Hierarchical Multi-Scale Gaussian Transformers for Stock Movement Prediction

Summary

This paper proposes a Transformer model for predicting stock price movements, with enhancements intended to capture both local structure and longer-range relationships in financial time series. A multi-scale Gaussian prior is added to encourage the attention mechanism to focus on nearby observations at different scales. Orthogonal regularization is used to reduce redundancy among attention heads, while a trading-gap segmentation component learns hierarchical features from high-frequency data.

The document reports that the approach compares favorably with recurrent neural networks such as LSTMs in modeling long-term dependencies. It gives no dataset details, evaluation metrics, numerical results, or implementation guidance in the supplied text. The summary is brief and partially garbled, so the claimed comparison cannot be assessed here; it should be treated as a description of the proposed architecture rather than evidence of trading performance.

Key ideas

• A multi-scale Gaussian prior is used to add locality to Transformer attention.
• Orthogonal regularization is intended to discourage redundant attention heads.
• Trading-gap segmentation is proposed to learn hierarchical features from high-frequency financial data.
• The model is designed to predict stock movements and capture long-range temporal relationships.
• The supplied summary makes a favorable comparison with LSTMs but provides no supporting evaluation details.

Tags

This summary was written by Stratmill's research agent from the original; it is not a copy of the source.