T. Sauer, "Time series Prediction by Using Delay Coordinate Embedding," in Predicting the Future and Understanding the Past, A. Weigend and N. Gershenfeld, Eds. redwood City, CA: Addison-Wesley. 14  Home/Search   Document Not in Database   Summary   Related Articles   Check

....noise is roughly uniform 8 bit sampling resolution quantization error, we can apply our method to evaluate the local quality of the manifold approximation. 4 The prediction task is easier if we have more points that lie on the manifold, thus better constraining its shape. In the competition, Sauer (1994) upsampled the 1000 available data points with an FFT method by a factor of 32. This does not change the effective sampling rate, but it fills in more points, more precisely defining the manifold. We use the same upsampling trick (without filtered embedding) and obtain 31200 full (x; d) ....

T. Sauer. (1994) "Time Series Prediction by Using Delay Coordinate Embedding." In Time Series Prediction: Forecasting the Future and Understanding the Past, A.S. Weigend and N.A. Gershenfeld, eds., Addison-Wesley, pp. 175-193.

....Competition. In particular, this section evaluates our query by query selection of the number of neighbors based on the consistency criterion (8) The number of neighbors is limited to range from 4 to 12. We adopt for the series A an embedding model having the same dimension m = 16 proposed in [8] and for the series D an embedding model with m = 20 as reported in [10] T able1 (left) compares the NMSE (Normalized Mean Squared Error) on the A test set of the local predictor based on the consistency criterion (CC) with the local method based on cross v alidation(Press) proposed in [4] and ....

.... = 20 as reported in [10] T able1 (left) compares the NMSE (Normalized Mean Squared Error) on the A test set of the local predictor based on the consistency criterion (CC) with the local method based on cross v alidation(Press) proposed in [4] and with the performance statistics reported by Sauer [8 ] and Wan [9] T able1 (righ t) compares the RMSE (Root Mean Squared Error) on the seriesD of the D Facto public. ISBN 2 930307 00 5, pp. 311 316 B orks 0, ES Netw r 0 A l 0 ug ra 2 NN Neu e l 2 l s i 000 icia pr Artif ( A p on B 8 ro m e 2 ce iu l edi pos g 6 ngs ym i ....

T. Sauer. Time series prediction by using delay coordinate embedding. In A. S. Weigend and N. A. Gershenfeld, editors, Time Series Prediction: forecasting the futur e and understanding the past, pages 175--193. Addison Wesley, Harlow, UK, 1994.

....0.06 Table 2: RMSE for market prediction, linear and nonlinear models, as well as estimated profit threshold for indirect LIBOR forecast To model time series using MLP s, we use simple feed forward networks with one hidden layer. The input is a straightforward time delay coordinate embedding [5]; one output unit provides a direct prediction. Independent parts of the data set are cross validated to determine stop training criteria and for model selection in the class of one hidden layer MLP s. We were able to achieve an RMSE of 0:08, which is well below the margin set by the market s ....

Tim Sauer. Time series prediction by using delay coordinate embedding. In Andreas S. Weigend and Neil A. Gershenfeld, editors, Time Series Prediction, volume 15. Addison-Wesley, 1994.

....whether they are deterministic or stochastic, of the inputs outputs type or with an internal parametric or non parametric representation. ffl deterministic models, non linear physical models and chaotic approach [7, 11, 32] ffl state models (internal representation) Hidden Markov models [27, 67] generalized Kalman s filter[60] recurrent neural networks [16] ffl non linear parametrics bi linear models [31] exponential type models (EAR [35] EXPAR [40] or NEAR [12] models with a non constant variance, autoregressive, conditionally heteroscedastic ARCH [5] ....

T. Sauer, Time series prediction by using delay coordinate embedding, in Time series prediction: Forecasting the future and understanding the past, A. S. Weigend and N. A. Gershenfeld, eds., Addison Wesley, Reading, MA, 1993, pp. 175--193.

