企业信用风险评估的统计模型——评级模型外文翻译资料

 2023-03-16 05:03

企业信用风险评估的统计模型——评级模型

作者:Aneta Ptak-Chmielewska

国籍:波兰

出处:Acta Universitatis Lodziensis. Folia Oeconomica

中文译文:

摘要:考虑到Altman提出的基于判别函数的模型(Z-score)在波兰经济条件下的不足之处,人们在90年代进行了一些尝试,以将这些模型调整到后共产主义经济的现实中。最初对多元判别分析模型的兴趣是由逻辑回归模型扩展而来,后来也由神经网络和决策树扩展而来。近年来,一些尝试也被应用于事件历史分析模型中。基于成熟破产风险模型的评级模型是信用风险评估的基本要素。本文着重于对所应用的统计方法进行批判性的评估,并且指出各种方法在模型估计中的优点和缺点。以企业为样本进行了实证比较分析,指出统计模型(评级模型)在企业信用风险评估中的可能应用。

关键词:统计模型; 评级模型; 事件历史分析

分类系统:G33, C58, C52, C45, C41, C34

1.引言

Altman于1968年引入了用于预测企业破产的多元模型,但是对这类模型的研究在波兰的历史中要短的多。

像我们在波兰那样,在转型经济条件下运作的企业市场上实施西方的模型并不成功。这些模型似乎在政治和经济变革的条件下不起作用。采用外国的模型对波兰条件的影响效果不足促进了有关本地模型的研究的发展。与国外的情况类似,最受欢迎的是基于判别分析的模型。

在90年代,开始建立和实施适应波兰经济具体情况的模型(Hadasik(1998)、Gajdka和Stos(1996)、Pogodzińska和Sojak(1995)等人的论文)。多元判别分析、回归模型和神经网络模型被使用到。我们可以列举出1990年之后发表有关破产模型的论文的作者:Appenzeller (2004)、Hołda (2001)、Michaluk (2000)、Gruszczyński(2005)、Mączyńska和Zawadzki (2006)、Hamrol和Chodakowski(2008)和其他的许多人。

传统的模型没有考虑到时间的变化也可能显著的。这种时间上的变化反映在生存模型中(即所谓的事件历史分析),其应用越来越多地出现在科学论文中(作者 2008)。

从历史角度来看,我们可以列举Gajdka和Stos(1996)提出并在1996年发表的破产预测模型。Hadasik 在她的模型中使用了多元判别函数(Hadasik 1998)。Hołda(2001)还利用多元判别分析模型,以80家企业(40家差、40家好)为样本,对模型进行了估计。Prusak(2009)创建的模型是基于用于实体财务比率分析的样本构建的。Stępień和Strąk(2004)使用了Logit模型。破产样本由39家公司组成,这些公司在1996-1998年间向Szczecin法院提交了合法的破产申请。

Appenzeller(2004)在她的作品中包括了多元判别模型中的动态变化。Mączyńska(2005)进行的研究也使用了判别函数。该样本基于1997-2002年间80个实体和FS数据。Korol(2005)揭示了神经网络相较于基于1998-2001年间180家企业和FS数据而言的判别函数的优势。Strąk ( 2005 )使用决策树证明了它是一种值得超越传统判别方法的动态性分析。Dębkowska(2012)以2009年的68家企业和FS为样本,将判别分析方法与逻辑回归和决策树进行了比较。作者(2012)将生存分析与逻辑回归和判别分析进行了比较。

本文(报告)的主要目的是比较不同统计、计量经济学和数据挖掘模型在破产预测中的有效性。本文的第二部分讨论了破产模型在评级模型开发中的应用。

2. 方法和模型

为此,本文描述并讨论了五种不同的模型和技术,并讨论了以下几种方法的优缺点:

‒ 判别分析(Fisher多元线性函数);

‒ 逻辑回归;

‒ 生存模型(半参数Cox回归模型);

‒ 决策树;

