Elsevier

Energy

Volume 216, 1 February 2021, 119299
Energy

Energy futures price prediction and evaluation model with deep bidirectional gated recurrent unit neural network and RIF-based algorithm

https://doi.org/10.1016/j.energy.2020.119299Get rights and content

Highlights

  • A novel model (RIF-DBGRUNN) is proposed to forecast crude oil futures price.

  • Deep bidirectional learning and random inheritance formula are combined with model.

  • Basic metrics and q-DSCID synchronization evaluation is used to measure accuracy.

  • Results show the RIF-DBGRUNN model outperforms the comparison models.

Abstract

Energy resources have firmly occupied an unshakable position, which is indispensable both in industrial field and daily life. More accurate prediction of energy futures price has always been a challenging issue. Motivated by this problem, a novel random deep bidirectional gated recurrent unit neural network is constructed to achieve more accurate forecasts of international crude oil futures prices. The random inheritance formula is proposed and integrated into the training process of the model, and it reflects the timeliness of historical data. Both the random inheritance formula and the deep bidirectional learning can effectively improve the model’s acquisition of effective information from historical data and improve the model’s accuracy. The proposed model is compared with SVM, GRU, ERNN, LSTM, DBGRUNN and RIF-GRUNN models, and a variety of evaluation indicators as well as a novel synchronization evaluation method of q-DSCID are used to measure accuracy. The empirical research results of four crude oil futures prices and coarse-grained moving absolute returns show that the proposed model outperforms the comparison models. For the Brent crude oil futures price prediction, its metrics R2, MAE, TIC, RMSE and SMAPE are 0.998, 0.200, 0.002, 0.267 and 0.283, which are the best in the comparison models.

Introduction

Crude oil is an indispensable source of energy in industrial and agricultural production, and it is mainly engrained in fuels, chemical raw materials and other purposes [1]. Crude oil has an irreplaceable impact on the daily life of the people as extensive as the country’s economy, politics and military [2]. Fluctuations in crude oil prices can trigger a country’s economic recession and even an economic crisis, causing wars and conflicts between nations [3,4]. Therefore, how to achieve a more accurate forecast of crude oil futures prices has become an increasingly focused issue. Kristjanpoller and Minutolo [5] proposed a hybrid model ANN-GARCH to predict oil price return volatility, and the results indicate that the ANN improves forecasting accuracy over the GARCH and ARFIMA model prediction. Cheng et al. [6] introduced a new hybrid vector error correction and nonlinear autoregressive neural network (VEC-NAR) model to predict future crude oil prices and verified the accuracy of the model.

Crude oil futures prices have nonlinearities, uncertainties and volatility [7,8]. Consequently, it is challenging to achieve high-precision prediction [[9], [10], [11]]. E tl al [9]. proposed a forecasting model based on VMD, ICA and ARIMA to predict crude oil price. Jammazi and Aloui [10] construct a hybrid model HTW-MPNN to achieve prominent prediction of crude oil price. Price series prediction is a typical time series forecast problem [12,13]. Support vector machine (SVM) and support vector regression (SVR) are employed earlier in settling time series forecast problems [14,15]. They have favorable theoretical basis but are not suitable for large sample price prediction. Therefore, optimization problems based on these two mediums are also increasing. Guo et al. [16] used the support vector machine to predict the oil price. In order to reduce the empirical component selection of the traditional algorithm, the genetic algorithm optimization parameters are used to construct a new hybrid model. The results show that this optimizes the accuracy of the model. Kao et al. [17] applied SVR to predict the stock price series and proposed the NLICA-SVR model. Experimental results show that the proposed model has better prediction accuracy, which indicates that ICA has better optimization ability for SVR. The powerful self-learning function of artificial neural network (ANN) provides a solution for the prediction of large sample sequences. The results of using artificial neural network to predict price series are often superior to traditional theoretical methods [[18], [19], [20]]. Ekonomou [21] developed ANN model to predict Greek energy consumption. The produced ANN results for years 2005–2008 are compared with the results produced by a linear regression method, a support vector machine method and with real energy consumption records showing a great accuracy. Keles et al. [22] proposed a power price prediction model based on artificial neural network. The clustering algorithm is used to optimize the screening data. Compared with the seasonal ARIMA model, the ANN-model achieved better results. Artificial neural networks possess exceptional manifestation in price forecasting, but lack the temporal correlation between data. The recurrent neural network (RNN) solves this conundrum favourably, and realizes the association between data at different times through the recurrent layer, and ameliorates the network architecture [[23], [24], [25]]. Berradi and Lazaar [26] applied a recurrent neural network to predict the stock price of Casablanca Stock Exchange, and used principal component analysis to reduce the number of features and improve the prediction accuracy of RNN. Wang and Wang [27] constructed a new predictive model ST-ERNN to forecast the crude oil prices as well as oil stock prices. The predicted values of the proposed model on the crude oil (and oil stock) prices are more in agreement with the real values than other models. In-depth research has found that recurrent neural networks have long-term dependencies in training and vanish gradient problems, leading to a decline in model prediction exactness.

