A quantitative comparison of models for univariate time series forecasting

Abstract

ARIMA is a popular method to analyze stationary univariate time series data, and nowadays it is considered the standard method for time series forecasting. We experimentally show that two machine learning-based approaches can be used for the forecasting of univariate time series. The experiments are made on ten public time series datasets and we report the results obtained by ARIMA, linear regression, and multilayer perceptron networks. The quantitative results show that linear regression and multi-layer perceptrons obtain more accurate predictions than the ones produced by ARIMA.