1、 外文翻译 原文 Financial Ratios and the Probabilistic Prediction of Bankruptcy Material Source: Journal of Accounting Research Vol. 18 No. 1 Spring 1980 Printed in U.S.A Author:JAMES A. OHLSON 1.Introduction This paper presents some empirical results of a study predicting corporate failure as evidenced by
2、 the event of bankruptcy. There have been a fair number of previous studies in this field of research, the more notable published contributions are Beaver 1966;1968a;1968b,Altman1968,1973,Altman and Lorries 1976, Altman and McGough 1974,Altman,Haldeman, and Narayanan 1977,Deakin 1972, Libby 1975,Blu
3、m 1974, Edmister 1972, Wilcox 1973, Moyer 1977, and Lev1971. Two unpublished papers by White and Turnbull 1975a,1975b and a paper by Santomero and Vinso 1977 are of particular interest as they appear to be the first studies which logically and systematically develop probabilistic estimates of failur
4、e. The present study is similar to the latter studies, in that the methodology is one of maximum likelihood estimation of the so-called conditional logit model. The data set used in this study is from the seventies (1970-76).I know of only three corporate failure research studies which have examined
5、 data from this period. One is a limited study by Altman and McGough1974 in which only failed firms were drawn from the period 1970-73 and only one type of classification error (misclassification of failed firms)was analyzed.Moyer1977 considered the period 1965-75,but the sample of bankrupt firms wa
6、s unusually small (twenty-seven firms).The third study,by Altman,Haldeman,and Narayanan1977, which “updates“the original Altman1968 study,basically considers data from the period 1969 to 1975.Their sample was based on fifty-three failed firms and about the same number of nonfailed firms.In contrast,
7、 my study relies on observations from 105 bankrupt firms and 2,058 no bankrupt firms. Although the other three studies differ from the present one so far as methodology and objectives are concerned, it is, nevertheless, interesting and useful to compare their results with those presented in this pap
8、er. However, one might ask a basic and possibly embarrassing question: why forecast bankruptcy? This is a difficult question, and no answer or justification is given here. It could perhaps be argued that we are dealing with a problem of “obvious“ practical interest. This is questionable since real-w
9、orld problems concern themselves with choices which have a richer set of possible outcomes. No decision problem I can think of has a payoff space which is partitioned naturally into the binary status bankruptcy versus non-bankruptcy.(Even in the case of a “simple“ loan decision, the payoff configura
10、tion is much more complex.)Existing empirical studies reflect this problem in that there is no consensus on what constitutes “failure“ with definitions varying significantly and arbitrarily across studies. In other words, the dichotomy bankruptcy versus no bankruptcy is, at the most, a very crude ap
11、proximation of the payoff space of some hypothetical decision problem. It follows that it is essentially a futile exercise to try to establish the relative decision usefulness of alternative predictive systems. Accordingly, I have not concerned myself with how bankruptcy (and/or failure)“ought“ to b
12、e defined, I also have refrained from making inferences regarding the relative usefulness of alternative models,ratios,and predictive systems (e.g.univariate versus multivariate).Most of the analysis should simply be viewed as descriptive statistics which may, to some extent, include estimated predi
13、ction error rates and no “theories“ of bankruptcy or usefulness of financial ratios are tested. Even so, there are a large number of difficult statistical and methodological problems which need to be discussed. Many important problems pertaining to the development of data for bankrupt firms have gon
