Multiple-cause discovery combined with structure learning for high-dimensional discrete data and application to stock prediction
详细信息 查看全文
- 作者：Weiqi Chen ; Zhifeng Hao ; Ruichu Cai ; Xiangzhou Zhang ; Yong Hu ; Mei Liu
- 关键词：Causal discovery ; High ; dimensional discrete data ; Structure learning ; Additive noise model ; Stock prediction
- 刊名：Soft Computing - A Fusion of Foundations, Methodologies and Applications
- 出版年：2016
- 出版时间：November 2016
- 年：2016
- 卷：20
- 期：11
- 页码：4575-4588
- 全文大小：919 KB
- 刊物类别：Engineering
- 刊物主题：Numerical and Computational Methods in Engineering
Theory of Computation
Computing Methodologies
Mathematical Logic and Foundations
Control Engineering - 出版者：Springer Berlin / Heidelberg
- ISSN：1433-7479
- 卷排序：20
文摘
Causal discovery in observational data is crucial to a variety of scientific and business research. Although many causal discovery algorithms have been proposed in recent decades, none of them is effective enough in dealing with high-dimensional discrete data. The main challenge is the complex interactions among large volume of variables, leading to numerous spurious causalities found. In this work, we propose a novel multiple-cause discovery method combined with structure learning (McDSL) to eliminate the spurious causalities. The method is carried out in two phases. In the first phase, conditional independence test is used to distinguish direct causal candidates from the indirect ones. In the second phase, causal direction of multi-cause structure is carefully determined with a hybrid causal discovery method. Validation experiments on synthetic data showed that McDSL is reliable in discovering multi-cause structures and eliminating indirect causes. We then applied this algorithm in discovering multiple causes of stock return based on 13-year historical financial data of the Shanghai Stock Exchanges of China, and established a stock prediction model. Experimental results showed that the McDSL discovered causes revealed changes of key risk factors of the stock market over 13 years, which indicated investors should change their investment strategy over time. Moreover, the causes discovered by McDSL have better performance in predicting stock return than that of other common filter-based feature selection algorithms.