A Review of DAN2 (Dynamic Architecture for Artificial Neural Networks) Model in Time Series Forecasting

recently, ghiassi, saidane and zimbra [int j forecasting, vol. 21, 2005, pp. 341-362] presented a dynamic-architecture neural network for time s er ie s p re dic t io n which p er fo r ms significantly better than traditional artificial neural networks and the arima methodology. the main objective of this article is to prove that the original dan2 model can be rewritten as an additive model. we show that our formulation has several advantages: first, it reduces the total number of parameters to estimate; second, it allows estimating all the linear parameters by using ordinary least squares or ridge regression; and, finally, it improves the search for the global minimum of the error function used to estimate the model parameters. to assess the effectiveness of our approach, we estimate two models for one of the time series used as a benchmark when the original dan2 model was proposed. the results indicate that our approach is able to find models with similar or better accuracy than the original dan2.