Pre-Model Structure Screening Across Non-Stationary Financial and Neural Time Series

Time series from diverse domains such as financial markets and neuroscience are frequently analyzed using statistical and machine-learning models that implicitly assume the presence of stable underlying structure. In practice, however, such signals are often highly non-stationary, regime-dependent, and noisy, raising the question of whether and when downstream modeling assumptions are valid. This work presents a cross-domain evaluation of a pre-model structure screening framework, applied to two fundamentally different classes of data: financial market indices and electroencephalographic (EEG) recordings. Using a magnitude-based coherence index computed in a sliding-window manner, the method assesses time-varying structural organization in raw signals without relying on task labels, frequency bands, prediction targets, or domain-specific modeling assumptions. Across both financial and neural time series, the analysis reveals intermittent and smoothly varying regimes of structural coherence that are not apparent from visual inspection of the raw data alone. Despite their distinct origins and interpretations, both domains exhibit analogous patterns of transient structural stability and instability, suggesting that pre-model structure screening captures domain-agnostic properties of complex, non-stationary systems. The contribution of this work is methodological rather than inferential. The proposed framework is intended as an upstream diagnostic and gatekeeping step, informing when modeling, inference, or prediction may be structurally admissible, and when such efforts may be unreliable. By synthesizing results across finance and neuroscience, this paper positions pre-model structure screening as a general tool for responsible modeling in complex data environments.