Causal Local States Infers Networks, Forecasts Dynamical Systems

Machine learning models excel at predicting real-world systems but often act as "black boxes," offering little insight into the underlying interactions driving the dynamics. Conversely, causal discovery methods can reconstruct interaction networks but typically don't prioritize predictive accuracy. Existing hybrid approaches struggle with heterogeneous systems, relying on global hyperparameters that fail to capture local variations. This research presents Causal Local States (CLS), a novel framework designed to simultaneously infer an approximate Granger-causal interaction network and forecast system dynamics. CLS operates by independently identifying the smallest set of neighbors for each node that enables near-optimal prediction of that node. These locally optimized neighborhoods are then combined to forecast the entire system. Evaluated on three benchmarks of increasing complexity, CLS demonstrated high fidelity in reconstructing the underlying networks. Furthermore, its forecasting performance was on par with models that were provided with the true network structure. This framework represents a significant step towards developing explainable and scalable forecasting methods for complex dynamical systems.