Neural Predictors’ Accuracy

We examine the performance of different predictors in the deterministic environment and test their robustness against noise. In particular, we mimic the classical case of measurement noise by adding a random Gaussian signal of different intensity to the deterministic output of some archetypal chaotic systems. Then, we examine the critical case of structural noise, represented by the slow variation of the growth rate parameter of the logistic map. In both cases, the presence of noise rapidly degrades all the performance indicators, but, interestingly, the best deterministic predictor, i.e., LSTM trained without teacher forcing, remains the best also in the stochastic and non-stationary environments. Finally, we examine solar irradiance and ozone concentration time series, and again the same predictor turns out to be the best and can also be reliably applied to similar datasets in the same domain (domain adaptation).