基于深度学习的中国原油期货价格波动预测研究

China’s crude oil futures market is affected by numerous factors with dynamic correlations and its price fluctuates dramatically. It is of great theoretical and practical significance to accurately predict the volatility of China’s crude oil futures and discover the evolutionary mechanism of the driving factors in China's crude oil futures market. Note that deep learning has a great capability of learning deep features and self-adapting parameters of high dimensional nonlinear data. Thus, it can help overcome the shortcomings of existing volatility models. This proposal aims to construct novel volatility models for China's crude oil futures based on deep learning methods. The contents are listed as the following: (1) Basing on the analysis of China’s crude oil futures market, we integrate the background knowledge into the deep neural networks and construct the variant structure models to forecast volatility. (2) We construct several hybrid models by incorporating the advantages of traditional volatility models and deep learning methods, so that we can reveal the transformation mechanism of parameter sequences. (3) Based on these hybrid models and variant structures, we further study the ensample strategies and construct deep ensample models for forecasting China's crude oil futures volatility. (4) Finally, all the new proposed deep models are applied to extract deep hidden features and evolution laws of the numerous factors, forecast future volatility of China's crude oil futures and design portfolio strategies. Hence, this proposal provides innovative approaches for modeling and forecasting China’s crude oil futures volatility, and promotes the theoretical progress of financial engineering. Furthermore, the research findings would provide policy advice for energy and finance related departments and also provide decision supports for investors’ portfolio optimization.

中国原油期货市场受众多动态关联因素的影响，其价格波动较为剧烈，精准预测中国原油期货波动率和挖掘驱动因素的深层演化规律具有重要的理论和现实意义。鉴于深度学习有高维非线性数据的深层特征学习能力和参数自适应能力，可以弥补现有波动率模型的不足，本项目拟基于深度学习构建中国原油期货波动率预测模型，具体研究分为四个部分：（1）面向中国原油期货市场，把背景知识融合到深度神经网络中，构建适用于波动率预测的深度网络变体结构模型；（2）再充分结合传统波动率模型和深度学习二者优势构建混合模型，并揭示传统波动率模型参数序列的变换机理；（3）进一步研究不同深度结构的集成策略，构建集成模型；（4）应用新构建的深度模型提取众多驱动因素的深层隐藏特征和演化规律，预测中国原油期货波动率和设计投资组合策略。本项目的研究将为波动率预测提供新的方法，推动金融工程的理论进展，为能源金融相关部门提供政策建议，也为投资者提供决策支持。

原油期货生态圈异常复杂，充斥着大量相互影响的因素，受众多动态关联因素的影响，其价格波动较为剧烈，因此，预测原油期货价格波动极具挑战性，特别是精准预测2018年推出的中国原油期货价格波动率和挖掘驱动因素的深层演化规律具有重要的理论和现实意义。然而，传统的波动率模型在处理金融市场高频数据及众多非线性关联关系的驱动因素时，其参数分布的设定还很局限，预测精度和稳健性等方面不能满足现实需求。本项目拟结合传统计量波动率模型和深度神经网络二者的优势，构建原油期货波动率模型，并应用于国际及中国原油期货市场的波动预测和风险管理。具体研究内容包括：（1）综合分析原油期货市场的价格时序特征、影响因素和各市场的特色因素等，用于各原油期货市场的价格波动预测；（2）充分结合传统波动率模型和深度学习二者优势构建混合模型，并揭示传统波动率模型参数序列的变换机理，以进一步提升预测精度；（3）进一步研究不同深度结构的集成策略，构建集成模型，预测中国原油期货波动率和设计投资组合策略。本项目的研究成果不仅为金融高频数据波动率预测提供新的方法，也为能源金融相关部门及金融市场投资者提供决策支持。目前，项目负责人与国内外科研人员合作正式发表期刊论文10篇、会议论文1篇、出版专著1部，发表期刊包括International Review of Financial Analysis、Energy Economics、Finance Research Letters、Journal of Forecasting等，合作指导研究生1名，指导相关主题的毕业论文9篇。