Researchers propose a simple, exploratory test to judge whether a model reconstructed from noisy time series actually reproduces the dynamics that generated the data. They point out that relying only on loss functions or standard metrics used for deterministic systems can be misleading for stochastic (random) dynamics. Their adequacy/reliability (AR) test instead asks whether the original observations “blend in” with data produced by the candidate model.

The AR test has two parts. The first uses Tukey’s interquartile range (IQR) rule for outliers. This compares the spread of values in the real data to values generated by the model and flags observations that fall far outside the typical range. The second part is a sign test, a simple nonparametric check that asks whether the direction of deviations between the two datasets is balanced in the way you would expect by chance. Together these checks assess both the coverage of trajectories in phase space (how well the model visits the same regions as the real system) and the statistical significance of any remaining differences. The authors use the common tolerance f = 1.5 for the IQR rule.

To illustrate and test their method, the team tried it on several well-known dynamical systems and on a data-driven recurrent neural network. The examples include chaotic three-dimensional models (Chua and Lorenz), two-dimensional oscillators (FitzHugh–Nagumo and the Lambda–Omega or Stuart–Landau oscillator), and a modified piecewise-linear recurrent neural network (PLRNN). Simulations added Gaussian noise and used standard numerical schemes (the Euler–Heun method for deterministic systems and its stochastic version for noisy systems). Training and experiments were implemented in Python with PyTorch.

This approach matters because reconstructing dynamics from noisy data is hard and ambiguous. Different reconstruction methods can give similar error numbers yet produce models that behave differently. The AR test avoids choosing an arbitrary error threshold tied to one metric. It is metric-agnostic, so it can help reveal when two models that look equally good by a loss function are actually different in the geometric structure of their trajectories.