Time Series Prediction Using Minimally-Structured Neural Networks: An Empirical Test

Artificial Neural Networks (ANN) have been much mentioned as a promising new tool for time series analysis and forecasting. However, to answer the question of how to determine the structure of ANN that can effectively capture the characteristics of the time series in a specific forecasting environment, it Is still required that ANN be rigorously analyzed, In terms of fitting capability and forecasting accuracy, using real world data that are often contaminated by noise and limited in the number of observations. In this paper, multilayered perceptrons (MLP) are adopted as approximators to time series generating processes. The information from ARIMA modeling is used to determine the Input units of MLP so that the designed MLP have minimal structures. The 111 series of Makridakls Competition Data are used to train the MLP and to analyze their performance. A comparative analysis with ARIMA models has been done to determine the factors that affect the forecasting performance of MLP. Examples of mese factors are the number of observations, observation intervals, seasonality, trend, and backpropagation learning parameters and procedure. The experimental results are expected to be used as a guideline for designing and training MLP.