for Comparing Accuracy of Time-Series Forecasting Methods

Abstract

The research and development of time-series forecasting requires a relative assessment of forecast accuracy, although determining which model or method to select is difficult. This study creates a simple experimental framework for selecting time-series forecasting methods, based on the methods employed as benchmarks in theM4 Competition and commonly used in machine learning competitions. We added gradient boosting and other methods used in this study. Our experimental results using M4 data confirmed the high accuracy of the combination and statistical models as in theM4 Competition.