Measures of serial data compressibility by neural network predictors

A time series or univariate random process is compressible if it is predictable. Experiments with a variety of processes readily show that adaptive neural networks are at least as effective as their linear counterparts in one-step-ahead prediction. The relationship between the predictive accuracy attained by the network, in the long run, and the closeness with which it can fit (and overfit) small segments of the same series in the course of many passes through the same data is examined. The findings suggest that the predictability of a process can be estimated by measuring the ease with which its increments can be overfitted