【作者】 张伟；

【导师】 王庆全；

【作者基本信息】 东北财经大学 ， 数量经济学， 2020， 硕士

【摘要】 金融市场是我国经济的重要组成部分,近年来随着人民生活水平的提高,越来越多的人参与金融投资,关于金融指数的预测研究也成为众多学者们研究的热门领域之一。金融指数预测对于国家经济宏观调控和投资者规避风险以获得最大收益,都有着十分重要的意义。然而,金融指数的发展受多种因素的影响,如政治、经济、投资者的心理活动等,因此金融指数序列作为一个复杂的动态非线性系统,使用线性方法很难预测准确。相对于个股数据,金融指数具有透明性好、抗操纵、仅具有系统风险等优点,是合适的研究对象。因此,本文中以上证综指为代表来完成预测方法的研究分析。基于神经网络模型对上证综指的预测展开了详细的分析,选用小波神经网络、径向基函数神经网络及长短期记忆模型这三个常用于时间序列预测分析的神经网络模型,并通过Stacking融合算法将三个改进的神经网络模型进行融合,从而得到预测精度更高、误差更小的融合模型。本文选用这三个神经网络模型,主要是因为其各具独特的特点,小波神经网络兼具了小波分析和神经网络的优点,可以捕捉到时域和频域的局部信号;径向基函数神经网络通过使用对中心点径向对称的非负非线性函数,较好的克服了传统神经网络容易陷入局部最优的缺点;长短期记忆模型作为循环神经网络中的典型代表,对处理时序数据具有独特的优势,能够充分利用之前的历史信息,发现数据间潜藏的复杂非线性关系,从而对未来的发展趋势进行更合理准确的预测。本文的主要内容可以分为以下几点:首先,对上证综指的数据进行了预处理,文中选用了 15个指标作为数据特征,并对归一化处理后的数据进行了主成分分析降维,在保留了原有的95.32%有效性息的情况下,从15维降到了 5维,极大的减少了模型的输入维度。其次,分别用三个神经网络模型进行建模,并从均方根误差和拟合优度的角度对模型结果进行了对比分析,各个神经网络模型均取得了较好的效果,其中相较于径向基函数神经网络和长短期记忆模型,小波神经网络模型的效果稍微差一点。虽然这三个神经网络模型都取得了一定的效果,但是这些模型在训练的过程中会表现出一些缺点,如无法寻找全局最优、梯度消失等,因此文中又引入了粒子群优化算法来优化改进这三个模型,根据预处理好的数据分别训练了APSO-WNN、APSO-RBF及APSO-LSTM模型,结果表明改进后的模型相对于改进前的模型确实取到了一定的提升效果,在模型误差及拟合优度上都优于未改进前的模型,极大的改善了原有模型容易出现梯度消失及陷入局部最优的问题。通过各模型间的横向和纵向对比分析,发现APSO-WNN模型的提升效果最为明显,模型的预测准确率和拟合优度也是最优的。最后,本文中根据Stacking的融合策略,将APSO-WNN、APSO-RBF、APSO-LSTM三个模型作为基学习器,将数据分析及挖掘中效果最好的XGBoost模型作为次级学习器,从而完成模型融合。所得的融合模型在各方面都有显著的提升,模型的预测准确率更高,拟合优度更是达到了 0.9960。综合来看,使用自适应粒子群优化算法改进WNN、RBF及LSTM神经网络模型,再将改进的模型使用Stacking融合算法进行融合,得到的最终模型极大地提升了原有模型的预测准确率,这对于预测精度要求较高的金融指数预测是非常有意义的。这样通过预测准确率更高的模型对上证指数进行预测时,得到的结果就更具说服力,可以使相关的政府部门、机构投资者根据预测的趋势对未来相关金融指数的发展有清楚的认知,从而可以有效的把控风险,制定出更加合理有效的发展策略。更多还原

