期刊
COMPUTATIONAL ECONOMICS
卷 57, 期 4, 页码 1041-1058出版社
SPRINGER
DOI: 10.1007/s10614-020-10008-2
关键词
Long short-term memory; Rectified forgetting gate; Multi-factor model portfolio; Recurrent neural network
资金
- National Social Science Foundation of China [15ZDC024]
- China Scholarship Council [201806495014]
The study introduces a new method called RFG-LSTM, which reduces the dimensionality and complexity of a neural network by using a rectified forgetting gate while maintaining the ability to process sequenced data. In real trading scenarios of China's A stock market, the model objectively learns market characteristics and rules, contributing to the implementation of portfolio investment strategies.
As has been demonstrated, the long short-term memory (LSTM) algorithm has the special ability to process sequenced data; however, LSTM suffers from high dimensionality, and its structure is too complex, leading to overfitting. In this research, we propose a new method, RFG-LSTM, which uses a rectified forgetting gate (RFG) to restructure the LSTM. The rectified forgetting gate is a function that can limit the boundary of an input sequence, so it can reduce the dimensionality and complexity of a neural network. Through theoretical analysis, we demonstrate that RFG-LSTM is monotonic, just as LSTM is; additionally, the stringency does not change in the new algorithm. Thus, RFG-LSTM also has the ability to process sequenced data. Based on the real trading scenario of China's A stock market, we construct a multi-factor alpha portfolio with RFG-LSTM. The experimental results show that the RFG-LSTM model can objectively learn the characteristics and rules of the A stock market, and this can contribute to a portfolio investment strategy.
作者
我是这篇论文的作者
点击您的名字以认领此论文并将其添加到您的个人资料中。
推荐
暂无数据