Geleneksel zaman serisi yöntemleri ve makine öğrenmesi yöntemlerinin öngörü performans karşılaştırması

One of the main objectives of the time series analysis is forecasting, and for this purpose both Machine Learning methods and statistical methods have been proposed in the literature. In this study, we use and compare some of these approaches in time series modelling and forecasting. In addition to traditional forecasting methods for time series data set which are namely Naive Method, Seasonal Naive Method, ARIMA, SARIMA, Exponential Smoothing, TBATS, Bayesian Exponential Smoothing Models with Trend Modifications and STL Decomposition, the forecasts are also obtained by using seven different machine learning methods. These methods are Random Forest, Support Vector Regression, XGBoosting, Bayesian Neural Network, Recurrent Neural Network, Long Short Term Memory Neural Network and Feed Forward Neural Network. It is also known that time series generally contain both linear and nonlinear patterns. In order to deal with this mixture data structure, a hybrid methodology which combines linear and nonlinear components was proposed by Zhang. According to him, predicted values of a time series can be obtained by summing both linear and nonlinear components. In this study, hybrid models are constructed by using machine learning methods for nonlinear pattern and statistical methods for linear pattern. Therefore, the forecasts are also obtained using hybrid models. The data set selected proportionally from different time frequencies in M4 Competition is used in this study. After observing the results of studies, the performance and impact of all methods are discussed. At the end of this discussion, most of the best models are mainly selected from machine learning methods for forecasting in this study. It is also seen that the forecasting performance of the model depends on both time frequency and forecast horizon. Lastly, the study proves that the hybrid approach is not always the best forecasting model for time series.