【作者】 佟伟民；

【导师】 李一军；

【作者基本信息】 哈尔滨工业大学 ， 管理科学与工程， 2008， 博士

【摘要】 中国证券市场已经初具规模,证券市场的国际化进程也大大加快。然而对一个成熟的证券市场而言,不仅要有成熟的市场参与者,良好的市场运作机制,也应该有多样化的投资工具,以满足投资者不同的投资需求。因此,股指期货就成为中国金融市场必不可缺的金融工具。股指期货不仅是市场的迫切需要,也是培育壮大机构投资者,稳定金融市场,促进中国金融市场国际化的需求。但同时也要意识到金融衍生工具是一把“双刃剑”,在发挥规避金融市场风险、增加金融体系风险应对能力的同时,也可能出现利用新的金融衍生工具进行违法违规的活动。因此,需要在发展股指期货的同时,加强股指期货监控,完善制度建设,丰富分析手段,确保其规范运作。目前国内外各大交易所对操纵行为的识别方法是基于经验和统计的简单划分,这些方法虽然得到广泛应用,但是实际起到的效果却十分有限,也无法适应日益复杂的监控需求。而基于数据挖掘技术的方法,不仅为股指期货操纵行为识别提供了新的实现手段,而且提高了股指期货操纵行为识别的能力,增强了监控的效率,使以前无法识别的操纵行为无处遁形。本文将从股指期货操纵行为问题的特点出发,探索基于数据挖掘的股指期货操纵行为发现的方法,为股指期货操纵行为识别提供新的支撑技术与应用方法。本文首先对股指期货操纵问题的识别进行了框架性研究,并针对股指期货交易数据特点,提出了股指期货操纵行为识别的数据-方法-指标体系模型。其次,本文从股指期货操纵行为识别的主要指标出发(即价格、成交、委托及资金),分析了其数据特点,指出由于交易数据具有噪声性、非稳定性、多粒度性、二重性和实体差异性,将现有的数据挖掘方法直接运用于股指期货操纵行为发现中时,存在很多不足。为适应交易数据的这些特点,本文在对原有数据挖掘技术改造的基础上,提出了针对股指期货交易主要指标的三种分析方法,即基于小波-BP神经网络-ARMAX/GARCH的股指期货合约价格的时间序列分析模型、成交数据的支持向量机分类模型和委托行为的多分类融合分类模型。这些新的方法分析的内容涵盖了股指期货交易数据的核心部分,同时弥补了现有方法存在的不足,提高了分析的准确性及识别的及时性。为适应股指期货时间序列的噪声性和非稳定性,并克服现有分析方法存在的缺点,本文在对股指期货时间序列分析方法改进的基础上,提出了一种基于小波-BP神经网络的分析构架。基于该模型的分析方法,采用了小波分解,并在分解后的各个尺度进行建模,最后通过小波重构对指数序列进行短期预测,提高了股指期货合约价格序列预测的准确性。针对股指期货成交数据的多粒度性以及成交数据分析的需求,本文提出在解决噪声、多粒度数据分类问题方面有突出优越性的支持向量机分类方法,引入几何求解方式,提出了支持向量机的切平面解法,根据成交数据直观地对投资者进行分类。为了能够客观描述股指期货委托行为所具有的二重性与实体差异性,本文探索建立了多分类器融合分类模型。一方面,该模型采用了具有良好二分类性能的决策树作为分类器,使得该方法能够反映委托行为的二重性特点;另一方面,该模型可以在委托数据空间中使用多个基本分类器,通过分类融合器对这些基本分类器的分类结果进行融合,得出综合的分类结果,从而提高了对投资者属性空间的覆盖能力。考虑到数据中存在的实体差异性,本文采用遗传算法对分类融合器进行优化,数据实验结果极好地反映了该方法的有效性。本文的研究成果解决了现有股指期货操纵行为发现中存在的一些问题和不足,同时也为数据挖掘方法在股指期货操纵行为识别的应用上开辟了更广阔的空间。更多还原

【Abstract】 China stock market has already begun to take shape and its internationalization advancement also is greatly accelerating. However, a mature stock market must not only have mature participants and good operation mechanism, but also must have diverse investment tools to meet investor’s different need. Stock index futures has become an indispensable investment tool in China financial market. It not only satisfies China capital market urgent need, simultaneously also cultivates organization investors, stabilizes financial market, and promotes China financial market towards internationalization. However, stock index futures, as an investment derivative product, is a“double-edged sword”. It can reduce market risk and increase finance system ability to cope with risk, also possibly aggravates destruct capability from financial risk. Therefore, during development of stock index futures, it is necessary to establish monitoring and managing system, to control risk management, and to standardize its operation. At the present, all large stock exchanges are commonly using monitoring methods based on experience and simple statistical division. Although these methods are widely utilized and obtain good effect, they cannot adapt to manipulation behavior and meet the complex analysis demand. A new method based on the data mining technique can carry out more thorough monitoring of stock index futures.Embarking from characteristics of manipulation problems of stock index futures, this article explores a method based on data mining to uncover manipulation behavior with stock index futures, and thus, provides a new support technology and application method to monitor stock index futures.This article first conducted frame research of manipulation recognition of stock index futures, proposed a data - method - index system model for recognition of stock index futures manipulation behavior.Next, starting from major criteria (price, deal, order and fund) for recognition of stock index futures manipulation behavior and analyzing data characteristic of stock index futures, this article has found that, because transaction data has noise, instability, multi-granularity, duality and entity difference, many shortcomings exist when the original data mining method is directly utilized to uncover manipulation behavior during stock index futures transaction. In order to adapt to these characteristics of transaction data, this article modifies the existing data mining technique and proposes three analysis methods when dealing with different transactions: namely (1) stock index futures time series analysis model based on wavelet - BP neural network - ARMAX/GARCH, (2) deal data support vector classification model and (3) order behavior multi-classifier fusion classification model. These new methods have made up the weakness of original method and hence, enhanced the analysis accuracy.In order to adapt to the noise and instability of stock index futures time series, and to overcome the shortcomings of the existing analysis method model, this article modified analysis method for stock index futures time series and put forward an analysis framework based on the wavelet - BP neural network. Based on this model, the analysis method has used the wavelet decomposition, and created a model for different wavelength and finally restructured all wavelet in order to carry out short-term forecast and therefore, to enhance series forecast accuracy.Regarding to the multi-granularity of and analysis demand of stock index futures transaction data, this article proposed a classification method based on support vector machines. Tangent plane solution of support vector machines is established by introducing geometrical solution. Therefore, investors can be classified based on straightforward transaction data.In order to objectively describe duality and entity difference of commission behavior of stock index futures, this article explored and established the multi-classifiers fusion classification model. On one hand, this model utilized decision tree with good performanceas classification method and hence, enabled this method to reflect the dual characteristics of order behavior. On the other hand, this model can use many basic classifiers in the commission data space, fuse their results through the classifier fusions, obtain comprehensive classification result, and therefore, enhance the capability to cover investor attribute space. Considering that entity difference exists in customer data, this article optimized classification fusion using the genetic algorithm. The experiment result has demonstrated that this method is extremely effective.Research results in the article solved some existing problems and the insufficiencies in uncovering manipulation behavior with stock index futures and therefore, opened a broader space for application of the data mining method in manipulation behavior recognition of stock index futures.更多还原