Adaptive Support Vector Regression Tuning Composite Model of BWCG and NGARCH for Applications of Time-Series Prediction

Abstract

Grey model (GM) has encountered the crucial problem of overshoot when applying to the non-periodic short-term prediction. At the same period, cumulated 3-point least squared linear prediction (C3LSP) alternatively confronts the opposite situation, i.e. underestimation. Nevertheless, a method of combining both preceding models is proposed for resolving the overshoot and underestimation phenomena significantly that is hybrid BPNN-weighted GREY-C3LSP prediction (BWGC) model. However, some predicted outcomes resulted from BWGC are not accurate enough as few observations deviate far away from both GM and C3LSP outputs. Thus, compensation is figured out to deal with the time-varying variance of the residuals in BWGC. That is, incorporating a non-linear generalized autoregressive conditional heteroscedasticity (NGARCH) into BWGC is applied, and then adaptive support vector regression (ASVR) is employed for tuning the appropriate coefficients for both BWGC and NGARCH to effectively improve the predictive accuracy.