基于DBN的汇率预测研究
发布时间：2018-09-07 12:19
【摘要】：汇率预测是一个重要的经济问题,已经引起了广泛的关注。然而,外汇市场是一个多变量的非线性系统,并且外汇市场中的各因素间的相关性错综复杂。因此,汇率预测是一项重要而富有挑战性的研究。神经网络作为非线性动力学系统,具有广泛的适应能力,学习能力,被成功地用于多变量非线性系统的建模和控制。 20世纪90年代以来,神经网络在经济、金融领域得到广泛的应用,已经成为汇率预测领域的有效工具之一。前馈神经网络(FFNN)是一种常用的汇率预测算法,但它的缺点是学习过程中易于陷入局部极小。深度信度网络(DBN)是2006年新提出的一种神经网络,能够收敛到全局最优,从而得到更精确的预测结果。 本文综述了汇率预测和深度信度网络的理论框架,研究了DBN的学习算法,并通过实验设计出DBN的最优网络结构。在此基础上,首次提出基于DBN的汇率预测方法,进行了相关实验,并对实验结果进行了分析。首先,对三种汇率序列数据做预处理,在训练阶段,我们将深度信度网络(DBN)与共轭梯度算法相结合,加快学习速度。测试阶段,使用四种评价指标来衡量算法的预测效果。最后将预测结果与前馈神经网络等几种经典算法的结果对比。实验结果表明,将DBN与共轭梯度法结合后,汇率预测的效果最好,具有良好的发展前景。
[Abstract]:Exchange rate forecasting is an important economic issue, which has attracted wide attention. However, the foreign exchange market is a multivariable nonlinear system, and the correlation between various factors in the foreign exchange market is complicated. Therefore, exchange rate forecasting is an important and challenging study. As a nonlinear dynamical system, neural network has wide adaptability and learning ability, and has been successfully used in modeling and control of multivariable nonlinear systems. The financial field has been widely used and has become one of the effective tools in the field of exchange rate forecasting. Feedforward neural network (FFNN) is a common exchange rate prediction algorithm, but its disadvantage is that it is easy to fall into local minima in the learning process. The deep reliability network (DBN) is a new neural network proposed in 2006, which can converge to the global optimum and obtain more accurate prediction results. In this paper, the theoretical framework of exchange rate prediction and depth reliability network is reviewed, the learning algorithm of DBN is studied, and the optimal network structure of DBN is designed through experiments. On this basis, the method of exchange rate forecasting based on DBN is put forward for the first time, and relevant experiments are carried out, and the experimental results are analyzed. Firstly, we preprocess the three kinds of exchange rate sequence data. In the training stage, we combine (DBN) with conjugate gradient algorithm to accelerate the learning speed. In the testing stage, four evaluation indexes are used to measure the prediction effect of the algorithm. Finally, the prediction results are compared with the results of several classical algorithms such as feedforward neural networks. The experimental results show that the combination of DBN and conjugate gradient method has the best effect and has a good prospect.
【学位授予单位】：南京大学
【学位级别】：硕士
【学位授予年份】：2012
【分类号】：TP18;F830.7
本文编号：2228247
[Abstract]:Exchange rate forecasting is an important economic issue, which has attracted wide attention. However, the foreign exchange market is a multivariable nonlinear system, and the correlation between various factors in the foreign exchange market is complicated. Therefore, exchange rate forecasting is an important and challenging study. As a nonlinear dynamical system, neural network has wide adaptability and learning ability, and has been successfully used in modeling and control of multivariable nonlinear systems. The financial field has been widely used and has become one of the effective tools in the field of exchange rate forecasting. Feedforward neural network (FFNN) is a common exchange rate prediction algorithm, but its disadvantage is that it is easy to fall into local minima in the learning process. The deep reliability network (DBN) is a new neural network proposed in 2006, which can converge to the global optimum and obtain more accurate prediction results. In this paper, the theoretical framework of exchange rate prediction and depth reliability network is reviewed, the learning algorithm of DBN is studied, and the optimal network structure of DBN is designed through experiments. On this basis, the method of exchange rate forecasting based on DBN is put forward for the first time, and relevant experiments are carried out, and the experimental results are analyzed. Firstly, we preprocess the three kinds of exchange rate sequence data. In the training stage, we combine (DBN) with conjugate gradient algorithm to accelerate the learning speed. In the testing stage, four evaluation indexes are used to measure the prediction effect of the algorithm. Finally, the prediction results are compared with the results of several classical algorithms such as feedforward neural networks. The experimental results show that the combination of DBN and conjugate gradient method has the best effect and has a good prospect.
【学位授予单位】：南京大学
【学位级别】：硕士
【学位授予年份】：2012
【分类号】：TP18;F830.7
【参考文献】
相关期刊论文 前3条
1 李丽;;基于ARMA-GARCH模型的股市量价动态关系研究[J];统计与决策;2011年04期
2 严太山;;基于遗传算法的神经网络学习算法研究[J];湖南理工学院学报(自然科学版);2007年01期
3 黄丹,黄伯元;人工神经网络及其在医疗卫生中的应用[J];卫生职业教育;2005年04期
，本文编号：2228247
本文链接：https://www.wllwen.com/guanlilunwen/huobilw/2228247.html
