【作者】 李林；

【导师】 方雯；

【作者基本信息】 北京交通大学 ， 金融硕士（专业学位）， 2021， 硕士

【摘要】 原油是一种重要的化工原料,其价格变动引起政府、企业和投资者的密切关注。原油期货价格与现货价格往往存在联动关系而且期货具有价格发现的功能,因此尽可能准确地预测原油期货价格具有现实意义。目前关于我国原油期货市场的研究较多采取的是低频数据,但是国际市场上高频数据的应用更为广泛,为了应对金融市场开放发展的要求,因此本文选取的实验数据为1分钟的收盘价。近年来随着计算机行业的快速发展,神经网络模型被运用在各个领域,使用循环神经网络对价格时间序列进行预测也是学界研究的热点。本文使用集成经验模态分解(EEMD)和门控循环神经网络(GRU)构建混合预测模型分别对上海原油期货SC和伦敦布伦特原油期货Brent进行预测。原油价格序列具有复杂的非线性,并且包含了较多的噪音,因此本文先使用EEMD将原始数据分解为代表不同含义的本征模块函数(IMF),分别对噪音序列和趋势序列建立神经网络进行训练预测。高频数据在实际应用中对模型效率有较高要求,要尽可能快速做出预测,因此本文选择了GRU神经网络,该网络相较LSTM结构更简洁、误差小,同时选取Nadam作为神经网络的优化器来提高模型的训练速度。本文基于所构建的预测模型,提出了适合于上海原油期货的量化交易策略,该策略结合实际价格走势以及模型的预测结果,制定了相应的交易信号,然后根据交易信号进行开仓和平仓的操作。本文先通过遍历的方法确定了具有代表性的开仓和平仓阈值,然后应用该结果,使用接下来一周的高频数据来进行回测。在实证研究中,本文选取了多种类型的误差评价指标,实验结果表明本文所提出的EEMD-GRU模型在预测准确性相较于不使用EEMD分解以及其他循环神经网络的模型和计量模型更加准确,预测时间和训练速度也得到了一定程度提升。同时交易回测结果表明交易策略在实现稳定收益的同时做到了风险控制,验证了本文预测模型的效果。更多还原

【Abstract】 Crude oil is an important chemical raw material,and its price changes arouse the close attention of governments,enterprises and investors.Crude oil futures prices and spot prices often have a linkage relationship and futures have the function of price discovery.Therefore,it is of practical significance to predict crude oil futures prices as accurately as possible.At present,the research on China crude oil futures market mostly uses low-frequency data,but the application of high-frequency data in the international market is more extensive.In order to meet the requirements of the open development of the financial market,the experimental data selected in this article is the closing price of 1minute.For the past few years,with the rapid improvement of the computer performance,neural network models are used in many different industries,and the use of recurrent neural networks to predict price time series is also a hot topic in academic research.This paper uses ensemble empirical mode decomposition(EEMD)and gated recurrent neural network(GRU)to construct a hybrid forecasting model to forecast Shanghai crude oil futures(SC)and London Brent crude oil futures(Brent)respectively.In terms of data frequency,crude oil price series are complex and non-linear,and contain a lot of noise.Therefore,this article first uses EEMD to decompose the original data into intrinsic module functions(IMF)representing different meanings.The sequence establishes a neural network for training prediction.High-frequency data has high requirements for model efficiency in practical applications,and it is necessary to make predictions as quickly as possible.Therefore,this paper chooses the GRU neural network,which has a simpler structure and less error than LSTM.At the same time,this paper chose Nadam as the optimization of the neural network.To improve the training speed of the model.Based on the constructed forecasting model,this article proposes a quantitative trading strategy suitable for Shanghai crude oil futures.This strategy combines actual price trends and the forecast results of the model to formulate corresponding trading signals,and then opens and closes positions based on the trading signals operating.In this paper,the representative hyperparameters are determined by the traversal method,and then the high-frequency data for the next week is used for back-testing.In the empirical research,this paper selects multiple types of error evaluation indicators.The experimental results show that proposed model in this paper is more precise than models that do not use EEMD decomposition and other recurrent neural networks.And training speed has also been improved.The results of the trading backtest show that the trading strategy achieves risk control while achieving stable returns,which verifies the effect of the prediction model in this article.更多还原