With the advent of the era of big data, machine learning algorithms have been better evolved. Gated recurrent unit (GRU) was proposed as a more sophisticated and reliable machine learning algorithm. The long-term dependence of RNN and the disappearance of gradients are solved magnificently [28]. GRU is a variant of Long short-term memory (LSTM) [29]. Compared with LSTM, GRU has a more compact configuration and higher efficiency. Studies have also shown that the effects of the two in practical applications are often not much different. Thanks to its superior performance, GRU has been widely used in various fields, and its hybrid optimization model has gradually increased. Wu et al. [30] applied GRU network to forecast short-term load considering impact of electricity price. The simulation study shows that this method can effectively improve the accuracy of short-term load forecasting compared with traditional methods. Chen et al. [31] proposed a general two-step method for residual life prediction, first applying KPCA to extract features and then applying GRU prediction. The results show that GRU is better than LSTM in training time and prediction accuracy, and can provide better prediction for nonlinear degradation process of complex systems.

Recently, many articles on energy price series forecasting have been published, providing more methods for energy price prediction. Hooman and Seyed [32] introduced a novel hybrid model employed ANFIS, ARFIMA, and Markov-switching to forecast Brent crude oil price and results show genetic algorithm weighted hybrid model outperforms the other models. Qiao and Yang [33] proposed WT-SAE-LSTM to forecast electricity prices of USA and results show that the proposed model outperforms other AI models, such as back propagation neural network, in forecasting accuracy. Cen and Wang [34] established a new prediction model with long short term memory based on prior knowledge data transfer to predict crude oil price and empirical results show that data transfer can greatly improve prediction accuracy of long short term memory. E et al. [35] proposed a novel hybridization of multi-scale model for predicting the energy price based on independent component analysis, gated recurrent unit neural network and support vector regression and experiments demonstrate the validity and reliability of the improved model. Most of them are based on improvements or mixtures of existing models and methods to achieve the goal of further improving the accuracy of model forecasts. Inspired by this, this paper introduces a novel energy price prediction model that integrates GRU, deep learning and bidirectional learning structure, and proposes a random inheritance formula to optimize the model training process.

In the process of applying neural network model to predict the price series of crude oil futures, it is necessary to provide sufficient historical data for model learning and training. Therefore, how to implement better data effective information extraction and more reasonable information application is a tough challenge. The extraction and utilization of input information directly affects the accuracy of the model prediction results. In order to counter this problem and improve the performance of the model, we construct a novel random deep bidirectional gated recurrent unit neural network (RIF-DBGRUNN) model. First, the deep learning method is applied to increase the depth and ameliorate the predictive property of the model. Then the bidirectional learning structure is established, which consists of two independent GRU units, input information is trained in both directions to obtain more effective information. This achieves a more comprehensive feature extraction of input information and improves the learning effect of the model. Some research found that the application value of the data points of the input sequence is different [36]. Future price fluctuations will inherit the characteristics of some historical data, more or less, depending on the time interval between the predicted sample points and the past sample points [37,38]. Based on this characteristic, a random inheritance formula (RIF) is proposed based on geometric Brownian motion [39,40]. In the process of model learning, the gradient correction is carried out for each learning, and the random inheritance formula is applied to the gradient correction. The gradient correction method based on RIF can correct the parameter matrix more reasonably and optimize the prediction result.

Aapply the RIF-DBGRUNN to predict four sets of energy futures price series, compare the prediction results with six generally utilized models, and perform linear regression analysis, basic accuracy measurement and q-order dyadic scales complexity invariant distance (q-DSCID) evaluation on the prediction results [41,42]. Comparing the model on a GRU basis and calculating the increase percentage (IP) in metrics. The proposed model is also applied to predict coarse-grained moving absolute return (CMAR) series. All test results demonstrate the excellent accuracy and stability of the proposed model.

The main contribution of this article can be summarized as four points. First, a novel RIF-DBGRUNN model is established to achieve more accurate energy futures prices prediction. Second, the random inheritance formula is introduced and integrated into the training process of GRU model to optimize the model training process. Third, deep bidirectional learning is applied to model construction in order to capture more effective historical information. Fourth, q-DSCID, a more reliable synchronization evaluation method, is used in the comparison and evaluation of model accuracy. The remaining part of this paper includes the following sections. Section 2 introduces the adopted method and the integrated framework of the proposed model. Section 3 describes the basic and complex metrics. Section 4 shows the experiment and comparative analysis results for crude oil futures price and coarse-grained moving absolute return series. Section 5 revels conclusions and expectations.

