Model Selection Approach for Time Series Forecasting

Abstract

The model selection aims to estimate the performance of different model candidates in order to choose the most appropriate one. In this study we suggest exploiting specific features of time series for the optimal forecasting model selection such as length, seasonality, trend strength and others. To demonstrate reliability of feature-based approach, forecasting error distribution of LSTM Recurrent Neural Network, Linear Regression model, Holt-Winters model and ARIMA model trained on 250 time series with various characteristics were compared. Results of statistical experiments have demonstrated a significant dependence of a forecasting model on the characteristics of a series. Proposed model selection approach allows formulating a priori recommendations for choosing the optimal forecasting model for the specific time series.