基于时间序列模型的上证指数拟合度比较分析研究

发布时间：2018-02-06 03:33

本文关键词： 上证指数时间序列模型状态空间神经网络实证分析　出处：《上海交通大学》2014年硕士论文　论文类型：学位论文

【摘要】：对于股票价格时间序列预测研究的必要性已经成为实务界和学术界的普遍共识，但是股票价格时间序列本身具有复杂性、多样性和善变性，并且有许多因素在在影响着股票的变化，由于这些影响股市的因素，有些可以度量，，有些却难以度量，因此很难将所有的因素都通过计算量化评价从而进行科学研究的计算和评价。现代统计学对于股票市场的研究存在的问题是往往看重样本内的拟合效果，研究优化各种模型从而达到对样本内数据的完美的回归，而对样本内拟合精度的日益严格要求往往使得人们忽略了研究模型的鲁棒性，无法将样本内优异的拟合度推广到样本外，尤其是面对股票市场长期走势不断出现的突变，不具有快速的反应能力，样本内外的拟合度相差过大反而使预测不具有实用性和前瞻性。而本文试图在对传统以及现代一些流行的统计模型进行比较分析的基础上选择出样本内样本外拟合度误差最小，且长期表现更为稳定的模型，使预测更具实用性。目前对股票价格的分析和预测尽管有大量的分析工具和模型，但总的大类可以分为基本面分析，技术分析和统计类分析，而本文将要讨论的就是其中的经济统计类分析。事实上，60年前就已经出现有关股票价格的时间序列预测研究，但因为影响股票价格的因素多变复杂且难以定义，每个人的理解和运用都不同，因此对应而使用的时间序列预测方法也多种多样，但总体来看可以分成两大类：第一类是较为传统的统计学方法，包括AR,MA,ARMA,ARIMA,ARCH和GARCH等，以及其后在其基础上衍生的统计方法包括本文将会采用的状态空间模型，而另一类是近几年来普遍流行的以人工神经网络模型为代表的计算智能方法。本文通过分别回顾并总结迄今为止现有的关于股票价格时间序列预测的两大类方法，并进而基于现有的关于股票价格时间序列预测的国内外研究现状进行现有研究的评述，指出当前研究存在的问题，并对1998年以来的股票数据进行实证分析，对，，状态空间模型和人工神经网络模型分别进行固定模型下的滚动预测和滚动更新模型下的滚动预测两种实证分析，论文在对四种模型进行拟合度的比较后得出简单模型的拟合情况虽然在短期内较差，但在长期，尤其是面对市场发生突变的情况下，拟合度相比复杂模型更好，因此更具有长期稳定性。复杂模型由于样本内过拟合，短期预测时波动性较大，长期预测在遇到市场波动时往往会出现不稳定的情况，预测准确度不如简单模型。此结论为未来实证模型的选择提供一定的参考价值。
[Abstract]:The necessity of stock price time series prediction has become a common understanding in the field of practice and academia, but the stock price time series itself has complexity, diversity and variability. And there are many factors in the impact of stock changes, because of these factors affecting the stock market, some can be measured, some difficult to measure. Therefore, it is difficult to calculate and evaluate all the factors through quantitative evaluation. The problem of modern statistics for stock market research is that the fitting effect in samples is often valued. Research and optimization of various models to achieve the perfect regression of the data in the sample, and the increasingly stringent requirements for the precision of fitting in the sample often make people ignore the robustness of the model. It is impossible to extend the excellent fitting degree in the sample to outside the sample, especially in the face of the sudden changes in the long-term trend of the stock market, so it does not have the ability to react quickly. The difference between the fitting degree inside and outside the sample makes the prediction not practical and prospective. However, this paper tries to select the sample outside the sample on the basis of comparing and analyzing some traditional and modern popular statistical models. Minimum coincidence error. And long-term performance more stable model, make the prediction more practical. Although there are a lot of tools and models for stock price analysis and prediction, the general categories can be divided into fundamental analysis, technical analysis and statistical analysis. What this paper will discuss is the economic statistical analysis. In fact, 60 years ago, there has been time series prediction of stock prices. However, because the factors affecting stock prices are complex and difficult to define, each person's understanding and application are different, so the corresponding time series prediction methods are also varied. But in general, it can be divided into two categories: the first is the more traditional statistical methods, including ARMAMAA ARIMAARCH and GARCH. The statistical methods derived from them include the state-space model which will be adopted in this paper. The other is the computational intelligence method which is popular in recent years, which is represented by artificial neural network model. In this paper, we review and summarize two kinds of methods about stock price time series prediction. . And then based on the existing research status quo of stock price time series prediction at home and abroad, pointed out the existing problems. And the empirical analysis of stock data since 1998, yes, The state space model and the artificial neural network model respectively carry on the rolling forecast under the fixed model and the rolling forecast under the rolling update model two kinds of empirical analysis. After comparing the fitting degree of the four models, it is concluded that the fitting condition of the simple model is worse in the short term, but in the long run, especially in the face of the market mutation, the fitting degree is better than the complex model. Because of the over-fitting in the sample, the volatility of the short-term prediction is large, and the long-term prediction will often be unstable in the event of market volatility. The prediction accuracy is not as good as the simple model. This conclusion provides a certain reference value for the choice of future empirical model.
【学位授予单位】：上海交通大学
【学位级别】：硕士
【学位授予年份】：2014
【分类号】：F832.51;F224

【参考文献】