Automatic model identification for time series analysis using deep learning

Abstract

Most time series data can be characterized by a linear process via the autoregressive integrated moving average model requiring a three-component vector which are the autoregressive, differencing, and moving average orders before fitting coefficients. A model identification which determines those orders is analyzed via the partial autocorrelation function to identify the autoregressive order, the autocorrelation function to identify the moving average order and an extended sample autocorrelation function to identify both orders which is a challenging problem for statisticians. Accordingly, the auto-ARIMA model was proposed to automatically vary those orders and estimates their corresponding coefficients. This thesis proposes three architectures of convolutional neural networks. They are widened to build the seasonal autoregressive integrated moving average model and the autoregressive conditional heteroskedasticity model. From the experiments, the proposed deep learning models outperform the auto-ARIMA model in the cases of identifying ARIMA order and the SARIMA order via precision, recall and f1-scores.