A novel hybrid approach for time series data forecasting using moving average filter and ARIMA-SVM

Abstract

Time series data consists of a variety of information in its patterns. It is composed of both linear and nonlinear parts. Depending on nature of time series data, either linear model or nonlinear model can be applied. Instead of applying linear time series model like Auto Regressive Integrated Moving Average (ARIMA) and nonlinear time series model like Support Vector Machine (SVM) and Artificial Neural Network (ANN) individually on time series data, the proposed hybrid model decomposes time series data into two parts using Moving Average Filter and applies ARIMA on the linear part of time series data and SVM on nonlinear part of time series data. The performance of the proposed hybrid model is compared using Mean Absolute Error (MAE) and Mean Squared Error (MSE) with the performances obtained by the conventional models like ARIMA, SVM, and ANN individually. The proposed model has shown efficient prediction results when compared with the results given by the conventional models of time series data having trended patterns.