[论文解读] Deep Learning Based on Generative Adversarial and Convolutional Neural Networks for Financial Time Series Predictions

Wilfredo Tovar|arXiv (Cornell University)|Aug 8, 2020

Stock Market Forecasting Methods参考文献 43被引用 7

一句话总结

本文提出了一种新颖的混合深度学习模型——Bi-LSTM-CNN，集成于生成对抗网络（GAN）框架中，用于生成保留真实股票数据关键趋势的合成金融时间序列数据。该模型在预测多个全球股指的股价走势方面优于现有方法，展现出在捕捉复杂市场动态方面的卓越准确性和鲁棒性。

ABSTRACT

In the big data era, deep learning and intelligent data mining technique solutions have been applied by researchers in various areas. Forecast and analysis of stock market data have represented an essential role in today's economy, and a significant challenge to the specialist since the market's tendencies are immensely complex, chaotic and are developed within a highly dynamic environment. There are numerous researches from multiple areas intending to take on that challenge, and Machine Learning approaches have been the focus of many of them. There are multiple models of Machine Learning algorithms been able to obtain competent outcomes doing that class of foresight. This paper proposes the implementation of a generative adversarial network (GAN), which is composed by a bi-directional Long short-term memory (LSTM) and convolutional neural network(CNN) referred as Bi-LSTM-CNN to generate synthetic data that agree with existing real financial data so the features of stocks with positive or negative trends can be retained to predict future trends of a stock. The novelty of this proposed solution that distinct from previous solutions is that this paper introduced the concept of a hybrid system (Bi-LSTM-CNN) rather than a sole LSTM model. It was collected data from multiple stock markets such as TSX, SHCOMP, KOSPI 200 and the S&P 500, proposing an adaptative-hybrid system for trends prediction on stock market prices, and carried a comprehensive evaluation on several commonly utilized machine learning prototypes, and it is concluded that the proposed solution approach outperforms preceding models. Additionally, during the research stage from preceding works, gaps were found between investors and researchers who dedicated to the technical domain.

研究动机与目标

• 解决在动态市场中预测高度复杂且混沌的金融时间序列的挑战。
• 通过开发一种更具适应性和准确性的模型，弥合研究人员与投资者在技术性金融预测之间的差距。
• 通过在GAN框架中引入结合Bi-LSTM与CNN的混合架构，超越现有单一模型方法（如独立的LSTM）的局限。
• 生成保留真实市场趋势显著特征的合成金融数据，以增强模型的泛化能力和训练效果。
• 在多样化的全球股票指数上，将所提模型与多种标准机器学习基线模型进行对比评估。

提出的方法

• 所提模型采用生成对抗网络（GAN）框架，其生成器基于混合的Bi-LSTM-CNN架构，用于合成逼真的金融时间序列数据。
• 生成器网络采用双向长短期记忆网络（Bi-LSTM），以捕捉股价序列中的长程时间依赖关系。
• 在Bi-LSTM之后集成卷积神经网络（CNN）组件，从序列数据中提取局部、类似空间的模式，增强特征表示能力。
• 判别器网络用于评估生成序列与真实金融时间序列的真实性，通过对抗训练提升数据质量。
• 模型在多个全球股指的历史数据上进行训练：TSX、S&P 500、SHCOMP 和 KOSPI 200，以实现跨市场适应性。
• 通过标准机器学习指标在多个基准模型上对系统进行评估，性能以趋势预测准确率为衡量标准。

实验结果

研究问题

• RQ1 在GAN框架中采用混合Bi-LSTM-CNN架构能否生成保留真实市场趋势特征的合成金融时间序列？
• RQ2 所提出的Bi-LSTM-CNN GAN模型在金融时间序列预测中的预测性能，相较于独立LSTM及其他传统模型表现如何？
• RQ3 该模型在S&P 500、TSX、SHCOMP和KOSPI 200等多样化全球股票市场中的泛化能力达到何种程度？
• RQ4 将对抗训练与混合深度学习生成器相结合，是否能提升股票趋势预测的鲁棒性与准确性？
• RQ5 在波动性高、非平稳的金融环境中，合成数据生成对模型泛化能力与性能有何影响？

主要发现

• 所提出的Bi-LSTM-CNN GAN模型在预测股价趋势方面优于基线模型，包括独立的LSTM网络。
• 混合架构有效捕捉了长期依赖关系（通过Bi-LSTM）与局部模式特征（通过CNN），从而提升预测准确性。
• 该模型在多个全球股票指数中表现出强大的泛化能力，表明其对多样化市场行为具有良好的适应性。
• 通过GAN框架进行对抗训练，显著提升了合成数据的质量，从而增强了预测模型的判别能力。
• 全面评估结果证实，所提模型在趋势预测任务中优于广泛使用的机器学习原型。
• 本研究识别出学术研究与实际投资者需求之间存在显著差距，而所提模型通过提升准确性和可解释性，有效弥合了这一鸿沟。