Discrimination of time series based on kernel method

Classical methods in discrimination such as linear and quadratic do not have good efficiency in the case of nongaussian or nonlinear time series data. In nonparametric kernel discrimination in which the kernel estimators of likelihood functions are used instead of their real values has been shown to have good performance. The misclassification rate of kernel discrimination is usually less than linear and quadratic methods because of its flexibility. However, the kernel estimates are depend on the bandwidth. This paper is concerned with the selection of bandwidth parameter to achieve an optimal discrimination with minimum rate misclassification. The methods obtained bandwidth examined via a simulation study.