Neural network based Monte Carlo simulation of random processes
Michael Beer, Pol D. Spanos · 2005
ABSTRACT: In this paper a procedure for Monte Carlo simulation of univariate stationary stochastic processes with the aid of neural networks is presented. As an alternative to traditional model-based simulation procedures, this one circumvents the difficulty of specifying a priori statistical properties of the process. This is particularly advantageous when only limited data are available. Neural networks operate model-free and learn directly from the data observed. They capture the pattern of short time series during the training procedure. The trained network can then generate a set of random process realizations which reflect the properties of the training data. In the present study a time lagged feed-forward network with logistic sigmoid activation functions is applied. The training of the network is realized by back-propagation with a generalized delta rule. An example demonstrates the usefulness of the proposed procedure. 1