Open AccessA Hybrid GSTARX-Jordan RNN Model for Forecasting Space-Time Data with Calendar Variation Effect
Author(s) -
F. Hikmawati,
Suhartono Suhartono,
Dedy Dwi Prastyo
Publication year - 2021
Publication title -
journal of physics. conference series
Language(s) - English
Resource type - Journals
SCImago Journal Rank - 0.21
H-Index - 85
eISSN - 1742-6596
pISSN - 1742-6588
DOI - 10.1088/1742-6596/1752/1/012013
Subject(s) - autoregressive model , time series , computer science , nonlinear system , series (stratigraphy) , linear model , recurrent neural network , econometrics , mathematics , artificial intelligence , artificial neural network , machine learning , paleontology , physics , quantum mechanics , biology
Generalized Space-Time Autoregressive (GSTAR) is one of the space-time models. The GSTAR model has its limitations of not being able to model a nonlinear time series, and this can be overcome by applying a hybrid model on GSTAR. This research aims to propose modeling hybrid Time Series Regression (TSR) and hybrid GSTARX-Jordan RNN, where TSR and GSTARX model as a linear component involving the predictor variable, which is an effect of calendar variation and Jordan-RNN as a nonlinear component. This research focused on a simulation study to evaluate the goodness of the model hybrid GSTARX-Jordan RNN. There were some scenarios experimented, i.e. simulation studies in data that have linear noise and non-linear noise. The results showed that a hybrid GSTARX-FFNN, GSTARX-DLNN, and GSTARX-Jordan RNN model is the best model for predicting simulation data containing trend, seasonality, calendar variations, and nonlinear noise series compared with TSR, and GSTARX models. In general, it is in line with the results of the 2018 M4 forecasting competition show that combined models or hybrid models tend to provide more accurate forecast performance than forecast results with individual models.