Reservoir computing is a machine learning paradigm inspired by the dynamics of recurrent systems, where a fixed nonlinear dynamical system—the reservoir—transforms input signals into high-dimensional representations. These representations allow a simple linear combination at the readout to approximate complex temporal relationships with minimal training effort. Since its emergence in the early 2000s, this approach has shown great potential in predicting chaotic behavior, particularly by exploiting the underlying structure of nonlinear dynamical systems. However, key challenges remain in making reservoir computing more robust, interpretable, and broadly effective across diverse application domains. This work investigates the predictive capabilities of a reservoir computing architecture based on untrained recurrent dynamics combined with a trained linear readout. The method is applied to two benchmark chaotic systems: the Duffing oscillator and the Rössler attractor. A validation strategy specifically designed for chaotic time series is employed to optimize key model hyperparameters. The results demonstrate accurate short-term predictions and highlight the method’s sensitivity to initialization, underscoring the importance of proper validation and hyperparameter tuning.
