A new model selection strategy in time series forecasting with artificial neural networks

Although artificial neural networks have recently gained importance in time series applications, methodological shortcomings still continue to exist. One of such shortcomings is the selection of the proper neural networks model for future use after finishing an experiment of neural networks. In other words, the number of inputs and hidden neurons must be determined, and when they are determined, the question of which trial among many trials caused by different initial weights will be selected must be answered. The general application is to divide the dataset into training, validation and test sets and selecting the neural network model that gives the smallest error value in the validation set. However, an overfitting neural networks model is likely to be selected in this case. A neural network model overfitting to validation set is likely to underperform in the test set, which shows the generalization ability of the model. A model selection strategy (GGD separation) for time series forecasting with neural networks is proposed within the scope of this dissertation. The proposed strategy initially determines the number of inputs and hidden neurons, and then selects a neural networks model from various trials by simultaneously considering performance of training and validation sets. A comparison of our proposed selection strategy with classic selection methods using simulation and real datasets shows that our selection model statistically improves neural networks performance. Simulation experiments indicate that the proposed selection strategy better converges the underlying data generalization process. Findings show that improvements in neural networks may give it a winning edge in the competition with ARIMA method. The proposed selection strategy, in general, exhibits more resistance to the validation size. Moreover, findings show that the number of trials is not a significant factor for the classic selection method; however it may influence the performance of the proposed selection strategy. Keywords: Neural Networks, ARIMA, Time Series, Forecasting, Model Selection.