基于时序性信息的财务报表欺诈识别

发布时间：2018-12-25 10:31

【摘要】：本文从财务欺诈识别领域的研究成果入手,在分析前人研究不足的基础上,针对不足,深入探索,提出了较为有效的解决办法。具体地,本文针对前人研究的两项不足做了三方面工作。第一,本文针对传统财务欺诈识别模型无法捕捉财务指标纵向年度异常的不足,将这种纵向异常提炼成时序指标并加至朴素贝叶斯分类模型,改善了模型的分类精度。其中,着重研究了差值、比值、相对值形式的时序指标哪种更为有效,以及不同年份间的时序指标如何合理加权的问题。实证研究的结论是比值形式的时序指标更为有效,且当分配0.8的权重给较近年份的时序指标,同时分配0.2的权重给较远年份的时序指标时,效果更好。第二,本文从聚类的角度验证了分类模型中构造的比值时序指标的有效性,同时,还挖掘了比值时序指标所能反映的欺诈特征。实证研究的结论是当财务欺诈行为存在时,由净资产收益率和每股收益各自衍生出的比值形式时序指标可能出现较大异常。第三,本文针对传统欺诈识别模型作为监督学习算法而固有的不足——对照样本的选择和标注工作存在潜在不合理性、冗余性,基于部分监督学习算法对传统模型进行了改造。实证研究的结论是经优化改造的模型,不仅能够排除不可靠对照样本的干扰,还能充分利用欺诈样本中包含的有利于识别欺诈的信息。优化后的模型在识别财务欺诈方面也拥有更为出色的表现。
[Abstract]:This paper starts with the research results in the field of financial fraud identification, and on the basis of analyzing the insufficiency of previous research, aiming at the deficiency, deeply explores, and puts forward some more effective solutions. Specifically, three aspects of work have been done in this paper in view of the two deficiencies of previous studies. Firstly, aiming at the deficiency of the traditional financial fraud identification model which can not catch the longitudinal annual anomaly of financial index, this paper abstracts this longitudinal anomaly into a time series index and adds it to the naive Bayes classification model, which improves the classification accuracy of the model. Among them, the difference value, the ratio, the relative value form time series index which is more effective, and the time series index between different years how to weigh reasonably are studied emphatically. The conclusion of the empirical study is that the time series index in the form of ratio is more effective, and the effect is better when the weight of 0.8 is assigned to the time series index of the recent year and the weight of 0.2 is assigned to the time series index of the farther year at the same time. Secondly, from the point of view of clustering, this paper verifies the validity of the ratio time series index constructed in the classification model, and also excavates the fraud characteristic that the ratio time series index can reflect. The conclusion of the empirical study is that when the financial fraud exists, the ratio forms derived from the return on net assets and earnings per share may be abnormal. Thirdly, aiming at the inherent deficiency of the traditional fraud identification model as a supervised learning algorithm-the selection and labeling of control samples is potentially irrational and redundant, the traditional model is modified based on the partially supervised learning algorithm. The conclusion of the empirical study is that the modified model can not only eliminate the interference of unreliable control samples, but also make full use of the information contained in fraud samples to identify fraud. The optimized model also has a better performance in identifying financial fraud.
【学位授予单位】：西南交通大学
【学位级别】：硕士
【学位授予年份】：2013
【分类号】：F234.4;F224

【参考文献】