【作者】 赵冠华；

【导师】 杜纲；

【作者基本信息】 天津大学 ， 管理科学与工程， 2010， 博士

【摘要】 财务困境分析与预测是财务管理和投资管理领域的一个重要研究方向,企业是否会陷入财务困境,这不仅关系到企业本身战略的制订与调整,而且还关系到投资者和债权人的利益。本文研究的目的,就是希望能够提出一种适合我国上市公司的、无企业规模限制、无行业局限、无股权结构限制,可以广泛应用的财务困境分析与预测方法。从而,向监管部门和广大投资者揭示,有哪些公司可能会陷入财务困境,使他们引起警觉,使监管部门维护市场稳定,为市场提供科学的决策信息。自从Altman对财务困境预测进行了开创性研究以来,财务困境分析与预测已经得到了突破性发展。近年来,也有不少学者在此领域做了许多有益的工作。但是,目前的研究总体上还缺乏系统的理论指导,尤其是在提高模型预测正确率的前提下,如何减少训练样本的数量、缩短模型运行时间、优化模型和核参数等方面,已有的成果还很少,有些方面的研究还处于起步和探索阶段。本文将遗传算法以及支持向量机理论应用于企业财务困境分析与预测,对支持向量机的算法改进以及模型参数优化等方面做了大胆的尝试,对改善模型的预测正确率、减少训练样本数量以及缩短模型运行时间等方面,进行了深入的分析和研究,主要工作和创新如下:第一,本文在对国内外已有财务困境概念定义的基础上,根据我国的实际情况,对财务困境的概念进行了界定;其次,通过对研究样本的统计分析,从财务报表项目以及财务和非财务指标三个方面,详细阐述了困境公司和正常公司在ST前不同时点上有着不同的特征。根据ST公司与正常公司的报表数据、财务指标数据的显著性差异检验结果以及均值变化趋势图,从统计学角度详细分析了哪些指标数据是导致企业出现财务困境的原因,寻找“警源”;最后,对企业发生财务困境的内外部因素进行了深入的分析,并给出了企业财务困境预测过程和预测方法框架。第二,提出了企业财务困境短期和中长期分析与预测应采用不同预测指标体系的观点。通过对ST公司和正常公司两组研究样本的指标数据分别进行正态分布检验、显著性差异检验以及因子分析处理后发现,对短期分析与预测有显著影响的指标较多,而对中长期分析与预测影响显著的指标明显减少。由于影响中长期分析与预测的指标减少,预测模型可利用的信息也就减少,从而导致与短期预测相比,中长期预测的预测正确率明显下降。另外,在指标的选取上,除了财务指标外,还选用了非财务指标,得出了股本结构和地域环境两个非财务指标对短期和中长期预测均有显著影响的结论。第三,提出了一种基于Renyi熵的最小二乘支持向量机的增长记忆算法。考虑到传统支持向量机对偶问题的求解过程相当于解一个线性约束的二次规划问题,计算矩阵的逆和存储核函数矩阵都需要较多的内存空间,同时,二次寻优算法也需要较多的运行时间。因此,本文独立推导出了一种适合企业财务困境预测的离散序列情况下的最小二乘支持向量机增长记忆算法,以避开求解矩阵的逆。同时,首次将信息熵引入增长记忆算法模型。实证结果表明,最小二乘支持向量机增长记忆算法确实节省了程序运行时间,而信息熵的引入,不但减少了训练样本的个数,而且,还提高了模型预测的正确率。第四,针对支持向量机及其改进算法中仅靠人工方法无法获得模型参数和核参数最优解的严重缺陷,本文将基于生物遗传机理的遗传算法参数优化技术应用于企业财务困境分析与预测。实证研究证实,遗传算法确实能在更大范围内自动寻优,能显著提高模型预测的正确率。尤其是将遗传算法应用于基于Renyi熵的最小二乘支持向量机增长记忆算法模型,使得在只有少量训练样本的情况下,也能获得较高的预测正确率。第五,用支持向量机及其改进算法作为工具,对短期及中长期分析与预测中多种预测模型进行了横向和纵向的比较。纵向比较结果表明,提前预测时间越短,预测正确率越高,而随着预测提前期的增加,预测的正确率显著下降;横向比较表明,支持向量机及其改进模型的预测正确率要好于传统预测模型,犯第Ⅰ类、第Ⅱ类错误的概率明显低于传统模型,进一步证实了支持向量机不但具有较好的拟合能力,而且,还有较好的泛化(预测)能力。实证结果还表明,使用高斯核函数后,其模型的预测效果要好于多项式核,但Renyi熵中的核函数只能使用多项式核,高斯核不适于做Renyi熵的核函数,这一点与其它应用领域不同,它与财务困境分析与预测的特殊性有关,这也是作者对本文做出的贡献。更多还原

