Employing local modeling in machine learning based methods for time-series prediction

• Propose a local modeling strategy for time series prediction.
• Consider the trend of a time series by the use of hybrid distance.
• Proper lags are selected by the use of mutual information.
• Develop algorithms to extract training patterns from historical data.
• Show the effectiveness of local modeling by experiments on real-world datasets.

Time series prediction has been widely used in a variety of applications in science, engineering, finance, etc. There are two different modeling options for constructing forecasting models in time series prediction. Global modeling constructs a model which is independent from user queries. On the contrary, local modeling constructs a local model for each different query from the user. In this paper, we propose a local modeling strategy and investigate the effectiveness of incorporating local modeling with three popular machine learning based forecasting methods, Neural Network (NN), Adaptive Neuro-Fuzzy Inference System (ANFIS), and Least Squares Support Vector Machine (LS-SVM), for time series prediction. Given a series of historical data, a local context of the user query is located and an appropriate number of lags are selected. Then forecasting models are constructed by applying NN, ANFIS, and LS-SVM, respectively. A number of experiments are conducted and the results show that local modeling can enhance the estimation performance of a forecasting method for time series prediction.