Forecasting multivariate time-series: Confidence intervals and comparison of performances of feed-forward neural network and statespace models

Summary form only given. A comparison of forecasting by nonlinear neural networks and by linear state-space models was performed. In both cases, individual solutions were obtained for time-series data sets and used to determine confidence intervals for parameters of the converged systems, Multivariate time-samples of the behavior of newborn human infants under each of two conditions of stimulation were analyzed. Using a multilayer feedforward network, connection weights for nonlinear forecasting were obtained from networks converged separately for each time-series data set by means of backpropagation. Confidence intervals for connection weights were determined from observed weights of the converged networks. Similarly, optimal parameters for linear forecasting were obtained from state-space analyses using Akaike´s information criterion, and confidence intervals for parameters were determined from observed parameters of the converged state-space models. Forecasting performances for networks were superior to state-space for 22 out of 23 comparisons