Self-Identification ResNet-ARIMA Forecasting Model

Abstract

The challenging endeavor of a time series forecast model is to predict the future time series data accurately. Traditionally, the fundamental forecasting model in time series analysis is the autoregressive integrated moving average model or the ARIMA model requiring a model identification of a three-component vector which are the autoregressive order, the differencing order, and the moving average order before fitting coefficients of the model via the Box-Jenkins method. A model identification is analyzed via the sample autocorrelation function and the sample partial autocorrelation function which are effective tools for identifying the ARMA order but it is quite difficult for analysts. Even though a likelihood based-method is presented to automate this process by varying the ARIMA order and choosing the best one with the smallest criteria, such as Akaike information criterion. Nevertheless the obtained ARIMA model may not pass the residual diagnostic test. This paper presents the residual neural network model, called the self-identification ResNet-ARIMA order model to automatically learn the ARIMA order from known ARIMA time series data via sample autocorrelation function, the sample partial autocorrelation function and differencing time series images. In this work, the training time series data are randomly simulated and checked for stationary and invertibility properties before they are used. The result order from the model is used to generate and fit the ARIMA model by the Box-Jenkins method for predicting future values. The whole process of the forecasting time series algorithm is called the self-identification ResNet-ARIMA algorithm. The performance of the residual neural network model is evaluated by Precision, Recall and F1-score and is compared with the likelihood basedmethod and ResNET50. In addition, the performance of the forecasting time series algorithm is applied to the real world datasets to ensure the reliability by mean absolute percentage error, symmetric mean absolute percentage error, mean absolute error and root mean square error and this algorithm is confirmed with the residual diagnostic checks by the Ljung-Box test. From the experimental results, the new methodologies of this research outperforms other models in terms of identifying the order and predicting the future values.