As the discrepancy between the model and the simulated data
(measured by the θ value used to simulate the data) rose, forecast
errors progressively increased for parametric forecasts, especially in
the low noise scenario (Fig. 3A–C). In contrast, nonparametric forecasts
were only marginally affected by the value of θ used in the simulated
datasets and were found to outperform parametric forecasts for large
value of θ (Fig. 3C and F), while the contrary was observed for smaller
θ values (Fig. 3A, D and E). SRMSE reached larger than one values for
both approaches, indicating that they both performed worse than a
simple average forecast in the long term. This result is not surprising
given the notorious sensitivity to initial conditions inherent to chaotic
systems, which prevents any reliable long-term forecast. No reliable
predictions could be achieved for more than 5 time steps in the future
in all cases. Parametric and nonparametric approaches were