<p>The prediction of time series is a fundamental problem in big data computing, image processing, computer vision, mathematical statistics, computing economy, and risk management. As a result of its volatility and complexity, time series prediction has been relatively slow in research for decades. Recently, machine learning methods have excelled in time series analysis, as well as advancing time series prediction, while time series prediction is still a difficult task due to the weak interpretability of machine learning and the complexity of time series. From the perspective of data fitting, this paper introduces a new fast and effective method for predicting time series. The method is mainly concerned with dividing the time series into short-, medium-, and long-term data based on the periodicity and trend of the series, which will allow the program to correctly predict the series by describing the three types of data. A cubic polynomial is used to approximate each local region of a time series, and the cubic polynomial coefficients are divided into three groups, which are used to achieve high-accuracy fitting of three types of data. Fitting accuracy of the new method is comparable to cubic polynomials. Extensive experiments on six real-world datasets. demonstrate that the proposed method has remarkable potential in time series prediction.</p>