1、 外文翻译 原文 A neural network model to predict long-run operating performance of new ventures MaterialSource:http:/ Bharat A. Jain and Barin N. Nag 1. Introduction In recent times, artificial neural network techniques have found a wide variety of applications in the finance area, addressing problems suc
2、h as corporate bond rating, credit evaluation and underwriting, bankruptcy, and thrift failure predictions.1) Neural networks attempt to model human intuition by simulating the process of adaptive biological learning. The special feature of neural networks is their ability to recognize patterns in t
3、he data, and thus classify data sets or predict data values. In financial applications, they provide a viable alternative to statistical models which have been widely used for prediction or classification tasks. Statistical models, however, suffer from a major drawback in that they often impose stri
4、ngent conditions which are not met in practice. Recent studies by Tam and Kiang 28 and Salchenberger et al. 26 provide evidence to suggest that neural networks outperform conventional statistical models such as discriminant analysis, logit models, K nearest methods, etc., in financial applications i
5、nvolving classification and prediction. These studies, however, have focussed on well-established firms where considerable financial and other information is available from financial reports, disclosure statements, analysts reports, market prices, and earnings performance. Few studies have focussed
6、on predicting the performance of new ventures, where publicly available information is limited and usually incomplete or imperfect. An interesting managerial decision problem in the domain of corporate finance applications involves forecasting the long-run performance of new ventures entering the ma
7、rket through an initial public offering (IPO). The firms issuing IPOs may either be start-up firms or may have operated for a few years as a private company. In either instance, the market has very little information about these companies in terms of past performance, current competitive position, a
8、nd future growth prospects. The lack of operating history and market prices adds to the uncertainty and makes it more difficult to forecast the future performance of these firms. Further, information asymmetry between managers and investors regarding the value of future growth opportunities provides
9、 incentives to window-dress performance measures. For instance, managers may attempt to window-dress their pre-issue performance measures prior to the IPO in order to condition investors beliefs favorably and thereby obtain inflated valuations. Although the true performance of these firms would be r
10、evealed subsequently, it is too late for investors who bought overpriced stock at the IPO. Thus, an important decision problem for the market, institutional investors, and fund managers is to predict the long-run operating performance of new issues based on incomplete and imperfect information and i
11、n the face of incentive problems such as adverse selection and moral hazard. This is the first study which attempts to use analytical tools to predict the longrun operating performance of new issues. Both neural networks and statistical models are used to ex-ante classify IPO firms as superior or in
12、ferior performers as measured by their long-run operating performance. The classification accuracy of the two methods are compared over different sample designs and classification decision rules. The IPO classification problem presents a considerable challenge since firm performance depends on sever
13、al complex and interrelated variables whose relationships are often ill defined or unclear. As a result, it becomes difficult to specify the parametric functional form of the relationships between variables. Additional problems include nonlinear relationships, correlated variables, and noisy data se
14、ts (different outputs for similar inputs). Due to certain inherent advantages, neural networks are expected to outperform statistical models in this kind of an application. Neural networks have been used extensively, and successfully, for modeling nonlinear statistical relations non parametrically.
