Optimized ARIMA–ANN hybrid model for time series and sample applications with financial time series

Time series analysis has become an important research area in recent years. Analyzing the past values and presenting the patterns in the past values of a variable, enable to foresight the future values of the variable and to forecast. Numerous methods have been developed to be able to model the time series. Each methods presented has different assumptions, different characteristics, strong and weak sides. For the purpose of improving the presented methods for time series analysis, lately, hybrid models are developed in which more than one methods are used. Hybrid models are presented with the aim of developing models that explain the past values better and forecast the future values more accurate by using strengths of different methods. One of the successful hybrid model classes in the literature is hybrid models using ARIMA and Artificial Neural Networks together. ARIMA models are strong for linear time series modelling, whereas Artificial Neural Network models give successful results on nonlinear time series. Because time series encountered in real life show linear and nonlinear characteristics together, hybrid models mostly gives better results than ARIMA and ANN. This thesis study aims to contribute in the time series literature by presenting Optimized ARIMA – ANN Hybrid Model. Optimized ARIMA – ANN Hybrid Model assumes that the time series is sum of linear and nonlinear components. Modelling the linear component is done by ARIMA models, and modelling the nonlinear component is done by ANN models. Model applies optimization to decompose the time series into linear and nonlinear components. Model errors of the linear component after optimization are added to nonlinear component, and the nonlinear part remodeled with ANN. In this way, hybrid model aims to get better model values and more accurate future forecasts by minimizing the model errors. In the study, Optimized ARIMA – ANN Hybrid Model is applied to three different financial time series; totally thirty models are constructed for weekly closing values of Spot Gold Prices, BIST 30 Index and US Dollar/Turkish Lira cross and totally thirty weekly closing values are forecasted with one step forward forecasting method. Model and future forecast results are compared with the results of ARIMA and ANN models, and the findings are discussed. Obtained findings shows that Optimized ARIMA – ANN Hybrid Model, models the time series better than ARIMA and ANN models and is superior to both models in future forecasting.