Large scale simulation experiments for the assessment of one-step ahead forecasting properties of stochastic and machine learning point estimation methods

The research in geophysical sciences often focuses on the comparison between stochastic and machine learning (ML) point estimation methods for time series forecasting. The comparisons performed are usually based on case studies. The present study aims to provide generalized results regarding the one-step ahead forecasting properties of several popular forecasting methods. This problem cannot be examined analytically, mainly because of the nature of the ML methods. Therefore, we conduct large-scale computational experiments based on simulations. Regarding the methodology, we compare a total of 20 methods among which 9 ML methods. Three of the latter methods are build using a neural networks algorithm, other three using a random forests algorithm and the remaining three using a support vector machines algorithm. The stochastic methods include simple methods, models from the frequently used families of Autoregressive Moving Average (ARMA), Autoregressive Fractionally Integrated Moving Average (ARFIMA) and Exponential Smoothing models. We perform 12 simulation experiments, each of them using 2 000 simulated time series. The time series are simulated using a stochastic model from the families of ARMA and ARFIMA models. The comparative assessment of the methods is based on the error and the absolute error of the forecast of the last observation.