基于优化回声状态网络的股价预测研究  

Stock Index Prediction Based on Optimized Echo State Network

作　　者：张宁致[1] 周佳丽[1] 孙武军[1] 

机构地区：[1]南京大学商学院,江苏南京210093

出　　处：《管理工程学报》2014年第1期94-101,共8页Journal of Industrial Engineering and Engineering Management

基　　金：国家自然科学基金资助项目(70972032);国家社科重点资助项目(08AJY029);江苏高校优势学科建设工程项目资助项目(PAPD)

摘　　要：股票市场是一个充斥着各种噪声的动态非线性系统,能够精确地对其进行预测是一项具有挑战性的任务。本文构建的BCC-ESN模型,是运用细菌群体趋药性算法(BCC)来优化回声状态网络(ESN)的权值结构,在继承ESN优良性质的同时,具有更高的模型预测能力。实验证明,BCC-ESN模型比前馈神经网络具有更好的学习和预测能力。经对上证指数进行短期价格预测,结果与BP网络、Elman网络和ESN网络进行比较,BCC-ESN模型精度明显优于其他三种网络预测。同时,在运算效率上,BCC-ESN模型继承了ESN的运算优势,明显优于其他神经网络预测模型,是一种切实可行、高效的预测算法,尤其在股票时间序列预测中有广泛的实用价值。针对BCC-ESN模型在训练预测中遇到的问题,比如耗费时间过长和过度拟合问题,本文亦提供了简单易行的思路和方法。Stock market analysis and forecasting have been the focus of economic research.Many research methods,such as fundamental analysis,technical analysis and time series analysis,have been introduced in the existing literature.Computational Intelligence is a set of nature-inspired computational methodologies,which incorporate the laws of nature into the optimization process through computer programs.Neural network has become one of primary computational intelligence means for price prediction in recent years.As a novel approach to recurrent neural network training,echo state network （ESN） has better chaotic series predictability and higher convergence rate than traditional networks.However,choosing initial transient is still based on experience rather than reliable method.The second part introduces ESN structure and bacterial colonychemotaxis （BCC） algorithm,and compares ESN with other neural network algorithms.We further introduce BCC-ESN structure and algorithm,discuss our proposed model using the BCC algorithm to optimize the weight structure of the ESN.The third part is a simulation experiment,using the BCC-ESN model to predict stock price.The purpose of this experiment is to analyze the prediction accuracy and efficiency under different circumstances.The experimental results show that BCC algorithm on the ESN optimization can improve prediction accuracy.The purposes of this paper are to address two primary issues.Firstly,due to the special structure of ESN reserve pool the training process is much more time-consuming when using BCC algorithm to optimize ESN.This is because a large number of weights need to be optimized by using the BCC algorithm.Secondly,the generalization ability is a very important aspect of neural network learning.When the BCC-ESN structure is too large,the training process may be faced with over-fitting problems,thereby making the fitness more than the optimal fit theory.Over-fitting model actually contains wrong information,and the predictive ability is very poor.Reflected in the spe

关 键 词：回声状态网络 细菌群体趋药性算法 股票价格预测 算法优化 

分 类 号：F830.91[经济管理—金融学]