Zaman serilerinin tahminlenmesinde klasik ve makine öğrenmesi yaklaşımlarına yönelik karşılaştırmalı bir çalışma: Türkiye'nin ihracatı üzerine deneysel bir analiz

Exports has become one of the main economic indicator for countries. Accordingly, an accurate forecasting for exports is an important step for decision making and finding the most appropriate forecasting model constitutes the main subject of many studies. By taking the popularity and success of the machine learning (ML) methods on time series forecasting tasks into consideration, they are utilized also in this study to observe their predictive performances on Turkish exports. In this respect, Long Short Term Memory (LSTM), Support Vector Machines (SVM) and Random Forest (RF) are applied and the results are compared with the most commonly used classical time series models such as Autoregressive Integrated Moving Average (ARIMA) and Exponential Smoothing (ETS) models. The analysis is conducted on Turkish monthly exports data taken from Turkish Statistical Institute (TURKSTAT) within the time interval of January 1997 – September 2019 and the main steps of the analysis are anomaly detection and cleaning, data preprocessing, model development, hyperparameter tuning and model selection and model comparison. The main findings can be summarized as follows; the anomaly detection and cleaning process improves the forecasting ability of the models, ETS is the best forecasting model and SVM model is the most promising among the ML models and the most competitive with the leading one. Besides, ARIMA has the poorest generalization ability among the others.