I would like to predict some final portion of a projectile's trajectory given a sequence of noisy measurements collected over some initial portion of its trajectory. An obvious solution would be to use a Bayes filter, i.e. Kalman or particle filter, to estimate the target's state during the initial portion of the trajectory and then using the motion model to predict the final portion. However, this has proven to be inaccurate in practice and it was suggested that a dynamic time-series neural network may be a better option. If this is a possible solution, then which time-series network in Matlab’s Neural Network toolbox is most appropriate? As a toy example, I’d like to produce output comparable to that of a Kalman filter: can this be achieved by using measurements as an input, and Kalman filter state estimates as a target value? Must the input values and target values be transformed in any way, such as whitening the data?