Section snippets

Gated recurrent unit

GRU model is an evolution of traditional RNN model. It solves the problem of long-term dependence in traditional RNN network, and has the outstanding characteristics of fast training and high precision [43]. Its structure unit is shown in Fig. 1. GRU can be regarded as a variant of LSTM. Based on LSTM network, GRU improves the activation function of hidden layer nodes and uses the activation function controlled by gate valve to calculate the output of hidden layer, so that each threshold

Basic accuracy metrics

To further study the predictive abilities of RIF-DBGRUNN, some regression model evaluation indicators are utilized in this work. Let otdenotes the true value, ytis the predicted value and o¯represents the mean of samples. The selection of indicators and their formulas are listed in Table 1 [[46], [47], [48], [49]]. Through these objective evaluation indicators, the comprehensive performance of the model is evaluated and the predictive ability of each model is compared.

Complex accuracy metrics

In this section, a

Crude oil futures price selection and processing

A total of four data sets were selected: the crude oil futures data of West Texas Intermediate crude oil (WTI), ICE Brent crude oil (BRE), Heating Oil (HO), Henry Hub Natural Gas (NG). Many investors regard the WTI spot contracts as the benchmark price for measuring crude oil price changes in the international energy market. Brent crude oil futures is the most widely used contract in global crude oil pricing. At present, about seventy percent of spot crude oil trading pricing refers to Brent

Conclusion

In order to predict the energy market futures price more accurately, a novel deep bidirectional GRU neural network model RIF-DBGRUNN based on the random inheritance formula is proposed. It integrates the method of deep learning and further ameliorates the prediction accuracy of the model through bidirectional learning structure and RIF-based training algorithm. The verification results signify that the proposed model stands out among comparison models with high prediction accuracy.

Applying the

Credit author statement

Bin Wang: Conceptualization; Methodology; Investigation; Software; Visualization; Writing – original draft and review & editing. Jun Wang: Conceptualization; Methodology; Investigation; Supervision; Writing – review & editing.

Declaration of competing interest

The authors declare that they have no known competing financial interests or personal relationships that could have appeared to influence the work reported in this paper.

Acknowledgment

The authors were supported by National Natural Science Foundation of China Grant No. 71271026.

References (50)

  • F.Z. Cheng et al.

    The vec-nar model for short-term forecasting of oil prices

    Energy Econ

    (2019)
  • Y. Yu et al.

    Lattice-oriented percolation system applied to volatility behavior of stock market

    J Appl Stat

    (2012)
  • E. Jw et al.

    Crude oil price analysis and forecasting based on variational mode decomposition and independent component analysis

    Phys Stat Mech Appl

    (2017)
  • R. Jammazi et al.

    Crude oil price forecasting: experimental evidence from wavelet decomposition and neural network modeling

    Energy Econ

    (2012)
  • R.H. Abiyev

    Fuzzy wavelet neural network based on fuzzy clustering and gradient techniques for time series prediction

    Neural Comput Appl

    (2011)
  • Y. Solgi et al.

    Variable structure fuzzy wavelet neural network controller for complex nonlinear systems

    Appl Soft Comput

    (2018)
  • S. Fu et al.

    Evolutionary support vector machine for RMB exchange rate forecasting

    Phys Stat Mech Appl

    (2019)
  • Y. Li et al.

    Subsampled support vector regression ensemble for short term electric load forecasting

    Energy

    (2018)
  • X.P. Guo et al.

    Improved support vector machine oil price forecast model based on genetic algorithm optimization parameters

    AASRI Procedia

    (2012)
  • L.J. Kao et al.

    Integration of nonlinear independent component analysis and support vector regression for stock price forecasting

    Neurocomputing

    (2013)
  • Y.Y. Hong et al.

    Short-term LMP forecasting using an artificial neural network incorporating empirical mode decomposition

    International Transactions on Electrical Energy Systems

    (2015)
  • B. Kordanuli et al.

    Appraisal of artificial neural network for forecasting of economic parameters

    Phys Stat Mech Appl

    (2017)
  • L. Ekonomou

    Greek long-term energy consumption prediction using artificial neural networks

    Energy

    (2010)
  • D. Keles et al.

    Extended forecast methods for day-ahead electricity spot prices applying artificial neural networks

    Appl Energy

    (2016)
  • A. Rahman et al.

    Predicting electricity consumption for commercial and residential buildings using deep recurrent neural networks

    Energy

    (2018)
  • Cited by (27)

    View all citing articles on Scopus
    View full text