财务危机是企业危机最综合、最显著的表现。随着我国证券市场机制和企业破产制度的完善，财务危机不但使企业遭受巨大损失，而且直接影响企业的生存和发展。如何提高企业财务危机预警效果，已成为企业的重要研究问题。然而，已有的财务危机预警研究主要围绕传统的单分类器财务危机预测方法展开，缺乏系统性。不但人工智能单分类器财务危机预测方法有待进一步扩充，而且忽视了以多分类器组合进行财务危机预测可能带来的好处，以及专家经验知识和非财务信息对财务危机预警的重要作用。在该研究背景下，当前正飞速发展的计算机科学、人工智能、数据挖掘和群决策理论，为从新的视角开展企业财务危机预警研究提供了思路。本书以财务管理理论和企业预警理论为基础，跟踪人工智能、数据挖掘以及群决策领域的前沿理论、方法和技术，采用定量和定性相结合、规范和实证相结合、机器学习和专家经验相结合的跨学科综合研究方法，对企业财务危机预警的智能决策方法体系展开系统的研究。本书从剖析企业财务危机问题的现象和本质入手，提出了包含财务指标实时监控、财务危机机器学习预警和财务危机专家经验预警三个方面的多层次财务危机预警体系框架，并抓住基于定量财务数据的机器学习预警和基于定性财务及非财务信息的专家经验预警两条主线来进行深入研究。首先，样本数据集和定量财务指标体系的构建。从上海和深圳股票市场收集135对上市公司各自的30个财务指标数据，经过排除残缺数据和异常数据预处理后，对其展开充分的统计描述和正态检验，证明用该样本对财务危机定量预测方法进行实验验证的适用性。然后，通过指标均值比较，逐步判别分析和共线性检验，精减得到财务危机预测的定量财务指标体系。其次，基于人工智能单分类器的财务危机预警方法研究。提出遗传算法动态优化决策树的财务危机预测方法，通过优化决策树输入属性集合来提高泛化预测能力。实验结果表明，该方法明显优于事先静态选定输入属性集合的决策树方法。同时，设计了支持向量机财务危机预测方法的流程算法，通过最大化软间隔来寻找最优分类超平面，并采用交叉验证和网格搜索相结合的方法确定支持向量机模型参数。实验结果表明，该方法在拟合能力、泛化能力和模型稳定性三个方面具有很好的综合性能。提出相似度加权投票组合k近邻案例的财务危机预测方法，采用知识引导和k近邻相结合的混合案例检索策略，并设计了相似度加权投票概率最大化原则来确定待预测案例的财务状况类别。实证实验不但分析了该方法参数的经验取值范围，而且证明它非常适用于财务危机的短期预测。再次，基于多分类器组合的财务危机预警方法研究。分别构建基于异构模型的财务危机并、串联组合预测方法和基于AdaBoost组合分类器的财务危机预测方法。提出财务危机的并联组合预测方法，建立了并联组合的加权多数投票模型和基本分类器的投票权重模型，并以差异性原则和个体优化原则来选择并联基本分类器。提出以类别相关经验为导向对财务危机进行串联组合预测的思想，通过单类择优算子和总体择优算子选择串联基本分类器，并设计了流程算法。提出财务危机的混合组合预测方法，用并联结构作为串联结构的基本模块，以弥补单纯串联组合易受某个基本分类器主导的缺陷。对基于各单分类器和三种组合结构的财务危机预测方法进行对比实验表明，并联组合财务危机预测方法在取得最高平均准确率的同时降低了离散程度，混合组合财务危机预测方法在取得最低离散程度的同时提高了平均准确率，两者都能起到不同基本分类器信息互补和扬长避短的作用；但是，串联组合财务危机预测方法容易受串联中的第一个基本分类器主导，预测性能没有明显提高。设计了基于AdaBoost组合分类器的财务危机预测框架和算法，分别采用单变量分析和决策树两种分类器算法训练基本分类器，实验结果表明，AdaBoost集成单变量分析的效果显著优于AdaBoost集成决策树、单决策树以及单支持向量机，弱分类器的选择对基于AdaBoost组合分类器的财务危机预测模型的性能有重要影响。最后，基于群决策的财务危机可能性评价预警方法研究。为了弥补单靠机器学习手段处理定量财务数据方式进行财务危机预警的不足，提出充分利用专家经验和知识处理财务及非财务定性信息方式来评价企业财务危机可能性的预警思想。为此，设计了包含财务及非财务信息的定性指标体系及其评分标准，以及定性指标权重的多专家协商机制，提出了专家期望协商因子的新概念，采用灰色综合评价方法进行财务危机可能性评价，并对灰色评价的灰类和白化权函数进行了具体设计。实例分析检验了该方法的有效性。对企业财务危机预警理论和方法体系的研究，丰富了该领域的理论研究成果，能够为实际企业开展财务危机预警提供理论指导和技术支持，具有重要的理论和实际意义。关键词 财务危机 预警 分类器 人工智能 群决策

Financial distress （FD） is the most synthetic and notable business distress.With the stock market mechanism and business bankruptcy law gradually being perfect，enterprises in FD not only suffer great loss，but also their survival and development are directly affected.So how to improve the effectiveness of business FD early warning has already become an important research problem.However，past FD early warning research only paid attention to traditional single classifier FD prediction methods，lack of systematism.Not only artificial intelligence single classifier FD preidciton methods should be further extended，but also they ignored the possible benefit of multi-classifier combination for FD prediction and the importance of experts’experiential knowledge and non-financial information for FD early warning.Nowadays，computer science，artificial intelligence，data mining and group decision-making are developing rapidly，which provides new idea for studying FD early warning from a new prospect.Based on financial management theory and business early warning theory，following the newest theories，methods and techniques in the fields of artificial intelligence，data mining and group decision，adopting the inter-subject methodology integrating qualitative and quantitative analysis，normative and empirical research，and expert experience and machine learning，this book systematically studies the theory and method system for business FD early warning.Starting with anatomizing the phenomena and essence of business FD，this book puts forward a multi-layer framework for FD early warning，which is consist of financial ratio’s real-time monitor，machine learning FD early warning，and FD early warning with experts’experience.For lucubration，it grasps two key clues of machine learning FD early warning based on quantitative financial data and FD early warning with experts’experience based on qualitative financial and non-financial information.Firstly，sample data set and quantitative financial ratios system were constructed.Data of135pairs of listed companies’30financial ratios were collected from Shanghai and Shenzhen stock exchange.After eliminating missing and abnormal data，statistical description and normality test proved that these data sets were suitable for empirical experiments，whose aim was to validate the effectiveness of quantitative FD prediction methods.Then，quantitative financial ratios system for FD prediction was refined through mean comparison，stepwise discriminant analysis and collinearity test.Secondly，FD early warning methods based on artificial intelligence single classifier were studied.FD prediction method based on genetic