【Abstract】 The financial market is an important part of my country’s economy.In recent years,with the improvement of people’s living standards,more and more people have participated in financial investment,and the research on the prediction of financial indexes has become one of the hot research fields of many scholars.Financial index prediction is of great significance for national economic macro-control and for investors to avoid risks to obtain maximum returns.However,the development of financial indexes is affected by many factors,such as politics,economy,and investors’psychological activities.Therefore,as a complex dynamic nonlinear system,the financial index sequence is difficult to predict accurately using linear methods.Compared with individual stock data,financial indexes have the advantages of good transparency,resistance to manipulation,and only systemic risks,and are suitable research objects.Therefore,in this article,the above-mentioned composite index is used as a representative to complete the research and analysis of forecasting methods.Based on the neural network model,a detailed analysis of the prediction of the Shanghai Stock Exchange Index was carried out.Three neural network models commonly used in time series forecasting and analysis,namely wavelet neural network,radial basis function neural network and long short-term memory model,were selected and integrated through Stacking.The algorithm fuses three improved neural network models to obtain a fusion model with higher prediction accuracy and smaller errors.These three neural network models are selected in this article mainly because of their unique characteristics.Wavelet neural network combines the advantages of wavelet analysis and neural network,and can capture local signals in the time domain and frequency domain;radial basis function neural network The network uses a non-negative non-linear function radially symmetric to the center point to better overcome the shortcomings of traditional neural networks that are easy to fall into local optimum;the long-short-term memory model,as a typical representative of recurrent neural networks,has advantages in processing time series data Unique advantage,it can make full use of previous historical information to discover the hidden complex nonlinear relationship between data,so as to make more reasonable and accurate predictions of future development trends.The main content of this article can be divided into the following points:First,the data of the Shanghai Stock Exchange Composite Index was preprocessed.In the article,15 indicators were selected as data features,and the normalized data was subjected to principal component analysis and dimensionality reduction,which retained the original 95.32%validity.In the case of information,the dimension is reduced from 15 to 5,which greatly reduces the input dimension of the model.Secondly,three neural network models were used for modeling,and the model results were compared and analyzed from the perspective of root mean square error and goodness of fit.Each neural network model has achieved good results,which are compared with Radial basis function neural network and long short-term memory model,wavelet neural network model are slightly less effective.Although these three neural network models have achieved certain results,these models will show some shortcomings during the training process,such as the inability to find the global optimum,the disappearance of the gradient,etc.,so the article also introduces the particle swarm optimization algorithm to optimize The three models were improved,and APSO-WNN,APSO-RBF and APSO-LSTM models were trained based on the pre-processed data.The results showed that the improved model did achieve a certain improvement effect compared with the model before the improvement.Both the error and the goodness of fit are better than the model before the improvement,which greatly improves the problem of the original model that the gradient disappears and falls into the local optimum.Through the horizontal and vertical comparison analysis between the models,it is found that the improvement effect of the APSO-WNN model is the most obvious,and the prediction accuracy and goodness of fit of the model are also the best.Finally,according to Stacking’s fusion strategy,this article uses the three models of APSO-WNN,APSO-RBF,and APSO-LSTM as the basic learner,and uses the XGBoost model with the best effect in data analysis and mining as the secondary learner to complete Model fusion.The resulting fusion model has been significantly improved in all aspects,the model’s prediction accuracy is higher,and the goodness of fit reaches 0.9960.On the whole,the adaptive particle swarm optimization algorithm is used to improve the WNN,RBF and LSTM neural network models,and then the improved model is fused using the Stacking fusion algorithm.The final model obtained greatly improves the prediction accuracy of the original model.It is very meaningful for financial index forecasts that require high forecast accuracy.In this way,when the Shanghai Composite Index is predicted by a model with higher prediction accuracy,the results obtained will be more convincing,allowing relevant government departments and institutional investors to have a clear understanding of the future development of relevant financial indexes based on the predicted trend.Therefore,we can effectively control risks and formulate more reasonable and effective development strategies.更多还原