Sliding-Window Reordering with Overlap Averaging: A Simple Time-Domain Augmentation for Multivariate Forecasting
The episode discusses a paper proposing a method called Temporal Patch Shuffle (TPS) for multivariate forecasting augmentation. TPS extracts overlapping temporal patches, shuffles them using a variance-based ordering heuristic, and averages them to create synthetic sequences. Hosts conclude that TPS is a simple, model-agnostic tool that increases sample diversity while preserving local temporal structure.