15、In addition, the nature of neural network memory leads to a reasonable response when presented with incomplete, noisy or previously unseen input. 2. Description of the IPO problem Private companies entrepreneurs in search of capital to finance their growth prospects decide to go public at some point
16、 in their corporate life cycle. By going public, they gain access to the capital markets, thereby enhancing their ability to raise capital to finance expansion or undertake positive net present value (NPV) projects. It would appear that the decision to go public is a favorable signal since it sugges
17、ts that issuers anticipate future growth and are turning to the capital markets to fund these opportunities. However, an equally likely alternative scenario can be constructed under which issuers have some unfavorable private information about their future prospects or see their growth levelling off
18、 and attempt to use the IPO as a means to divest their holdings at favorable prices. These two different motivations for going public lead to vastly different expectations regarding long-run operating performance. Issuers who go public to fund growth projects are likely to experience an increase in
19、their performance measures relative to the period before going public. On the other hand, issuers who go public to divest their holdings due to unfavorable private information are likely to experience a decline in performance relative to the pre-IPO period. In the absence of information on operating
20、 history or market prices, investors are unable to distinguish between these two types of IPO firms. Information asymmetry provides an opportunity to issuers with poor future prospects to window-dress their performance measures prior to going public in an attempt to condition investor beliefs favora
21、bly and thereby obtain higher valuations. Thus, attempts at ex-ante classification of firms as long-run superior or inferior performers based purely on pre-issue measures is likely to lead to high error rates. The nature of the IPO classification problem raises doubts on the suitability of standard
22、statistical methodology to model the problem. The basic objective is to use information available at the time of the IPO to ex-ante classify firms as superior or inferior performers. Since managers have considerable discretion in reporting financial numbers, attempts at window-dressing reduce the in
23、formation content of financial measures. Other problems include that of noisy data sets (i.e., the same input can provide different outputs) and correlated variables. Another problem in applications such as IPOs is that incentive problems can result in non-monotonic relationships. For instance, cons
24、ider the decision by entrepreneurs on the extent of ownership to retain in the post-IPO firm. Generally, higher ownership retention by entrepreneurs managers signals good news since they are credibly conveying their confidence in the future of the firm by investing their own money. Low post-IPO mana
25、gement ownership, on the other hand, signals a lack of confidence on the part of insiders regarding the firms future prospects. However, empirical evidence suggests that the relationship between ownership retention and firm performance is U-shaped with performance declining at very low or very high
26、ownership retention. One explanation suggested in the literature for this form of relationship is that very high levels of management ownership leads to a high degree of risk aversion. As a result, managers often forego positive NPV projects that increase firm value. Thus, noisy data, non-monotonic
27、relationships and correlated variables present major obstacles in modeling the IPO problem using standard statistical methodology. Further, the long-run performance of new issues depends on the interaction of many factors whose relationships are often ill defined or unclear thus making it difficult,
28、 if not impossible, to specify the functional form of the relationship between variables. Parametric statistical methods which depend on several restrictive assumptions such as lognormality or sample path continuity may not be suitable for the IPO problem since they are not robust to specification e
