A denoising carbon price forecasting method based on the integration of kernel independent component analysis and least squares support vector regression

• Published in: Neurocomputing (Amsterdam) Vol. 434; pp. 67 - 79
• Main Authors: E, Jianwei, Ye, Jimin, He, Lulu, Jin, Haihong
• Format: Journal Article
• Language: English
• Published: Elsevier B.V 28.04.2021
• Subjects: Carbon price; Extreme-point symmetric mode decomposition; Intrinsic mode function; Kernel independent component analysis; Least squares support vector regression; Carbon price; 00–01; 99–00; Kernel independent component analysis; Intrinsic mode function; Extreme-point symmetric mode decomposition; Least squares support vector regression
• ISSN: 0925-2312; 1872-8286
• DOI: 10.1016/j.neucom.2020.12.086

•The potential independent components (ICs) are separated out via KICA.•KICA transforms the problem of joint estimation into univariate estimations of ICs.•LSSVR method is utilized to generate the forecasting results of carbon prices.•Experiments indicate the effectiveness and robustness of the proposed model. During the past few decades, accurately forecasting carbon price has become a significant research field and aroused concerns from both scholars and policymakers, which contributes the Organized Exchange to scientifically and rationally allocate a fixed-quantity of carbon emissions among prospective polluters. The conventional forecasting approaches, however, suffer from the poor prediction accuracy due to the nonlinearity and non-stationarity of the carbon price series. Meanwhile, monitoring and filtering the inherent noise in carbon price series, which are the main steps in the forecasting model, are perceived as the challenging tasks to work in. To address these obstacles, a denoising-hybridization procedure, which is a hybrid model of extreme-point symmetric mode decomposition (ESMD), kernel independent component analysis (KICA) and least squares support vector regression (LSSVR), is put forward for predicting the carbon price. Firstly, the carbon price is decomposed into several intrinsic mode functions (IMFs) via the ESMD method. Secondly, independent components (ICs), which reflect the internal formation mechanism, are separated out from the IMFs via KICA method. Further, the IC comprised of the noise is eliminated according to the results of noise monitoring. Finally, the LSSVR method is applied to the remaining ICs for achieving the forecasting results of carbon price, wherein the particle swarm optimization (PSO) algorithm is employed to synchronously optimize the hyper parameters in LSSVR. The empirical results on four carbon futures prices from European Union Emissions Trade System (EU ETS) demonstrate the effectiveness and robustness of the promoted denoising-hybridization procedure. Comparative experiments illustrate the superiority of the proposed method from the perspective of statistical performance criteria.