Connor, J. T. 1993. Bootstrap methods in neural network time series prediction. In International Workshop on Applications of Neural Networks to Telecommunications, J. Alspector, R. Goodman and T. X.  Home/Search   Document Not in Database   Summary   Related Articles   Check

....of the data vs. the randomness of initial conditions of the network. This is not the first article that uses the bootstrap in a connectionist context. Weigend et al. : 1992) used the bootstrapping of residuals to evaluate the forecasting power of a neural net for exchange rate forecasts, and Connor (1993) also bootstrapped residuals to obtain error bars for the iterated time series predictions. The goals were different from the goal of the work reported here. In this article we resample pairs which will be clarified in Section II A. Resampling pairs was first suggested by Efron (1982) and first ....

....one model to obtain a distribution, called bootstrapping residuals. The latter method was used in single step prediction by Weigend et al. : 1992) in the context of foreign exchange rate predictions. One model was built on one split of the data. Similarly, in an application to load forecasting, Connor (1993) trained one single step prediction network on one split of the data, then resamples from the empirical distribution of the single step errors and adds these to the inputs in order to obtain estimates of the errors of iterated forecasts. In this residuals bootstrap, the residuals obtained from one ....

Connor, J. T. 1993. Bootstrap methods in neural network time series prediction. In International Workshop on Applications of Neural Networks to Telecommunications, J. Alspector, R. Goodman and T. X.

....to the original predictions to create a synthetic training set on which the learner is retrained. By creating a number of bootstrapped predictions and comparing their average prediction with that of the original predictor, one arrives at a first order bootstrap estimate of the predictor s bias [Connor 1993; Efron and Tibshirani, 1993] It is known that this estimate is itself biased towards zero; a standard heuristic is to divide the estimate by 0.632 [Efron, 1983] A disadvantage of the bootstrap method is that, because it requires repeated fitting, it is computationally expensive. 3.2.2 ....

Connor, J. (1993). Bootstrap Methods in Neural Network Time Series Prediction. In J. Alspector et al., eds., Proc. of the Int. Workshop on Applications of Neural Networks to Telecommunications, Lawrence Erlbaum, Hillsdale, N.J.

....techniques described in this paper) the bias will constitute a larger and larger portion of the remaining error. Bias is not as easily estimated as variance; it is usually estimated by expensive cross validation, or by running ensembles of learners in parallel (see, e.g. Geman et al. 1992] and Connor [1993]) Future work will need to include methods for efficiently estimating learner bias and taking steps to ensure that it too is minimized in an optimal manner. Acknowledgements I am indebted to Michael I. Jordan and David J.C. MacKay for their help in making this research possible. Thanks are ....

J. Connor. (1993) Bootstrap methods in neural network time series prediction. In J. Alspector et al., eds., Proceedings of the International Workshop on Application of Neural Networks to Telecommunications, Lawrence Erlbaum, Hillsdale, NJ.

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