Summary: This short article presents a new algorithm of time series prediction: {\ssf PerKE}. It implements the kernel regression for the time series directly without any data transformation. This method is based on a new type of kernel functions ‒ periodic kernel functions ‒ two examples of which are also introduced in this paper. This new algorithm belongs to the group of semiparametric methods as it needs the initial step that separates the trend from the original time series.