Multi-WRNN model for pricing the crude oil futures market

• Published in: Expert systems with applications Vol. 182; p. 115229
• Main Authors: Hajiabotorabi, Zeinab, Samavati, Faramarz F., Maalek Ghaini, Farid Mohammad, Shahmoradi, Akbar
• Format: Journal Article
• Language: English
• Published: New York Elsevier Ltd 15.11.2021; Elsevier BV
• Subjects: B-spline multiresolution; Crude oil; Decomposition; Deep recurrent neural network (DRNN); Discrete Wavelet Transform; Discrete wavelet transform (DWT); Empirical analysis; Neural networks; Pricing; Pricing future market; Recurrent neural networks; Refineries; Time series; Time series forecasting; Volatility; Volatility predicting; Wavelet analysis; Wavelet transforms; Deep recurrent neural network (DRNN); Time series forecasting; Discrete wavelet transform (DWT); Volatility predicting; B-spline multiresolution; Pricing future market
• ISSN: 0957-4174; 1873-6793
• DOI: 10.1016/j.eswa.2021.115229

•Proposing a multi-factor wavelet-based recurrent neural network model for crude oil futures pricing.•Using multiresolution decomposition in the model to predict non-stationary time series.•Introducing a flexible and versatile model to include new key factors.•Analyzing several local wavelets for constructing the wavelet-based model for the pricing. In this paper, we introduce a multi-factor wavelet-based deep recurrent neural network (Multi-WRNN) model for more accurate pricing of the crude oil future market. This model is capable of including several key factors (e.g. stock-change and refinery capacity utilization rate) flexibly. The Multi-WRNN model enables us to classify the time series of the key factors into stationary and non-stationary. Also, the model provides a dynamical system for predicting the variation mean, volatility, and value of non-stationary key factors which have high influences on market in crisis time, and consequently, can be applied in pricing of the crude oil market. The model uses a decomposition method to efficiently predict the volatilities time series. We compare our results with empirical mode decomposition (EMD) and the discrete wavelet transforms (DWT). Then, according to locality of support property of DWTs, including B-spline reverse subdivision, we use them to decompose the original high resolution volatility time series into a lowresolution time series and several details in various resolution levels. Then all decomposed time series from the DWT multiresolution process are fed to a deep recursive neural network (DRNN) model. Moreover, we analyze and compare different DWTs, including B-splines with various degrees, and two of Daubechies wavelets. Since the locality of the operations is essential, we consider only compact supported wavelets. Also, we analyze the effect of all given key factors in the performance of our pricing model. Finally, the results of Multi-WRNN model is compared with the conventional models, such as stochastic two-factor, ARIMA, FFNN-GARCH, and GARCH which show that the Multi-WRNN model with all seven key factors as inputs of the network outperforms the other models.