Design of Experiment to Optimize the Architecture of Deep Learning for Nonlinear Time Series Forecasting

The neural architecture is very substantial in order to construct a neural network model that produce a minimum error. Several factors among others include the input choice, the number of hidden layers, the series length, and the activation function. In this paper we present a design of experiment in order to optimize the neural network model. We conduct a simulation study by modeling the data generated from a nonlinear time series model, called subset 3 exponential smoothing transition auto-regressive (ESTAR ([3]). We explore a deep learning model, called deep feedforward network and we compare it to the single hidden layer feedforward neural network. Our experiment resulted in that the input choice is the most important factor in order to improve the forecast performance as well as the deep learning model is the promising approach for forecasting task.