Automatic SARIMA Order Identification Convolutional Neural Network

Abstract

Deep learning concept is popularly used to solve a classification problem including the model identification in time series analysis. For the time series models such as the autoregressive integrated moving average (ARIMA) model and the seasonal autoregressive integrated moving average (SARIMA) model, statisticians mostly identify the ARIMA/SARIMA orders before building the forecasting model. To identify them, they investigate sample ACF and PACF to extract p, d, q while most researchers automate this step using the likelihood based-method via AIC ignoring the residual diagnostic. This paper uses ACF, PACF and differencing time series images as inputs to the convolutional neural network architecture that automatically identifies the ARIMA/SARIMA orders, called the automatic ARIMA order identification convolutional neural network (AOC). The performance of AOC outperforms the likelihood based-method in terms of identifying ARIMA order via precision, recall and f1-score. Moreover, AOC is extended to identify the SARIMA order, called the automatic SARIMA order identification convolutional neural network (ASOC). The performance of the ASOC model provides better performance than the likelihood method via precision, recall and f1-score.