|
Abstract:
|
Unit root tests (URTs) are used to determine whether time series are likely stationary (I(0)) or nonstationary (I(1)). Many URTs have been proposed in the literature and none are optimal under every scenario. In this study, we combine multiple existing URTs using machine learning techniques--deep learning neural networks (DLNNs) and support vector machines (SVMs.) These combined URTs are compared to a number of standard URTs under various scenarios using Monte Carlo simulation studies. Separate simulated training and testing datasets were used. The DLNN combined URTs outperform all of the other URTs with the simulated test data, not necessarily with the training data. The SVM combined URTs outperform all of the other URTs during training--but not during testing. The DLNN algorithm uses the dropout methodology which reduces overfitting by randomly dropping a number of the units with their connections with associated weights during training. The DLNN technique over- samples the I(1) training cases, as there are less than I(0) cases, to balance the classes. The SVM tested did not balance the I(0) and I(1) classes.
|
