PENENTUAN UNIT HIDDEN OPTIMAL PADA MODEL NEURAL NETWORK DENGAN ANALISIS KONTRIBUSI INCREMENTAL SEL

Abstract

This paper discusses about choosing the optimal number of hidden units at neural network models which is applied to the Composite Stock Price Index data in Surabaya Stock Exchange. One of the problem in fitting NN models is  an NN model which fits well may give poor out-of-sample forecasts. Thus it is required traditional methods such as ACF and PACF to select a good NN model, e.g. to select appropiate lagged variables as the ‘inputs’. The Incremental Contribution of Cells methods which belong to the general-to-specific procedure is used to choose the optimal number of hidden units. The size and topology of networks is selected by reducing the size of the network through the use of kuadratic correlation coefficients and graphical analysis of network output for every  hidden layer cell. The resulted of NN model is compared with models with different architecture those obtain from the Box-Jenkins methods. The Akaike’s Information Criterion and Schwartz Bayesian Criterion are used for comparing different models.