14、e mostly unnoticed in the literature. 2.Some Comments Regarding Methodology and Data Collection The econometric methodology of conditional logit analysis was chosen to avoid some fairly well known problems associated with Multivariate Discriminant Analysis (MDA,for short).The MDA approach has been t
15、he most popular technique for bankruptcy studies using vectors of predictors. Among some of the problems with these studies are:(i)There are certain statistical requirements imposed on the distributional properties of the predictors. For example, the variance-covariance matrices of the predictors sh
16、ould be the same for both groups (failed and non-fa iled firms);moreover, a requirement of normally distributed predictors certainly mitigates against the use of dummy independent variables. A violation of these conditions, it could perhaps be argued, is unimportant(or simply irrelevant)if the only
17、purpose of the model is to develop a discriminating device. Although this may be a valid point, it is nevertheless clear that this perspective limits the scope of the investigation. Under many circumstances, it is of interest to go through more traditional econometric analysis and test variables for
18、 statistical significance, etc.(ii)The output of the application of an MDA model is a score which has little intuitive interpretation, since it is basically an ordinal ranking(discriminatory) device. For decision problems such that a misclassification structure is an inadequate description of the pa
19、yoff partition, the score is not directly relevant. If, however, prior probabilistic of the two groups are specified, then it is possible to derive posterior probabilities of failure. But, this Bayesian revision process will be invalid or lead to poor approximations unless the assumptions of normali
20、ty, etc. are satisfied.(iii)There are also certain problems related to the “matching“ procedures which have typically been used in MDA. Failed and non-failed firms are matched according to criteria such as size and industry, and these tend to be somewhat arbitrary. It is by no means obvious what is
21、really gained or lost by different matching procedures, including no matching at all. At the very least, it would seem to be more fruitful actually to include variables as predictors rather than to use them for matching purposes. The use of conditional logit analysis, on the other hand, essentially
22、avoids all of the problems discussed with respect to MDA.The fundamental estimation problem can be reduced simply to the following statement: given that a firm belongs to some prespecified population, what is the probability that the firm fails within some prespecified time period? No assumptions ha
23、ve to be made regarding prior probabilities of bankruptcy and/or the distribution of predictors. These are the major advantages. The statistical significance of the different predictors are obtained from asymptotic (large sample) theory. To be sure, as is the case in any parametric analysis, a model
24、 must be specified, so there is always room for misspecification of the basic probability model. 3.Evaluation of Predictive Performance There is no way one can completely order the predictive power of a set of models used for predictive (decision) purposes. As a minimum, this requires a complete spe
25、cification of the decision problem, including a preference structure defined over the appropriate state-space. Previous work in the area of bankruptcy prediction has generally been based on two highly specific and restrictive assumptions when predictive performance is evaluated. First, a (mis)classi
