Evaluation of the Machine Learning Techniques for Forecasting the Seasonal Time Series

Time series with seasonal variability is widely used to describe processes in various fields such as trade, analysis of financial markets, forecasting of passenger air transportation, description of climatic changes, and others. Modern machine learning methods can be effectively used to construct predictive models of such series. In this case, only historical data accumulated over several seasonal observations are used as input data. Knowledge of other parameters is not required. The article considers constructing a time series forecasting model with seasonal variability, describing a technological process. The general methodology of model building is described, as well as the requirements for the input data sets. A comparison of several methods is given, among which are both classical: SARIMA, Holt-Winters Exponential Smoothing and modern methods: Facebook Prophet, XGBoost. Comparative analysis of the accuracy of models is made based on calculated statistical metrics. Also, the performance of the methods was analyzed. Based on the totality of estimates, the best strategy in terms of accuracy and speed has been chosen to solve the problem. In conclusion, recommendations for improving prediction accuracy are given, and future research directions are outlined.