Feedforward Neural Nets as Models for Time Series Forecasting

Abstract

We have studied neural networks as models for time series forecasting, and our research compares the Box-Jenkins method against the neural network method for long and short term memory series. Our work was inspired by previously published works that yielded inconsistent results about comparative performance. We have since experimented with 16 time series of differing complexity using neural networks. The performance of the neural networks is compared with that of the Box-Jenkins method. Our experiments indicate that for time series with long memory, both methods produced comparable results. However, for series with short memory, neural networks outperformed the Box-Jenkins model. Because neural networks can be easily built for multiple-step-ahead forecasting, they may present a better long term forecast model than the Box-Jenkins method. We discussed the representation ability, the model building process and the applicability of the neural net approach. Neural networks appear to provide a promising alternative for time series forecasting. INFORMS Journal on Computing, ISSN 1091-9856, was published as ORSA Journal on Computing from 1989 to 1995 under ISSN 0899-1499.

Hybrid Water Demand Forecasting Model Associating Artificial Neural Network with Fourier Series

“…However, it was generally recommended that data should not be normalized to the limits of activation function, because these nonlinear functions have asymptotic limits, reaching the limits only for infinite values (Tang and Fishwick, 1993), and also in the case the future values are slightly outside the range reached by the observed data so far. A linear normalization for the interval [0.25, 0.75] was used in this study…”

Neural networks for regional employment forecasts: are the parameters relevant?

“…We focus on forecasting biannual growth rates, that is, forecasting 2 years ahead (t, t ? 2), and use panel data for the periods 1987-2004and 1993-2004 The panel nature of the data is indeed the most important aspect of our experiments. Differently from conventional panel models (see, for example, Baltagi 2001), a standard NN does not include temporal correlation.…”

“…Whether or not our findings match these considerations relies on whether our data should be considered 'complex'. Generally, Tang and Fishwick (1993) state that, for each series of data, a set of NN parameters can be found that performs significantly better than the rest. 9 For example, a momentum value set at 0.5 means that 50 percent of the weight adjustment, at each stage, will be on the basis of the current error, while the remaining 50 percent will be due to the adjustment applied in the previous iteration.…”

Approximate k-NN delta test minimization method using genetic algorithms: Application to time series

“…There is abundant literature on nonlinear models for time series forecasting [3,4,5,6,7,8]. Among the existing methods are neural networks [9,10,11,12,13,14,15], radial basis function networks [11,16,17,18], support vector machines [19,20,21,22], self organizing maps [23,24] and other variants of these models [11,25,26,27,28]. However, building these models takes considerable computational time compared to linear models.…”