# MASTER: Market-Guided Stock Transformer for Stock Price Forecasting

**Authors:** Tong Li, Zhaoyang Liu, Yanyan Shen, Xue Wang, Haokun Chen, Sen Huang  
**arXiv:** [2312.15235](https://arxiv.org/abs/2312.15235)  
**Date:** 2023-12-23  
**Summary page:** https://www.summarizepaper.com/en/paper/2312.15235v1/

## Key points

- Stock price forecasting is challenging due to the volatile nature of the stock market
- The paper introduces MASTER, a model that leverages market information for feature selection and information aggregation
- Ablation study shows that most settings of MASTER outperform baselines
- Influence of temperature β in gating mechanism is studied, smaller β results in stronger feature selection and larger β reduces gating effect
- Performance of MASTER evaluated on CSI300 dataset with similar results on CSI800 dataset
- Visualizations of attention maps show how MASTER captures momentary and cross-time stock correlations
- Cross-time correlation analysis reveals asymmetric patterns that change slowly as lookback window slides to forecast on different dates
- Existing methods overlook time-specific details and fail to capture momentary and cross-time stock correlations adequately
- MASTER addresses these issues by incorporating market-guided feature selection and effective information aggregation
- Experimental results demonstrate superiority over previous works, visualizations provide insights into realistic stock correlation patterns

## Summary

Stock price forecasting has long been a challenging task due to the volatile nature of the stock market. To address these limitations, this paper introduces MASTER (MArkert-Guided Stock TransformER), a novel model that leverages market information for automatic feature selection and effectively models both intra-stock and inter-stock information aggregation. The authors conduct an ablation study to evaluate different combinations of parameters and find that most settings consistently outperform the baselines. Additionally, the influence of temperature β in the gating mechanism is studied. A smaller β results in stronger feature selection, while a larger β reduces the gating effect. The performance of MASTER is evaluated on the CSI300 dataset, with similar results observed on the CSI800 dataset. The paper also presents visualizations of attention maps to demonstrate how MASTER captures momentary and cross-time stock correlations. Cross-time correlation analysis reveals that correlation patterns are asymmetric and change slowly as the lookback window slides to forecast on different dates. In conclusion, existing methods often overlook time-specific details of stock sequences and fail to adequately capture momentary and cross-time stock correlations. MASTER addresses these issues by incorporating market-guided feature selection and effectively modeling intra-stock and inter-stock information aggregation. Experimental results demonstrate its superiority over previous works, while visualizations provide valuable insights into realistic stock correlation patterns. Overall, this refined summary provides a comprehensive overview of MASTER's contributions in addressing key challenges in stock price forecasting and highlights its superior performance compared to existing methods through experimental evaluations and visualizations of captured stock correlations.