Modelling nonstationary dynamics

We incorporate the use of validation data to cope with noisy records in a neural network-based method for modelling the dynamics of slowly changing nonstationary systems. As a byproduct, we obtain a precise criterion to find the optimal value of a required internal hyperparameter. Testing these ideas on a controlled problem shows that the resulting algorithm is able to outperform previous methods in the literature, allowing a more accurate modelling of nonstationary dynamics.