STOCK-JEPA: Prior-Anchored Latent Revision Representation Learning in Equity Markets

Battle Count: 50

Highly relevant. The paper directly addresses stock return prediction and portfolio construction, demonstrating significant improvements in RankIC and Sharpe ratios over strong baselines in both US and China markets. The approach of combining interpretable financial priors with deep learning is a key trend in quantitative finance.

High complexity. Requires implementing a custom JEPA architecture with separate branches for anchor and revision, managing EMA updates for target encoders, and handling large-scale financial data pipelines with point-in-time constraints.

The paper provides detailed architectural specifications, hyperparameters, and data processing steps in the appendices. However, the specific code implementation is not explicitly linked in the main text, though the methodology is described with sufficient detail for re-implementation.