algorithm dynamically optimizing decision tree （DT） was brought forward.By optimizing the input attributes set of DT，it can improve the generalization ability of FD prediction.Experimental result showed that this method was much better than traditional DT which statically chooses input attributes in advance.The workflow of support vector machine （SVM） FD prediction method was designed to find the optimal classification hyper-plane by maximizing the soft margin.Parameters of SVM model were determined by cross validation and grid search.Experimental result showed that this method had very good synthetic performance in fitting ability，generation ability，and model stability.FD prediction method based on similarity weighted voting CBR was proposed.A hybrid case retrieval on knowledge guided strategy and k nearest neighbor principle was adopted.The principle of maximum similarity weighted voting probability was designed to determine the financial condition class of target case.Empirical experiment not only analyzed the empirical value range of parameters，but also proved that this method was very suitable for short-term FD prediction.Thirdly，FD early warning mehod based on multi-classifier combination was studied.Multi-classifier combination methods based on heterogeneous models and multi-classifier combination methods based on AdaBoost were respectively designed.Parallel combination FD prediction method was brought forward.Parallel combination’s weighted majority voting model and basic classifier’s voting weight model were constructed.Diversity principle and individual optimization principle were adopted to select basic classifiers.Sequential combination FD prediction method guided by apriori class-wise knowledge was proposed.Single best selection operator and whole best selection operator were adopted to choose basic classifiers for sequential combination，and the workflow of sequential combination FD prediction was designed in detail.Hybrid combination FD prediction method was thought out.Parallel structure was used as the basic module of sequential structure，so as to make up the limitation of pure sequential combination，which is easily dominated by certain basic classifier.Comparative experiment among single classifier FD prediction methods and three combination FD prediction methods indicated that parallel combination FD prediction method got the highest mean accuracy as well as decreased the variation degree，and hybrid combination FD prediction method got the lowest variation degree as well as improved mean accuracy.However，sequential combination FD prediction method was easily dominated by the first basic classifier and did not get evident improvement in prediction performance.The framework and algorithm of multi-classifier combination method based on AdaBoost for FD prediction was designed，and the single attribute test and decision tree were respectively used for base classifier training.Experimental results showed that the AdaBoost ensemble based on single attribute test outperformed the AdaBoost ensemble based on decision tree，the single decision tree，and the single support vector machine.The choice of weak learner has improtant impact on the FD prediction performance of AdaBoost ensemble.Finally，FD possibility evaluation early warning method based on group decision was studied.To remedy the shortcoming of FD early warning which only processes quantitative financial data by means of machine learning，this paper advanced the thought of FD early warning that experts’experience and knowledge should be fully utilized to process financial and non-financial information and evaluate enterprise’s FD possibility.Therefore，qualitative evaluation measures system consist of financial and non-financial information was designed，as well as its scoring criterion.Given definition of the new concept of expert’s expected negotiation factor，multi-expert negotiation mechanism for weighting qualitative measures was designed in detail.Grey synthetic evaluation method was used to evaluate enterprise’s FD possibility，and the grey classes and their whitenization weight functions were designed specifically.Case study validates the effectiveness of this method.Theory and method system of business FD early warning proposed in this paper can greatly enrich the theoretical research fruit in this field.It also can guide and support particular enterprises to implement FD early warning.Therefore，this research is theoretically and practically important.Key words financial distress early warning classifier artificial intelligence group decision making