There is a way to determine whether a signal can support reliable prediction before building a model. In this work, we break this down step-by-step using real-world data, from raw signal behavior to actual prediction performance. Many predictive modeling projects fail not because of model limitations, but because the underlying data does not contain sufficient and reproducible structure to support prediction. This work introduces a practical method to assess predictive feasibility before investing significant time and resources in model development. Using the Predictive Feasibility Index (PFI), based on structural behavior, cross-run consistency, and prediction stability, signals can be classified into three categories: GO, LIMITED, or NO-GO. This provides a clear decision framework for determining whether predictive modeling is viable. The approach is validated across multiple real-world datasets, including industrial vibration data, NASA turbofan degradation datasets (FD001, FD002), NASA battery aging data, and IBM quantum calibration systems. Results show that only signals with consistent and reproducible structure lead to stable and reliable predictions. Applying this method allows organizations to: Avoid months of wasted modeling effort Reduce development costs Identify non-viable data early Focus resources on signals with real predictive value This work provides a practical pre-model decision step, enabling teams to determine whether predictive modeling is worth pursuing in the first place.
Jos Aben (Tue,) studied this question.