....change that occurred immediately after these similar portions of points. Previous studies haveshown that forecasting methods based on local models produce predictions that are better than or comparable to competing models and they haveanumberoffavorable properties not shared by other methods [5,8,9,25,34]. # Published in ########### ## ### ############# ######## ## ######## ######### ########## ### ######### ########, KatholiekeUniversiteit Leuven, Belgium, 112 128, July,1998 # Permanent email: mcnames alumni.stanford.org To use local models for time series prediction there are many decisions ....

....estimates y### from x# which is used in turn to estimate the rst component of the input vector x### . x### = # y### ;y ##### ; y ########## # (11) This process is iterated for p steps nally producing the prediction y### . There has been much debate about which method is superior [4, 5, 8, 9, 34]. There is strong empirical evidence that iterated prediction performs better on short term forecasts for a variety of nonlinear models [4, 5, 8] Sauer has suggested averaging the direct and iterated predictions to reduce variance [34] Direct prediction is questionable because for nonlinear ....

[Article contains additional citation context not shown here]

Tim Sauer. Time series prediction by using delay coordinate embedding. In Andreas S. Weigend and Neil A. Gershenfeld, editors, Time Series Prediction, Santa Fe Institue Studies in the Sciences of Complexity, pages 175-193. AddisonWesley, 1994.

....presence of noise in the data blurs the trajectories and breaks this invariance. Filtered embedding has been developed to reduce the effect of noise. An idea is to use a linearly transformed version of the lag vector. The singular value decomposition is such a method (Broomhead and King, 1986; Sauer, 1992). It consists in finding out the principal direction of the reconstructed attractor. In order to take into account the nonlinear nature of the data, nonlinear schemes have been introduced as well (Fraser and Swinney, 1986; Farmer and Sidorowich, 1988; Grassberger et al. 1991; Schreiber, 1993) On ....

Sauer, T. (1992) Time series prediction by using delay coordinate embedding, in A.S. Weigend and N.L. Gershenfeld (eds), Time Series Prediction, Santa Fe Institute, Proceedings Vol. XV. Addison-Wesley, pp. 175--93.

....associated training algorithm (temporal backpropagation) are more able to handle the time dependent nature of the data. Direct methods of class 3 often require high functional complexity in order to emulate the system. An example of combination of local techniques of type 1 and 3 is provided by Sauer (Sauer, 1994). In the next section we will present our local technique as a member of the second class of predictors. 3 A LOCAL METHOD FOR ITERATED PREDICTION We propose a locally weighted regression method to estimate a one step ahead predictor trained and selected according to a k step ahead criterion. ....

....is made according to the iterated PRESS. The horizon of the iterated criterion is h = 5 for the series A and h = 25 for the series D. The number of neighbors is limited to range from 4 to 12 for both series. We adopt for the series A an embedding model having the same dimension m = 16 proposed in (Sauer, 1994) and for the series D an embedding model with m = 20 as reported in (Zhang Hutchinson, 1994) Each prediction of the local model, inclusive of the modeling phase, takes about one second of computation on a Pentium machine. Table 1 compares the NMS (Normalized Mean Squared) prediction errors on ....

[Article contains additional citation context not shown here]

Sauer T. 1994. Time series prediction by using delay coordinate embedding. Pages 175--193 of: Weigend A. S. & Gershenfeld N. A. (eds), Time Series Prediction: forecasting the future and understanding the past. Harlow, UK: Addison Wesley.

....training algorithm (temporal backpropagation) are more able to handle the time dependent nature of the data. Direct methods of class 3 often require high functional complexity in order to emulate the system. An example of combination of local techniques of type 1 and 3 is provided by Sauer [9]. In the next section we will present our local technique as a member of the second class of predictors. 3 A local method for iterated time series prediction We propose a locally weighted regression method to estimate a one stepahead predictor trained and selected according to a k step ahead ....

....We adopt a local learning iterated prediction method where the selection of neighbors is made according to the iterated PRESS with horizon h = 2. The number of neighbors is limited to range from 4 to 12. We adopt for the series A an embedding model having the same dimension m = 16 proposed in [9] and for the series D an embedding model with m = 20 as reported in [15] Each prediction of the local model, inclusive of the modeling phase, takes about one second of computation on a Pentium machine. Table 1 compares the NMS (Normalized Mean Squared) prediction errors on the A test set of the ....

[Article contains additional citation context not shown here]

T. Sauer. Time series prediction by using delay coordinate embedding. In A. S. Weigend and N. A. Gershenfeld, editors, Time Series Prediction: forecasting the future and understanding the past, pages 175--193. Addison Wesley, Harlow, UK, 1994.

....3000 inputs produced the error 0.054, but for direct prediction of x(t 85) Regarding laser data, it is difficult to compare our results with known results from the literature, since they involve sophisticated preprocessing of primary data and or high complexity of computation. Sauer s approach [23] involves spline interpolation, low pass filtering of data, PCA projection, weighted regression and a huge number of LLMs even for various time prediction horizons. Wan s method [24] is based on a quite sophisticated and complex FIR feedforward network model with 25 dimensional input ....

T. Sauer. Time series prediction by using delay coordinate embedding. In A. S. Weigend and N. A. Gershenfeld, editors, Time Series Prediction: Forecasting the Future and Understanding the Past, volume XV, pages 175--193, Reading, MA, 1994. Santa Fe Institute, Addison-Wesley.

....by the organizers of the Santa Fe Institute competition in time series prediction [GW94] We used as test problem one of these time series, consisting of output of a CH 3 far infrared laser. This nonlinear chaotic time series has been predicted successfully with a locally linear model by Sauer [Sau94] and with a timedelay neural network by Wan [Wan94] Wan used an embedding dimension of 8; this value for the dimensionality of the embedding space was adopted in our experiments too, which means that our networks had 8 inputs each. The original data were measured as integers in the range from 0 ....

Sauer, T. (1994). Time series prediction by using delay coordinate embedding. In Weigend, A.S, and N.A. Gershenfeld (Eds.) (1994) Time Series Prediction. Reading, MA: Addison-Wesley, 175-193.

....and phase between all reasonably fitting training sequences (and there are even more than three) these sequences have to be sampled at a higher rate. Thus, if the training data set is upsampled by a factor of 10 (by filling in points using an interpolation technique, as suggested by Sauer [5]) and the data sequences to be compared include the upsampled data points, then the desired training sequence with continuation 545 617 is actually the best match for the last 13 or more (up to 24) data points of the original training set. 13 24 points include 1 3 periods of the waveform, which ....

....set. 13 24 points include 1 3 periods of the waveform, which represents a reasonable range of sequence lengths resp. embedding dimensions for Data Set A. However, larger sequence sizes yield again other best matches. 3. Comparing Pattern Matching with Other Methods Several elaborate predictions [4, 5, 6, 7, 8] have been carried out for Data Set A so far. The best forecast within the Santa Fe Competition was achieved by Wan [7] who assembled a network of Finite Impulse Response (FIR) linear filters. The network basically reproduced the training set sequence 545 619 as forecast for the first 75 time ....

Sauer, T., Time Series Prediction by Using Delay Coordinate Embedding, in [9], pp. 175--193, 1994.

No context found.

T. Sauer, "Time series Prediction by Using Delay Coordinate Embedding," in Predicting the Future and Understanding the Past, A. Weigend and N. Gershenfeld, Eds. redwood City, CA: Addison-Wesley. 14

No context found.

T. Sauer. Time Series Prediction by Using Delay Coordinate Embedding(Data Set A), pages 175--193. Addison Wesley, 1993.

No context found.

Sauer, T. (1993) Time series prediction by using delay coordinate embedding, In Time Series Prediction: Forecasting the Future and Understanding the Past, Ed. Weigend, A.S. and Gershenfeld, N.A., AddisonWesley, New Jersey, pp. 175-193.

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