PhyCR: Physics-Guided Causal Retrieval for Time Series Generalization Under Deployment Constraints

Time-series foundation models promise broad generalization but remain cross-channel blind and physically ungrounded, capping accuracy below decades-old statistical baselines at a much higher deployment cost. Domain-specific deep learning instead learns joint structure by brute force, at the cost of interpretability and sample efficiency. We propose PhyCR (Physics-Guided Causal Retrieval), an alternate paradigm: a discovered causal graph partitions signals into interacting subsets, each resolved via a physics-chosen operator and non-parametric summary statistics into a compact descriptor. This leads to simplified inference by retrieval rather than a learned fusion head. On four widely different, challenging benchmarks (NATOPS, WADI, SMD, FRP), PhyCR demonstrates competitive or superior results compared to representative methods from each method class. Careful ablation studies confirm the causal-discovery hypothesis, including a negative case with no causal graph available where performance degrades as predicted. The resulting representation is orders of magnitude smaller than deep or foundation-model baselines, positioning physics-grounded representation, not model scale, as a practical route to satisfy accuracy and deployment constraints.