STOCK-JEPA: Prior-Anchored Latent Revision Representation Learning in Equity Markets
STOCK-JEPA learns better financial representations by combining a structured financial prior with a deep learning model that predicts incremental revisions. It uses a fixed prior to anchor predictions and another branch to capture new, predictable information from historical context. This method improves prediction accuracy in equity markets by explicitly separating what the prior knows from what historical data adds.