When Predictive Signals Fail: Why Some Time-Series Data Cannot Support Prediction
Year: 2026
Type: preprint
Abstract: In this work, we show that predictive feasibility depends on three observable properties of the signal: the presence of structural behavior the consistency of that structure across observations and the predictability of the signal itself Across widely used benchmark datasets (IMS, C-MAPSS) and independent vibration data, we show that signals can be: predictable but not informative structurally active but inconsistent or consistent and suitable for prediction These cases reveal two fundamental failure modes: inconsistent structure, where patterns do not generalize non-informative signals, where prediction is possible but meaningless This leads to a key conclusion: Predictability is not equivalent to predictive usefulness. The results suggest that the primary limitation in predictive systems is not the model, but the signal itself.
Author: Aben, Jos
Institution: TARH (Portugal)
Language: English
Subfield: Control and Systems Engineering
Field: Engineering
Domain: Physical Sciences
Open Access status: green
OpenAlex ID: W7154412841
Cites:
Cited by:
Related to: