Causal Local States: Achieving Scalable Simultaneous Causal Network Inference and Forecasting for Dynamical Systems

While this thesis demonstrates the CLS framework using reservoir computing (RC) and nextgeneration reservoir computing (NGRC) as models and transfer entropy (TE) and convergent cross mapping (CCM) as causal measures, it is important to emphasize that the approach itself is both model-and causal-measure agnostic.Even though there are a lot of causal discovery algorithms [3], only a limited number of studies that use RC (or variants) to directly connect network inference to predictive performance have been conducted.Li et al. introduced an RC-based framework that uses one-step prediction error to evaluate different neighborhood configurations and subsequently encode the local coupling structure directly into the reservoir [4].Further, Srinivasan et al. provide a brief example of combining transfer entropy with their parallel RC scheme, where they [1] This property depends on the specific form of the underlying system equations; In general this does not hold for every system