A HYBRID TRANSFER FUNCTION AND ARTIFICIAL NEURAL NETWORK MODEL FOR TIME SERIES FORECASTING

In �nance and banking, the ability to accurately predict the future cash requirement is fundamental to many decision activities. In this research, we study time series forecasting of cash in ow and out ow requirements as the output series of Indonesian Central Bank (BI) at the representative o�ces in Aceh Province, Indonesia. We use a Consumer Price Index (CPI) as the leading indicator of the input series to predict the output series. A CPI measures change in the price level of a market basket of consumer goods and services purchased by households. CPI has been used in time series modeling as a good predictor for response. In this study, we propose a hybrid approach to forecast the cash in ow and out ow by combining linear and non linear models. This methodology combines both Transfer Function and Radial Basis Function (RBF) Neural Network models. The idea behind this model is the time series are rarely pure linear or nonlinear parts in practical situations. The RBF neural network is used to develop a prediction model of the residual from Transfer Function model. The RBF Neural Networks model is trained by Gaussian activation function in hidden layer. The main concept of the proposed hybrid model approach is to let the Transfer Function forms the linear component and let the neural network forms the nonlinear component and then combine the results from both linear and nonlinear models. This combination model provides a better forecast accuracy than the individual linear or non linear model.