【Abstract】 Financial distress prediction is an important research direction of financial management and investment management, since whether the enterprise will be in financial difficulties or not is not only related to the business strategy formulation and adjustment of its own but also related to the interests of investors and creditors. The purpose of this study is to put forward a method that can be widely used in financial distress prediction and suitable for China’s listed companies, with no restrictions of firm size, no limitations of industry, and no restrictions of ownership structure. Thus, it can reveal which companies will get into financial difficulties to the regulatory authorities and investors so that they may be alerted, and then maintain market stability and provide scientific information for decision making.Financial distress prediction has been a breakthrough since Altman carried out pioneering research on it. In recent years, many scholars in this field have done a lot of useful work .However, generally speaking , the current study is still lack of systematic theoretical guidance, especially in how to reduce the number of training samples, how to shorten the run-time of models, and how to optimize model and the nuclear parameters under the premise of improving the model prediction accuracy. Plus, the pre-existing achievments are very few, and some of them are still in start-up and exploratory stage.This article applies genetic algorithms theory and support vector machine to corporate financial distress prediction, and does a bold attempt to ameliorate spport vctor mchine algorithms and model parameters . It also deeply analyzes and studies in aspects of improving the prediction accuracy of models, reducing the number of training samples, and shortening running time of models, etc. To conclude, the main work and innovations are as follows:Firstly, based on the definitions of the concept of financial distress both at home and abroad,this paper puts forward a definition of the concept of financial distress according to China’s actual situation.Secondly, by statistical analysis of study samples, the paper explicitly states the different characteristics of ailing companies and normal companies in varied timepoints before ST in three aspects: financial reports items, financial indexes and non-financial indexes. According to results of significant difference tests and tendency charts of mean changes of ST and normal companies’reports data and financial index data, the paper dissects index data which lead to companies’financial distress from the statistics point of view, looking for the“Warning Resources”. Finally, the paper makes in-depth analysis of internal and external factors which result in corporate finacial distress and introduces the process of corporate financial distress prediction and forecasting methodological framework.Secondly, this article proposes that short-term and long-term financial distress predictions of corporate should use different indicator systems. After carrying out normal tests, significant difference tests and the treatment of factor analysis on two study samples’s indicator data of ST and normal companies separately , it shows that indicators which have a significant influences on short-term forecasts is much more, while indicators which have a significant effects on long-term forecasts reduce obviously. Because of the reduction of indicators which have a significant effects on long-term forecasts, the information which forecasting models can use reduces, and then compared to short-term forecasts the accuracy of forecast long-term forecast accuracy has declined markedly. In addition, it uses non-financial indicators for the first time in the aspect of indicators selection, and concludes that the two non-financial indicators—the geographical environment and the capital structure—affect both short- time forecasts and long-term forecasts significantly .Thirdly, this article provides a growth memory algorithm of least squares support vector machine which is based on the Renyi-entropy. Considering that the solving process of the dual problem of the traditional support vector machine is equivalent to solving a linear constrained quadratic programming problem, the inverse matrix calculation and storage of nuclear function matrix require more memory spaces, and at the same time quadratic optimization algorithm also requires more running time, this article therefore derives independently a growth memory algorithm of least squares support vector machine which is suitable for enterprise financial distress prediction of discrete sequences in order to avoid solving the inverse matrix. Meanwhile, it introduces the information entropy to growth memory algorithm model for the first time. Empirical results show that growth memory algorithm of least squares support vector machine does indeed save the running time of program, while the introduction of information entropy not only reduces the number of training samples but also improves the model prediction accuracy.Fourthly, considering the serious shortcomings that support vector machine algorithm and its improvement by artificial means alone will not have access to model parameters and the nuclear parameters of the optimal solution, this article introduces genetic algorithm parameter optimization technology which is based on bio-genetic mechanisms to corporate financial distress prediction. Empirical studies confirm that genetic algorithms can indeed optimize automatically in a wider range, which can significantly improve the model prediction accuracy. In particular, by applying genetic algorithm to growth memory algorithm of least squares support vector machine which is based on the Renyi-entropy , it is also able to obtain a higher prediction accuracy in the case of a small number of training samples.Fifthly, this article takes horizontal and vertical comparison on many forecast mode of short-term and long-term prediction with the support vector machine and its improved algorithm as a tool. Longitudinal comparison shows that the shorter the forecast ahead period is, the higher the forecasting accuracy will be, and with the increase of forecast ahead time, forecast accuracy drops significantly; horizontal comparison shows that prediction accuracy of support vector machine and its improved model are better than the traditional forecasting models, the probability of making mistakes of first classⅠ,Ⅱis obviously lower than the traditional model, which further confirms that the support vector machine has a good fitting capability as well as a good prediction capability. The empirical results also show that after using Gaussian kernel function, its effect of model prediction is better than the polynomial kernel. But the kernel function of the Renyi entropy can only use polynomial kernel, Gaussian kernel is not suitable to be the kernel function of Renyi entropy .This is different from other application areas, resulting from the special nature of financial distress prediction, which is also a major contribution of this study.更多还原