26、fication matrix is assumed to be an adequate partition of the payoff structure. Second, the two types of classification errors have an additive property, and the “best“ model is one which minimizes the sums of percentage errors. Both of these assumptions are arbitrary, although it must be admitted t
27、hat the first assumption is of some value if one is to describe at least one implication of using a model. Much of this discussion will therefore focus on such a (mis)classification description. Nevertheless, the second assumption will also be used at some points, since it would otherwise be impossi
28、ble to compare the results here with those of previous studies. The comparison cannot be across models because the time periods, predictors, and data sets are different. Rather, the question of interest is one of finding to what extent the results conform with each other. 4 Conclusions There are two
29、 conclusions which should be restated. First, the predictive power of any model depends upon when the information (financial report) is assumed to be available. Some previous studies have not been careful in this regard. Second, the predictive powers of linear transforms of a vector of ratios seem t
30、o be robust across (large sample) estimation procedures.Hence,more than anything else, significant improvement probably requires additional predictors. 译文 财务比率与破产概率预报 资料来源 :会计研究杂志, 18 卷第 1 号, 1980 年春季美国印刷 作者: 詹姆士 A. 欧森 1.说明 本文介绍了通过实证研究来预测公司破产的一些研究结果。在这一领域一直有相当数量的人在研究,值得注意的研究贡献有比弗( 1966, 1968a; 1968b
31、),奥尔特曼( 1968; 1973),奥尔特曼和洛里斯( 1976),奥尔特曼和麦高夫( 1974) ,奥尔特曼,霍尔德曼和纳拉亚南( 1977)迪金( 1972),利比( 1975) , 布卢姆( 1974) , 埃德米斯特( 1972),威尔科克斯( 1973),莫耶( 1977)和列弗( 1971)。瓦特和特恩布尔的两篇未发表的论文( 1975a; 1975b)和一篇圣美罗和威萨( 1977)的文章特别有意义,因为他们是第一次从逻辑和系统上估计失败机率的研究。本研究是类似后者的研究,该方法是对所谓的条件 Logit 模型的最大似然估计。 用于这项研究的数据集为七十年代 ( 1970 1
32、976) 。据我所知,只有三个对企业失败的研究已审查了这一期间的数据。一个是奥特曼和麦高夫( 1974)的研究,对来自 1970 年至 1973 年期间,且只有一种分类误差(错误分 类的失败公司)的失败公司样本进行了分析。莫耶( 1977)考虑了 1965 年至 1975 年期间的数据 , 但破产公司的样本是非常小的 (27 家公司 )。第三项研究 , 奥特曼,霍尔德曼和纳拉亚南( 1977),是在奥特曼( 1968)研究基础上的进一步探讨,数据来自 1969 年至 1975 年期间。他们的样本为五十三家破产企业和相同数目的为破产企业。与此相反,我的研究是基于对 105 家破产企业和 2,05
33、8 家未破产的公司的研究。尽管这三项研究不同于目前所关注的方法和目标,不过,在本文中,他们的结论是有意义和有用的。 然而,人们可能会问一个 可能令人尴尬的问题:为什么预测破产 ? 这是一个很难回答的问题,在这里我也不能给出答案或理由。也许,可以说我们正在处理 “ 明显的 ” 实际利益的问题。在现实世界中,人们可以通过这项研究来及时了解企业情况,并成为人们关注和选择企业的依据。 但目前我能想到的解决该问题的方法没有将盈利空间自然分为非破产与二元状态破产。(即使在 “ 简单的 ” 贷款政策下,回报的配置也是很复杂的。)现有的实证研究与研究跨变极大,且其都任意地定义来研究此问题,对何为 “ 失败 ”
34、 还没有共识。换句话说,没有破产与二分法破产,充其量是近似对盈利空间的一些假设性的回答。因此尝试建立相对决策有用性的替代预测系统基本上是徒劳无益。因此,我并不关心自己如何来对破产(和 /或失败)的定义界定。 我有作出替代模式,分析比率的相对有用性,以及预测系统(例如:单变量与多元)。其中,大部分的分析应视为描述性统计,这可能在某种程度上,包括估计预测错误率,并没有对财务比率的效用进行测试。即便如此,也有一大批较难的统计方法问题需要讨论。破产企业数据的发展方面的许多重要问题,在大多文献中被忽视。 2.方法和数据收集 一些意见 选择计量方法的条件进行 logit 分析,目的是为了避免一些与多元判别
35、分析( MDA 的简称)相关众所周知的问题。 MDA 方法是破产预测研究中最受欢迎的技术。一些该方法研究的问题: ( 1) 有某些统计的要求,预测的分布需具备某些属性。例如,预测的协方差矩阵应为相对应的两组(失败和非失败企业);此外,要求一般分布式预测对虚拟的独立变量要可用。违反这些条件,如果模型的唯一目的是制定一个判别的设备,它或许可以辩称,是重要的(或只是不相关的),这虽然可能是一个有效点,但这种观点显然限制了调查范围。在许多情况下,它是通过 传统计量测试变量,去分析和统计意义等。( 2) MDA 模型应用程序的输出是一个直观解释,因为它基本上是序号排列 ( 无区分 ) 的分数。对于解决,
36、例如一个错误分类结构,不能描述分区的问题,得分不直接相关。但是,如果事先指定两组的统计概率,则可能得出失败后的概率。此贝叶斯修订过程将会无效或接近无效,除非正常的假设等能够被满足。( 3)亦有某些通常在 MDA 的 “ 配对 ” 程序产生的相关问题。失败和非失败的公司将根据条件如资产规模和行业来进行匹配,这些往往有点武断。这明显没有意义,通过不同的匹配程序,包括完全没有匹配,不会增加或者减 少对结论的影响。最起码,在预测的变量上它似乎卓有成效,而不是将它用于匹配。 使用条件 logit 分析,另一方面,基本上避开所有前面讨论的对 MDA 的问题。只是,下面可以减少基本估计问题:鉴于公司属于一些
37、预先指定的人口,公司内的失败的概率是属于什么预先的时间段?没有假设要提出破产的先验概率和(或)分布的预测。这些都是主要的优势。不同预测的统计意义来自渐近( 大样本 ) 理论。当然,正如在任何参数的分析中必须指定模型一样,因此总是会存在偏差导致基本概率模型的假设错误。 3预测性能的评价 有没有办法可以完全以一个用于预测 (决定)用途。来设置模型的预测能力。这就要求一个完整的决策问题,其中包括定义适当的状态空间的偏好结构。在破产预测领域以往的研究,一般是基于两个非常具体、严格的假设时,对预测性能进行了评价。第一,(错误)分类矩阵被假定为适当的分区。第二,两种类型的分类错误有添加剂的财产,和 “ 最
38、佳 ” 的模型是最大限度减少的百分比错误的款项。这两项假设是任意的的,但必须承认,第一种假设是有价值,如果需要描述,至少需要使用模型的一个假设。因此,这次讨论的很多将重点放在(错误)分类描述的问题上。不过,第二个假设还将用在一些观点上,因为否则可能与以前的研究的结果无法进行比较。比较不能跨模型,因为时间段、预测和数据集是不同的。相反,感兴趣的问题是寻找到何种程度上结果符合彼此。 4.结论 有两个必须重申的结论。第一,任何模型的预测能力取决于当前信息(财务报告),假设为可用。一些以前的研究没有考虑。第二,预测比率的线性变换的能力是在(大样本)估计过程中能力更显著。因此,重大改进模型可能需要额外的预测。