Cross-Domain Generalization of a Regime-Aware Forecasting Framework Across Financial, Meteorological, and Physiological Time Series

Abstract

The infrastructure style benchmarks are a necessary set, but they do not necessarily give a good idea of how an architecture will perform on data that has a very different statistical nature. This paper does just that here. We use a five-stage regime-aware forecasting framework proposed in an accompanying paper (Causal-Regime Tokeniser, a Wavelet–State-Space Dual-Path Predictor, evidence-weighted fusion, counterfactual cross-regime validation and drift aware continual adaptation) and apply it to three public datasets without any changes specific to the domains: a 5-year sample of daily closing prices for the S&P 500, a dataset of 3-year NOAA hourly temperature records, and the WESAD physiological stress dataset. It has the lowest MAE and RMSE across all three domains, the lowest CRPS and highest coverage nearer the nominal value in the probabilistic calibration task, and the highest F1 score on the physiological stress-classification task, when compared to the three baselines used in the original validation of the framework: a progressive-learning transformer, a Taylor-series/reinforcement-learning hybrid, and a kernel-based non-Markovian predictor. The largest coverage difference appears on the financial series, an increase of some 6 percentage points compared to the next highest baseline, which is related to the dynamics of the framework's conformal recalibration when abrupt volatility regimes occur. Combined, these findings suggest regime discovery, multi-scale forecasting and reliability-weighted fusion as a general tool for non-stationary structurally discontinuous data which is not tailored specifically to the infrastructure domain the framework was originally developed to address — a feature that would directly impact anyone with a forecasting system running on multiple statistically different streams of data.