G. Dorffner, "Neural Networks for Time Series Processing," Neural Network World, vol. 6, pp. 447-468, 1996.  Home/Search   Document Details and Download   Summary   Related Articles   Check

....network. This method is often called the sliding window technique as the N tuple input slides over the full training set. Figure 1 gives the basic architecture. Figure 1: The standard method of performing time series prediction using a sliding window of, in this case, three time steps. As noted in [4] this technique can be seen as an extension of auto regressive time series modelling, in which the function f is assumed to be a linear combination of a fixed number of previous series values. Such a restriction does not apply with the MLP approach as MLPs are general function approximators. 3 ....

Dorffner, G. 1996, Neural Networks for Time Series Processing. Neural Network World 4/96, 447468.

....of the network. This method is often called the sliding window technique as the N tuple input slides over the full training set. Figure 1 gives the basic architecture. Figure 1: The standardmetho o perfodF5E time seriespredictio using a slidingwindo on in this case, three time steps. As noted in [4] this technique can be seen as an extension of auto regressive time series modelling, in which the function f is assumed to be a linear combination of a fixed number of previous series values. Such a restriction does not apply with the non linear neural network approach as such networks are ....

Dorffner, G. 1996, Neural Networks for Time Series Processing. Neural Network World 4/96, 447-468.

....z f using a neural net is to use a feedforward net with one input for each member of the window and one output for z f (N t) 5] An obvious extension of this model is to have multiple outputs corresponding to multiple lead times. These models are often called sliding window nets. As discussed in [3] a more sophisticated predictor may be produced by buffering either the hidden units or the output units and recurrently adding these activations to the input vector, as in respectively Elman or Jordan type nets. These predictors are particularly useful when the data is inherently noisy [3] In ....

....in [3] a more sophisticated predictor may be produced by buffering either the hidden units or the output units and recurrently adding these activations to the input vector, as in respectively Elman or Jordan type nets. These predictors are particularly useful when the data is inherently noisy [3]. In the work reported here we have used a single hidden layer feedforward net with a sliding window input, trained with a scaled conjugate gradient algorithm. A sliding window feed forward neural network is shown below using a window size of 4. z z f Input Layer Output Layer Time Encoding ....

G. Dorffner, Neural Networks for Time Series Processing,Neural Network World 4/96, 447-468, 1996.

....Time series prediction has been successfully used to support the decision making in several real world application areas. Several techniques can be deployed for modeling and predicting time series, among which we highlight the Box Jenkins approach [1] and the Artificial Neural Networks (ANNs) [2]. The former approach provided an advance in the field of time series prediction but it is only capable of constructing linear models. In contrast, ANNs are able to approximate non linear functions, however problems have to be faced concerning the choice of an appropriate network architecture to ....

....with few free parameters. Consequently, they are easier and faster to be designed than non linear models. However, linear models are not capable of adequately describing all existing phenomena. An alternative approach which implements non linear models is via the use of Artificial Neural Networks [2]. In [5] the authors point out two characteristics of neural networks that make them very attractive for time series prediction: the ability to approximate functions, and the direct relationship with classical statistical models, such as Box Jenkins models. Despite these advantages, we can quote ....

[Article contains additional citation context not shown here]

G. Dorffner, Neural Networks for Time Series Processing, Neural Network World, 6(4), 1996, 447-468.

....series prediction has been successfully used to support the decision making in several application areas. A number of techniques has been developed for modeling and predicting time series, among which we highlight the Box Jenkins approach [Box et al. 1994] and the Artificial Neural Networks (ANNs) [Dorffner 1996]. The former approach, although widespread, is only capable to construct linear models. In contrast, ANNs are able to model non linear functions, however problems have to be faced concerning the choice of an appropriated network architecture to model the series to be predicted. A more unifying ....

....with few free parameters. Consequently, they are easier and faster to be designed than nonlinear models. However, linear models are not capable of adequately describing all existing phenomena. An alternative approach which implements non linear models is via the use of Artificial Neural Networks [Dorffner 1996]. In [Drossu, Obradovic 1995] the authors point out two characteristics of neural networks that make them very attractive for time series prediction: the ability to approximate functions, and the direct relationship with classical statistical models, such as Box Jenkins models. Despite these ....

[Article contains additional citation context not shown here]

Dorffner, G.(1996), "Neural Networks for Time Series Processing", Neural Network World, 6(4), pages447-468.

....Hidden Input Context 1 Investments in different media categories TOM Figure 4: SRN architecture with input and output for this problem. A tapped delay line is a simple form of a short term memory and TDNN s have therefore often been used in prediction studies e.g. 4 7] As noted in [8] this technique can be seen as an extension of auto regressive time series modeling, in which the output is assumed to be a linear combination of a fixed number of previous series values. Such a restriction does not apply with the MultiLayer Perceptron (MLP) approach, as MLPs are general function ....

Dorffner, G., "Neural Networks for Time Series Processing", Neural Network World 4/96, 1996, pp. 447-468.

....This method is often called the sliding window technique as the N tuple input slides over the full training set. Figure 1 gives the basic architecture. Figure 1: The standard method of performing time series prediction using a sliding window of, in this case, three time steps. As noted in [4] this technique can be seen as an extension of auto regressive time series modelling, in which the function f is assumed to be a linear combination of a fixed number of previous series values. Such a restriction does not apply with the MLP approach as MLPs are general function approximators. III. ....

Dorffner, G. 1996, Neural Networks for Time Series Processing. Neural Network World 4/96, 447-468.

....nonlinear time series. These empirical results can be taken as valuable hints with respect to the practical application of neural networks in prediction tasks. 1 Introduction Feedforward and recurrent multilayer perceptrons (e.g. of the so called Jordan and Elman type see [ Bengio, 1995 ] or [ Dorffner, 1996 ] for an overview) are popular neural networks for complex time series processing tasks, such as forecasting in financial applications [ Weigend et al. 1991; Refenes et al. 1994; Trippi Turban, 1993 ] They are applied due to their strength in handling non linear functional dependencies ....

....limits are obviously limits of this kind of learning (compare [ Horne Giles, 1995 ] p.700) One can conclude from these results that for unknown underlying nonlinear characteristics of a time series, the feedforward NAR model appears to be most likely to lead to satisfying results. In [ Dorffner, 1996 ] it has been shown that Jordan type neural networks resemble a version of a nonlinear autoregressive moving average (NARMA) model (see also [ Connor et al. 1992 ] In our case, both Jordan type networks implement NARMA(1,1) models (with the extension of selfrecurrent loops at the state ....

[Article contains additional citation context not shown here]

Dorffner G.: Neural networks for time series processing, Neural Network World, 6(4)447-468, 1996.

No context found.

G. Dorffner, "Neural Networks for Time Series Processing," Neural Network World, vol. 6, pp. 447-468, 1996.

No context found.

G. Dorffner. Neural networks for time series processing. Neural Network World, 6(4):447--468, 1996.

No context found.

G. Dorffner. Neural networks for time series processing. Neural Network World, 6(4):447--468, 1996.

Online articles have much greater impact   More about CiteSeer.IST   Add search form to your site   Submit documents   Feedback  

CiteSeer.IST - Copyright Penn State and NEC