. Therefore, we proposed a novel methodology which applied ARIMAX time series and machine learning models, combined with feature selection, for the price forecasting. This paper proposes a novel approach, which involves four steps of data gathering, factor identification, feature selection and model construction, to forecasting the weekly and monthly prices of 24 petrochemical products, ranging from the upstream to the downstream of the petrochemical industrial chain. Among the various relevant factors which might affect the product prices, the most significant ones were identified by applying feature selection. The product prices were modelled and predicted using ARIMAX time series model and various machine learning models, including random forest (RF), support vector machine (SVM), gradient boosted decision tree (GBDT), etc. The data were classified into training set and test set. The results were assessed by mean absolute percentage error (MAPE) - a measure of forecasting accuracy, and direction statistics (Dstat), which evaluates the forecasting performance in terms of a downward/an upward trend in prices. Taking the price forecast of LLDPE in East China market as an example, it was shown by applying feature selection that, among the various relevant factors considered in this paper, the ones affecting LLDPE price the most were brent price, PE futures price and Purchasing Managers’ Index (PMI); additionally, the historical values of LLDPE price were also found to contribute to accurate forecasts. For LLDPE weekly price forecasting, the minimum MAPE of 0.7% was obtained using RF method, with Dstat being 64.1%; and the highest Dstat of 84.2% was achieved by applying GBDT and Multi-Layer Perceptron (MLP) methods, with MAPE being 1.3% and 1.4%, respectively. For LLDPE monthly price forecasting, a MAPE value of 1.3% and a Dstat value of 90.0% were achieved with ARIMAX algorithm. In general, considering all 24 petrochemical products studied in this work, good weekly and monthly forecasts were obtained regarding accuracy and tendency, by applying ARIMAX and machine learning models. The contents in this paper provide the following benefits: first, a wide range of petrochemical products were studied, filling the gaps in the literature and enriching the database; second, the applications of feature selection with a number of machine learning models, as well as ARIMAX model, to price forecasts, were evaluated and the methodology is applicable to other related industries; last but not least, the price forecasts provide guidance for petrochemical production, achieving economical and sustainable industrial development.