‒ 神经网络。

2.1.判别分析

判别分析用于将分类目标分为好和坏两组。这种分类是基于功能的。在这种方法中,变量集(区间)被用来构造一个以最好的方式区分好坏的规则。

主要目的是正确地分组。该函数最大化子种群之间的距离。

这种方法有一定的局限性。变量必须是正态分布的,这经常被违反。下一个假设是组间方差相等。在估计最终模型之前,必须验证这些假设。

为了得到好的分类结果并且校正误差估计,需要一个平衡的样本。

多元模型检测系统针对破产发出警告的判别函数可以采用多种形式——它们可以是线性函数或平方函数等。线性判别函数通常采用以下形式(作者2012):

其中:

Z ‒ 因变量,

‒ 截距,

‒ 判别系数(权重),

‒ 解释变量(财务比率)。

所呈现的判别函数形式中被称为Fisher判别函数的,它的参数a,被称为判别系数/比值(权重)。

在确定判别函数的形式后,确定临界值,从而使得以明确的方式将个体分类为有财务风险和无风险。

在具体分组中判别函数的平均值和中间的阈值是最经常被确定的。如果给定企业的Z值低于Z的临界值,则该个体被归类到容易受到破产风险的类别中去,而如果高于临界值,那么该个体就被认为是健康安全的。

线性判别函数的主要优点是易于理解和应用于任何IT系统和应用程序。由于应用了权重,该模型同时考虑了许多变量。它通过变换从多变量空间得到一维,以便在选定测度的基础上对情况进行评价。特殊解释变量对因变量的影响是可以确定的(尽管并非总是如此)。该模型可以被用于小样本中,在企业破产风险分析方面的分类非常精确,在动态分析方面也有适用的可能性。判别模型在许多流行的程序中都可用(Ptak-Chmielewska 2012)。

线性判别函数的基本函数是稳定性差。由于经济和空间条件的变化,这些模型可能会过时。它不可能应用于对企业状况有重大影响,比如人类活动等的定性变量。解释变量的正态性分布这一强假设往往经常被违背,而且关于个体的矩阵方差—协方差组相等的假设也很难实现。我们有必要了解总体人口的意见概率以及犯第一类错误和第二类错误的成本,只有利用这些信息才能正确估计出犯第一类错误和第二类错误的可能性。下一个限制条件是必须有夫妻和独立的日期,如果条件缺失就会使分类变得不可能。此外,比率值与实体财务状况之间的线性相关性(尽管在现实中通常是非线性的)使估计值受到限制。判断破产概率缺乏直接可能性(如果用于构建分类模型)使得判别分析在评级模型中的直接应用受到限制。企业信用风险评估的评级模型需要对个人客户层面的违约概率进行估计(Ptak-chmilewska 2012)。

2.2.逻辑回归分析

Logit模型是当今非常流行的用于预测企业破产案例的方法。二项式模型中的logit函数采用以下形式(Gruszczyński 2001, Matuszyk 2015):

其中:

‒ 因变量,通常决定破产概率,

‒ 截距,

‒ 权重(系数),

‒ 解释变量 ‒ 财务比率。

假定值来自lt;0;1gt;,其中0是“好”企业而1是“坏”企业。二项式模型的估计中一个重要的问题是要确定正确的“阈值”点。对于基于平衡样本估计的模型,该点的值通常等于0.5。组别的结构对这一点的取值是有影响的(好坏企业的情况)。

所谓的优势比在解释Logit分析的结果中起着重要作用,这个比率是根据事件发生的可能性与不发生的可能性之间的关系来计算的。

Logit模型需要满足许多假设。其中最重要的是:样本的随机性、大样本量、变量之间无共线性、观测的独立性。

逻辑回归的优点是不需要假设解释变量呈正态性分布,也不需要假设方差-协方差组矩阵相等。解释变量可以是名义变量。lt;0;1gt;是以分析事件发生概率的形式收到的结果,这个结果给出了评级模型中违约概率的直接估计。逻辑回归分析的结果以优势比的形式出现更便于解释和理解。由于应用了权重,同时考虑了许多变量,从而保证了模型的多元特性。与其它方法相比,逻辑回归模型具有更高的分类精度,也高于线性多元判别分析。此外,这种方法的优势还在于在许多统计程序中都具有可用性。

