金融时序数据的遗忘理论研究

The relationship between the investment behavior of investors and the time-varying data of market transactions forms a mutual mapping. Human's recognition and forgetting of market information will be reflected in the time-varying data characteristics of the market. How to reveal this feature is a major challenge in the cross-field of financial engineering and computer algorithms. Aiming at the problem of insufficient use of time dimension information of financial data, slow conversion of prediction model with market style, and difficulty in generalizing short-term results in different cycles, this project aims to explore forgetting theory in machine learning through bionic forgetting technology, and to predict the future trend of financial market by combining forgetting characteristics of financial time-varying data. Analysis. This project is based on the research of new derivative variable construction based on frequency domain analysis and financial text; induces the market operation mode by factor dimensionality reduction and feature extraction algorithm; and develops a learning algorithm with forgetting mechanism to fit the generation of human investment experience on the premise of the obtained features and taking in-depth learning as the prototype. This project solves the problem of using time dimension information when machine learning is applied to financial data forecasting, and also provides technical support for the analysis and supervision of human-led financial market.

金融市场投资人的投资行为与市场交易时变数据形成相互映射的关系，人类对市场信息的认知和遗忘特征一定会反映在市场的时变数据特征中，如何揭示这种特征是对于金融工程和计算机算法交叉领域的一大挑战性研究。针对研究中广泛存在的金融数据时间维度信息利用不充分，预测模型随市场风格转换缓慢，短期成果难以在不同周期泛化的问题，本项目旨在通过仿生遗忘等技术来探索机器学习中的遗忘理论，并和金融时序数据的遗忘特征结合来对金融市场的未来走势进行预测分析。本项目以基于频域分析和金融文本的新型衍生变量构造研究作为研究的基础；通过因子降维和特征提取算法来归纳市场运行模式；并以所得到的特征为前提，以深度学习为原型开发具有遗忘机制的学习算法来拟合人类投资经验的产生。本项目解决了机器学习应用于金融数据预测时时间维度信息的利用问题，也为由人主导的金融市场的分析和监管提供技术支持。

本项目以基于频域分析和金融文本的新型衍生变量构造研究作为研究的基础；通过因子降维和特征提取算法来归纳市场运行模式；并以所得到的特征为前提，以深度学习为原型开发具有遗忘机制的学习算法来拟合人类投资经验的产生。本项目解决了机器学习应用于金融数据预测时时间维度信息的利用问题，也为由人主导的金融市场的分析和监管提供技术支持。研究思路主要是针对研究中广泛存在的金融数据时间维度信息利用不充分，预测模型随市场风格转换缓慢，短期成果难以在不同周期泛化的问题，通过仿生遗忘等技术来探索机器学习中的遗忘理论，并和金融时序数据的遗忘特征结合来对金融市场的未来走势进行预测分析。形成了系列成果：发表论文篇,9包括SCI论文2篇，SSCI论文2篇，EI论文3篇，CSCD期刊3篇，申请国家发明专利4 项。