Multitimeframe Drummond Patches and JEPA Pretraining for Short-Term Retail OHLC Series Forecasting

We propose a method for constructing scale-invariant representations of retail revenue time series based on three-bar Drummond geometry (DG) computed over three adjacent periods, extended with a multitimeframe context (day, partial calendar week, and a rolling 7-day window). Self-supervised pretraining on these “patches” is performed using a joint-embedding predictive architecture (JEPA) with spatiotemporal masking, followed by fine-tuning with output heads that quantify predictive uncertainty for next-day and next-week forecasts. The work analyzes the properties of affine invariance of the features and the identifiability of the weekly phase; empirical improvement over strong baseline models on real-world data is demonstrated.