在实际应用中逻辑回归模型也存在一些不足之处。在满足正态性假设的情况下,采用线性多元判别法可获得更高的分类精度。解释变量不能具有相关性,因为解释变量的相关性会导致模型的高波动性。该模型也在解释变量的分布显著偏离正态分布的情况下具有高度的不稳定性。逻辑回归的分类结果要比人工神经网络的分类结果差。它要求必须有完整的数据,数据缺失就无法分类。对于具有高比例坏公司的大样本,理论上说可以获得良好的分类结果,但是在实践中很难获得(Matuszyk 2015)。

2.3.事件历史分析—Cox回归模型

事件历史分析—生存分析被描述为一套统计技术,旨在描述和研究个体的生命周期,即某些事件的频率、它们的顺序、分布、个体在不同状态下花费的时间。

由于可能发生的事件数量不同,我们划分了单集分析和多集分析,而单集分析是事件历史跟踪分析的最基本模型。

随机过程是分析的主题,分为三个基本领域(Blossfeld和Rohwer 2002):

‒ 预计发生区分的状态(时间)时间,

‒ 不同状态之间的转换强度,

‒ 事件的数量和顺序。

事件历史分析的基本分析结构是状态空间和时间轴。状态空间是离散的,时间度量可以是连续的,也可以是离散的。时间轴本身可以用两种方式定义:日历时间或相对时间。进入状态对于所研究群体中的所有个体来说都是共同的,这是由所有个体在一个事件于时刻发生时的共同经历来定义的(那么它是一个队列),破产事件的发生是将一个企业从一批活跃企业的队列中淘汰掉,然后它就是单集模型的最终事件(退出状态)。

– 区间因变量—Y—事件发生的时间(天或月)

– 强度模型,该值可能超出范围 [0–1]

示意图1.单集模型

附:外文原文

STATISTICAL MODELS FOR CORPORATE CREDIT RISK ASSESSMENT – RATING MODELS

Abstract. Taking into consideration the weakness of the models based on discrimination function (Z-score) proposed by Altman within the conditions of polish economy some attempts were taken in the 90s to adjust these models to the reality of post-communist economy. The initial interest in the models of multivariate discriminant analysis was extended by logistic regression models and then also by neural networks and decision trees. In the recent years some attempts were also taken to apply models of the event history analysis. Rating models based on developed bankruptcy risk models are basic element in credit risk management. Paper focuses on the critical assessment of statistical methods applied and points out the advantages and disadvantages of various approaches toward the estimation of models. Empirical comparative analysis were conducted based on the sample of enterprises. The possible application of statistical models in credit risk assessment of enterprises (rating models) was pointed out.

Keywords: statistical models, rating models, event history analysis

JEL: G33, C58, C52,

剩余内容已隐藏,支付完成后下载完整资料


STATISTICAL MODELS FOR CORPORATE CREDIT RISK ASSESSMENT – RATING MODELS

Abstract. Taking into consideration the weakness of the models based on discrimination function (Z-score) proposed by Altman within the conditions of polish economy some attempts were taken in the 90s to adjust these models to the reality of post-communist economy. The initial interest in the models of multivariate discriminant analysis was extended by logistic regression models and then also by neural networks and decision trees. In the recent years some attempts were also taken to apply models of the event history analysis. Rating models based on developed bankruptcy risk models are basic element in credit risk management. Paper focuses on the critical assessment of statistical methods applied and points out the advantages and disadvantages of various approaches toward the estimation of models. Empirical comparative analysis were conducted based on the sample of enterprises. The possible application of statistical models in credit risk assessment of enterprises (rating models) was pointed out.

Keywords: statistical models, rating models, event history analysis

JEL: G33, C58, C52, C45, C41, C34

1. INTRODUCTION

Multivariate models for forecasting bankruptcy of the enterprises were introduced by Altman in 1968 but the works on those types of models have much shorter history in Poland.

