Fraudulent Financial Statements Derection Based On Time Series Information

Starting from the research achievements in the field of financial fraud identification, this paper analyzes the weaknesses of previous research. On this basis, this paper explores deeply and puts forward some effective solutions about the weaknesses. More specifically, this paper conducts the research from three aspects aiming at the two weaknesses of previous research. Firstly, aiming at the weakness that traditional financial fraud identification model can not capture the vertical time series abnormality of financial indexes, this paper refines the abnormality into time series indexes and adds the indexes to the Naive Bayes classification model, which improves the accuracy of the previous classification model. Meanwhile this paper discusses the time series indexes in form of difference, ratio and relative value which are more effective, as well as how to weight the time series indexes among different years reasonably. The conclusion of empirical research is that the time series indexes in form of ratio value are more effective. Besides, when we assign0.8weights to the closer time series indexes and0.2weights to the distant time series indexes, the result can be better. Secondly, from the perspective of clustering, this paper verifies the effectiveness of time series indexes in form of ratio value, which are constructed in classification model. And the paper also mines the time series indexes in form of ratio value, which can reflect characteristics of fraud. The conclusion of empirical research is that when financial fraud exists, the time series indexes in form of ratio value, derived from rate of equity and earnings per share, can be abnormal. Thirdly, aiming at the inherent weakness of traditional model as a supervised learning algorithm, that is, the work of selecting and labeling control samples may be potentially unreasonable or redundant, this paper optimizes the traditional model based on partially supervised learning algorithm. The conclusion of empirical research is that the optimized model can exclude interference from unreliable control samples. At the same time, the new model can make full use of information in fraudulent samples, which is just useful for identifying fraud. Also, the optimized model has a better performance in identifying financial fraud.