MASTER is a stock price forecasting model that uses a Transformer architecture guided by market information to capture both momentary and cross-time stock correlations. It is designed for researchers and practitioners in quantitative finance and algorithmic trading.

How It Works

MASTER employs a Transformer model enhanced with market information to improve stock price prediction. It models correlations between stocks across time and uses market data to guide feature selection, aiming for more robust forecasting. The approach incorporates specific preprocessing steps like RobustZScoreNorm for features and a custom DropExtremeLabel and CSZscoreNorm for labels during training.

Quick Start & Requirements

• Install: pip install pandas==1.5.3 torch==1.11.0 pyqlib
• Data: Download and unpack data into the data/ directory.
• Run: Execute python main.py.
• Prerequisites: Python 3.x, pandas, PyTorch, Qlib.
• Data: Requires processed stock data (e.g., CSI300, CSI800) and market information. Links to updated datasets are provided.
• Note: The original codebase was complex; this repo offers a streamlined version.

Highlighted Details

• AAAI-2024 paper publication.
• Models momentary and cross-time stock correlations.
• Guides feature selection with market information.
• Includes pre-trained models for CSI300 and CSI800.

Maintenance & Community

The project is associated with SJTU-DMTai. The README notes that original authors have moved on, and contributions to a Qlib implementation are from volunteers. Users are directed to this repository for MASTER-specific questions.

Licensing & Compatibility

The repository does not explicitly state a license. The code is provided as supplementary material for an academic paper. Commercial use or linking with closed-source projects would require clarification on licensing terms.

Limitations & Caveats

The project has experienced significant data issues post-publication, with original test/validation data being problematic and corrected versions provided. The authors' access to original data/codebase has expired, limiting further updates or verification. The implementation of DropExtremeLabel is described as "clumsy" and integrated directly into the training loop. A separate Qlib implementation exists but differs in data sources and preprocessing.