The implementation of the western models to the market of enterprises functioning in the conditions of transitional economy as we had in Poland was not successful. It appeared that those models are not working in conditions of political and economic changes. Insufficient effects of adopting foreign models to Polish conditions contributed to the development of research concerning local models. The biggest popularity, similarly to the situation abroad, was gained by

the models based on discriminant analysis.

In the 90s the activities were started to build and implement the models adjusted to the specifics of the Polish economy (papers i.e. by Hadasik (1998), Gajdka amp; Stos (1996), Pogodzińska amp; Sojak (1995)). The multivariate discriminant analysis, the regression models and neural networks models were used. Amongst the authors of the bankruptcy models who published their papers after 1990 we can enumerate: Appenzeller (2004), Hołda (2001), Michaluk (2000), Gruszczyński (2005), Mączyńska amp; Zawadzki (2006), Hamrol amp; Chodakowski (2008), Jagiełło (2005) and many others.

The traditional models do not take into consideration changes in time which can be significant. Such changes in time are reflected in the survival models (so called Event History Analysis), the application of which is more and more present in scientific papers (Author 2008).

From the historical perspective we can enumerate that Gajdka and Stos (1996) developed models to predict bankruptcy and presented them in 1996. Hadasik in her models used multivariate discriminant function (Hadasik 1998). Hołda (2001) also estimated models based on the sample of 80 enterprises (40 bad and 40 good) using multivariate discriminant analysis method. Models created by Prusak (2009) were built based on the sample used for financial ratio analysis of entities. Logit models were used by Stępień and Strąk (2004). The sample of bankruptcies consisted of 39 companies with legal bankruptcy applications submitted in the Court of Szczecin in years 1996–1998.

In her works Appenzeller (2004) included dynamic changes in the multivariate discriminant model. Research conducted by Mączyńska (2005) also used discriminant functions. The sample was based on 80 entities and FS data from years 1997–2002. Korol (2005) revealed the advantage of neural networks comparing to discriminant function based on 180 enterprises and FS from period 1998–2001. Strąk (2005) used decision trees convincing that it is worth moving analysis beyond the traditional discriminant approach. Dębkowska (2012) compared the discriminant analysis method with the logistic regression and decision trees based on the sample of 68 enterprises and FS from 2009. Author (2012) compared the survival analysis with the logistic regression and the discriminant analysis.

The main goal of this paper (report) was to compare the effectiveness of different statistical, econometric and data mining models in bankruptcy prediction. Those models are frequently used in rating models for credit risk assessment. In second part of this paper the use of bankruptcy models in rating models development was discussed.

2. METHODS AND MODELS

For the purpose of this paper five different models and techniques were described and discussed. Advantages and disadvantages of following methods were described and discussed:

‒ Discriminant analysis (Fisher multivariate linear function);

‒ Logistic regression;

‒ Survival models (semiparametric Cox regression model);

‒ Decision trees;

‒ Neural netoworks.

2.1. Discriminant analysis

The discriminant analysis is used for classification purposes into two groups: good and bad. This classification is based on the function. In this method the set of variables (interval) is used to construct a rule that distinguishes between good and bad in the best possible way.

The main purpose is correct classification into groups. The function maximizes the distance between subpopulations.

There are some limitations of this method. The variables must be normally distributed which is quite often violated. The next assumption is the equality of variances between groups. Those assumptions must be verified before the final model is estimated.

A balanced sample is required to obtain good classification and to correct errors estimation.

The discriminant functions, on which the multivariate models detection systems warn against the bankruptcy can assume various forms – th

剩余内容已隐藏,支付完成后下载完整资料


资料编号:[595637],资料为PDF文档或Word文档,PDF文档可免费转换为Word

您需要先支付 30元 才能查看全部内容!立即支付

课题毕业论文、文献综述、任务书、外文翻译、程序设计、图纸设计等资料可联系客服协助查找。