29、rrors. Thus, it is not surprising that prior empirical studies that applied statistical models to the IPO market exhibited extremely low explanatory power. Neural network models appear to be a promising alternative to statistical techniques in modeling the IPO problem. They provide a data-driven non
30、parametric approach to modeling the IPO problem and do not depend on restrictive parametric assumptions. Unlike parametric statistical models, they do not need to, a priori, specify the functional relationship between variables. They are adaptive and respond to structural changes in the data generat
31、ing process in ways that parametric models cannot. The pattern recognition capabilities of neural networks allows them to match large amounts of input information simultaneously and then generate a categorical or generalized output. The nature of network memory leads to a reasonable network response
32、 when presented with incomplete, noisy or previously unseen input, a property referred to as generalization. Further, neural networks can also be used for noise filtering. They are able to preserve a greater depth of structure and detail compared to traditional filters while still removing the noise
33、. The IPO market has over the years become an important source of capital in the economy. As the predictive power of statistical methods in the IPO type of problem is likely to be low, any alternative method which produces even a small improvement in the predictive ability of decision-makers is like
34、ly to be of considerable value. 译文 以 神经网络模型 预测 新建 企业 长期运营业绩 资料来源 :/ 作者： 巴拉特 .A.简 ； 柏琳 N.耐格 当今社会，人工神经网络技术已经在金融领域广泛应用，解决诸如企业债券评级，信用评估及包销，破产预测问题。神经网络模型试图通过人类大脑神经运作的模拟。神经网络的特点是具有较好的识别能力，从而预测分类的数据集或数据值。在财务管理的研究中，它提供了一个可行的统计模 型用于预测或分类任务，解决了已被广泛使用却有一个严重的缺点的统计模型所存在的问题，即在实践中经常不满足严格的条件。 Tam、 Kiang 和 Salchenbe
35、rger 等最新研究发现神经网络优于传统的统计模型，如判别分析， Logit 回归模型， K 最近的方法，那些涉及的分类和预测的财务应用方法。不过，这些研究都是关于那些拥有完善的财务报告、披露声明和审计报告的企业。很少有研究是关于预测新企业的，因为他们的公开的信息有限，信息通常是不完整或不完善的。 现在，出现了企业财务管理决策研究的新领域，预测首次公开发行（ IPO）企 业的长期发展业绩，这可能是一家新创立的公司也可能是一家已经经营了几年的公司。市场上有关这些公司的资料很少，如公司过去的业绩、目前的竞争地位和将来的发展前景。因为缺少经营的历史资料和市场价格，导致不确定性的增加，使人们更难预测这
36、些公司将来的表现。管理者也许意图在信息监管者的面前掩盖问题，通过有利条件使投资者错误的高估企业的价值。虽然，这些公司的真实情况会在其后会被发现，但是对于那些在首次公开招股时购买高估股票的投资者来说，已经为时已晚。因此，投资机构和基金监管者者的一个重要的市场决策问题是：在存在逆向选择和道 德风险的情况下，如何通过不完整的信息来预测新公司的发展情况。 这是第一个试图利用分析工具来预测新企业的长期发展业绩的研究。神经网络和统计模型都是用于事前判别上市公司的长期经营情况的优劣。针对不同的样本和分类决策方法设计，这两种方法的判别精确度不同。 IPO 分类是一个相当大的挑战，因为公司的业绩只能通过几个复杂
37、、相关的变量体现，这些变量往往具有关联和不明确性。因此，确定这些参数变量之间的关系函数形式变得十分困难。其他问题，例如，非线性关系、相关变量和矛盾的数据集（类似的输入得出不同的结果）。由于一些先天的优势， 神经网络将优于某类类的应用统计模型。在非线性非参数化建模的统计关系中，神经网络已经被广泛的使用并且取得成功。此外，由于神经网络的记忆性质，当输入了不完整的、有错误或不具有前后相关的信息时，产生合理的反应。 2、上市问题的说明 民营企业为了企业的长远发展，在他们企业生命周期的某个时点，在公开的资本市场上筹集资金。通过上市，他们进入资本市场从而提高筹集资金的能力，从而可以采取积极的财务政策或积极
38、的净现值（ NPV）项目。这样看来，上市是一个有利的信号，表明发行人预期企业未来的增长 ，并正在通过资本市场筹资，以抓住增长的机会。然而，同样有可能是替代方案，发行者有一些对未来前景不利的私人信息或者认为企业成长将趋于平缓，尝试通过使用首次公开发行为手段，以优惠的价格剥离其持有的股份。这两种不同动机的上市导致截然不同的预期长远经营业绩。前一种新股的发行人，很可能在上市前提高企业的业绩。后一种，由于不利的私人信息，通过上市实现剥离其股份的发行人，很可能在上市前期降低企业的业绩。在没有经营历史和市场价格信息的情况下，投资者很难区分这两类首次上市的公司的类型。信息不对称使未来前景不佳的发行者有一个机
39、 会粉饰企业的业绩，在公开上市前，提前采取措施试图让投资者相信企业将来的发展，从而获得更高的估价。因此，纯粹基于预先发行措施，试图在事先将企业分类为长期业绩优或劣，很可能导致高错误率。 预测首次公开发行股票性质的优劣成为了一个问题，人们怀疑根据标准的统计方法来解决这个问题的适用性。预测基本目的是：在首次发行前，利用信息可以事先划分为优或劣业绩的企业。但是，由于企业经理们采取减少财务数据信息的相关措施，试图粉饰这一情况。预测结果出现问题，包括矛盾的数据集（即相同的输入得到了不同的输出）和相关变量；另一个问题是： 首次公开招股激励问题可能会导致非单调的关系。例如，考虑在上市后，公司的所有权保留的问
40、题。一般而言，较高的企业经理人保留企业的所有权是一个好信息，因为他们用自己的钱进行投资，说明看好企业的前景。另一种情况，上市后的低管理权，预示着部分内部人士对企业将来的前景缺乏信心。不过，经验证据表明，当所有权保留程度在很低或很高时，所有权保留程度和企业业绩之间的关系是 U 型的。对这种关系的形式，一种在文献中的解释认为这是因为非常高的管理水平导致了风险厌恶程度高。因为，管理者往往放弃能增加公司价值的正净现值的项目。 因此，使用 标准的统计方法对首次公开招股进行建模会出现一个严重的问题：错误数据、非单调关系和相关变量。此外，新发行股票的长期业绩取决于相互作用且不明确或者不清晰的许多因素，因此预
41、测是很困难的。不过不是不可能，可以通过指定变量之间关系的函数关系这一形式来预测。取决于几个限制性假设的参数统计方法可能不适合新股发行上市的问题，因为他们提供的数据不够健全。因此，之前应用于新股市场的实证研究呈现出极低的解释力，这一结果并不奇怪。 神经网络模型似乎是一个有希望替代原有建模统计的技术，解决首次公开发行股票的预测问题，提供了一个数据驱 动的非参数建模方法，这种方法不依赖于严格的参数假设。与传统的模型参数不同，他们并不需要先检验指定变量之间的函数关系。他们能适应和回应数据结构变化的方式，而参数化模型不能。由于神经网络模式的识别能力，他们能够将同时输入的大量信息进行匹配，然后产生一个明确或广义的输出。当输入错误或者输入的信息使模型以前不曾了解的时，该网络存储器的性质会使网络模型有一个合理的调整，这个属性被称为泛化。此外，神经网络业可以用于噪声的滤波。他们能够保持一定的结构和细节与传统的过滤器进行更深入的比较，同时还消除了噪声。多年前， IPO 市场已经成为资本经济的重要来源。对于预测首次公开上市发行股票的类型问题来说，统计方法的预测能力可能很低，任何其他方法的小小改进，对决策者的预测能力很可能具有极大的价值。
