基于EEMD-SVR的预测模型与应用

发布时间：2018-08-18 13:02

【摘要】：针对金融时间序列的复杂性，本文将经验模态分解（EMD）引入金融时间序列预测框架中进行研究。EMD多应用于通信、气象领域的数据处理，而应用于金融领域则不多，但是它具有明显的优点：能根据数据本身的时间尺度特征准确反映原时间序列的物理特性而不造成信号损失，而且无需预先设定任何基函数，这与小波分析、傅里叶变换等方法有本质的区别。但EMD存在模态混叠问题，因此需要对EMD方法进行改进和优化以提高预测的有效性。本文首先构建了总体经验模态分解（EEMD）模型，基于获取的原始数据，模拟产生多条路径的修正数据，每一次修正的数据中加入不同的白噪声以抵消原始数据中的噪声，对每个修正的序列进行EMD分解，而后取多次分解的平均值作为最后的分解序列，从而升了序列的信噪比，解决模态混叠问题。之后，将支持向量回归（SVR）模型引入到金融时间序列分析，同时，采用多种群遗传算法（MPGA）进行SVR的参数寻优。MPGA引入多个种群同时进行遗传进化搜索，对不同的种群设置相应的控制参数，并在不同种群之间依靠移民算子完成信息交互，最终在多个种群协同进化下得到最优解，，可以有效地防止早熟，大大提高收敛速度。最后，基于前文构建的EMD与SVR的改进模型，在趋势交易中进行应用。构建EEMD-MPGA-SVR预测模型。应用的结果表明：其一，对比不同参数寻优的SVR模型发现，不论是否进行EEMD分解，与网格搜索法SVR、标准遗传算法SVR相比，多种群遗传算法SVR的参数估计及其预测效果都是最好的。其二，对比进行EEMD分解前后的预测效果发现，基于EEMD分解的SVR预测效果明显优于直接采用原始序列的SVR预测（偏差较小），而且能较快地捕捉市场信息，由此所触发交易的累计收益率也更高，平均累计收益率也更高，从而收益的稳定性更强。
[Abstract]:In view of the complexity of financial time series, empirical mode decomposition (EMD) is introduced into the framework of financial time series prediction. EMD is mostly used in communication, meteorological data processing, but not in financial field. But it has obvious advantages: it can accurately reflect the physical characteristics of the original time series according to the time scale characteristics of the data without causing signal loss, and it does not need to set any basis function in advance, which is similar to wavelet analysis. Fourier transform and other methods have essential differences. However, EMD has the problem of modal aliasing, so it is necessary to improve and optimize the EMD method to improve the effectiveness of prediction. In this paper, a global empirical mode decomposition (EEMD) model is constructed. Based on the original data obtained, the modified data of multiple paths are simulated, and different white noises are added to each modified data to offset the noise in the original data. Each modified sequence is decomposed by EMD, and then the average value of multiple decomposition is taken as the final decomposition sequence, which increases the SNR of the sequence and solves the problem of mode aliasing. Then, the support vector regression (SVR) model is introduced into the financial time series analysis. At the same time, the multi-population genetic algorithm (MPGA) is used to optimize the parameters of SVR. By setting corresponding control parameters for different populations and interacting information among different populations by means of immigration operators, the optimal solution can be obtained under the coevolution of multiple populations, which can effectively prevent premature maturing and greatly improve the convergence rate. Finally, based on the improved model of EMD and SVR, it is applied in trend trading. EEMD-MPGA-SVR prediction model is constructed. The application results show that: firstly, compared with the SVR model with different parameters, it is found that, whether or not EEMD is decomposed, it is compared with the grid search method and the standard genetic algorithm (SVR). Multi-population genetic algorithm (SVR) is the best in parameter estimation and prediction. Secondly, comparing the prediction results before and after EEMD decomposition, it is found that the prediction effect of SVR based on EEMD decomposition is obviously better than that of SVR prediction using the original sequence directly (the deviation is small), and it can capture market information quickly. The cumulative yield and the average cumulative yield are higher, so the return is more stable.
【学位授予单位】：华南理工大学
【学位级别】：硕士
【学位授予年份】：2013
【分类号】：F830.91;F224;TP18

【参考文献】